Automated Hip Osteoporosis Screening Using YOLOv9 and Data-Driven Models


<!-- metadata commented in wiki content


<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<big>'''Automated Hip Osteoporosis Screening Using YOLOv9 and Data-Driven Models'''</big></div>

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
'''Sandeep Kaur<sup>1</sup>, Jaspreet Singh<sup>2</sup>'''</div>

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
'''Department of Computer Science & Engineering, Chandigarh University, Mohali, India<sup>1,2</sup>'''</div>

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
'''Department of Computer Engineering & Technology, GNDU Amritsar, India<sup>1</sup>                                        [mailto:Sandeep.dcet@gndu.ac.in Sandeep.dcet@gndu.ac.in]<sup>1</sup>, '''[mailto:Jaspreet.e10279@cumail.in '''Jaspreet.e10279@cumail.in''']'''<sup>2</sup>'''</div>
-->

==Abstract:==

'''Objective:''' This study aimed to develop and evaluate a deep learning- based approach for automated estimation of hip bone mineral density (BMD), detection of osteoporosis and evaluation of fracture risk from routine hip radiographs. Further, the performance of multiple YOLOv9 variants was systematically compared with established transfer learning models to evaluate their potential for supporting osteoporosis as an image-based screening approach, where access to dual-energy X-ray absorptiometry (DXA) is limited.

'''Methods:''' Routine hip x-ray Images were processed & analysed using multiple variants of YOLOv9 and established transfer learning models. The models were compared using standard evaluation metrics like accuracy, precision, recall, F1-score and Matthews Correlation Coefficient (MCC) for evaluating the model’s performance. In addition, patient-specific BMD measurements and the distribution of osteoporosis cases were further analysed to characterize bone health within the study cohort and determine the applicability of the proposed framework for automated osteoporosis screening.

'''Results: '''Comparative evaluation showed that the YOLOv9 variants demonstrated superior predictive performance than the investigated transfer-learning models. Among, YOLOv9 variants the overall best performance was obtained by YOLOv9-E with an accuracy of 97.02%, precision of 96.71%, recall of 96.35%, F1-Score of 96.53% and MCC of 95.6%. the Strongest transfer learning model was efficientNetB0, achieving accuracy of 94.1% and MCC of 88.1%. At Patient level analysis mean BMD values ranged between 0.72 and 0.77g/cm<sup>2</sup> with osteoporosis prevalence varied from 20% to 44% across the analysed population.

'''Discussion:''' the strong performance of YOLOv9-E indicates the potential of object detection models may effectively learn radiographic patterns associated with low bone mass and osteoporosis. The complex hip anatomy and bone-related structural features explored & increased the model capacity with the improvement of YOLOv9 variants. These results support further investigation of routine hip radiographs for opportunistic osteoporosis screening although further external and clinical validation is required.

'''Conclusion: '''The results suggest that YOLOv9 based analysis of routine hip radiographs demonstrate the effective and scalable approach for automated osteoporosis assessment. YOLOv9-E shows the strongest and balanced predictive performance relative to evaluated models. This approach may facilitate the early osteoporosis screening and clinical decision-making in healthcare settings where access to DXA is limited.

''Keywords:'' Osteoporosis, Hip X-rays, Bone Mineral Density (BMD), Hip Fracture Identification, Fracture Risk Evaluation, Transfer Learning Techniques, Object Detection Models, Medical Imaging Analysis

=1. Introduction=

Osteoporosis is an extreme skeletal disorder which makes the bones to lose their strength and significantly increases the chances of fracture, particularly at bone weight bearing areas like hip. Hip fracture is considered one of the most disabling consequences of decreased bone strength, frequently resulting into long-term immobility, loss of independence and death among the elderly [1], [2]. Since hip fractures pose a big burden to patients, care givers and the health care systems, early detection of people at risk is among the priority areas in clinical practice and preventive care [3], [4].

Measurement of hip bone mineral density by the Dual energy X ray absorptiometry (DXA) is the current clinical standard of diagnosing osteoporosis as well as estimating the risk of fractures, used to make treatment decisions and also long-term risk assessment. Cost, infrastructure need, and geographical inequalities of healthcare facilities however hinder the availability and use of DXA in most regions of the globe [5], [6]. This means that most people with low BMD do not get proper screening or prophylaxis and this exposes them to hip fractures that can be prevented [7], [8] .

To address these shortcomings, there has been an impetus towards the use of opportunistic strategies involving the use of routinely obtained hip radiographs. Standard hip X rays are cheap, commonly available, and cause low radiation to the patients and hence make it an attractive alternative to large scale bone health [9], [10]. The radiographic characteristics of trabecular pattern and cortical integrity have been reported to be associated to intrinsic bone quality which argues in their possible application in analysing fracture risks [11], [12]. It is also shown by recent research that deep learning models can be used to reliably detect osteoporosis and risk of a fracture using hip radiographs and perform just as well as DXA based tests [13], [14].

Deep learning has emerged as a powerful technique for computer-aided medical image analysis, enabling to learn complex structural and textural features automatically directly from medical images [15]. In particular, Convolutional Neural Networks (CNN) are specialized deep learning architectures that progressively extract hierarchical spatial features, discriminative imaging features required for accurate classification and widely applied to medical image classification [16]. Transfer learning has become an effective strategy in medical image analysis which enables knowledge learned from large-scale datasets by pretrained CNN architectures and adapting their learned feature representation for specific medical imaging tasks, particularly in scenario when only limited amount of domain-specific training data is available. This study investigates the effectiveness of transfer for osteoporosis classification by employs VGG16, EfficientNetB0, Xception, InceptionResNetV2 and DenseNet201 for differentiate between normal and osteoporotic hip radiographs [17].

In medical imaging the important factors affecting the performance of deep learning model are image quality and consistency. Accordingly, preprocessing and enhancement techniques are employed to standardize radiographic inputs, improve image characteristics and facilitate the extraction of clinically relevant features. The improvement in local contrast and to enhance the usability of anatomical structures Contrast Limited Adaptive Histogram equalization (CLAHE) is used and image intensity values standardized by normalization. Similarly, the morphological gradient enhances the structural contours by computing the difference between dilated and eroded images [18].

To improve the model generalization, data augmentation increases the variability within the training dataset and reduce the overfitting. To improve the representation of osteoporosis- related structural features in hip radiographs these methods, combined with the image enhancement techniques, to form the preprocessing pipeline [19].

Beyond image-level classification, object detection provides the spatial information by identifying the localizing clinically relevant regions of interest (ROIs). YOLO family integrates localization and classification simultaneously, enabling efficient analysis of radiographic images. This study evaluates the YOLOv9 variants for identifying osteoporosis related radiographic patterns within relevant anatomical regions and compare their performance against established transfer learning models [20].

The methodological basis for the proposed automated hip osteoporosis screening model is provided by combining the image enhancement, transfer learning and YOLO based object detection.

This paper is organized in the following way: Section 2 is the review of relevant literature on hip fracture detection, BMD evaluation, and AI in medical imaging, Section 3 outlines the methodology including image preprocessing, augmentation, and model training, Section 4 outlines experimental findings and performance comparing between YOLO based models and conventional approaches and Section 5 explains clinical implications, limitations and research directions.

=2. Literature Review=

The Bone mineral density of the hip bone is an important parameter that can be used to measure the health of the skeletal system and assess the risk of fractures due to the fracture risk, especially in the case of osteoporosis, which is the main cause of hip fractures on the global scale [21]. Bone Mineral Density of hips is an important parameter because it can not only be used to determine debilitating outcomes, decrease medical expenses and enhance the quality of life of patients but also provide early intervention. Although Dual energy X-ray Absorptiometry is necessary to measure Bone Mineral Density in the hips, its restricted availability and high cost in most clinical environments have motivate researchers to find Dual energy X-ray Absorptiometry as a means to computational tools that could yield good, automated measurements based on routine images [22].

The evaluation of hip bone mineral density assessment and fracture risk prediction has undergone a notable transformation progressing from traditional statistical techniques to advanced deep learning-based methodologies. Initial research was largely aimed at establishing its ability to reproduce clinically accepted diagnostic outcomes using automated standard radiographs. Due to advancement in computational capabilities and imaging technologies, they encouraged the adoption of extracting the complex machine learning techniques for improving the predictive performance and diagnostic accuracy. Ahmed and Lee [23] showed that Support Vector Machines over hip radiographs were capable of suggesting an accuracy of 91.8, which demonstrates that classical algorithms could be used to provide automated evaluation. The Acosta-Batlle J. et al. [24] tested ANNs on a joint hip DXA and radiograph data, attaining 90.7 percent accuracy, and was able to affirm that the neural networks could be used to estimate BMD.

The architectural revolution emerged with Convolutional neural networks started to dominate the field with the invention of deep learning which eliminate the requirement of manual feature selection. Yaska et al. [25] used CNNs on hip radiographs with a 92.2% accuracy, Liu et al. [26] used the method on CT-derived hip dataset and yield the 92.5% accuracy. ResNet-based models improved to 93.8. DenseNet models that were used reached 93.1 indicating the benefit of deeper networks in their ability to capture finer anatomical details [27].

