HoLy-Net: Segmentation of histological images of diffuse large B-cell lymphoma
Hussein Naji,Lucas Sancere,Adrian Simon,Reinhard Büttner,Marie-Lisa Eich,Philipp Lohneis,Katarzyna Bożek
DOI: https://doi.org/10.1016/j.compbiomed.2024.107978
IF: 7.7
2024-01-13
Computers in Biology and Medicine
Abstract:Over the last years, there has been large progress in automated segmentation and classification methods in histological whole slide images (WSIs) stained with hematoxylin and eosin (H&E). Current state-of-the-art (SOTA) techniques are based on diverse datasets of H&E-stained WSIs of different types of predominantly solid cancer. However, there is a scarcity of methods and datasets enabling segmentation of tumors of the lymphatic system (lymphomas). Here, we propose a solution for segmentation of diffuse large B-cell lymphoma (DLBCL), the most common non-Hodgkin's lymphoma. Our method applies to both H&E-stained slides and to a broad range of markers stained with immunohistochemistry (IHC). While IHC staining is an important tool in cancer diagnosis and treatment decisions, there are few automated segmentation and classification methods for IHC-stained WSIs. To address the challenges of nuclei segmentation in H&E− and IHC-stained DLBCL images, we propose HoLy-Net - a HoVer-Net-based deep learning model for lymphoma segmentation. We train two different models, one for segmenting H&E− and one for IHC-stained images and compare the test results with the SOTA methods as well as with the original version of HoVer-Net. Subsequently, we segment patient WSIs and perform single cell-level analysis of different cell types to identify patient-specific tumor characteristics such as high level of immune infiltration. Our method outperforms general-purpose segmentation methods for H&E staining in lymphoma WSIs (with an F1 score of 0.899) and is also a unique automated method for IHC slide segmentation (with an F1 score of 0.913). With our solution, we provide a new dataset we denote LyNSeC (lymphoma nuclear segmentation and classification) containing 73,931 annotated cell nuclei from H&E and 87,316 from IHC slides. Our method and dataset open up new avenues for quantitative, large-scale studies of morphology and microenvironment of lymphomas overlooked by the current automated segmentation methods.
engineering, biomedical,computer science, interdisciplinary applications,mathematical & computational biology,biology