Supporting the interpretation of cell measurements
Flow cytometry measures the properties of many individual cells. To classify B cell non-Hodgkin lymphomas, experts need to identify characteristic cell populations and assess their significance. Rare entities, different measurement conditions and limited training data complicate automated classification.
IAP developed an explainable AI prototype for Beckman Coulter and the PLAIT web platform for expert evaluation. FlowXAI combines classification with explanations and a reliability assessment for each individual sample. The scientific methods were developed with clinical and scientific partners.
Identifying unusual samples before training
The unsupervised Tile Mining method examines the structure of the measurements without requiring assigned diagnoses. It identifies samples whose data structure differs from that of other samples. These cases are flagged for targeted investigation by experts.
This structural assessment also helps select representative training examples. It supports investigation of which diagnostic groups are distinguishable in the available measurements and which examples should be used to train the subsequent classifier.
Cell populations, diagnostic levels and reliability
The supervised ALPODS method learns descriptions of diagnostically relevant cell populations. Their features and frequencies form the basis for classification. The diagnostic levels reflect clinical priorities in lymphoma diagnostics and combine several specialised decisions.
The output includes a proposed classification, the relevant cell populations and a case-level self-assessment. Experts can examine which features support a classification and which samples the AI itself considers less reliable. Explanations and self-assessment are integral to the method.
PLAIT as an interface for expert evaluation
PLAIT stands for “Phenotyping Lymphomas with an Artificial Intelligence Toolbox”. After registration and login, experts can submit anonymised FCS files through a ticket. The platform presents the analysis results in its ticket system.
Annotated cell populations can be downloaded as FCS files and investigated further in tools such as Kaluza or FlowJo. The application thereby connects automated analysis with the tools used by experts. IAP also provided the server for integration into the university network.
Scientific evaluation and implementation
The study published in PLOS Medicine in 2026 evaluated FlowXAI on 19,493 peripheral blood samples and additionally on an external dataset of 638 samples using a different antibody panel. Evaluation examined classification performance at several diagnostic levels and the information provided by case-level self-assessment.
An additional experiment investigated training with 512 representative samples: 256 normal controls and 256 lymphoma samples. Evaluation used the remaining samples. This addresses a practical research question: how can an explainable classifier be developed when only a small number of training cases is available for individual lymphoma groups?
The collaboration connects method development, published evaluation and a usable platform for expert assessment. Beckman Coulter’s original statement appears below.
