Histological images (WSIs) of tumor samples are routinely available in pathology labs and reflect the physiological phenotype of tumor cells and their genomic aberrations. Deep learning techniques offer the potential to extract from these WSIs hidden morphological features that associate with molecular properties. WSIs thus contain a largely untapped source of valuable information on the molecular properties and their spatial organization. Here, we want to develop a deep learning framework that allows exploiting WSIs as a cost-efficient proxy for spatially resolved molecular profiling of tumor samples.
As a case study we will train deep learning models that can use histological images to predict the presence of molecular properties that associate with metastatic prostate cancer. Applying these models on an in-house cohort with a dedicated multifocal design will allow identifying the models with prognostic value that allow predicting lesions with high metastatic potential.