Deep Representation Learning of High-Resolution Whole Slide Histopathology Images

Whole slide tissue histopathology images (WSIs) play a crucial role in tissue specimen assessment and diagnosis of associated diseases. Recent technological progress in image acquisition systems has led to an increasing accumulation of high-resolution histopathology images. Nevertheless, employing these images to develop clinical decision support systems has been hampered by the need for manual examination of WSIs, a subjective, labor-intensive, time-consuming, and error-prone process. This creates a burgeoning demand for new analytic approaches to analyze/pre-process such images. Furthermore, a content-based representation allows the integrated study of histopathology images with other data modalities enabling holistic and multi-modal analysis of human diseases. Deep learning approaches have shown promising performance on feature extraction from images. However, dealing with WSIs introduces new challenges, demanding more efficient approaches to learn an informative representation of these images. This research aims to employ deep learning approaches for representation learning of WSIs focusing on Barrett's Esophagus (BE). In this setting, three different approaches will be considered: Bag of Visual Words (BoVW), Neural Image Compression (NIC), and Graph Neural Networks (GNN).