Due to the limitation of Convolution Neural Network in capturing complex anatomical patterns and need for more precise analysis of pelvic imaging increased, leads to adoption of deep learning architectures like ResNet and DenseNet. For improving, the gradient propagation and feature reuse, which leads to significant gains for classification accuracy, predictive performance, more consistent diagnostic performance and model reliability as compared with earlier approaches [13]. Augusto et al. [28]obtained 92.9%, and on clinical BMD datasets, Random Forest methods obtained 91.6% [29] , and EfficientNet models obtained 94.0% [30]. The balance between accuracy and generalizability and clinical applicability was 93.7% with hybrid CNN-SVM [31].

While prior/ traditional CNN models have demonstrated the high diagnostic performance, the classification is made on the entire image which leads to introduce the irrelevant anatomical structures and background noise. Consequently, recent studies have shifted towards object detection approaches which are capable of identifying specific regions of interest [32], [33]. Comparative analysis made indicates that VGG models achieve better feature representation with accuracy as high as 97.5%, while YOLO based models provide the additional advantage of fast inference, achieving 93.5% accuracy for real-time clinical applications [34].

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">Table 1:Literature Review for Hip Bone Mineral Density, Fracture Prediction & Osteoporosis screening of AI Models</span></div>

{| style="width: 100%;border-collapse: collapse;" 
|-
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Authors'''</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Insights'''</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Methods/Models'''</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Dataset'''</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Findings'''</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Outcomes'''</span>
|-
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Erickson et al., ''' ''RadioGraphics, ''2017 [23]</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Comprehensive review of machine learning applications in medical imaging.</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">ML, CNNs, Feature Engineering, Deep Learning</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Review article</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Discussed strengths, limitations, evaluation metrics, and applications of ML in medical imaging.</span>
|  style="border: 1pt solid black;text-align: center;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Established foundational understanding of AI techniques for medical image analysis.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Xing et al., ''' ''Frontiers in Medicine, ''2022 [30]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Developed a machine learning model to predict 1-year postoperative mortality in geriatric hip fracture patients.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Random Forest, LASSO, Logistic Regression</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Retrospective geriatric hip fracture surgical cohort</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Seven key predictors identified: age, surgery delay, COPD, albumin, hemoglobin, malignancy history, blood transfusion. RF achieved AUC=0.813 on test set.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Random Forest outperformed Logistic Regression for mortality prediction.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Golestan et al., ''' ''Journal of Medical Artificial Intelligence, ''2023 [21]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Investigated whether femoral neck BMD can be estimated from hand, knee, and pelvis X-rays.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Deep CNN with age and sex covariates</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Osteoarthritis Initiative (OAI), 553 patients</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Achieved >75% sensitivity and specificity for low BMD detection.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Demonstrated potential for opportunistic osteoporosis screening from non-hip radiographs.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Ong et al., ''' ''Bioengineering, ''2023 [27]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Systematic review of AI-based osteoporosis classification using CT imaging.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Review of ML/DL approaches on CT images</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">39 studies from PubMed, MEDLINE, Web of Science, ClinicalTrials.gov</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Reported accuracy range 61.8–99.4%, AUC range 0.582–0.994.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">AI-enabled CT analysis can support opportunistic osteoporosis diagnosis without DXA.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Kang et al., ''' ''Frontiers in Physiology, ''2023 [22]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Predicted BMD directly from CT images using explainable AI.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Deep Learning + Explainable AI (XAI)</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">CT scans and corresponding DXA measurements</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Correlation coefficient with DXA BMD = 0.90; maximum F1-score = 0.875.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Demonstrated reliable BMD estimation from CT and provided explainable predictions.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Wang et al., ''' ''Quantitative Imaging in Medicine and Surgery, ''2024 [26]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Developed a fully automated deep learning system for osteoporosis screening from chest CT scans.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">VB-Net, DenseNet, ResNet-18, ResNet-50</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">2,274 chest CT examinations</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Segmentation Dice >0.93; ResNet-18 achieved osteoporosis AUC=0.96.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Demonstrated effective opportunistic osteoporosis screening </span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Acosta-Batlle et al., ''' ''Skeletal Radiology, 2025 ''[24]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Proposed a fully automated AI pipeline for osteoporosis/osteopenia screening using hip radiographs.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">CNN-based preprocessing pipeline, classification network</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">2,691 hip radiographs from 1,654 patients</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Preprocessing accuracy >99%; osteoporosis/osteopenia prediction AUC=86.6% (test set).</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Demonstrated feasibility of opportunistic osteoporosis screening using routine hip X-rays.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Yilmaz et al., ''' ''Scientific Reports, 2025 ''[35]'' ''</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Developed an automated hip fracture detection and classification system and compared it with clinicians.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">YOLOv5, MobileNetV2, Xception, InceptionResNetV2</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Pelvic radiographs collected from three hospitals</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Classification accuracy ranged from 94.66%–97.67%, outperforming clinicians (84.53%).</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">AI-assisted diagnosis improved fracture detection speed and accuracy.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Muhammad et al., ''' ''PLOS One, ''2025 [33]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Proposed BONE-Net, a hybrid deep learning framework for osteoporosis detection.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">DenseNet169 + Vision Transformer + Attention Model</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Kaggle Osteoporosis Knee X-ray Dataset</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Accuracy=86.11%, Specificity=94.74%, Precision=92.86%.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Hybrid CNN-transformer architecture improved osteoporosis classification performance.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Krogue et al., ''' ''Radiology: Artificial Intelligence, 2020''</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Automated identification and subclassification of hip fractures using deep learning.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">DenseNet + Object Detection Model</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">1,118 radiographic studies, 3,026 labeled hips</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Fracture detection accuracy=93.7%, multiclass classification accuracy=90.8%.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Achieved expert-level performance and improved clinician diagnostic accuracy when used as an aid.</span>
|}


====2.1 Research Gaps====

Although the estimation of hip bone mineral density (BMD) and the risk of fractures have made a huge progress, the modern body of research has a number of limitations:

'''1 Limited Attention to sub-regions and Bone Microarchitecture: '''Detailed analysis of femoral shaft, distal femur, trabecular bone and cortical microarchitecture remains largely unexplored, most of the studies focus on the hip, femoral neck or total hip BMD using the DXA, CT or radiographs [9], [11], [13], [21], [24]. There is a critical gap in current AI models that integrate both sub-regions anatomy and 3D bone microarchitecture analysis.

'''2 Limited Explainable AI (XAI) Integration: ''' The deep learning models are finding the broader application [13], [14], [22], [24], [26] most of the models behave as “black boxes”. Discussion on Explainability have done by few studies [22]. A gap exists for the research into explainable AI (XAI) models because existing deep learning models are failed to provide the interpretable diagnostic pathways.

'''3 Scarcity of Large Multicentre and Diverse Datasets: '''Most of the investigations depends on single-centre and moreover, they are focused on population which is region specific [9], [13], [24], [31].  Only few studies have incorporated with the multicentre validation [28], [29]. The recent AI models are trained on small, uniform datasets rather than different group of patients and diverse hospital scanners.

'''4 Emphasis on Osteoporosis classification rather than Early Prediction: '''The osteoporosis/non-osteoporosis categories were used in most of the studies for the classification [13], [24], [31], [33]. For fracture detection few models focus on the predicting the progression from normal bone to osteopenia and osteoporosis.

'''5 Lack of integration between bone Quality and Biomechanical Strength: '''Early diagnosis rely on bone mineral density [10], [19], [20]. Rarely incorporated the bone strength and mechanical properties. Finite Element analysis, Stress distribution and Biomechanical simulations are absent in AI Models.

'''6 Limited investigation of recent YOLO-Based Object detection Models: '''CNNs, EfficientNet, Random Forest and hybrid architectures are employed in many studies [13], [24], [30], [31], [33]. The advanced object detection architectures like YOLOv8, YOLOv9, RT-DETR, Vision Transformers and Swin Transformers are lacking behind for the exploration of osteoporosis detection and localization.

====2.2 Contributions of this Research====

To overcome the shortcomings of earlier investigations, the study is a solid deep learning architecture of automated hip BMD estimations, hip fractures, and hip fracture risk estimations through routine hip radiographs. The study provides clinically relevant and highly accurate object detection model by combining modern YOLOv9 and various transfer learning models.

The relevant methodological improvements were made to ensure that performance and stability were as maximized as possible. Preprocessing of hip radiographs was done to enhance clarity of features and reduce noise through resizing, normalization, filtering, gamma correction and provided more stable learning. Various performance measures such as accuracy, precision, recall, F1-score and Matthews Correlation Coefficient were used to evaluate the models as they offer a complete view of the quality of the prediction.

The YOLOv9-E model was found to be the most effective model with 97.02 up to 96.53 accuracy, precision, recall and F1-score alongside an MCC of 0.956 and highly balanced predictions with high confidence. Confusion matrix analysis has verified that there are insignificant cases of misclassification implying the strength of the model and its application in clinical practice.

The main contributions of the research are:

:* '''Automated Hip BMD and Fracture Assessment:''' Left an AI-driven system that would classify the levels of hip BMD and forecast to risk a fracture using only routine radiographs, without needing to use manual interpretation.

:* '''High-Performance YOLOv9-E Model:''' YOLOv9-E model Proposed and tested the YOLOv9-E architecture, which outperforms all of its key metrics, outperforming sub traditional deep learning models.

:* '''Clinical Relevance:''' A low false-negative rate was maintained, which was beneficial in early detection of osteoporosis, informative clinical decision-making and offered an effective solution in settings with limited access to DXA.

