Self-explaining artificial intelligence for the classification of B cell non-Hodgkin lymphoma: A diagnostic decision support study
by Michael C. Thrun, Jörg Hoffmann, Stefan W. Krause, Peter Krawitz, Quirin Stier, Andreas Neubauer, Cornelia Brendel, Alfred Ultsch Background Multiparameter flow cytometry is a cornerstone of B cell non-Hodgkin lymphoma (B-NHL) diagnostics, but interpretation requires substantial expertise and is complicated by high-dimensional data, variable sample quality, limited data for rare entities, and evolving clinical classification systems. Current artificial intelligence approaches often require large training datasets and provide limited insight into the rationale behind individual diagnostic decisions. Methods and findings We developed FlowXAI, a self-explaining artificial intelligence system designed to support B-NHL classification while explicitly reporting case-level diagnostic trustworthiness. FlowXAI combines unsupervised structural analysis with a clinically motivated, multi-level diagnostic framework reflecting routine diagnostic priorities. An unsupervised Tile Mining (TM) procedure performs pre-diagnostic sample-quality assessment by identifying structurally atypical samples. TM is applied to filter training data, enabling substantial reduction of training requirements while preserving unbiased evaluation on independent test samples.FlowXAI was evaluated using repeated cross-validation on 19,493 peripheral blood samples and further assessed on an independent external benchmark dataset generated at a separate diagnostic center using a different antibody panel. Across diagnostic levels, FlowXAI achieved performance comparable to a deep learning–based system despite requiring approximately two orders of magnitude fewer training samples. When predictions were classified as confident by the system’s internal self-assessment, diagnostic performance exceeded that of the neural network baseline. Unsupervised structural analysis demonstrated clear separation between normal controls and selected lymphoma entities such as chronic lymphocytic leukemia–like lymphomas and hairy cell leukemia, while other entities were not clearly separable using the antibody panels studied. Conclusions FlowXAI provides accurate, data-efficient, and transparent support for B-NHL immunophenotyping from nonstandardized flow cytometry data. By combining interpretable decision logic with explicit self-assessment, FlowXAI offers a clinically meaningful framework for diagnostic support and training, particularly in settings with limited expert availability or rare lymphoma subtypes. The main limitation is the retrospective evaluation using specific antibody panels, and FlowXAI requires prospective validation as a decision-support tool within integrated diagnostic workflows.