Scipediacontent: Scipediacontent moved page Review 167263214755 to Peng et al 2026a

2026-03-18T09:52:43Z

Scipediacontent moved page Review 167263214755 to Peng et al 2026a

Scipediacontent: Scipediacontent moved page Draft content 418798471 to Review 167263214755

2026-03-18T08:59:05Z

Scipediacontent moved page Draft content 418798471 to Review 167263214755

Scipediacontent: Created page with " == Abstract ==
Deep learning (DL)-based models have demonstrated significant advancements in medical image classification. However, the scarcity of accurately labeled tra..."

2026-03-18T08:58:46Z

Created page with " == Abstract == <p>Deep learning (DL)-based models have demonstrated significant advancements in medical image classification. However, the scarcity of accurately labeled tra..."

New page

== Abstract ==

<p>Deep learning (DL)-based models have demonstrated significant advancements in medical image classification. However, the scarcity of accurately labeled training data and the prevalence of label noise remain critical obstacles to further performance improvements. Although semi-supervised learning and learning with noisy labels (LNL) methods each offer partial remedies, their independent application often leads to suboptimal outcomes. To address this, we propose a unified framework termed the Semi-supervised Adaptive Distillation Network (SADNet), which synergistically integrates semi-supervised training with noise-robust distillation. SAD-Net consists of three core components. First, a semi-supervised learning framework is employed to generate pseudo-labels from unlabeled data, thereby augmenting the training set. Subsequently, a Noise Filtering Module (NF-Module) is introduced, which combines a Convolutional Neural Network (CNN) with an Improved Fuzzy C-Means (IFCM) algorithm using a weighted average distance metric. This module produces weighted soft labels from both models and filters out noisy samples based on a confidence threshold. Finally, an Adaptive Weighted Distillation Module (AWD-Module) is designed, incorporating the IFCM along with two CNN architectures. It processes the high-confidence samples selected by the NF-Module and performs classification via dynamically weighted soft labels derived from all three models. Extensive experiments on two medical image datasets show that SAD-Net achieves superior performance compared to state-of-the-art semi-supervised methods, attaining the highest scores in accuracy, sensitivity, specificity, and F1-score. Moreover, it outperforms leading LNL approaches across all evaluated metrics. These results validate the efficacy of the proposed SAD-Net in simultaneously mitigating the problems of limited labeled data and noisy labels in medical image classification.OPEN ACCESS Received: 23/09/2025 Accepted: 03/12/2025 Published: 20/03/2026</p>

== Document ==
<pdf>Media:Draft_content_418798471-8286-document.pdf</pdf>

← Older revision	Revision as of 09:52, 18 March 2026
(No difference)

← Older revision	Revision as of 08:59, 18 March 2026
(No difference)

Peng et al 2026a - Revision history

Scipediacontent: Scipediacontent moved page Review 167263214755 to Peng et al 2026a

Scipediacontent: Scipediacontent moved page Draft content 418798471 to Review 167263214755

Scipediacontent: Created page with " == Abstract == Deep learning (DL)-based models have demonstrated significant advancements in medical image classification. However, the scarcity of accurately labeled tra..."

Scipediacontent: Created page with " == Abstract ==
Deep learning (DL)-based models have demonstrated significant advancements in medical image classification. However, the scarcity of accurately labeled tra..."