Multi-modal AI foundation models approach for diagnosis of myelodysplastic syndromes across whole bone marrow aspirate and biopsy slides.

Abstract

6571 Background: Morphologic diagnosis of myelodysplastic syndromes (MDS) from bone marrow aspirates and biopsies is subjective and error prone. Compared to narrow/specific predictive AI methods used in the literature, recently developed pathology foundation models pretrained on large, diverse histopathology corpora have the potential to improve generalizability while reducing task-specific annotation requirements. Here, we evaluate and compare multiple pretrained foundation models across bone marrow aspirate and biopsy whole-slide images (WSI), incorporate a cell-level bag-of-cells (BoC) foundation model, and assess tri-modal ensemble strategies. Methods: Digitized bone marrow aspirate (Asp) and biopsy (Bx) WSIs from patients with MDS and non-MDS cytopenias were analyzed using four pretrained pathology foundation models (GigaPath, H-Optimus-0, Virchow2-CLS, MUSK) for tile-level feature extraction. Cell-level features from aspirate-derived BoCs were extracted using DinoBloom-B, a foundation model pretrained for hematologic cell representations. Tile- or cell-level features were aggregated using attention-based multiple instance learning (ABMIL). Models were evaluated on 645 patients (319 MDS, 326 non-MDS) from the NHLBI MDS Natural History Study (NCT02775383) using 5-fold cross-validation (CV) and a held-out test set (n=129). For each modality, the best-performing model was selected based on test performance. Predictions were evaluated individually and combined using soft-probability ensembles. An oracle ensemble estimated the upper bound of multimodal complementarity. Results: Performance of the foundation models is summarized in Table 1. A soft-vote tri-modal ensemble of the selected models improved performance beyond any single modality (AUC 0.83; accuracy 0.73). Error analysis demonstrated partially non-overlapping failure modes across modalities, and an oracle ensemble achieved substantially higher accuracy (0.94), indicating unrealized multimodal complementarity. Conclusions: Foundation models enable robust AI-based diagnosis of MDS across bone marrow specimen types. Tri-modal ensembling improves diagnostic performance, while oracle analysis motivates TriPath, a unified framework for joint modeling of aspirate, biopsy, and cellular morphology to support standardized and generalizable MDS diagnosis. Performance summary. Model Bx WSI​ CV AUC Bx WSI​ Test AUC Asp WSI CV AUC Asp WSI Test AUC Asp-derived BoC CV AUC​ Asp-derived BoC Test AUC​ GigaPath+ABMIL​ 0.80+/-0.05 ​ 0.82 ​ 0.77+/-0.06​ 0.66​ ​ H-Optimus-0+ABMIL​ 0.82+/-0.04​ 0.79​ 0.77+/-0.07​ 0.68​ Virchow2-CLS+ABMIL​ 0.81+/-0.04​ 0.76​ 0.78+/-0.05 ​ 0.74 ​ MUSK+ABMIL​ 0.77+/-0.05​ 0.81​ 0.76+/-0.04​ 0.72​ DinoBloom-B+ABMIL​ ​ 0.83+/-0.05 ​ 0.78 ​ Tri-modal Soft Ensemble Test AUC 0.83 Oracle ensemble (upper bound) Test Accuracy 0.94