What question did this study set out to answer?

This study aims to evaluate the effectiveness of fine-tuned convolutional neural networks in estimating colorectal cancer histological grade and pathological stages from biopsy samples.

July 5, 2026Open Access

Region-of-interest histopathology with fine-tuned CNNs enables colorectal cancer grading and pathologic stage estimation

Key Points

This study aims to evaluate the effectiveness of fine-tuned convolutional neural networks in estimating colorectal cancer histological grade and pathological stages from biopsy samples.
Utilized 1.5-mm colorectal cancer tissue microarray cores to assess histological features.
Fine-tuned several ImageNet-pretrained CNN models (e.g., ResNet, DenseNet) to classify tumor grades and T categories.
Analyzed model performance using area under the curve metrics for various classification tasks.
Achieved a maximum AUC of 0.991 for tumor histological grade classification across CNNs.
Predicted early versus advanced pathological T category with a maximum AUC of 0.888.
Distinguished early versus advanced pathological stage with a maximum AUC of 0.721.

Abstract

Abstract While AI-based diagnostic support using whole slide images of pathological specimens has advanced in recent years, the development of systems that work with small biopsy specimens remains underexplored. In this study, we evaluated 1.5-mm-diameter colorectal cancer tissue microarray cores to estimate histological grade, to predict the resection-assigned pathological T (pT) category from microscopic morphology, and to estimate pathological TNM stage. We fine-tuned multiple ImageNet-pretrained convolutional neural network (CNN) backbones (ResNet, WideResNet, DenseNet, MobileNet, and EfficientNet) and evaluated individual models and ensemble learning. A maximum area under the curve for two-class classification of tumor histological grade was 0.991 across CNNs. A Nemenyi multiple-comparison test indicated significant differences in performance across tumor grading task definitions. Furthermore, the models predicted early versus advanced pT category (resection-assigned) with a maximum AUC of 0.888, and the Nemenyi test again indicated differences across task definitions. Early versus advanced pathological stage was separated with a maximum AUC of 0.721. Gradient-weighted class activation mapping visualization confirmed that the CNN model focused on histological features characteristic of Grade 3 tumors, such as tumor budding and poorly differentiated clusters, validating its classification performance. These results suggest that pathologist-selected biopsy-scale region of interests (ROIs) enable robust grading and pT-associated risk stratification with modest computational requirements, although they do not directly measure anatomical depth of invasion. They also suggest the potential of CNN-based diagnostic support systems as practical tools for treatment planning that can be implemented with limited computational resources, including constrained image data and storage capacity.

AIに質問

Bookmark

View Full Paper