=3. Dataset Acquisition=

The dataset of hip radiographs has a total of 443 PNGs of which 234 are the normal cases and 209 are the osteoporosis cases as shown in <span id='cite-_Ref230176178'></span>[[#_Ref230176178|''Figure 1'']]. The data was systemically arranged so that it could be used to create an automated and intelligent system of evaluating bone health through hip X-ray pictures [36]. Instead of reducing the task to the simple classification, the proposed system is designed in such a way that it can assist in the more detailed assessment of hip bone condition. It does not only differentiate between the cases of Normal and Osteoporotic but helps to identify areas that might indicate the low bone mineral density [35], [37]. This combination method improves the accuracy of diagnosis and makes the system more clinically significant. <span id='cite-_Ref230176302'></span>[[#_Ref230176302|''Figure 2'']] also shows the proportional representation of the two classes with the Normals cases taking 52.8 and the Osteoporosis cases taking 47.2 percent of the dataset. Fair learning in both groups can be achieved by the near-balanced distribution, as well as risk of biasy models, and eventually lead to more stable and reliable performance in automated osteoporosis detection of hip radiographs.


[[Image:Draft_kaur_609399513-image1.png|center|600px]]

<div id="_Ref230176178" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">Figure 1: Collected Dataset</span></div>

[[Image:Draft_kaur_609399513-image2.png|600px|alt=' '| ]]
''' '''

<div id="_Ref230176302" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">Figure 2:Dataset Distribution per Class</span></div>

==3.1 Proposed Methodological ==

<span id='cite-_Ref233208888'></span>[[#_Ref233208888|'''Figure 3''']] represents the full procedural plan which has been designed to measure hip bone mineral density automatically. It starts with organized retrieval of hip radiographic images which are then accurately classified before further processing. All images are resized to 200 200 pixels so that they have the same dimensions and have the same level of compatibility with deep learning models. Normalization of pixel intensity is then used to normalize image data to enable the training of the stable model [38]. Adaptive histogram equalization is used to better the contrast and finer details of the trabecular bone in areas of critical anatomical interest, e.g. the neck and the femur head, to increase clarity [39]. Morphological gradient techniques are also used to refine structural edges and highlight bone edges which can be more easily extracted. In order to add strength and avoid overfitting, data augmentation methods such as rotation, horizontal flipping, zoom distortions, and mild brightness modifications are introduced to experience imaging diversity in the real world. The resulting processed data is further subdivided into training, validation, and testing sets with a 70:15:15 ratio in order to achieve balanced learning and unbiased evaluation of the obtained information. Lastly, state of the art transfer learning models and object detection architectures are developed and evaluated based on standard performance measures and the best and most trustworthy model is chosen to perform automated hip BMD classification.

[[Image:Draft_kaur_609399513-image3.jpg|600px|alt=' '| ]]
''' '''

<div id="_Ref233208888" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">'''Figure 3:''' ''Flowchart of proposed Work''</span></div>

==3.2 Resizing and Normalization Process==

The <span id='cite-_Ref233209010'></span>[[#_Ref233209010|'''Figure 4''']] shows the pre-processing processes of resizing and normalization applied on hip X-ray images that are utilized in the BMD analysis. First, resizing is used to adjust the size of pictures with different sizes which were initially of 224 x 224 and 331 x 331 pixels to a standard size of 200 x 200 pixels. This guarantees equal spatial measures in the dataset without any loss of the critical anatomical features of the hip area. Normalization follows resizing so as to standardize the pixel intensity values, e.g., to a common range, e.g. 01; or by modifying them relative to mean and standard deviation. Unlike resizing, normalization does not change the size of the image but increases the stability of numbers and learns the model better. These preprocessing steps combined form a standardized and optimized input format, which makes BMD prediction and classification more reliable and accurate [40], [41].

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:Draft_kaur_609399513-image4.png|600px]] </div>

<div id="_Ref233209010" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">'''Figure 4:''' '' Resizing and Normalization Process''</span></div>

====3.3 Image Enhancement Techniques====

====3.3.1 Contrast Limited Adaptive Histogram Equalization (CLAHE)====

Image enhancement methods were used to increase the quality of the hip radiographs in the BMD dataset and improve the clarity and quality of the diagnostic results before they were trained in models. Contrast Limited Adaptive Histogram Equalization (CLAHE) as illustrated in <span id='cite-_Ref233209090'></span>[[#_Ref233209090|'''Figure 5''']] was employed to enhance the local contrast of the images of hip X-rays. In contrast to the conventional global histogram equalization, CLAHE operates on the image in small contextual areas (tiles) allowing to better visualise the fine anatomical structures (e.g., the neck of the femur and the pattern of trabecular bone) which are important when the BMD is to be determined accurately [42], [43].

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:Draft_kaur_609399513-image5.png|600px]] </div>

<div id="_Ref233209090" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">''Figure 5: CLAHE Pre-Processing''</span></div>

===3.3.2 Morphological Gradient Technique===

The result of morphological gradient method on the hip radiographs in the BMD dataset is shown in <span id='cite-_Ref233209184'></span>[[#_Ref233209184|'''Figure 6''']]. The method increases the margins and contours of significant body parts in the hips. The morphological gradient can be used by calculating the difference between the eroded and dilated images of the image to highlight the outlines of objects like the femoral head, the femoral neck and the surrounding cortical bone [44], [45], [46]. Consequently, the structural outlines are clear and clearer. This enhanced edge representation leads to better identification of areas in which bone density changes can be and therefore better feature extraction and analysis towards BMD prediction and osteoporosis diagnosis.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:Draft_kaur_609399513-image6.png|600px]] </div>

<div id="_Ref233209184" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">''Figure 6: Morphological Gradient Process''</span></div>

====3.4 Data Augmentation====

The comparison between the class distributions in the hip BMD data under the condition of a set of data without data augmentation and the data augmentation with the data is shown in <span id='cite-_Ref233209327'></span>[[#_Ref233209327|'''Figure 7''']]. Any pre-augmentation analysis indicated that the dataset had a class imbalance; there were 234 images of normal hips and 209 images of osteoporosis. The difference between the two classes was not very significant, but the minor majority of the Normal one also may result in bias when training the model and, consequently, influence the classification.

Following the use of data augmentation methods, the dataset size has grown significantly and it has become balanced. The number of normal images had increased to 3,963 whereas Osteoporosis images had grown to 3,927. Such close balance between the two classes helps in providing a more balanced and objective training process. Due to this, the model also attains high stability, increased reliability and increased generalization of hip BMD prediction and osteoporosis classification.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:Draft_kaur_609399513-image7.png|468px]] </div>

<div id="_Ref233209327" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">''Figure 7:Data Augmentation Before and After Augmentation''</span></div>

==3.5 Dataset Splitting==

<span id='cite-_Ref233210331'></span>[[#_Ref233210331|'''Figure 8''']] The hip Bone Mineral Density data set was randomly used but evenly split into a training, a validation and a testing dataset in 70,15,15 percent proportion in order to have a balanced and reliable model development. Under the Normal group, 2,774 images were assigned to the training, 594 images to the validation and 595 images to the testing. On the same note, 2748 images were utilized in the Osteoporosis category, 589 images in validation, and 590 tests. This almost equal distribution of the two classes assists in terms of the balance of the classes across each subset.

The 70 percentage of training is large enough to offer adequate data to the model to learn significant features of changes in the hip bone density and structural integrity. With the exposure of the model to a variety of Normal and Osteoporotic hip images, the model is now able to detect fine details that serve to distinguish between healthy bone and the situation of reduced bone density. The 15 percent validation set is important in the training process since it allows monitoring of performance and tuning of hyperparameters which prevents overfitting and contributes to a higher level of stability of the model. Lastly, the 15 percent testing set is held very independent and is tested only after the training is finished in order to determine the actual generalization capability of the model to unknown hip BMD images.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:Draft_kaur_609399513-image8.png|600px]] </div>

<div id="_Ref233210331" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">''Figure 8:Dataset Splitting Class-wise''</span></div>

==3.6 Models Deployments==

===3.6.1 Transfer Learning Models===

In the <span id='cite-_Ref233212475'></span>[[#_Ref233212475|''Table 2'']], the highly-trained convolutional neural networks, i.e. VGG16, EfficientNetB0, Xception, InceptionResNetV2 and DenseNet201, were used to extract meaningful features based on hip bone mineral density (BMD) images. These architectures have been chosen wisely as they have demonstrated the ability to learn complex structural and textural patterns in medical imaging information. Using the transfer learning, the models were made to use representations obtained in the past to make feature extraction more efficient. Therefore, they were able to produce slight differences between normal and osteoporotic hip conditions resulting in better overall classification and model reliability.

