Scholarlink

<p>Artificial Intelligence (AI), particularly in the domains of computer vision and deep learning (DL), is increasingly being integrated into medical diagnostics. Although it has not yet entirely transformed clinical practice, its role in medical image analysis is rapidly expanding, offering the potential for more accurate, efficient, and accessible diagnostic capabilities that contribute to improved patient outcomes. Currently, DL techniques are employed in a variety of medical imaging tasks, such as anomaly detection, pathology classification, and anatomical structure segmentation across modalities, including endoscopy, MRI, and CT scans. Despite these advancements, the widespread clinical adoption of DL models remains constrained by several limitations, including the opacity of model decision-making, data imbalance, and lack of interpretability. To address these challenges, this study proposes an <i>explainable knowledge distillation (KD)</i> framework that integrates Gradient-weighted Class Activation Mapping (Grad-CAM) with a ResNet-based teacher-student architecture. The proposed methodology was evaluated on a class-imbalanced dataset comprising ten categories of gastrointestinal (GI) abnormalities derived from the <i>Capsule Vision 2024 Challenge</i>. The resulting student model achieved a classification accuracy of 96.52% and an F1-score of 96.37%, thereby surpassing previously reported methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A knowledge distillation framework integrating Grad-CAM in ResNet for imbalanced gastrointestinal abnormality classification in capsule endoscopy

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月16日

查看全文
  • Tejashwini K,
  • Karthik K,
  • Jayakumar Jeganathan

摘要

Artificial Intelligence (AI), particularly in the domains of computer vision and deep learning (DL), is increasingly being integrated into medical diagnostics. Although it has not yet entirely transformed clinical practice, its role in medical image analysis is rapidly expanding, offering the potential for more accurate, efficient, and accessible diagnostic capabilities that contribute to improved patient outcomes. Currently, DL techniques are employed in a variety of medical imaging tasks, such as anomaly detection, pathology classification, and anatomical structure segmentation across modalities, including endoscopy, MRI, and CT scans. Despite these advancements, the widespread clinical adoption of DL models remains constrained by several limitations, including the opacity of model decision-making, data imbalance, and lack of interpretability. To address these challenges, this study proposes an explainable knowledge distillation (KD) framework that integrates Gradient-weighted Class Activation Mapping (Grad-CAM) with a ResNet-based teacher-student architecture. The proposed methodology was evaluated on a class-imbalanced dataset comprising ten categories of gastrointestinal (GI) abnormalities derived from the Capsule Vision 2024 Challenge. The resulting student model achieved a classification accuracy of 96.52% and an F1-score of 96.37%, thereby surpassing previously reported methods.