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

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Oncology

Background:

• Accurate breast cancer diagnosis relies on precise histopathological image analysis.
• Identifying mitotic figures is crucial for grading and prognosis.
• Existing methods face challenges with complex image data and annotations.

Purpose of the Study:

• To develop and evaluate a two-stage algorithm for breast cancer histopathology image classification using hybrid Convolutional-Transformer models.
• To assess the performance of ConvMixer and Coatnet models in mitosis detection.
• To address challenges posed by incomplete cell annotations in real-world datasets.

Main Methods:

• A two-stage algorithm was implemented, combining training and mitosis detection.
• Hybrid Convolutional-Transformer models (ConvMixer, Coatnet) were utilized.
• Advanced color normalization and data augmentation were applied during training.
• Models were trained and validated on the MITOS-ATYPIA-14 and Cherbourg Hospital datasets.

Main Results:

• Coatnet exhibited slightly superior performance compared to ConvMixer in mitosis detection.
• The hybrid architecture effectively captured both local and global image features.
• The two-stage approach demonstrated robustness in handling complex histopathological data.

Conclusions:

• Hybrid Convolutional-Transformer models show significant potential for automated breast cancer diagnosis.
• Coatnet's architecture is promising for accurate mitotic figure detection in histopathology.
• This methodology offers a viable solution for improving medical image analysis in oncology.