Revealing Morphometric Differences between Hematopoietic and Leukemic Stem Cells in Acute Myeloid Leukemia: AI-Based Analysis of Imaging Flow Cytometry Data

Acute myeloid leukemia (AML) is a heterogenous blood cancer originating from genetic transformation of hematopoietic stem cells (HSCs) into aberrant leukemic stem cells (LSCs), causing accumulation of immature leukemic blasts and suppression of healthy hematopoiesis in the bone marrow (BM). The frequency of LSCs holds strong prognostic value in AML, and LSC surveillance during treatment improves minimal residual disease monitoring. However, LSC detection and differentiation from healthy HSCs remain challenging due to substantial heterogeneity in immunophenotypes. Therefore, new techniques are warranted for HSC-LSC discrimination. In this regard, imaging flow cytometry (IFC) combined with artificial intelligence has proven a novel tool for morphometric cell characterization. Recently, we have described the application of deep learning on IFC data to classify stem cells as either CLEC12A⁺ aberrant AML LSCs or CLEC12A^- healthy HSCs, achieving up to 93% accuracy using brightfield (BF), side scatter (SSC), and DNA imagery (Hybel et al. 2024). In the present study, we have analyzed the image data to identify the discriminative cellular features that may have contributed to the previously observed results. Using the IDEAS software, a total of 85 morphometric feature values, including 31 BF, 36 SSC, and 18 DNA features, were generated for 44,277 CLEC12A⁺ LSCs and 8,817 CLEC12A^- HSCs acquired by IFC from five AML BM samples and ten healthy donor BM samples, respectively. A logistic regression model was trained for HSC-LSC differentiation and revealed significant differences in 17 BF features (e.g., cellular area, compactness, and contrast), 22 SSC features (e.g., intensity, gradient root mean square, and perimeter), and 15 DNA features (e.g., nuclear area, circularity, and lobe count). Following class balancing by oversampling, an overall accuracy of 78%, sensitivity of 76%, and specificity of 81% were achieved. Moreover, unsupervised principal component analysis was performed on LSCs and HSCs, both separately and combined. However, HSCs and LSCs were not segregated into distinct groups. The analyses revealed subtle differences in LSC morphometry between different AML patients, whereas HSCs from healthy donors displayed morphometric similarity. Our results support the use of IFC as a technique for characterizing pathological changes within stem cells in AML. The morphometric differences described between HSCs and LSCs may facilitate future development of innovative methodologies for monitoring minimal residual disease within AML.

No relevant conflicts of interest to declare.