<div id="_Ref233212475" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">Table 2: Performance Oriented Comparison of Transfer Learning Models for Osteoporosis detection</span></div>

{| style="width: 100%;border-collapse: collapse;" 
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Model'''</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Architecture'''</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Suitability for Medical Imaging'''</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Strengths'''</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Limitations / Computational Cost'''</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''VGG16 '''[47]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">16-layer CNN with 3×3 convolutions</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Widely adopted baseline for medical feature extraction.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Highly structured and well-defined; proven effectiveness in transfer learning.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Massive parameter count; high memory and computational overhead.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''EfficientNetB0 '''[48]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Compound scaling of depth, width, and resolution</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Ideal for resource-constrained hip BMD classification.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Highly optimized; excellent accuracy-to-parameter ratio.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">May require careful hyperparameter tuning for niche datasets.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''Xception '''[49]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Depthwise separable convolutions</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Excellent for identifying fine-grained structural differences in tissues.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Highly efficient; excellent at capturing intricate spatial patterns.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Can be sensitive to overfitting on small datasets without strong regularization.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''InceptionResNetV2 '''[50]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Inception modules with residual connections</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Capable of mapping highly complex medical image structures.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Extremely deep feature extraction; highly stable training via residual paths.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Extremely high computational demand and resource-intensive training.</span>
|-
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">'''DenseNet201 '''[51]</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Dense connections between layers</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Excellent for detecting fine-grained details in complex diagnostic tasks.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">Maximizes feature reuse; mitigates vanishing gradient problem.</span>
|  style="border: 1pt solid black;vertical-align: top;"|<span style="text-align: center; font-size: 75%;">High GPU memory consumption due to continuous feature concatenation.</span>
|}

===3.6.2 Object detection Models ===

The latest versions of object detectors such as YOLOv9-T, YOLOv9-S, YOLOv9-M, YOLOv9-C and YOLOv9-E were used in this study to detect hip bone mineral density images. These models have been chosen due to their high ability to localize and identify relevant parts in medical images and they have high computational efficiency. The models were effective in capturing significant structural attributes of the hip because of their optimized architectures and better feature extraction mechanisms. Consequently, they could accurately detect and distinguish the patterns of normal and osteoporotic states, therefore the detection performance and the system reliability were increased.

:* '''YOLOv9-T:''' YOLOv9-T is the smallest model in the YOLOv9 family and was created to be able to detect quickly with a minimum of hardware needed. It also has a lightweight structure that allows it to perform effectively in real time; particularly in systems with minimal processing capabilities. It is also less efficient in its speed but the accuracy of detection is relatively poor compared to the bigger models because of minimal network complexity [52].

:* '''YOLOv9-S (Small):''' This is the best computation-detection tradeoff in YOLOv9-S due to its smaller size. It has a relatively larger architecture than the Tiny version, and is able to fetch richer and more detailed features. This model has been applied to the issues that need to be stable, but with a decent processing speed [53].

:* '''YOLOv9-M (Medium):''' YOLOv9-M goes even further to add depth and capacity to the network to enhance feature learning and detection reliability. It provides greater accuracy compared with the versions Tiny and Small but is still able to work within viable computational constraints. It is the best version to use in processes that require a higher level of accuracy and average hardware capabilities [54].

:* '''YOLOv9-C (Compact/Custom):''' YOLOv9-C includes architectural improvements in improving the feature representation and detection accuracy. It is designed in such a way that it offers effective learning and simultaneously high levels of accuracy. This version strikes a good balance between the model complexity and computational requirement [55].

:* '''YOLOv9-E (Enhanced):''' YOLOv9-E is the latest and most sophisticated model in the YOLOv9 family, which is aimed at achieving the best level of detection. Its highly advanced and rich architecture allows it to extract features better and more effectively predict. This improved capability is however associated with greater computational needs as well as training time [56].

3.'''7 Performance Metrics'''

The execution of the postulated models in categorizing the hip bone mineral density level was evaluated by a set of statistical evaluation measures based on their respective confusion matrices. A confusion matrix is an organized Table that compares the actual class of the data with the model-generated prediction results giving detailed numbers of true positives (TP), true negatives (TN), false positives (FP) and false negatives (FN). Such fundamental elements are the basis of computing the various performance indicators. Through the examination of those values a number of quantitative data may be estimated to determine the accuracy, sensitivity, specificity and the general reliability of the models in the classification of hip Bone Mineral Density [57]. <span id='cite-_Ref233212561'></span>[[#_Ref233212561|'''Table 3''']] is a mathematical expression of these metrics of evaluation.

<div id="_Ref233212561" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">'''Table 3''':Performance Parameters</span></div>

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:Draft_kaur_609399513-image9.png|600px|alt=''''4.Results and Discussion''''|'''4. Results and Discussion''']] </div>

'''4. Results and Discussion'''

'''4.1 Performance Evaluation of Transfer Learning Models'''

<span id='cite-_Ref233212692'></span>[[#_Ref233212692|Table 4]] and <span id='cite-_Ref233210613'></span>[[#_Ref233210613|'''Figure 9''']] show that the five pre-trained models have a high and robust performance in the hip bone mineral density classification. EfficientNetB0 has the best overall results, as its accuracy is 94.1, the precision is 94.5, the recall is 93.8, and the F1-score is 94.1. Its MCC was 88.1% which means that it has a very balanced and strong predictive potential in both classes.

The DenseNet201 also gave quite competitive results as the accuracy of the model was 94.0, precision 93.7, recall 94.2 and the F1-score of the model was 93.9 and the MCC was 87.9. These values are roughly similar to those of the EfficientNetB0, which demonstrates its high capability to obtain meaningful patterns of hip BMD images. Xception obtained a stable and balanced classification performance, with an accuracy of 93.5, precision of 93.0, recall of 94.0, and of F1-score 93.5, and MCC of 86.7.

VGG16 achieved a result of 92.3, 91.8, 92.7 and 92.2 in terms of accuracy, precision, recall and F1-score respectively, with a MCC of 84.4 that is reliable but relatively moderate. InceptionResNetV2 achieved the lowest performance with an accuracy of 92.0, precision of 91.5, recalls of 92.3 and F1-score of 91.9 with a Matthews Correlation Coefficient of 84.0 indicating that it has a low predictive power despite being rated as satisfactory.

<div id="_Ref233212692" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">''Table 4:Performance Evaluation of TL Models''</span></div>

{| style="width: 100%;border-collapse: collapse;" 
|-
|  style="border: 1pt solid black;vertical-align: top;"|'''S.No'''
|  style="border: 1pt solid black;vertical-align: top;"|'''Model'''
|  style="border: 1pt solid black;vertical-align: top;"|'''Accuracy (%)'''
|  style="border: 1pt solid black;vertical-align: top;"|'''Precision (%)'''
|  style="border: 1pt solid black;vertical-align: top;"|'''Recall (%)'''
|  style="border: 1pt solid black;vertical-align: top;"|'''F1-Score (%)'''
|  style="border: 1pt solid black;vertical-align: top;"|'''MCC (%)'''
|-
|  style="border: 1pt solid black;vertical-align: top;"|1
|  style="border: 1pt solid black;vertical-align: top;"|VGG16
|  style="border: 1pt solid black;vertical-align: top;"|92.3
|  style="border: 1pt solid black;vertical-align: top;"|91.8
|  style="border: 1pt solid black;vertical-align: top;"|92.7
|  style="border: 1pt solid black;vertical-align: top;"|92.2
|  style="border: 1pt solid black;vertical-align: top;"|84.4
|-
|  style="border: 1pt solid black;vertical-align: top;"|2
|  style="border: 1pt solid black;vertical-align: top;"|EfficientNetB0
|  style="border: 1pt solid black;vertical-align: top;"|'''94.1'''
|  style="border: 1pt solid black;vertical-align: top;"|'''94.5'''
|  style="border: 1pt solid black;vertical-align: top;"|93.8
|  style="border: 1pt solid black;vertical-align: top;"|'''94.1'''
|  style="border: 1pt solid black;vertical-align: top;"|'''88.1'''
|-
|  style="border: 1pt solid black;vertical-align: top;"|3
|  style="border: 1pt solid black;vertical-align: top;"|Xception
|  style="border: 1pt solid black;vertical-align: top;"|93.5
|  style="border: 1pt solid black;vertical-align: top;"|93.0
|  style="border: 1pt solid black;vertical-align: top;"|94.0
|  style="border: 1pt solid black;vertical-align: top;"|93.5
|  style="border: 1pt solid black;vertical-align: top;"|86.7
|-
|  style="border: 1pt solid black;vertical-align: top;"|4
|  style="border: 1pt solid black;vertical-align: top;"|InceptionResNetV2
|  style="border: 1pt solid black;vertical-align: top;"|92.0
|  style="border: 1pt solid black;vertical-align: top;"|91.5
|  style="border: 1pt solid black;vertical-align: top;"|92.3
|  style="border: 1pt solid black;vertical-align: top;"|91.9
|  style="border: 1pt solid black;vertical-align: top;"|84.0
|-
|  style="border: 1pt solid black;vertical-align: top;"|5
|  style="border: 1pt solid black;vertical-align: top;"|DenseNet201
|  style="border: 1pt solid black;vertical-align: top;"|94.0
|  style="border: 1pt solid black;vertical-align: top;"|93.7
|  style="border: 1pt solid black;vertical-align: top;"|'''94.2'''
|  style="border: 1pt solid black;vertical-align: top;"|93.9
|  style="border: 1pt solid black;vertical-align: top;"|87.9
|}


<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:can't find package drawp|600px]] </div>

<div id="_Ref233210613" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">'''Figure 9''':Performance Parameters of TL Models</span></div>

==4.2 Quantitative indicators demonstrating the performance of the proposed models in hip BMD estimation==

The baseline features of the hip bone mineral density data in the training, validation and testing sets also indicate consistency as well as some significant differences as evidenced in table 4. In the training set, there were 5,522 patients, 1,183 in the validation set and 1,185 in testing set. The percentage of women in the three groups were 79.1 (4,370) in the training set, 79.1 (936) in the validation set and 77.6 (920) in the testing set with no statistically significant difference observed (p = 0.18).

The overall age of the participants in both the training and the validation group was 68.1/12.2 years and, in the testing, group was 72.4/11.1 years, and the difference is statistically significant (p < 0.001). The median Dual energy X-ray Absorptiometry and radiograph time interval was 22 days (IQR: 18-30) in training and validation group, and 25 days (IQR: 8-70) in testing group, which also demonstrates a great variation (p < 0.001).

