Before the algorithm can be implemented, they’ll have to determine how to integrate it with their image management system. “That would be step one, from the technical standpoint,” Dr. Jeck says. The next step would be the validation study. They would compare the algorithm’s output to the original pathologist diagnosis, “and possibly to a clinical diagnosis over time,” he says. Given that there may be examples in which a diagnosis wasn’t initially made on the biopsy but came later, “we want to devise what the gold standard would be in the full, proper context of the patient history.” Or the algorithm may be able to correctly diagnose a case that the pathologist did not. “We wouldn’t want to ding the algorithm for the fact that it’s doing better than a pathologist. Trying to anticipate that possibility is important when dealing with these algorithms.”
Because the algorithm requires human review of the slides, another validation step would involve comparing to the gold standard a pathologist’s interpretation of the slides in context with the algorithm’s results. They would want to do that for at least 60 cases, he says, but more would help to get a full sense of the algorithm’s breadth. They would also need to look at a diverse set of cases over a significant period of time.
Once the validation is complete, they would have to build out the test in production and create a system to order it, as well as one for documenting that it was used and one for ongoing monitoring, to ensure that all is working properly. And all of this, he concedes, is somewhat speculative. “That’s just the loose sketch of everything.”
Mayo Clinic in Rochester is evaluating the potential applications of AI-enhanced pipelines for analysis of measurable residual disease flow cytometry data in order to accelerate interpretation and throughput in the cell kinetics clinical laboratory.
Dr. Seheult and his colleagues first introduced AI into clinical MRD testing in 2024, when a deep neural network (DNN)-assisted, human-in-the-loop workflow was deployed for the CLL MRD assay (Seheult JN, et al. Cancers (Basel). 2025;17[10]:1688). The laboratory incorporated that model into the existing CLL workflow using a scripting interface to call the AI model from the software already in use. “Our technologists were familiar with that software solution and they didn’t want to move away from it at that time,” he says. The manual approach for the CLL assay takes about 15 minutes. With the AI-enhanced workflow, “we dropped that analysis time to under two to four minutes,” Dr. Seheult says.
The cell kinetics laboratory in 2025 began work on the next-generation CCADDAS (Clustering and Classification of All events, Dimensionality reduction, Downsampling, and Aberrancy Scaling) pipeline for multiple myeloma MRD, B-cell acute lymphoblastic leukemia (B-ALL) MRD (Seheult JN, et al. Blood Adv. Published online Sept. 18, 2025. doi: 10.1182/bloodadvances.2025016126) and acute myeloid leukemia (AML) MRD. They have also investigated the improvements that this new multistage pipeline could offer over the DNN-assisted approach for CLL MRD.
The traditional approach to MRD assessment by flow cytometry involves identifying a small number of clonal cells out of millions of events, typically by visualizing and identifying different cell populations on scatterplot matrices. The CCADDAS pipeline simplifies this approach by clustering and downsampling event populations. “Instead of a million events, you’re left with about 250,000 events, which is easier for a human to visualize on two-dimensional dot plots and a lot easier for a human to analyze,” Dr. Seheult says. It also allows the laboratory to accommodate more specimens. (Dr. Seheult highlights the vision and leadership of Pedro Horna, MD, that inspired the pipeline’s development, and acknowledges the significant contributions of Min Shi, MD, PhD; Horatiu Olteanu, MD, PhD; Gregory E. Otteson; and others.)
For the latest iteration of the pipeline, which is extensible to multiple MRD assays in the laboratory, the group made more substantial workflow changes. “We deployed the pipeline on the cloud and changed the analysis software we’re using to make it more user-friendly and interpretable and to allow easier inspection of the analysis file by multiple individuals.” Adjusting to the new workflow and addressing its limitations took more than a year, he says, before the clinical launch with multiple myeloma in late 2025 and subsequent rollout to CLL.
With the MRD assays already validated in the laboratory as laboratory-developed tests using a manual workflow, they treated the AI implementation as a test modification. “We compared the performance characteristics we observed using the AI pipeline to the previous validation to make sure all the performance claims were still met,” he says. The goal was to consistently identify and quantify MRD populations in negative, low-positive, and high-positive cases. They also verified precision and the lower limit of quantification to ensure the same sensitivity as the manual workflow.
Another step involved conducting specificity studies. “We looked at different immunophenotypes and other preanalytical variables,” Dr. Seheult says. “If you have issues with your flow cytometer, like the voltages are off, or if there are reagent issues such as lot-to-lot changes, how does that affect the outputs we’re seeing from the pipeline?”