Deep learning pipeline predicts breast cancer subtypes from histopathology images

Researchers have developed a new deep learning framework that can predict PAM50 breast cancer subtypes using H&E-stained whole-slide images, which reduces reliance on costly molecular tests. This method combines a multi-objective optimization approach with the non-dominated sorting genetic algorithm II (NSGA-II) and Monte Carlo dropout to evaluate uncertainty, while optimizing various factors like patch informativeness and spatial diversity. The model utilizes a ResNet18 backbone for feature extraction alongside a custom CNN head for classification, and it was evaluated using the internal TCGA-BRCA dataset. The PAM50 gene signature helps categorize breast cancer into specific subtypes, paving the way for tailored treatment strategies. This efficient pipeline selects a small but informative set of patches, potentially lowering costs and speeding up clinical processes.