Regarding the body mass index, the average Body Mass Index of the training and validation sets was 24.7 + 3.7 kg/m 2, whereas the average Body Mass Index of the testing set was 24.0 + 3.8 kg/m 2 (p = 0.02). The average hip Bone Mineral Density of the training and validation groups was 0.774 + 0.168 g/cm 2 which compared to the testing set was lower and had 0.719 + 0.150 g/cm 2 (p < 0.001). On the same note, the median T-score was -2.2 (IQR: -3.2 -1.3) in training and validation sets but was -1.4 (IQR: -2.3 -0.5) in testing set indicating statistically significant difference (p < 0.001).

In the cases of disease distribution, 43.9% of participants in both training (2,425 replications) and validation (519 replications) set had osteoporosis, whereas 20.8% (246 replications) of the testing set had the same (p < 0.001). In general, demographic factors such as sex distribution did not differ between the datasets but a variety of clinical parameters such as age, Body Mass Density, T-score, and prevalence of osteoporosis were significantly different, though it can affect model evaluation and generalization in prediction of hip Body Mass Density.

<div id="_Ref233213536" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">Table 5:Quantitative indicators demonstrating the performance of the proposed models in hip BMD estimation</span></div>

{| style="width: 100%;border-collapse: collapse;" 
|-
|  style="border: 1pt solid black;"|'''Variable'''
|  style="border: 1pt solid black;"|'''Training Set (n = 5,522)'''
|  style="border: 1pt solid black;"|'''Validation Set (n = 1,183)'''
|  style="border: 1pt solid black;"|'''Testing Set (n = 1,185)'''
|  style="border: 1pt solid black;"|'''p-value'''
|-
|  style="border: 1pt solid black;"|Number
|  style="border: 1pt solid black;"|5,522
|  style="border: 1pt solid black;"|1,183
|  style="border: 1pt solid black;"|1,185
|  style="border: 1pt solid black;"|—
|-
|  style="border: 1pt solid black;"|Female, n (%)
|  style="border: 1pt solid black;"|4,370 (79.1%)
|  style="border: 1pt solid black;"|936 (79.1%)
|  style="border: 1pt solid black;"|920 (77.6%)
|  style="border: 1pt solid black;"|0.18
|-
|  style="border: 1pt solid black;"|Mean age (SD), years
|  style="border: 1pt solid black;"|68.1 (12.2)
|  style="border: 1pt solid black;"|68.1 (12.2)
|  style="border: 1pt solid black;"|72.4 (11.1)
|  style="border: 1pt solid black;"|<0.001
|-
|  style="border: 1pt solid black;"|Median time (IQR) between DXA and radiographs
|  style="border: 1pt solid black;"|22 (18, 30)
|  style="border: 1pt solid black;"|22 (18, 30)
|  style="border: 1pt solid black;"|25 (8, 70)
|  style="border: 1pt solid black;"|<0.001
|-
|  style="border: 1pt solid black;"|Mean BMI (SD), kg/m²
|  style="border: 1pt solid black;"|24.7 (3.7)
|  style="border: 1pt solid black;"|24.7 (3.7)
|  style="border: 1pt solid black;"|24.0 (3.8)
|  style="border: 1pt solid black;"|0.02
|-
|  style="border: 1pt solid black;"|Mean BMD (SD), g/cm²
|  style="border: 1pt solid black;"|0.774 (0.168)
|  style="border: 1pt solid black;"|0.774 (0.168)
|  style="border: 1pt solid black;"|0.719 (0.150)
|  style="border: 1pt solid black;"|<0.001
|-
|  style="border: 1pt solid black;"|Median T-score (IQR)
|  style="border: 1pt solid black;"|−2.2 (−3.2, −1.3)
|  style="border: 1pt solid black;"|−2.2 (−3.2, −1.3)
|  style="border: 1pt solid black;"|−1.4 (−2.3, −0.5)
|  style="border: 1pt solid black;"|<0.001
|-
|  style="border: 1pt solid black;"|Osteoporosis, n (%)
|  style="border: 1pt solid black;"|2,425 (43.9%)
|  style="border: 1pt solid black;"|519 (43.9%)
|  style="border: 1pt solid black;"|246 (20.8%)
|  style="border: 1pt solid black;"|<0.001
|}

==4.3 Object Detection Models==

The <span id='cite-_Ref233213779'></span>[[#_Ref233213779|''Table 6'']] and <span id='cite-_Ref233210723'></span>[[#_Ref233210723|'''Figure 10''']] shows the results of various YOLOv9 models, such as the YOLOv9-T, and the YOLOv9-E, to detect hip bone conditions and estimate bone mineral density to determine osteoporosis. Accuracy, precision, recall, F1-score and Matthews Correlation Coefficient are used to evaluate the models.

The baseline in this case was YOLOv9-T which had an accuracy of 91.62 percent, meaning that the model was able to identify hip X-rays or Bone Mineral Density-based cases of osteoporosis in 92 out of 100 cases. The fact that its precision of 91.08 percent shows that most of its positive prediction was accurate, with not many false alarms and the recall of 90.75 percent shows that the model was able to identify most of the patients who had low Bone Mineral Density or changes due to osteoporosis. As a result, the F1-score 90.91 percent is a balanced result and the MCC of 84.6 percent indicates that the overall correlation between the predicted and actual labels is strong even when there are class imbalances.

The YOLOv9-S model demonstrates better results in all measures. The accuracy increases to 93.48 percent, whereas the precision and the recall are 93.02 percent and 92.66 percent respectively, and the F1-score is 92.84 percent. These enhancements show that YOLOv9-S is more effective in detecting osteopenia and osteoporosis in hip X-rays with minimum errors. Its predictive consistency is also confirmed by an MCC of 88.4 percent.

Even better is the YOLOv9-M that has an accuracy of 94.96 percent. The precision and recall stand at 94.55 percent and 94.12 percent respectively, indicating that the model is efficient in identifying normal and low Bone Mineral Density hip with a few errors. The F1-score of 94.33 percent indicates its balanced performance to detect true positives and minimize false positives and the MCC of 91.4 percent validates that there is a high level of agreement between the predicted and actual diagnoses.

Even better results are obtained with the YOLOv9-C model with 96.18 percent accuracy. The resultant precision at 95.84 percent and 95.41 percent recall gives an F1-score of 95.62 percent, a perfect indication of osteoporotic changes in hip Bone Mineral Density images. Its MCC of 93.6 percent exhibits the ability to make reliable predictions of various cases of patients.

Lastly, YOLOv9-E has the best performance of all the models. The accuracy stands at 97.02 percent indicating that almost all the hip Bone Mineral Density cases are correctly identified. Recall and precision are 96.71 percent and 96.35 percent, respectively, and it can be concluded that the model is quite efficient in recognition of the true osteoporotic cases and low false positives. The F1-score of 96.53 percent proves an ideal compromise between the precision and the recall, and the MCC of 95.6 percent provides the very high correlation between the allegations and the actual Bone Mineral Density based diagnosis.

<div id="_Ref233213779" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">Table 6:Performance Evaluation of Object Detection Models</span></div>

{| style="width: 100%;margin: 1em auto 0.1em auto;border-collapse: collapse;" 
|-
|  style="border: 1pt solid black;vertical-align: top;"|'''Model'''
|  style="border: 1pt solid black;vertical-align: top;"|'''Accuracy (%)'''
|  style="border: 1pt solid black;vertical-align: top;"|'''Precision (%)'''
|  style="border: 1pt solid black;vertical-align: top;"|'''Recall (%)'''
|  style="border: 1pt solid black;vertical-align: top;"|'''F1-Score (%)'''
|  style="border: 1pt solid black;vertical-align: top;"|'''MCC (%)'''
|-
|  style="border: 1pt solid black;vertical-align: top;"|YOLOv9-T
|  style="border: 1pt solid black;vertical-align: top;"|91.62
|  style="border: 1pt solid black;vertical-align: top;"|91.08
|  style="border: 1pt solid black;vertical-align: top;"|90.75
|  style="border: 1pt solid black;vertical-align: top;"|90.91
|  style="border: 1pt solid black;vertical-align: top;"|84.6
|-
|  style="border: 1pt solid black;vertical-align: top;"|YOLOv9-S
|  style="border: 1pt solid black;vertical-align: top;"|93.48
|  style="border: 1pt solid black;vertical-align: top;"|93.02
|  style="border: 1pt solid black;vertical-align: top;"|92.66
|  style="border: 1pt solid black;vertical-align: top;"|92.84
|  style="border: 1pt solid black;vertical-align: top;"|88.4
|-
|  style="border: 1pt solid black;vertical-align: top;"|YOLOv9-M
|  style="border: 1pt solid black;vertical-align: top;"|94.96
|  style="border: 1pt solid black;vertical-align: top;"|94.55
|  style="border: 1pt solid black;vertical-align: top;"|94.12
|  style="border: 1pt solid black;vertical-align: top;"|94.33
|  style="border: 1pt solid black;vertical-align: top;"|91.4
|-
|  style="border: 1pt solid black;vertical-align: top;"|YOLOv9-C
|  style="border: 1pt solid black;vertical-align: top;"|96.18
|  style="border: 1pt solid black;vertical-align: top;"|95.84
|  style="border: 1pt solid black;vertical-align: top;"|95.41
|  style="border: 1pt solid black;vertical-align: top;"|95.62
|  style="border: 1pt solid black;vertical-align: top;"|93.6
|-
|  style="border: 1pt solid black;vertical-align: top;"|YOLOv9-E
|  style="border: 1pt solid black;vertical-align: top;"|'''97.02'''
|  style="border: 1pt solid black;vertical-align: top;"|'''96.71'''
|  style="border: 1pt solid black;vertical-align: top;"|'''96.35'''
|  style="border: 1pt solid black;vertical-align: top;"|'''96.53'''
|  style="border: 1pt solid black;vertical-align: top;"|'''95.6'''
|}


