Scholarlink

Class imbalance is a critical challenge in medical image classification, particularly in gastroscopy, where accurately detecting underrepresented pathological conditions is essential for clinical decision-making. This paper investigates algorithm-level techniques’ effectiveness in addressing severe class imbalance in gastroscopy image classification. A curated data set comprising four clinically relevant categories was constructed: normal, esophagitis, Barrett’s esophagus, and gastric polyps by consolidating and reclassifying samples from two open-access sources: the HyperKvasir and GastroVision datasets. The resulting dataset exhibits a pronounced imbalance ratio of approximately \( \rho \approx 73 \) , closely reflecting real-world diagnostic scenarios. To mitigate the adverse effects of imbalance, we evaluate the performance of two algorithm-level methods: Focal Loss and a cost-sensitive learning strategy (CoSen). These methods were integrated into deep learning pipelines based on a convolutional neural network and transformer architectures. Experimental results demonstrate that both approaches substantially improve the F1-score, particularly for the minority polyp class. The best model achieved a 40.4% increase in F1 score for one class compared to the baseline, underscoring the value of algorithmic imbalance mitigation strategies for gastrocopy image classification.

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

Classification of Gastroscopy Images Under Extreme Class Imbalance: A Deep Learning Pipeline

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Adrian Krenzer,
  • Tobias Friedetzki,
  • Max Dietsch,
  • Frank Puppe

摘要

Class imbalance is a critical challenge in medical image classification, particularly in gastroscopy, where accurately detecting underrepresented pathological conditions is essential for clinical decision-making. This paper investigates algorithm-level techniques’ effectiveness in addressing severe class imbalance in gastroscopy image classification. A curated data set comprising four clinically relevant categories was constructed: normal, esophagitis, Barrett’s esophagus, and gastric polyps by consolidating and reclassifying samples from two open-access sources: the HyperKvasir and GastroVision datasets. The resulting dataset exhibits a pronounced imbalance ratio of approximately \( \rho \approx 73 \) , closely reflecting real-world diagnostic scenarios. To mitigate the adverse effects of imbalance, we evaluate the performance of two algorithm-level methods: Focal Loss and a cost-sensitive learning strategy (CoSen). These methods were integrated into deep learning pipelines based on a convolutional neural network and transformer architectures. Experimental results demonstrate that both approaches substantially improve the F1-score, particularly for the minority polyp class. The best model achieved a 40.4% increase in F1 score for one class compared to the baseline, underscoring the value of algorithmic imbalance mitigation strategies for gastrocopy image classification.