SuperCRF

SuperCRF is a computational pathology framework that improves cell classification in H&E-stained tumor slides by incorporating multi-resolution contextual information. It combines a deep learning spatially constrained-convolutional neural network (SC-CNN) for cell detection and classification with a conditional random field (CRF) that integrates cellular neighborhood information and tumor regional classification from lower-resolution images. The CRF is informed by a superpixel-based machine learning framework applied to the lower-resolution images. By leveraging this hierarchical approach inspired by how pathologists perceive tissue architecture, SuperCRF achieved an 11.85% overall improvement in cell classification accuracy compared to the state-of-the-art SC-CNN classifier alone.