<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:can't find package drawp|600px]] </div>

<div id="_Ref233210723" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">'''Figure 10''':Performance Evaluation of Object Detection Models</span></div>

===4.3.1 Confusion Matrix of Proposed Models===

<span id='cite-_Ref233210871'></span>[[#_Ref233210871|'''Figure 11''']] is a confusion matrix of Hip Bone Mineral Density osteoporosis classification model shows a high degree of accuracy in diagnosis between osteoporotic and normal users. Out of the actually normal patients, 2670 of them were correctly identified as normal (true negatives), whereas 91 of them were erroneously diagnosed with osteoporotic (false positives). This shows that this model is very good in identifying healthy cases with few unnecessary positive predictions. Even though false positives can lead to further clinical tests, they are not as serious as the failure to diagnose the actual osteoporosis cases. The model has a high capacity to identify low bone mineral density in the hip region, as 2659 cases were correctly identified as osteoporosis (true positives) in relation to osteoporotic patients. A total of 102 osteoporotic patients were however erroneously forecasted as normal (false negatives). These lost cases are of clinical significance as untreated osteoporosis may predispose hip fractures and other associated complications. Nevertheless, the relatively low false negative per 100000 true positive is indicative of high sensitivity and equal classification.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
 [[Image:Draft_kaur_609399513-image10.png|600px]] </div>

<div id="_Ref233210871" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">'''Figure 11''':Confusion Matrix for Proposed Model</span></div>

===4.3.2 Training and Loss Curve for Proposed Model===

<span id='cite-_Ref233211000'></span>[[#_Ref233211000|'''Figure 12''']] shows the training performance of the YOLOv9-E model after 50 iteration cycles of the hip bone mineral density (BMD) classification problem, subfigures (a), (b), (c) and (d) display various measures of evaluation, which can be used altogether to describe the learning process and stability of the model.

Starting with Figure 12(a), the curve of training loss gradually goes down with the increase in the epochs. This decreasing tendency is a sign that this model is gradually decreasing the errors of prediction on the training set. The continuous and gradual decrease signifies the consistency of optimization and good adaptation of model parameters as the model is learning. To supplement this, Figure 12(b) depicts the validation loss curve which also gradually falls and stabilizes at a low value towards the later stages of training. This same pattern in both training and validation loss curves indicates that the model can be generalized well to unseen data and that there is no major overfitting happening during the process of learning the hip BMD patterns.

Following the analysis of the loss, Figure 12(c) shows the accuracy curves of the training and validation dataset. The less the loss, the more is the accuracy with the epochs and finally the plateau will be high and stable. This negative correlation between loss and accuracy proves that the YOLOv9-E model becomes more efficient at the ability to classify hip BMD conditions correctly with each training session. Furthermore, the high compatibility between training and validation accuracy also proves the strength and high level of reliability of the predictive abilities of the model.

At last, Figure 12(d) provides the top-5 accuracy measure, which provides a more detailed analysis of classification results. This measure is based on a prediction being correct provided a prediction that is made that belongs to the correct hip BMD classification coming within the top five confident predictions made by the model. This larger evaluation is especially beneficial in medical image activities, in which subtle visual resemblance between categories of hip BMD can add to the complexity of classification. A large top-5 accuracy means that the model is not only making correct predictions, but also it is ranking correctly the most likely category among its own predictions.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
''' [[Image:Draft_kaur_609399513-image11.png|468px]] '''</div>

<div id="_Ref233211000" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">'''Figure 12''':a–d Accuracy and Loss curves of the YOLOv9-E Model</span></div>

The <span id='cite-_Ref233211151'></span>[[#_Ref233211151|'''Figure 13''']] displays the automated hip osteoporosis detection based on a hip radiographic using the proposed YOLOv9-E object detection model. The picture has colour-enhanced picture of hip joint excluding the femur head, neck, proximal neck and shaft of the femur. The main area of interest (ROI) that is identified by the proposed YOLOv9-E model is the large yellow rectangular bounding box. This area is the proximal femur, which is the normal anatomical location of hip bone mineral density (BMD) during clinical DXA-based evaluation. The localization of this vital anatomical structure is done accurately by the model where the model predicts that the further analysis is confined to the site of clinical interest to assess osteoporosis which minimizes background noise and influence of irrelevant features.

The primary ROI includes a smaller bounding box marked with Osteoporosis to show where exactly the YOLOv9-E model identifies the osteoporotic changes. YOLOv9-E is an object detector more based on the leverage of larger feature accumulation, better gradient flow, and enhanced detection heads, making it more accurate in localization and classification even in the medical image with subtle structural changes compared to generic object detectors. The model classifies the osteoporosis diagnosis in accordance with the acquired radiographic patterns that are linked to the biomineral density loss.

<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">
''' [[Image:Draft_kaur_609399513-image12.jpeg|306px]] '''</div>

<div id="_Ref233211151" class="center" style="width: auto; margin-left: auto; margin-right: auto;">
<span style="text-align: center; font-size: 75%;">'''Figure 13''':Proposed Model Detecting Osteoporosis Region</span></div>

=5. Conclusion and Future scope=

This study presented an automated osteoporosis screening and hip bone mineral density estimation using routine hip radiographs. The transfer learning models achieved the highest accuracy of 94.1% with efficientNetB0 when evaluated experimentally. But comparison analysis shows that YOLOv9 Object Detection Model performs well over the traditional transfer learning models. YOLOv9-E performs excellently with the highest accuracy of 97.02%, precision 96.7%, recall 96.35%, F1 Score 96.53% and Matthews Correlation Coefficient.

The osteoporosis related structural changes in hip radiographs can be effectively identified after achieving the better result from advanced Object Detection Models. For contribution to robust model generalization and improved diagnostic performance, the proposed pre-processing pipeline, adds the CLAHE enhancement, morphological gradient filtering and data augmentation.

Further, investigation lays emphasis on the multicentre validation across larger and more heterogeneous population and integration of longitudinal clinical records to improve predictive modelling. The early osteoporosis prediction may be improved by exploring the imaging biomarkers with biomechanical analysis, fracture risk assessment and explainable artificial intelligence techniques which leads to the development of clinical decision support systems that improve diagnostic accuracy.

==Declarations==

'''Conflict of interest:''' The author’s Sandeep Kaur and Jaspreet Singh declared no possible conflicts of interest related to research, authorship and/or publication of this article.

'''Funding:''' Author’s Sandeep Kaur and Jaspreet Singh have not received any financial support   for research, and publication of this article.

'''Author's contribution:''' Sandeep Kaur performed all the experiments and prepared the results, Jaspreet Singh supervised the preparation and development of the manuscript.

'''Availability of data and material:''' Han-Wei Zhang. (2022), ”Automated Osteoporosis Classification and T-score Prediction Using Hip Radiographs via Deep Learning Algorithm”  Mendeley. [https://doi.org/10.17632/GMG3VVMVJ4 https://doi.org/10.17632/GMG3VVMVJ4].

'''Code Availability:''' Code used is tailor-made by the author.

'''Ethical approval:''' This article does not have any studies with human and animal performed by the authors.

'''Human Ethics and Consent to Participate statements:''' Not applicable.

