Molecular Subtype Prediction for Breast Cancer Using H&E Specialized Backbone
Proceedings of the MICCAI Workshop on Computational Pathology, PMLR 156:1-9, 2021.
Identifying the molecular markers to categorize breast cancer is one of the key steps in determining the prognosis and treatment strategy. The standard clinical practice is to do this analysis based on multiple immunohistochemistry (IHC) stainings for each biomarker, which is expensive and inconsistent when lacking resources. In this work, we investigated the predictiveness of morphological characteristics of Hematoxylin and Eosin (H&E) stained tissues for molecular subtype analysis, as an initial step for direct treatment response prediction based on H&E whole slide images (WSI). Transfer learning using backbones pre-trained on natural images is a common practice to deal with the challenge of lack of large and precisely annotated datasets. However, using pre-training on natural images is not optimal for clinical images. To deal with this challenge and leverage large pools of unlabeled data, we propose to use a specialized backbone pre-trained on H&E WSI in a self-supervised setting, i.e. without any labels. Our experiments show that this backbone is capable of learning discriminating morphological characteristics from H&E images which are well predictive of the molecular subtypes in weakly supervised settings. Also, the network performs better in terms of generalization to unseen data from new scanner types, despite the relatively small size of the dataset used for pre-training the backbone.