Towards Comprehensive Quantification of Morphological and Stromal Features in Bone Marrow Using AI-Driven Image Analysis
D. Sage, R. Sarkis, L.-F. Celma, M. Krause, J. Carretti, C.R. Chardon, N. Dewarrat, T. Smirnova, L. de Leval, O. Naveiras
Colloque Français d'Intelligence Artificielle en Imagerie Biomédicale (IABM'26), Lyon, French Republic, March 9-11, 2026.
The bone marrow microenvironment is central to the persistence and recurrence of acute myeloid leukemia (AML). Recent evidence indicates that measurable residual disease could be associated with specific stromal abnormalities immediately after induction chemotherapy, indicating incomplete remodeling of the leukemic niche. However, stromal morphology is complex, heterogeneous, and spatially intertwined across very large whole-slide images, making an exhaustive and reproducible quantification impractical with manual assessment alone.
DeepMarrow is a newly funded project aiming to establish an image-analysis pipeline for the systematic and quantitative characterization of stromal features in H&E-stained bone marrow sections. We hypothesize that unexplored morphological attributes—together with spatial relationships among stromal elements, hematopoietic cells, and anatomical structures such as bone trabecular—act as phenotypic integrators of molecular cues governing response to chemotherapy. Extracting such features at single-cell scale may provide robust prognostic indicators of high-risk AML.
Methodologically, the project uses recent advances in computational pathology, including AI tools for image analysis. Building on our experience with MarrowQuant (1), the workflow combines (i) generalist deep networks for tissue and structure detection, (ii) specialized cell-level models, and (iii) custom-trained modules when suitable pre-trained solutions are unavailable. Spatial organization is quantified using graph-based and distance-based descriptors, enabling systematic evaluation of microenvironmental contexts such as the proximity of specific cell types to bone.
The entire pipeline is integrated into the open-software package, QuPath, with access to anonymized images through an OMERO database, together with interfaces for expert annotation, model fine-tuning, training, and feature extraction. This infrastructure ultimately enables translation of image-derived features into clinically actionable predictions. As an initial case study, we developed MegaQuant (2), focused on megakaryocytes—precursors of platelets and a diagnostically relevant but morphologically challenging cell type. In H&E-stained sections, megakaryocytes display broad morphological variability and complex nuclear structures, complicating precise segmentation. To facilitate ground-truth creation, annotations were performed both manually and with assistance from the Segment Anything Model (SAM). SAM drastically accelerated annotation while maintaining quality, with all outputs validated by pathologists. These annotations were then used to train a customized Cellpose model, incorporating iterative refinement between detections and expert corrections. Nuclei detection using an adapted StarDist model enabled extraction of single-cell descriptors aligned with clinical assessment practices. MegaQuant thus provides the foundation for comprehensive quantification of stromal features.
@INPROCEEDINGS(http://bigwww.epfl.ch/publications/sage2601.html,
AUTHOR="Sage, D. and Sarkis, R. and Celma, L.-F. and Krause, M. and
Carretti, J. and Chardon, C.R. and Dewarrat, N. and Smirnova, T. and
de Leval, L. and Naveiras, O.",
TITLE="Towards Comprehensive Quantification of Morphological and Stromal
Features in Bone Marrow Using {AI}-Driven Image Analysis",
BOOKTITLE="Colloque Fran{\c{c}}ais d'Intelligence Artificielle en
Imagerie Biom{\'{e}}dicale ({IABM'26})",
YEAR="2026",
editor="",
volume="",
series="",
pages="",
address="Lyon, French Republic",
month="March 9-11,",
organization="",
publisher="",
note="")