Detection of Acute Myeloid Leukemia from Bone Marrow Aspirate Smears using Self-Supervised Cell Representations

Brendan O'Fallon, Muir Morrison, Alexandra Rangel, Lauren M. Zuromski, Paul English, David Ng

Keywords: Bone marrow, acute myeloid leukemia, self-supervised learning, deep learning, multiple instance learning

TL;DR: We demonstrate that a cell image representations generated by self-supervised learning are useful for identifying Acute Myeloid Leukemia in bone marrow aspirate smears

Abstract:

Automated detection of hematopathological disorders such as Acute Myeloid Leukemia (AML) from bone marrow images has the potential to improve the speed and accuracy of diagnosis. However, most deep learning methods for AML detection require large numbers of labeled cell images, which can require many hours of expert annotation and review. Here we compare a self-supervised model trained using DINOv2 with a fully supervised model trained on a large dataset of labeled cell images, and evaluate both models on an AML detection task. We find that the self-supervised model marginally outperforms the fully supervised model (AUROC 99.3 vs 98.4), suggesting that cell representations generated via self-supervised methods may lead to improved disease classification while obviating the need for manual annotation of thousands of cell images.

Submission Number: 32