2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Saturday, December 12, 2026
05:30 p.m. - 07:30 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Imaging, Molecular Testing, Omics Technologies, Pathology, Bioinformatics, Blood Banking, Computational Biology, Emerging Technologies, Profiling, Treatment Considerations, Drug Development, Supportive Care, Therapy Sequence, Biological Therapies, Chimeric Antigen Receptor, (CAR)-T Cell Therapies, Women’s Health, Workforce, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Generative AI, Machine Learning, Natural Language Processing, Biological Processes, Apoptosis, Genomics, Hematopoiesis, Molecular Biology, Clinical Practice (Health Services and Quality), Clinical Guidelines, Quality Improvement, Quality of Care, Systems-Based Hematology, Diseases, Bleeding and Clotting, Thromboembolism, Bone Marrow Failure Syndromes, Inherited Marrow Failure Syndromes, Paroxysmal Nocturnal Hemoglobinuria, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Plasma Cell Disorders, Myeloid Malignancies, Acute Myeloid Malignancies, Chronic Myeloid Malignancies, Metabolic Disorders, Iron Deficiency, Measurable Residual Disease, Non-Hodgkin Lymphoma, B Cell Lymphoma, Aggressive Lymphoma, Research, Fundamental Science, AML, APL, Translational Research, MDS, MPN, Clinical Research, Clinical Trials, Real-World Evidence, Study Population, Human, Adult, Pediatric
PAPER

2560 Bloodbytes: A bone marrow biopsy pathology foundation model for automated diagnostic, morphologic, and molecular assessment

Steven Huang, Reyhan Keser, PhD2, Yanglan Ou, D.Sc., M.E.3, Prasad Bandodkar, Ph.D.4, Gregory Otteson5, Zachary Larson6, Xi Zhang, MD, PhD7, Clarissa E Jordan, MD8, Jess Peterson, M.D.9, Pedro Horna, M.D.10, Aref Al-Kali, M.D.11, Min Shi, M.D.12, Horatiu Olteanu, M.D., Ph.D.13, Dragan Jevremovic, MD PhD14, Tariq Kewan, MD15, Eric Hsi, M.D.16, Dong Chen, MD, PhD17, Jansen N Seheult, MBBCh, BAO, MSc, MS, MD18, Wenchao Han, PhD19
1 Mayo Clinic, Rochester, Minnesota, United States
2 Mayo Clinic, Rochester, Minnesota, United States
3 Mayo Clinic, Rochester, Minnesota, United States
4 Mayo Clinic, Rochester, Minnesota, United States
5 Mayo Clinic, Rochester, Minnesota, United States
6 Mayo Clinic, Rochester, Minnesota, United States
7 Mayo Clinic, Rochester, Minnesota, United States
8 Mayo Clinic, Rochester, Minnesota, United States
9 Mayo Clinic, Rochester, Minnesota, United States
10 Mayo Clinic, Rochester, Minnesota, United States
11 Mayo Clinic, Rochester, Minnesota, United States
12 Mayo Clinic, Rochester, Minnesota, United States
13 Mayo Clinic, Rochester, Minnesota, United States
14 Mayo Clinic, Rochester, Minnesota, United States
15 Mayo Clinic, Rochester, Minnesota, United States
16 Mayo Clinic, Rochester, Minnesota, United States
17 Mayo Clinic, Rochester, Minnesota, United States
18 Mayo Clinic, Rochester, Minnesota, United States
19 Mayo Clinic, Rochester, Minnesota, United States

Background: Morphologic assessment of the bone marrow core biopsy is central to the diagnostic, prognostic, and theragnostic evaluation of hematologic diseases. Unlike aspirate smears, core biopsies preserve tissue architecture, cellular relationships, and the marrow microenvironment, and their interpretation requires integrating information across spatial scales. General-purpose pathology foundation models (FMs) are predominantly trained on solid-organ specimen pathology, in which bone marrow is underrepresented or absent. We hypothesized that domain-specialized training with bone marrow specimens would yield more clinically useful representations for disease-specific applications than generalist FMs trained with substantially larger amounts of training data and developed BloodBytes on that basis. 

