AI Model Aims to Significantly Simplify Blood Cancer Diagnosis

“This will allow us to evaluate blood samples—which are collected in clinical settings anyway—in a much more automated manner in the future,” says Dr. Florian Ingelfinger, first author of the study and researcher at the Department of Internal Medicine I at the University Medical Center Freiburg and junior research group leader at the German Consortium for Translational Cancer Research (DKTK), Freiburg partner site. “The model reduces the workload on hospitals while also helping to discover new cell populations that are of interest as diagnostic markers or potential targets for therapies.” Ingelfinger heads the Laboratory for Generative Single Cell Immunology at the University Medical Center Freiburg.

From Research to Routine Diagnostics

CytoVI analyzes single-cell data generated by flow cytometry. No images are captured in the process. Laser measurements are used to record the properties of individual cells, such as specific proteins, cell size, and internal structure. This creates a digital profile for each cell. CytoVI recognizes typical patterns in these profiles and can, for example, automatically distinguish leukemia cells from other blood cells. A key strength of the model is that it mathematically accounts for technical differences between various measurements—such as those caused by different devices, measurement times, or laboratories—and helps distinguish them from biological differences. This enables a more comparable analysis.

“Above all, patients need a reliable diagnosis and clarity as quickly as possible about what disease they have. CytoVI could help us with this in the future,” says Prof. Dr. Robert Zeiser, head of the Department of Tumor Immunology and Immunoregulation at the Division of Internal Medicine I at the University Medical Center Freiburg, who contributed to the publication. In 63 female patients with various B-cell lymphomas, CytoVI automatically identified T-cell groups associated with specific forms of the disease. Among these was a rare cell population that had not been described in the original manual analysis. Such cell groups can provide clues to diagnostic biomarkers and potential therapeutic targets.

The analysis of CLL data from routine diagnostics was particularly clinically relevant. “We were able to see that the automatic identification corresponds very well with the manual evaluation,” says Ingelfinger. The reliable identification of very few cancer cells—for example, following cancer treatment—remains a challenge at present. The reliability of the analysis is provided to physicians in each case, enabling targeted reviews. “CytoVI is intended to support the physician’s assessment, not completely replace it,” Ingelfinger adds.

In this way, the large volumes of flow cytometry data generated over many years in routine diagnostics can be more easily compared and evaluated collectively. This unlocks a treasure trove of data for research that was previously of only limited use. Researchers at the University Medical Center Freiburg now aim to apply the method to other types of blood cancer. CytoVI is freely available as open-source software at https://scvi-tools.org. Researchers can use the method to analyze their own single-cell data.