AUTHORS
Prasad Bandodkar, Ph.D., Reyhan Keser, PhD2, Steven Huang3, Zachary Larson4, Yanglan Ou, D.Sc., M.E.5, Vincent Olteanu6, Wilson Gonsalves, MD7, Saurabh Zanwar, MBBS, MD8, Pedro Horna, M.D.9, Clarissa E Jordan, MD10, Linda Baughn, PhD11, Richard Kandasamy, PhD12, Nicholas Jeffrey Braun, PhD13, Xi Zhang, MD Phd14, Jess Peterson, M.D.15, Moritz Binder, MD, MPH16, Eric Hsi, M.D.17, Min Shi, M.D.18, Dong Chen, MD, PhD19, Horatiu Olteanu, M.D., Ph.D.20, Dragan Jevremovic, MD PhD21, Shaji Kumar, MD22, Wenchao Han, PhD23, Jansen N Seheult, MBBCh, BAO, MSc, MS, MD24
AFFILIATIONS
1 Mayo Clinic, Rochester, MN, United States
2 Mayo Clinic, Rochester, MN, United States
3 Mayo Clinic, Rochester, MN, United States
4 Mayo Clinic, Rochester, MN, United States
5 Mayo Clinic, Rochester, MN, United States
6 Tufts University, Medford, Massachusetts, United States
7 Mayo Clinic, Rochester, Minnesota, United States
8 Mayo Clinic, Rochester, Minnesota, United States
9 Mayo Clinic, Rochester, MN, United States
10 Mayo Clinic, Rochester, MN, United States
11 Mayo Clinic, Rochester, MN, United States
12 Mayo Clinic, Rochester, MN, United States
13 Mayo Clinic, Rochester, Minnesota, United States
14 Mayo Clinic, Rochester, MN, United States
15 Mayo Clinic, Rochester, MN, United States
16 Mayo Clinic, Rochester, Minnesota, United States
17 Mayo Clinic, Rochester, MN, United States
18 Mayo Clinic, Rochester, MN, United States
19 Mayo Clinic, Rochester, MN, United States
20 Mayo Clinic, Rochester, MN, United States
21 Mayo Clinic, Rochester, MN, United States
22 Mayo Clinic, Rochester, Minnesota, United States
23 Mayo Clinic, Rochester, MN, United States
24 Mayo Clinic, Rochester, MN, United States
ABSTRACT
Background: Multiple myeloma has a highly heterogeneous clinical course. Risk stratification depends largely on genetics: high-risk lesions include del(17p)/TP53 mutation and t(4;14), among other primary IGH translocations, while standard-risk t(11;14) predicts sensitivity to targeted BCL2 inhibitors. S-phase fraction identifies high-risk disease independently of these molecular alterations. Nearly all require specialized assays, principally plasma cell-sorted fluorescence in situ hybridization (FISH) and DNA-content flow cytometry. S-phase fraction is not routinely available at most institutions, given assay complexity and competing demands on limited aspirate material. Immature plasma cell morphology has been associated with high proliferative rate and inferior outcome. An AI morphology-based surrogate derived from routine bone marrow images could provide immediate prognostic value to patients who currently lack access to this marker.
Methods: We evaluated Wright-Giemsa-stained bone marrow aspirates (BMA) and hematoxylin-and-eosin (H&E)-stained core biopsy sections from approximately 300 multiple myeloma cases with ground-truth labels from flow cytometry and FISH. For BMA, cell crops were captured from 60x oil full field bone marrow aspirate scans (Bionovation CSFA800-30) and encoded using DinoBloom (Koch et al., MICCAI 2024). For core biopsies, we used an in-house DINOv2 image encoder (BloodBytes) pretrained on H&E images of archival bone marrow biopsies (Pramana Spectral HT and Leica GT450). Per-case predictions were generated using attention-based multiple instance learning (MIL). We trained separate models for: (1) S-phase status (high ≥2% vs. low <2%), and (2) select cytogenetic abnormalities [t(4;14), t(11;14), and TP53 deletion]. To validate that the model could recognize morphologic features of plasma cells, we evaluated semiquantitative plasma cell assessment as a binary task (low ≤10% vs. high >10% PC). Performance was evaluated using 4-fold cross-validation without cohort-specific calibration, reporting AUC, accuracy, F1 score, and sensitivity.
Results: Using BMA morphology alone, the strongest discriminative performance was observed for S-phase status (n=70 test cases, 28 ≥2%; AUC 0.84 ±0.05, F1 0.76 ±0.02) and t(11;14) detection (n=29, 9 positive; AUC 0.84 ±0.05, F1 0.75 ±0.04), with sensitivities of 82±8% and 83±10%, respectively. TP53 deletion status was predicted with an AUC of 0.81 ± 0.03 and an F1 score of 0.69 ± 0.07 in 50 cases, including 12 positive cases. Plasma cell quantification (n = 71; 33 ≤10% and 38 >10% plasma cells) achieved an AUC of 0.87 ± 0.05, an F1 score of 0.78 ± 0.06, and a sensitivity of 0.79 ± 0.06. Performance for t(4;14) detection was more modest (n = 29; 5 positive; AUC, 0.70 ± 0.11; sensitivity, 0.77 ± 0.18). Attention analysis showed that the aspirate S-phase model assigned high attention almost exclusively to plasma cells.
Using biopsy morphology, t(11;14) detection showed the strongest performance (AUC, 0.78 ± 0.09; F1, 0.66 ± 0.07; sensitivity, 0.67 ± 0.06). Performance was lower than that of the corresponding BMA models for S-phase status (AUC, 0.76 ± 0.07; F1, 0.69 ± 0.07), TP53 deletion status (AUC, 0.66 ± 0.10; F1, 0.64 ± 0.07), plasma cell quantification (AUC, 0.70 ± 0.02; F1, 0.69 ± 0.01), and t(4;14) detection (AUC, 0.63 ± 0.06; F1, 0.55 ± 0.08). The core biopsy model preferentially attended to regions with dense plasma cell infiltration.
Conclusions: Bone marrow aspirate morphology enabled prediction of S-phase status, t(11;14), and TP53 deletion with excellent discrimination. Similarly strong performance for semiquantitative plasma cell assessment (AUC 0.87) and attention map inspection support that the model learned meaningful plasma cell morphology. More modest performance for t(4;14) suggests that some alterations may lack distinctive cytomorphologic correlates on routine stains, although sample size was limited. Across all tasks, aspirate-based models outperformed core biopsy models, highlighting the value of cellular detail for inferring myeloma-associated biologic features. Because S-phase fraction is not routinely measured owing to assay complexity, a morphology-based surrogate could yield rapid prognostic information from BMA smears. This approach may serve as a low-cost, sample-sparing triage tool to identify patients warranting confirmatory molecular or cytogenetic testing, particularly when comprehensive testing may be unavailable.
DISCLOSURES
Prasad Bandodkar, Ph.D.: No relevant conflicts of interest to declare.