Automated analysis of flow cytometry data with minimal training files: Research evaluation of an elastic image registration algorithm for TBNK, stem cell enumeration, and lymphoid screening tube assays
Allison Irvine,Suhail Tahir,Vishnu Tripathi,Farzad Oreizy,Moen Sen,Anthony Giuliano,Anna Lin,Angela Chen,Chih‐Hung Lai,Imelda Omana‐Zapata,Yang Zeng,Paresh Jain,Scott J. Bornheimer
DOI: https://doi.org/10.1002/cyto.b.22210
2024-10-19
Cytometry Part B Clinical Cytometry
Abstract:Automated analysis of flow cytometry data can improve objectivity and reduce analysis time but has generally required work by software and algorithm experts. Here, we investigated the performance of BD ElastiGateTM Software (hereafter ElastiGate), which allows users to automate gating by selecting gated training files, then uses elastic image registration to gate new files. Three assays of increasing complexity were examined: TBNK, stem cell enumeration (SCE), and lymphoid screening tube (LST). For TBNK analysis, 60 peripheral blood (PB) samples from normal, HIV+, and controls were tested with ground truth analysis by an existing automated method. For SCE, 128 samples including bone marrow (BM), cord blood (CB), and apheresis were tested with analysis by multiple manual analysts. For LST, 80 PB and 28 BM samples were tested with manual analysis. For ElastiGate, a minimal number of training files was selected. Results were compared by Bland–Altman or F1 score analysis. For TBNK, ElastiGate using three training files (1 control, 1 normal, 1 HIV+) showed mean %bias across all reported populations between −1.48% and 7.13% (average 2.08%). For SCE, ElastiGate using three BM and two CB training files showed median F1 scores >0.93 in comparison to >0.94 and >0.92 for two other manual analysts. For LST, ElastiGate using four training files for each of PB and BM showed median F1 scores >0.945 for 13 of 14 PB populations and 10 of 14 BM populations, with generally similar or better performance for normal samples compared to abnormal; populations with lower scores were often associated with lower agreement between manual analysts. Based on analysis of three assays with four sample types of increasing complexity, ElastiGate with minimal training files may perform as an automated gating assistant. The results reported here are for research use only, not for use in diagnostic or therapeutic procedures.
pathology,medical laboratory technology