==References==

{| style="width: 100%;" 
|-
|  style="vertical-align: top;"|[1] 
|  style="vertical-align: top;"|J. O. and A. K. J., “An estimate of the worldwide prevalence and disability associated with osteoporotic fractures,” ''Osteoporosis International, ''vol. 17, no. 12, pp. 1726-1733, Sept. 2006. 
|-
|  style="vertical-align: top;"|[2] 
|  style="vertical-align: top;"|G. B., J. O. and А. K. J., “World-wide Projections for Hip Fracture,” ''Osteoporosis International, ''vol. 7, no. 5, pp. 407-413, Sept. 1997. 
|-
|  style="vertical-align: top;"|[3] 
|  style="vertical-align: top;"|H. P., M. J, C.-E. CS, V. D., M. K, V. B and B. S, “Meta-analysis: Excess Mortality After Hip Fracture Among Older Women and Men,” ''Annals of Internal Medicine, ''vol. 152, no. 6, pp. 380-390, March 2010. 
|-
|  style="vertical-align: top;"|[4] 
|  style="vertical-align: top;"|S. E. K. T. H. J. K. J. Magaziner J, “Survival experience of aged hip fracture patients,” ''American Journal of Public Health, ''vol. 80, no. 3, pp. 279-283, 1990. 
|-
|  style="vertical-align: top;"|[5] 
|  style="vertical-align: top;"|M. D, A. K, B. DE, B. ER, B. S, B. ML, M. AR, M. PJ, S. JE, W. DA and I. C. F. W. group, “Coordinator-based systems for secondary prevention in fragility fracture patients,” ''Osteoporosis International, ''vol. 22, no. 7, pp. 2051-2065, Jul 2011. 
|-
|  style="vertical-align: top;"|[6] 
|  style="vertical-align: top;"|K. JA, J. O, O. A, S. I, R.-J. I, D. A, D. L. C and J. B, “Long-term risk of osteoporotic fracture in Malmö,” ''Osteoporosis International, ''vol. 11, no. 8, pp. 669-674, 2000. 
|-
|  style="vertical-align: top;"|[7] 
|  style="vertical-align: top;"|G. D, S. HT, B. VE, S. B, G. WA, G. ER, L. A, S. B, H. AJ and S. RB, “Low Rates of Bone Health Evaluation Before and After Primary Fragility Fractures,” ''The Journal of Bone and Joint Surgery, ''vol. 10, no. 4, 14 Oct 2025. 
|-
|  style="vertical-align: top;"|[8] 
|  style="vertical-align: top;"|B. D, N. ND, M. VE, N. TV, E. JA and C. JR, “Mortality risk associated with low-trauma osteoporotic fracture and subsequent fracture in men and women.,” ''Journal of the American Medical Association, ''vol. 301, no. 5, pp. 513-521, 4 Feb 2009. 
|-
|  style="vertical-align: top;"|[9] 
|  style="vertical-align: top;"|P. RJ, C. JJ, B. W, K. CJ, T. C, S. N and B. JM, “Opportunistic assessment of osteoporosis using hip and pelvic X-rays with OsteoSight™: validation of an AI-based tool in a US population,” ''Osteoporosis International, ''vol. 36, no. 6, pp. 1053-1060, Jun 2025. 
|-
|  style="vertical-align: top;"|[10] 
|  style="vertical-align: top;"|P. PJ, P. BD, L. T, R. d. AM, B. RJ and B. N, “Opportunistic screening for osteoporosis using abdominal computed tomography scans obtained for other indications.,” ''Annals of internal medicine, ''vol. 158, no. 8, pp. 588-595, Apr 2013. 
|-
|  style="vertical-align: top;"|[11] 
|  style="vertical-align: top;"|L. L, K. S, X. Y, Z. C, Z. Y, H. W, R. Y and Z. L, “Predictive value of a prediction system based on bone mineral density and proximal femoral imaging indexes for hip fracture risk in osteoporotic patients.,” ''American Journal of Translation Research, ''vol. 18, no. 2, pp. 1278-1286, 15 Feb 2026. 
|-
|  style="vertical-align: top;"|[12] 
|  style="vertical-align: top;"|G. Chang, L. Wang, G. Liang, J. Babb, G. Wiggins, P. Saha and R. Regatte, “Quantitative assessment of trabecular bone micro-architecture of the wrist via 7 Tesla MRI: preliminary results,” ''Magnetic Resonance Materials in Physiscs, Biology and Medicine, ''vol. 24, no. 2, pp. 191-199, Aug 2011. 
|-
|  style="vertical-align: top;"|[13] 
|  style="vertical-align: top;"|Y. N, S. S, K. A, G. R, N. T, N. K, T. K, K. H, N. H, K. K, F. Y and O. T, “Deep Learning for Osteoporosis Classification Using Hip Radiographs and Patient Clinical Covariates,” ''Biomolecules, ''vol. 10, no. 11, 2020. 
|-
|  style="vertical-align: top;"|[14] 
|  style="vertical-align: top;"|H. CS, C. YP, F. TY, K. CF, Y. TY, L. YC and P. YC, “Application of deep learning neural network in predicting bone mineral density from plain X-ray radiography,” ''Archives of Osteoporosis, ''vol. 16, no. 1, Oct 2021. 
|-
|  style="vertical-align: top;"|[15] 
|  style="vertical-align: top;"|L. G, K. T, B. BE, S. AAA, C. F, G. M, v. d. L. JAWM, v. G. B and S. CI, “A survey on deep learning in medical image analysis,” ''Medical Image Analysis, ''vol. 42, pp. 60-88, July 2017. 
|-
|  style="vertical-align: top;"|[16] 
|  style="vertical-align: top;"|D. Shen, G. Wu and H.-I. Suk, “Deep Learning in Medical Image Analysis,” ''Annual review of biomedical engineering, ''vol. 21, pp. 221-248, Jun 2017. 
|-
|  style="vertical-align: top;"|[17] 
|  style="vertical-align: top;"|R. Yamashita, M. Nishio, R. Do and K. Togashi, “Convolutional neural networks: an overview and application in radiology,” ''Insights Imaging, ''vol. 9, no. 4, pp. 611-629, Aug 2018. 
|-
|  style="vertical-align: top;"|[18] 
|  style="vertical-align: top;"|H. ,. K. Pishtiwan, S. b. Sebastian, P. b. Sergey, W. b. Guangyao, J. b. Arthur, R. b. Turkey, I. b. Abdalla, v. H. a. Luca, L. b. Philippe and P. a. Martijn, “Deep learning in fracture detection: a narrative review,” ''Acta Orthopaedica, ''vol. 91, no. 2, pp. 215-220, Jan 2020. 
|-
|  style="vertical-align: top;"|[19] 
|  style="vertical-align: top;"|T. Vokes and D. Gillen, “Using clinical risk factors and bone mineral density to determine who among patients undergoing bone densitometry should have vertebral fracture assessment,” ''Osteoporosis International, ''vol. 21, no. 12, pp. 2083-2091, Dec 2010. 
|-
|  style="vertical-align: top;"|[20] 
|  style="vertical-align: top;"|A. R, A. H. M, R. S and B. D, “CT determination of bone mineral density and structural investigations on the axial skeleton for estimating the osteoporosis-related fracture risk by means of a risk score,” ''The British Journal of Radiology, ''vol. 72, no. 858, pp. 569-578, Jun 1999. 
|-
|  style="vertical-align: top;"|[21] 
|  style="vertical-align: top;"|G. Keyvan, A. S. Catriona, B. Alexander, Z. Saba, V. Maksims, P. Tomi and D. C. Mark, “Approximating femoral neck bone mineral density from hand, knee, and pelvis X-rays using deep learning,” ''journal of medical artificial intelligence, ''vol. 6, june 2023. 
|-
|  style="vertical-align: top;"|[22] 
|  style="vertical-align: top;"|K. Jeong-Woon, P. Chunsu, L. Dong-Eon, Y. Jae-Heung and K. MinWoo, “Prediction of bone mineral density in CT using deep learning with explainability,” ''Frontiers in Physiology, ''vol. 13, 2022. 
|-
|  style="vertical-align: top;"|[23] 
|  style="vertical-align: top;"|B. Erickson, P. Korfiatis, Z. Akkus and T. Kline, “Machine Learning for Medical Imaging,” ''Radiographics, ''vol. 37, no. 2, pp. 505-515, Mar-April 2017. 
|-
|  style="vertical-align: top;"|[24] 
|  style="vertical-align: top;"|J. Acosta-Batlle, D. Coronado-Gutiérrez, J. Soto, J. Moujir, C. López, B. C. Suevos, D. M. Vázquez, D. M. Orduña, D. Fernández and S. J. Blázquez, “An AI-based pipeline for osteoporosis/osteopenia prediction using hip radiographs,” ''Skeletal Radiology, ''vol. 55, no. 2, pp. 415-424, Feb 2025. 
|-
|  style="vertical-align: top;"|[25] 
|  style="vertical-align: top;"|K. Yasaka, H. Akai, A. Kunimatsu, S. Kiryu and O. Abe, “Deep learning with convolutional neural network in radiology,” ''Japnese journal of Radiology, ''vol. 36, no. 4, pp. 257-272, April 2018. 
|-
|  style="vertical-align: top;"|[26] 
|  style="vertical-align: top;"|S. Wang, X. Tong, Q. Cheng, Q. Xiao, J. Cui, J. Li, Y. Liu and X. Fang, “Fully automated deep learning system for osteoporosis screening using chest computed tomography images,” ''Quantitative Imaging in Medicine and Surgery, ''vol. 14, no. 4, pp. 2816-2827, March 2024. 
|-
|  style="vertical-align: top;"|[27] 
|  style="vertical-align: top;"|O. Wilson, W. L. Ren, M. Andrew, Z. L. Xi, J. S. Weizhong, H. T. Jiong, K. Naresh and T. P. D. H. James, “Artificial Intelligence Applications for Osteoporosis Classification Using Computed Tomography,” ''Bioengineering, ''vol. 10, no. 12, 27 Nov 2023. 
|-
|  style="vertical-align: top;"|[28] 
|  style="vertical-align: top;"|S. S. A, M. Straus Takahashi, J. d. M. F. EM, S. Bennett, N. Khandwala, M. Major, J. Paparian, Garozzo, F. Velloni, S. Rothenberg, B. L. Kayat, R. Filice, J. Mongan and K. F. Campos, “Advancing osteoporosis opportunistic screening: multicenter validation of a deep learning algorithm using abdominal CT scans,” ''Abdominal Radiology (New York), ''vol. 51, no. 5, pp. 2631-2641, May 2026. 
