Improving diagnosis, classification, and prognosis of MDN with an AI-based model

A ‘foundation model’, extracting features from images, was associated with improvements in the diagnosis, classification, and prognostication of patients with myelodysplastic neoplasm (MDN). Currently, the team that designed this successful model is scaling up to facilitate implementation in clinical practice.

“Bone marrow cytology and histopathology images are highly important for the classification and prognostication of myeloid neoplasms,” according to Mr Gianluca Asti (IRCCS Humanitas Research Hospital, Italy) [1]. “Since these data are difficult to analyse, we explored the potential of an artificial intelligence (AI)-based model to improve the situation.” The team derived whole slide images from 1,167 patients with myeloid neoplasms and looked at diagnostic accuracy and various features in the context of personalised risk assessment.

The diagnostic accuracy of the model was high for the MDN training set (AUROC 0.96) and the test set (AUROC 0.90). Interestingly, the accuracy of the personalised prediction of overall survival was lower if the team only used clinical features (C-index 0.78) than if they added genomic and karyotype features (C-index 0.82), or even imaging features (C-index 0.88). The corresponding C-indexes for leukaemia-free survival were 0.68, 0.80, and 0.90, displaying the improved prognostication that can be achieved with the model.

“Our Foundation model, designed to extract complex features from whole slide images, provided significant advancements in the diagnosis, classification, and prognostication of MDN,” Mr Asti summarised the findings. “Currently, we are working on the PATHroclus federated platform to improve our model, develop a comprehensive database, and create a virtual atlas for haematological malignancies. It is our goal to design next-generation classifications and improve diagnostic standards and reproducibility across Europe.”

1. Asti G, et al. A fine-tuned foundation model for digital pathology in myelodysplastic syndromes: advancing personalized medicine. S331, EHA2025 Congress, 12–15 June, Milan, Italy.

Copyright ©2025 Medicom Medical Publishers



Posted on 2 September 20252 September 2025