Deep learning for subtyping classical hodgkin lymphoma on histopathology images: A comprehensive lymphoscope

Classical Hodgkin lymphoma (CHL) is an uncommon form of lymphoma that mainly affects young adults and adolescents. Accurate morphologic assessment of CHL is critical for precise subtyping and effective treatment planning. Despite its clinical importance, there is a notable research gap in the automated subtyping of CHL, in contrast to the extensive developments in more common cancers. To address this gap, we curated a comprehensive dataset of 1247 whole-slide images (WSI) for four CHL subtypes from three medical centers from Türkiye, ensuring a reliable basis for our research. We developed a deep-learning pipeline for Classical Hodgkin lymphoma subtyping. Our weakly supervised model, Instance and Embedding Fused Multiple Instance Learning (IEF-MIL), utilizes a multiscale dual-stream network, outperforming state-of-the-art MIL models in the existing literature. Furthermore, we have incorporated state-of-the-art self-supervised learning foundation models trained on hundreds of histopathology whole slide images into our weakly supervised pipeline through transfer learning. We have demonstrated the generalizability and limitations of these models on our out-of-distribution dataset. Additionally, our model enhances interpretability with heatmaps at three magnification levels, providing deeper insights into its predictions.