Normal and Atypical Mitosis Image Classifier using Efficient Vision Transformer

2025-08-28

AI Summary

This study addresses the fine-grained classification of typical versus atypical mitoses in histopathological images for the MIDOG 2025 challenge, aiming to improve cancer grading and diagnostic accuracy. We propose a hybrid CNN–ViT architecture based on EfficientViT-L2, integrated with chromatic deconvolution for stain enhancement, five-fold ensemble learning, and a unified nuclear image dataset spanning multiple cancer types. Domain generalization is rigorously evaluated via leave-one-cancer-type cross-validation. To our knowledge, this is the first work to adapt a lightweight, efficient vision transformer for mitotic abnormality recognition. In the preliminary phase, our method achieves a balanced accuracy of 0.859, an ROC AUC of 0.942, and a raw accuracy of 0.85—demonstrating superior and well-balanced performance over baseline methods. These results validate the model’s robustness and strong generalization capability across diverse cancer types.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Multi-instance/Multi-view Learning
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

We tackle atypical versus normal mitosis classification in the MIDOG 2025 challenge using EfficientViT-L2, a hybrid CNN--ViT architecture optimized for accuracy and efficiency. A unified dataset of 13,938 nuclei from seven cancer types (MIDOG++ and AMi-Br) was used, with atypical mitoses comprising ~15. To assess domain generalization, we applied leave-one-cancer-type-out cross-validation with 5-fold ensembles, using stain-deconvolution for image augmentation. For challenge submissions, we trained an ensemble with the same 5-fold split but on all cancer types. In the preliminary evaluation phase, this model achieved balanced accuracy of 0.859, ROC AUC of 0.942, and raw accuracy of 0.85, demonstrating competitive and well-balanced performance across metrics.

Problem

• Classifying normal versus atypical mitosis images
• Using EfficientViT-L2 for accuracy and efficiency
• Assessing domain generalization across cancer types

Innovation

• Hybrid CNN-ViT architecture for mitosis classification
• Leave-one-cancer-type-out cross-validation for generalization
• Stain-deconvolution augmentation for improved performance