A vision foundation model approach to quantitative bone marrow fibrosis assessment
Justin Johnson1, Mohamed E Salama2,3
1New York University Grossman School of Medicine, Department of Pathology, New York City, NY, United States of America.
Background:
Bone marrow reticulin fibrosis is prognostic in myeloproliferative neoplasms (MPNs). However, fibrosis assessment is limited by interobserver variability. Computational approaches address this, but remain anchored to qualitative assessment, have limited scalability, and none have explicitly evaluated marrow adiposity. Here we developed a pipeline to predict fibrosis scores using vision-transformer (ViT) models.
Methods:
Slide from 643 bone marrow cores from 391 patients with a MPN were digitized and analyzed. Cores were tiled, and Merz grids were overlaid on whole-slide images. Fiber-grid intersections were automatically counted, and ViT models were developed using grid scores as labels. Fat-excluded models were also developed, and sampling strategies were swept. All-length score reflects reticulin intersections relative to all eligible marrow-core grid length, whereas the fat-excluded score reflects reticulin intersections relative to the non-fat eligible marrow compartment. Performance was tested on an unseen patient cohort.
Results:
Systematic-random sampling of 64 (SR64) tiles per slide had the strongest validation performance across sampling methods, whereas random sampling of 16 tiles showed the strongest correlation with the prior validated manual Merz-grid method. On testing, the SR64 all-length model achieved root mean squared error (RMSE) 2.97, R 2 0.86, and Pearson r 0.93. Meanwhile, SR64 fat-exclusion resulted in RMSE 2.47, R 2 0.87, and Pearson r 0.93. Compared with manual Merz-grid measurements, all-length predictions showed stronger correlation (Pearson r = 0.745) than fat-excluded predictions (r = 0.439) and automated all-length Merz counting (r = 0.687). Cluster visualization demonstrated plausible tile phenotypes associated with model predictions.
Conclusions:
ViT-based models can predict automated Merz-derived bone marrow fibrosis. The all-length model strongly correlated with the prior validated manual Merz-grid method. Meanwhile, limited tile sampling preserved strong performance. These findings support a scalable and reproducible framework for fibrosis quantification in myeloproliferative neoplasms.
