Scholarlink

Accurate classification of endometrial tissue regions in hysteroscopic imaging is vital for early detection of abnormalities. This study evaluates two deep learning models—a custom Convolutional Neural Network (CNN) and a fine-tuned Data-Efficient Image Transformer (DeiT)—for binary classification of normal versus abnormal endometrial regions of interest (ROIs). Both models were trained and validated on 443 labelled ROIs collected from 40 patients using five-fold cross-validation. The CNN achieved an average accuracy of 77% (AUC = 0.83), while the DeiT yielded 74% (AUC = 0.82). A classical approach using classical texture features and a Support Vector Machine (SVM) achieved higher accuracy (81%), indicating that deep models may be limited by the small dataset size and input resolution. These results highlight the challenges of applying deep learning to data-scarce medical imaging tasks and motivate future work on dataset expansion, transfer learning, and hybrid feature integration for improved diagnostic performance.

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

Deep Learning Architectures for Endometrial Abnormality Detection via Hysteroscopic ROI Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Andreas C. Anastasiou,
  • Vasilios Tanos,
  • Marios Neofytou,
  • Ioannis Constantinou,
  • Panayiotis Tanos,
  • Eirini Schiza,
  • Marios S. Pattichis,
  • Constantinos S. Pattichis,
  • Andreas Panayides

摘要

Accurate classification of endometrial tissue regions in hysteroscopic imaging is vital for early detection of abnormalities. This study evaluates two deep learning models—a custom Convolutional Neural Network (CNN) and a fine-tuned Data-Efficient Image Transformer (DeiT)—for binary classification of normal versus abnormal endometrial regions of interest (ROIs). Both models were trained and validated on 443 labelled ROIs collected from 40 patients using five-fold cross-validation. The CNN achieved an average accuracy of 77% (AUC = 0.83), while the DeiT yielded 74% (AUC = 0.82). A classical approach using classical texture features and a Support Vector Machine (SVM) achieved higher accuracy (81%), indicating that deep models may be limited by the small dataset size and input resolution. These results highlight the challenges of applying deep learning to data-scarce medical imaging tasks and motivate future work on dataset expansion, transfer learning, and hybrid feature integration for improved diagnostic performance.