# Deep learning-based classification of colorectal cancer in histopathology images for category detection

## TL;DR

This paper uses deep learning to classify colorectal cancer in histopathology images, achieving high accuracy with models like ResNet-50 and a two-stage framework.

## Contribution

A two-stage prediction framework is introduced to improve classification of underrepresented CRC classes using moderate-depth deep learning models.

## Key findings

- ResNet-50 achieved a micro-averaged ROC AUC of 0.9933 and F1-score of 87.51% for CRC classification.
- The two-stage framework improved High-grade dysplasia classification accuracy from 53.57% to 71.43%.
- Swin Transformer V2 showed high accuracy for Hyperplasia polyp (95.83%) and Adenocarcinoma (93.33%) detection.

## Abstract

Accurate and timely diagnosis of colorectal cancer (CRC) is essential for effective treatment and better patient outcomes. This study explores the application of deep learning (DL) for automated CRC categories classification using hematoxylin and eosin-stained histopathology (H&E) images. Among the models, ResNet-34 demonstrated a strong balance of performance and complexity, achieving an overall accuracy of 85.04%, with top-2 and top-3 classification accuracies of 96.68% and 99.23%, respectively. ResNet-50 exhibited the highest micro-averaged ROC AUC of 0.9933 and F1-score of 87.51%. Swin Transformer V2 model also showed competitive results, with Swin v2-t-w8 achieving particularly high accuracy in Hyperplasia polyp detection (95.83%) and Adenocarcinoma (93.33%), alongside strong ROC AUCs (0.9926 for Hyperplasia polyp and 0.9864 for Adenocarcinoma), though at the cost of increased computational demands. We further developed a two-stage prediction framework comprising a binary abnormal detection stage followed by a multiclass cancer classifier. This approach substantially improved classification robustness, particularly for underrepresented and morphologically complex classes. Particularly, High-grade dysplasia classification accuracy improved from 53.57% with ResNet-34 to 71.43% in its two-stage extension. These results suggest that moderate-depth architectures can effectively capture the morphological diversity of colorectal cancer stages and provide an interpretable, efficient deep learning-based diagnostic tool to support pathologists.