Learning‐Based Classification of B‐ and T‐Cell Lymphoma on Histopathological Images: A Multicenter Study

This is the authors' abstract. We don't add key points for this paper.

Lymphoma, a clonal malignancy from lymphocytes, includes diverse subtypes requiring distinct immunohistochemical stains for accurate diagnosis. Limited biopsy specimens often restrict the use of multiple stains, complicating diagnostic workflows. Lymphomas are typically classified into B-cell and T-cell types, each with specific markers. This study represents the first feasibility study in deploying deep learning models for B- and T-cell lymphoma classification on histopathological images. We analyzed 1510 H&E-stained sections (750 B-cell, 760 T-cell) with CNN models (Xception, NASNetL, ResNet50, EfficientNet), enhanced by Convolutional Block Attention Modules (CBAMs). All models demonstrated strong classification capabilities, with EfficientNet achieving the highest accuracy at 91.5% and the best precision at 91.9%, while Xception performed the best recall at 91.5%. Furthermore, the deep learning models significantly outperformed human pathologists in classification accuracy and inference speed, processing images in milliseconds compared to the several seconds required for manual diagnosis. These findings underscore the effectiveness of advanced CNN models in improving diagnostic precision while reducing dependency on manual staining and interpretation, and the integration of AI-driven classification can provide valuable support for pathologists.