Bone Marrow Fibrosis Grading Using Attention-Based Multiple Instance Learning

Lauren M. Zuromski, Alexandra E. Rangel, Muir J. Morrison, Paul English, Nicholas C. Spies, Brendan O'Fallon, David P. Ng

Keywords: Bone marrow fibrosis, multiple instance learning, whole slide imaging, digital pathology, large language model, DINOv2

TL;DR: We demonstrate that an attention-based multiple instance learning approach improved bone marrow fibrosis grading from whole slide images of bone marrow core biopsies.

Abstract: An attention-based multiple instance learning approach is used to improve bone marrow fibrosis (BMF) grading from whole slide images of bone marrow core biopsies. Slide-level labels were parsed from biopsy reports using a large language model, and features were extracted using our fine-tuned DINOv2-based backbone. The model achieved good agreement between BMF predictions and labels $(R^2 = 0.72,\ \kappa = 0.58)$. Attention maps showed the model focused on diagnostic regions, highlighting its accuracy and interpretability.

Submission Number: 36