Deep learning predicts IDH1 mutations from leukemia bone marrow images
Published in Artificial Intelligence, Journals, Personalized medicine, Scanners and Imaging.
Published October 29, 2026.
A deep-learning model predicts IDH1 mutation status from digitized bone marrow smears in acute myeloid leukemia (AML). Eckardt and colleagues report the multicenter study today in npj Precision Oncology.
The researchers used whole-slide images from 436 patients across four centers in Germany and the United States. Their attention-based multiple-instance learning approach achieved a reported mean pooled ROC-AUC of 0.81.
This metric describes how well the model separates mutation-positive and mutation-negative cases across decision thresholds; it is not equivalent to 81% diagnostic accuracy.
Analysis of image regions receiving high attention linked the predictions to immature leukemia phenotypes. The findings suggest that routine cell morphology contains signals associated with a therapeutically relevant genetic alteration.
The study supports further evaluation of image-based molecular screening. It does not establish that images can replace confirmatory molecular testing, or show that using the model improves treatment outcomes. Prospective assessment and validation in additional clinical settings remain important before routine deployment.
Source: Eckardt et al., npj Precision Oncology. Published October 29, 2026. DOI: 10.1038/s41698-026-01762-x.