Two-Stage Hybrid Convolutional-Transformer Models for Breast Cancer Histopathology

Abstract

In this study, we present a two-stage algorithm utilizing hybrid Convolutional-Transformer models, specifically ConvMixer and Coatnet, for breast cancer histopathology image classification. Our methodology focuses on two critical stages: training and mitosis detection. The training stage incorporates advanced color normalization and data augmentation techniques, enhancing the models' ability to process complex histopathological images. This stage prepares the models to accurately identify mitotic figures using the MITOS-ATYPIA-14 and a real-world dataset from Cherbourg Hospital. In the mitosis detection stage, the trained models are applied to detect and classify mitotic figures, leveraging the hybrid architecture's strengths in capturing both local and global features. Our results demonstrate that Coatnet slightly outperforms ConvMixer, owing to its efficient combination of convolutional and transformer layers. This two-stage approach addresses the challenges of incomplete cell annotations and showcases the hybrid models' potential in advancing medical image analysis, particularly in breast cancer diagnosis.