Methods: BloodBytes is a 1.1-billion-parameter vision transformer (i.e., ViT-G/14) trained with the DINOv2 self-supervised framework on 11 million 256x256-pixel patches sampled at 5x, 10x, 20x, and 40x from 9,457 H&E-stained bone marrow core biopsy whole-slide images (one slide per case, 2023 to 2025, two scanner platforms). Patches were selected by content-aware sampling within U-Net-derived tissue regions. This sample size is roughly 35-fold smaller than UNI-v2 and 310-fold smaller than Virchow-v2 in slide count. Case-level performance was evaluated with an attention-based multiple-instance learning on six tasks for classification: five-category diagnostic classification [acute myeloid leukemia (AML), plasma cell neoplasm (PCN), myeloproliferative neoplasm (MPN), myelodysplastic syndrome (MDS) or MDS/MPN, negative/ reactive], megakaryocyte morphology classification, fibrosis grading, cellularity assessment, blast percentage dichotomized at 20%, and bi-allelic/multi-hit TP53 alterations in AML. Classification labels were derived from pathology reports. Each task was evaluated across four randomized train/validation/test splits (50:25:25%), with pretraining slides excluded from held-out evaluation. The state-of-the-art general-purpose pathology FMs, UNI-v2 and Virchow-v2, were benchmarked under identical patch extraction, aggregation, and splits. 

Results: Performance was summarized as area under the receiver operating characteristic curve (AUC) and macro-weighted F1 score, averaged across all folds. BloodBytes achieved the highest AUC and F1 on every task. On the unbalanced five-category classification task at 20x (n = 558), BloodBytes reached a mean AUC (F1) of 0.92 (0.74), versus 0.90 (0.67) for UNI-v2 and 0.89 (0.66) for Virchow-v2. For fibrosis grading at 10x (n = 268), BloodBytes achieved a mean AUC (F1) of 0.89 (0.72) versus 0.84 (0.59) and 0.86 (0.67) for Uni-v2 and Virchow-v2, respectively, and it likewise outperformed both FMs on identifying megakaryocyte dysplasia at 5x [0.88 (0.83) versus 0.83 (0.80) and 0.82 (0.77), n = 562] and cellularity assessment at 5x [0.93 (0.79) vs 0.92 (0.75) and 0.92 (0.77), n = 587]. Blast quantification at 20x, evaluated in 445 biopsies (171 <20%, 274 ≥20%), showed comparable discrimination across all FMs, with BloodBytes reaching an AUC (F1) of 0.95 (0.92) against 0.95 (0.90) and 0.95 (0.89). The largest gain was for detecting bi-allelic/multi-hit TP53 alterations at 40x evaluated with 171 cases (normal karyotype versus bi-allelic/multi-hit TP53 alterations): 0.95 (0.85) for BloodBytes vs 0.75 (0.69) and 0.70 (0.67) for Uni-v2 and Virchow-v2.  

Conclusion: BloodBytes, a bone marrow-specific FM, generated representations that transferred more effectively than general-purpose pathology FMs across diagnostic, morphologic, and molecular endpoints, despite pretraining on an order of magnitude fewer slides. This was most notable for multi-hit TP53 alterations, which define an aggressive, treatment-resistant disease with dismal survival and have no established morphologic correlate. BloodBytes recovered an image signal from routine H&E under identical patch extraction and splits, whereas generalist FMs largely did not. These findings indicate that domain specificity and multi-resolution pretraining can matter as much as corpus scale, and position BloodBytes as a foundation for scalable AI-assisted bone marrow assessment. External validation and multimodal integration are ongoing.

Steven Huang: No relevant conflicts of interest to declare.