Self-explaining artificial intelligence for the classification of B cell non-Hodgkin lymphoma: A diagnostic decision support study.

Self-explaining AI combining unsupervised structural analysis and a multi-level diagnostic framework from flow cytometry data

The authors developed FlowXAI, a self-explaining AI to support B cell non-Hodgkin lymphoma (B-NHL) classification from multiparameter flow cytometry, explicitly reporting case-level diagnostic trustworthiness. The system combines unsupervised structural analysis (a Tile Mining procedure for pre-diagnostic sample-quality assessment) and a multi-level diagnostic framework reflecting routine priorities. Evaluated by repeated cross-validation on 19,493 peripheral blood samples and on an independent external dataset generated with a different antibody panel, FlowXAI matched a deep learning system while requiring roughly two orders of magnitude fewer training samples. When predictions were flagged confident by internal self-assessment, performance exceeded the neural network baseline. The main limitation is retrospective evaluation on specific antibody panels.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The main clinical value lies in transparency: an AI that reports its own uncertainty is far more deployable in routine diagnostics than a black box, especially in centers with limited expertise or for rare entities. Data-efficient training is a concrete advantage. External validation exists but remains retrospective and antibody-panel-dependent; prospective validation in integrated workflows is essential before adoption.

Analysis by Dr Thibaut Benquey