|-
|  style="vertical-align: top;"|[29] 
|  style="vertical-align: top;"|L. M. Wu, B. B. Wang, B. M. Lin, M. M. Li, Y. M. Wu, H. M. Si, Y. M. Zeng, L. M. Lu, L. M. Gao, Z. M. Chen, R. M. Yu, L. P. Zhao, Y. P. Nie, K. P. Li and B. M. Shen, “A Deep Learning-Based Clinical Classification System for the Differential Diagnosis of Hip Prosthesis Failures Using Radiographs A Multicenter Study,” ''The Journal of Bone and Joint Surgery, ''vol. 107, no. 16, pp. 1798 - 1809, 20 August 2025. 
|-
|  style="vertical-align: top;"|[30] 
|  style="vertical-align: top;"|X. Fei, L. Rong, L. Ming, Z. Zongke, X. Zhou and D. Xin, “A New Random Forest Algorithm-Based Prediction Model of Post-operative Mortality in Geriatric Patients With Hip Fractures,” ''Frontiers in Medicine, ''vol. 9, 11 May 2022. 
|-
|  style="vertical-align: top;"|[31] 
|  style="vertical-align: top;"|B. S. S. Dias, R. Querrer, P. T. Figueiredo, A. F. Leite, N. S. d. Melo, L. R. Costa, M. F. Caetano and M. C. Farias, “Osteoporosis screening: Leveraging EfficientNet with complete and cropped facial panoramic radiography imaging,” ''Biomedical Signal Processing and Control, ''vol. 100, 2025. 
|-
|  style="vertical-align: top;"|[32] 
|  style="vertical-align: top;"|N. Vedant, K. Virja and S. Sakshi, “A COMPARATIVE ANALYSIS OF DEEP LEARNING MODELS AND CONVENTIONAL APPROACHES FOR OSTEOPOROSIS DETECTION IN HIP X- RAY IMAGES,” in ''Futuristic Trends in Artificial Intelligence V3B3'', vol. 3, Iterartive International Publishers IIP, 2023, pp. 23-30.
|-
|  style="vertical-align: top;"|[33] 
|  style="vertical-align: top;"|I. Muhammad, S. Routhu and B. Lee, “BONE-Net: A novel hybrid deep-learning model for effective osteoporosis detection,” ''PLOS One, ''vol. 20, no. 10, October 2025. 
|-
|  style="vertical-align: top;"|[34] 
|  style="vertical-align: top;"|K. V, N. V, B. V and C. S, “A Comparative Analysis of Deep Learning Models and Conventional Approaches for Osteoporosis Detection in Hip X-Ray Images,” in ''2023 World Conference on Communication & Computing (WCONF)'', RAIPUR, 2023. 
|-
|  style="vertical-align: top;"|[35] 
|  style="vertical-align: top;"|A. Yilmaz, K. Gem, M. Kalebasi, R. Varol, Z. Gencoglan, Y. Samoylenko, H. Tosyali, G. Okcu and H. Uvet, “An automated hip fracture detection, classification system on pelvic radiographs and comparison with 35 clinicians,” ''Scientific Reports, ''vol. 15, no. 1, 8 May 2025. 
|-
|  style="vertical-align: top;"|[36] 
|  style="vertical-align: top;"|h.-W. Zhang, ''Automated osteoporosis classification and T-score prediction using Hip radiographs via deep learning algorithm, ''Mendeley data, 2022. 
|-
|  style="vertical-align: top;"|[37] 
|  style="vertical-align: top;"|H. O. El-Sisi, F. E.-Z. A. El-Gamal and N. A. Hikal, “A non-invasive computer-aided personalized diagnosis system for Osteopenia and Osteoporosis,” ''Egyptian Informatics Journal, ''vol. 29, 2025. 
|-
|  style="vertical-align: top;"|[38] 
|  style="vertical-align: top;"|“Normalization (image processing),” Grokipedia, 2026.
|-
|  style="vertical-align: top;"|[39] 
|  style="vertical-align: top;"|B. EM, C. P, B. K, B. P and M. CR, “Adaptive histogram equalization in digital radiography of destructive skeletal lesions,” ''Radiology, ''vol. 166, no. 3, 1988. 
|-
|  style="vertical-align: top;"|[40] 
|  style="vertical-align: top;"|T. P. Nguyen, D.-S. Chae, S.-J. Park and J. Yoon, “A novel approach for evaluating bone mineral density of hips based on Sobel gradient-based map of radiographs utilizing convolutional neural network,” ''Computers in Biology and Medicine, ''vol. 132, 2021. 
|-
|  style="vertical-align: top;"|[41] 
|  style="vertical-align: top;"|R. Jang, J. H. Choi, N. Kim, J. S. Chang, P. W. Yoon and C.-H. Kim, “Prediction of osteoporosis from simple hip radiography using deep learning algorithm,” ''Scientific reports, ''vol. 11, Oct 2021. 
|-
|  style="vertical-align: top;"|[42] 
|  style="vertical-align: top;"|F. Muttaqin, Jamari, R. R. Isnanto, T. I. Winarni and A. P. Bayuseno, “Application of Contrast Enhancement Method on Hip X-ray Images as a Media for Detecting Hip Osteoarthritis,” ''International Journal of Advanced Computer Science and Applications, ''vol. 15, no. 11, pp. 760-765, 2024. 
|-
|  style="vertical-align: top;"|[43] 
|  style="vertical-align: top;"|W. Huang, P. Pan, K. Liu, Y. Xu, S. Cheng, H. Wang, L. Han, Y. Qi, L. Ren and J. Chu, “Intelligent identification of osteoporosis on hip X-rays using vision transformer,” ''Bone Reports, ''vol. 28, 16 Jan 2026. 
|-
|  style="vertical-align: top;"|[44] 
|  style="vertical-align: top;"|A.-B. J, C.-G. D, S. J, M. J, L. C, S. B. C, V. D. M, O. D. MDP, F. D and B. S. J, “An AI-based pipeline for osteoporosis/osteopenia prediction using hip radiographs,” ''Skeletal Radiology, ''vol. 55, no. 2, pp. 415-424, Feb 2026. 
|-
|  style="vertical-align: top;"|[45] 
|  style="vertical-align: top;"|L. Zhou, T. Nguyen, S. Choi and J. Yoon, “U-Net-Based Deep Learning Hybrid Model: Research and Evaluation for Precise Prediction of Spinal Bone Density on Abdominal Radiographs,” ''Bioengineering, ''vol. 12, no. 4, 3 April 2025. 
|-
|  style="vertical-align: top;"|[46] 
|  style="vertical-align: top;"|M. Brandl, A. Jahnke, C. Fölsch, M. Rickert and B. Ishaque, “Mid- to long-term periprosthetic bone density changes after cementless short stem hip arthroplasty in elderly: A clinical and radiological analysis,” ''Journal of Orthopaedics, ''vol. 57, pp. 17-22, 4 June 2024. 
|-
|  style="vertical-align: top;"|[47] 
|  style="vertical-align: top;"|S. Tammina, “Transfer Learning Using VGG-16 with Deep Convolutional Neural Network for Classifying Images,” ''International Journal of Scientific and Research Publications, ''vol. 9, no. 10, pp. 143-147, 2019. 
|-
|  style="vertical-align: top;"|[48] 
|  style="vertical-align: top;"|A. Altaf, A. Altaf and F. Rehman, “Modified EfficientNet-B0 Architecture Optimized with Quantum-Behaved Algorithm for Skin Cancer Lesion Assessment,” ''Diagnostics, ''vol. 15, no. 24, 2025. 
|-
|  style="vertical-align: top;"|[49] 
|  style="vertical-align: top;"|F. Chollet, “Xception: Deep Learning with Depthwise Separable Convolutions,” ''Computer Vision and Pattern Recognition, ''2017. 
|-
|  style="vertical-align: top;"|[50] 
|  style="vertical-align: top;"|S. Christian, I. Sergey, V. Vincent and A. Alex, “Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,” in ''Proceedings of the AAAI Conference on Artificial Intelligence '', 2016. 
|-
|  style="vertical-align: top;"|[51] 
|  style="vertical-align: top;"|H. Gao, L. Zhuang, v. d. M. Laurens and Q. W. Kilian, “Densely Connected Convolutional Networks,” in ''IEEE Conference on Computer Vision and Pattern Recognition (CVPR)'', 2017. 
|-
|  style="vertical-align: top;"|[52] 
|  style="vertical-align: top;"|C. Ming, W. Chunping, Y. Ying, L. Yirui, L. Yifan and C. Xianglong, “Improved YOLOv9-based remote sensing image detection method,” ''Scientific Reports, ''vol. 15, 29 December 2025. 
|-
|  style="vertical-align: top;"|[53] 
|  style="vertical-align: top;"|Y. Muhammad, “What is YOLOv9: An In-Depth Exploration of the Internal Features of the Next-Generation Object Detector,” ''arXiv, ''12 September 2024. 
|-
|  style="vertical-align: top;"|[54] 
|  style="vertical-align: top;"|M. Shaoni, “YOLOv9 Object Detection Features, Benefits, and Use Cases,” Digital Ocean, 2025.
|-
|  style="vertical-align: top;"|[55] 
|  style="vertical-align: top;"|I.-H. Y. a. H.-Y. M. L. Chien-Yao Wang, “YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information,” in ''Computer Vision – ECCV 2024 (European Conference on Computer Vision)'', 2024. 
|-
|  style="vertical-align: top;"|[56] 
|  style="vertical-align: top;"|A. S. Habeb and A. M. I. Nor, “You Only Look Once (YOLO): A comprehensive review of its evolution and applications in electronic component defect detection,” ''Results in Control and Optimization, ''vol. 23, 2026. 
|-
|  style="vertical-align: top;"|[57] 
|  style="vertical-align: top;"|M. Je, S. Hwang, S. Lee and Y. Kim, “Development and evaluation of a machine learning model for osteoporosis risk prediction in Korean women,” ''BMC Women's Health, ''vol. 25, no. 1, 2025. 
|}