A Deep-Learning Pipeline for Diagnosis of Myelodysplastic Syndromes/Neoplasms Using Bone Marrow Smears

Background: A diagnosis of myelodysplastic syndromes/neoplasm (MDS) is predicated on visual review of peripheral blood and bone marrow (BM) smears by a specialized hematopathologist to identify morphologic dysplasia. This process can be subjective and prone to diagnostic challenges. Prior analyses from the MDS Natural History Study (NHS) demonstrated discordance between local and central pathology review, which may impact treatment decisions (Gorak et al. Blood Advances 2023). This is increasingly problematic in the current era with other causes of cytopenias and pre-MDS conditions such as Idiopathic Cytopenia of Unknown Significance (ICUS) and Clonal Cytopenia of Unknown Significance (CCUS) where morphology may be abnormal but does not reach the level of dysplasia required for formal MDS diagnosis. Automation of BM smear examination has been limited by the high complexity of maturing BM cellular components and wide spectrum of dysplastic morphologies. In this study, we propose the first deep learning pipeline to discriminate MDS from ICUS and CCUS using digitized BM aspiration smears. The AI-augmented BM smear review may aid hematopathologists through standardization of the diagnostic process while increasing precision by enabling analysis of a larger number of nucleated cells. Methods: We developed a deep-learning pipeline to discriminate MDS from other cytopenia conditions. We evaluated 645 patients from the MDS NHS (ClinicalTrials.gov Identifier: NCT02775383). There are 319 MDS, 276 CCUS, and 50 ICUS cases in the dataset. CCUS and ICUS were combined into one category as “non-MDS” given either lack of morphologic dysplasia or <10% dysplasia and <5% blasts. We also evaluated the method on a small set of 31 patients (26 non-MDS and 5 MDS) from Moffitt Cancer Center as an external test set. This solely examined morphology in these patients without input of other clinical or molecular features. Details of the deep learning pipeline are as follows: The first step is a DenseNet model for region of interest (ROI) selection from BM aspiration smears. Next, a YOLO model was deployed for nucleated cell detection from the ROI tiles. We used pre-trained DenseNet and YOLO models (Tayebi et al. Communications medicine 2022) to avoid manual annotation. A bag of 500 nucleated cells was created for each patient following pathological guidelines to gauge the percentage of dysplastic and/or blast cells. The automatic pipeline leveraged a hybrid Vision-Transformer (ViT) model for aggregation of information across the cells in a bag-of-cells to gauge the dysplastic and/or blast cells percentage. Transfer-learning from the ImageNet dataset was employed to address the need for a larger training-set for ViT model. The use of the ViT model for aggregation eliminates the need for cell level annotation of dysplastic vs normal cell which is one of the major bottlenecks in development of machine learning methods for BM smear examination. Results: The pipeline achieved 5-fold cross-validation AUROC of 0.78+/-0.02 and test AUROC of 0.77 with 68.99% test accuracy for the MDS NHS dataset. The sensitivity and specificity on the MDS NHS test set was 0.63 and 0.75, respectively. The AUROC was 0.90 with an accuracy of 64.52% for the 31 patients in the external test set. The sensitivity and specificity for the same was 0.80 and 0.62 in that order. Conclusion: In summary, an AI pipeline is presented for automatic examination of BM aspiration smears for discrimination of MDS from pre-MDS conditions. We show promising results on a multi-institutional dataset. Future direction involves validation of this method on a larger external test set, aggregation of information from BM aspiration and biopsy, and investigation of high-risk vs low/intermediate risk CCUS cases.