Scholarlink

In recent years, deep learning has shown significant success in automatic detection and classification of gastrointestinal (GI) diseases from endoscopy images. However, the complex and non-transparent nature of deep neural networks questions their clinical applicability due to limited interpretability. This study investigates the performance and explainability of nine state-of-the-art deep learning architectures on three endoscopic image datasets. Through a comparative analysis of various explainable AI (XAI) techniques, including GradCAM, ScoreCAM, LayerCAM, LIME, SmoothGrad, Guided Backpropagation (Guidedbp), Integrated Gradients (IG), and Contrastive Explanation (CE) this study enhances model interpretability and provide insights into the decision-making process. Our findings illustrate the interpretability trade-offs between different architectures and suggest ways for improving diagnostic reliability in clinical settings. The study’s outcomes show the necessity of model transparency and interpretability in medical diagnostics.

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

Enhancing Interpretability in Gastrointestinal Disease Detection: A Comparative Analysis of Deep Learning Architectures and Explainable AI

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Muhammad Fahad,
  • Saif Hassan,
  • Faouzi Alaya Cheikh,
  • Mohib Ullah

摘要

In recent years, deep learning has shown significant success in automatic detection and classification of gastrointestinal (GI) diseases from endoscopy images. However, the complex and non-transparent nature of deep neural networks questions their clinical applicability due to limited interpretability. This study investigates the performance and explainability of nine state-of-the-art deep learning architectures on three endoscopic image datasets. Through a comparative analysis of various explainable AI (XAI) techniques, including GradCAM, ScoreCAM, LayerCAM, LIME, SmoothGrad, Guided Backpropagation (Guidedbp), Integrated Gradients (IG), and Contrastive Explanation (CE) this study enhances model interpretability and provide insights into the decision-making process. Our findings illustrate the interpretability trade-offs between different architectures and suggest ways for improving diagnostic reliability in clinical settings. The study’s outcomes show the necessity of model transparency and interpretability in medical diagnostics.