Optimizing Foundation Models for Histopathology: A Continual Learning Approach to Cancer Detection

Accurate analysis of histopathology images is a key step in cancer diagnosis and treatment planning. Specialized computational models for histopathology image analysis may be needed to better capture the unique characteristics of histopathological images that could be missed by models pretrained on generic data, such as ImageNet. We employed an incremental learning approach to enhance the performance of foundation models using the EfficientNet B0 architecture. Our study comprised three key experiments. First, we established a baseline accuracy of 84.4% by training the model on 50% of our dataset. Second, we investigated the impact of retraining different numbers of top blocks during incremental learning phases, finding that retraining three to six blocks provided the most significant accuracy improvements. Third, we analyzed the trade-off between accuracy improvement and training time, determining that retraining three to six blocks offered the best balance between performance gains and computational efficiency. Our results demonstrate the effectiveness of specialized models in capturing the details specific of histopathological images. Despite computational limitations, our findings underscore the importance of tailored histopathological models.