Scholarlink

The precise and automated classification of various tissue types in histopathological images is a vital component in advancing pathology and enhancing clinical decision-making. In the context of colorectal cancer (CRC) diagnosis, texture analysis plays a significant role in quantifying tumor-to-stroma ratios and identifying morphological patterns associated with disease progression. Despite its importance, most existing research has predominantly focused on binary tissue classification, leaving multiclass segmentation underexplored. This study aims to address this gap by utilizing a publicly available dataset comprising 5,000 histological images of human CRC tissue, which are annotated into eight distinct tissue categories. Our research involved a systematic evaluation of convolutional neural network (CNN) and Vision Transformer (ViT) for multiclass tissue classification, comparing their performance. The findings indicate that both models effectively extract discriminative histological features. Notably, the ViT achieved superior precision, recall, and overall accuracy, attaining a rate of 91.2% compared to 87.2% for the CNN. These results underscore the potential of transformer-based architectures to enhance automated CRC tissue recognition, thereby supporting more reliable computational pathology systems.

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

Evaluating Convolutional and Transformer-Based Deep Learning Models for Colorectal Cancer Tissue Recognition

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Marina Adriana Mercioni,
  • Andreea-Luiza Crețu,
  • Nina Ivanovic,
  • Raluca Dumache

摘要

The precise and automated classification of various tissue types in histopathological images is a vital component in advancing pathology and enhancing clinical decision-making. In the context of colorectal cancer (CRC) diagnosis, texture analysis plays a significant role in quantifying tumor-to-stroma ratios and identifying morphological patterns associated with disease progression. Despite its importance, most existing research has predominantly focused on binary tissue classification, leaving multiclass segmentation underexplored. This study aims to address this gap by utilizing a publicly available dataset comprising 5,000 histological images of human CRC tissue, which are annotated into eight distinct tissue categories. Our research involved a systematic evaluation of convolutional neural network (CNN) and Vision Transformer (ViT) for multiclass tissue classification, comparing their performance. The findings indicate that both models effectively extract discriminative histological features. Notably, the ViT achieved superior precision, recall, and overall accuracy, attaining a rate of 91.2% compared to 87.2% for the CNN. These results underscore the potential of transformer-based architectures to enhance automated CRC tissue recognition, thereby supporting more reliable computational pathology systems.