Myelodysplastic syndrome / neoplasm (MDS) presents a diagnostic challenge due to the need of expert morphological analysis, and the reliance on genomic analysis of collected bone marrow material for the definite diagnosis. This study aimed to facilitate this process by developing a computer vision AI-based model that is capable of diagnosing MDS from images from peripheral blood smears (PBS). We used a cohort of 43,371 neutrophils from 84 MDS and 60 non-MDS samples to train a neutrophil classifier to differentiate between dysplastic and non-dysplastic cells. The model was initially fed with PBS images from patients with prominent MDS (pMDS) and further refined to detect non-prominent MDS (npMDS), i.e., without clear-cut dysplastic features in their neutrophils. The model learning was only based on the single-cell annotation of the neutrophils from pMDS, without human-generated morphological features as input. The trained neutrophil classifier achieved an overall accuracy of 94%, with a sensitivity and specificity of 0.95 and 0.94, respectively. On a patient-level, the model correctly identified 91 out of the 94 samples, with a sensitivity and specificity of 0.98 and 0.96, respectively, and AUC of 0.999. In npMDS, 43 out of the 44 samples were correctly identified. Our study demonstrates the potential of AI-based models to improve the efficiency of MDS diagnostics. Our model runs on standard CPUs, offering an accessible solution that can be integrated into existing clinical workflows and potentially reduces the dependence on specialized morphologists and genomic analysis from bone marrow punctures.
Romano et al. (Tue,) studied this question.