Attention-Based Multiple Instance Learning for Cellularity Estimation in Bone Marrow Core Biopsies

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

Keywords: Bone Marrow, Core Biopsy Cellularity, Attention-Based Multiple-Instance Learning, DINOv2

TL;DR: Using an ABMIL based regressor with custom fine-tuned DINOv2 ViT, we're able to achieve good performance at estimating bone marrow core cellularity.

Abstract:

Estimation of cellularity in bone marrow tissue specimens is a routine task, useful in the diagnosis and monitoring of hematologic diseases. In this work we train a weakly supervised regressor on a large dataset of bone marrow biopsy slides with LLM parsed cellularity
estimates from case reports. We utilize attention-based multiple instance learning and a DINOv2 vision transformer model. The model is trained and validated on a set of over $5600$ bone marrow biposy slide scans achieving a $R^2$ of $0.776$ on our test set.

Submission Number: 37