Self-supervised learning can distinguish myelodysplastic neoplasms from clinical mimics using bone marrow biopsies.

The diagnosis of myelodysplastic neoplasms (MDS) requires examination of the bone marrow for morphologic evidence of dysplasia. We sought to determine whether a self-supervised learning (SSL) artificial intelligence-based image analysis approach can be used to reliably distinguish MDS from its clinically relevant mimics using bone marrow biopsies (BMBx). The whole-slide images (WSIs) of hematoxylin and eosin (H&E)- and reticulin-stained BMBx sections from 243 unique patients (89 MDS, 55 non-MDS cytopenic controls [NMCCs], and 99 negative control [NC] cases) were partitioned into image tiles for analysis. These image tiles were then processed using the Barlow Twins SSL model to identify histomorphologic phenotype clusters (HPCs). Review of the HPCs revealed the clusters enriched in MDS cases captured known histopathologic features of the disease, including hypercellularity (characterized by enrichment in hypercellular image tiles), dysplastic and loosely clustered megakaryocytes, increased immature hematopoietic cells, increased vascularity, fibrosis, and cell streaming patterns. For external validation, 95 MDS BMBx WSIs from an independent institution were analyzed using the trained model. The model demonstrated consistent HPC enrichment patterns, supporting its robustness and generalizability. The trained ensemble model using H&E- and reticulin-stained slides distinguished MDS from NCs with an area under the curve (AUC) of 0.82, and from age-matched NMCCs with an AUC of 0.80. These findings demonstrate the potential of SSL approaches to capture diagnostically relevant morphologic patterns and to improve the reproducibility of MDS diagnosis.