Abstract 3870: Deep learning-powered morphological analysis of acute myeloid leukemia.

Abstract In this study we demonstrate that deep learning-based analysis of single cell images can distinguish acute myeloid leukemia (AML) bone marrow (BM) mononuclear cells (MNCs) from healthy BM MNCs based on morphological features. We further show that differences in ex vivo drug responses to BCL2 inhibitor venetoclax links to shifts in morphology of venetoclax-sensitive (ven-sen) and -resistant (ven-res) AML. Our sample cohort consisted of cryopreserved BM MNC samples from 6 healthy donors and 12 donors with AML. The AML samples were selected based on ex vivo sensitvity to venetoclax (6 ven-sen + 6 ven-res). Moreover, 2 longitudinal sample pairs (diagnosis-relapse) from 2 AML patients were included in the cohort. The AML samples were from patients recruited to the NCT04267081 trial and healthy donor samples from patients undergoing hip replacement surgery. Samples were thawed, fixed with 2% paraformaldehyde, and stored at +4°C until the time of imaging. The samples were prepared in a single cell suspension and 25,000-50,000 brightfield images collected per sample using the high-resolution single cell imaging and sorting platform, REM-I (Deepcell). Imaging data were synced to REM-I’s data suite Axon. Morphological analysis through Axon was carried out based on 115 dimensions of cell morphology; 51 human-interpretable, and 64 deep-learning features. Analysis parameters were set to random sampling with equal number of data points per sample. Differential morphology analysis on Axon yielded divergence scores (Range: 0-1) as a quantified measure of morphological distinction.Comparison of healthy samples (n=6) against pooled AML samples (n=12) demonstrated that healthy and AML BM cells are distinct in their morphology and map into different clusters in a uniform manifold approximation and projection (UMAP) graph. Differential morphology analysis between these sample groups highlighted deep-learning embeddings in the top 10 differential morphology features, with divergence scores ranging from 0.55-0.65. Additionally, an analysis comparing healthy samples (n=6) to ven-sen (n=6) and ven-res (n=6) AML, revealed that the two different AML sample groups were distinct in morphology from each other as well as from the healthy samples. Finally, in the analysis of two longitudinal AML patient sample pairs for morphological analysis of AML progression, we observed changes in morphology between the diagnosis and relapse samples. This anaysis also revealed a substantial morphological shift in the patient that had a higher change in ex vivo ven response between the diagnosis and relapse samples. Our study demonstrates that cutting-edge methods powered by deep learning can facilitate the analysis of morphological features of AML with high-dimensionality that cannot be achieved by methods that depend only on human perception. These results indicate that drug response and disease progression in AML are reflected by changes in cell morphology. Citation Format: Ezgi June Olgac, Meghan Burr, Minna Suvela, Daniela Mendoza-Ortiz, Ella Sinervuori, Heikki Kuusanmäki, Mika Kontro, Caroline Heckman. Deep learning-powered morphological analysis of acute myeloid leukemia abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 3870.