Deep learning predicts gene rearrangements from histopathology in large B-cell lymphoma

Abstract Large B-cell lymphomas are molecularly heterogeneous mature B-cell neoplasms in which assessment of MYC , BCL2 , and BCL6 rearrangement status contributes to integrated diagnosis, risk stratification, and treatment planning. Fluorescence in situ hybridization (FISH) remains the standard method for detecting these rearrangements but can be time-consuming, tissue-consuming, and resource-intensive. Here, we present HE2FISH, a weakly supervised deep-learning framework for slide-level prediction of FISH-defined MYC , BCL2 , and BCL6 rearrangement status from routine haematoxylin and eosin (H&E)-stained whole-slide images of cases diagnosed in routine practice as large B-cell lymphoma and showing a diffuse growth pattern. Across a multicenter cohort of 1377 patients from five hospitals with paired H&E and FISH data, HE2FISH demonstrated robust cross-center generalization, achieving a mean external area under the receiver operating characteristic curve exceeding 0.81 for single-gene prediction, and also showed evaluable performance for FISH-defined co-rearrangement patterns. In the CHCAMS cohort with survival follow-up, HE2FISH-predicted rearrangement status stratified overall and disease-free survival comparably to FISH. Attention-based analyses provided visual summaries of high-attention image regions associated with model predictions, and incorporation of 13 structured clinicopathological variables improved performance in selected settings, particularly for co-rearrangement prediction. By generating rearrangement-probability estimates from routine H&E-stained slides without additional tissue use at the prediction stage, HE2FISH provides an H&E-based prescreening approach that may support prioritization for confirmatory molecular testing within integrated diagnostic workflows.