AUTHORS
Bahga Katamesh, MD, Yuanlong Wang2, Najla Al Ali, MS3, Rory Shallis, MD4, Onyee Chan, MD5, Samuel Benjamin Reynolds, MD6, Alison Walker, MD7, Andrew T Kuykendall, MD8, Seongseok Yun, MD, PhD9, Jeffrey Lancet, MD10, Eric Padron, MD11, David Sallman, MD12, Rami Komrokji, MD13, Ping Zhang14, Ling Zhang, MD15, Andrew Srisuwananukorn, MD16, Zhuoer Xie, MD,MS17
AFFILIATIONS
1 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
2 Ohio State University Medical Center, Columbus, Ohio, United States
3 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
4 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
5 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
6 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
7 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
8 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
9 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
10 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
11 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
12 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
13 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
14 Ohio State University, Columbus, Ohio, United States
15 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
16 Ohio State University Medical Center, Columbus, Ohio, United States
17 H. Lee Moffitt Cancer Center, Tampa, Florida, United States
ABSTRACT
Background
Differentiating hypoplastic MDS (hMDS) from aplastic anemia (AA) remains a major challenge in hematopathology. Both entities share overlapping clinical and bone marrow histopathological features, leading to significant variability among hematopathologists and clinicians in rendering a diagnosis. The distinction is clinically critical due to the inherently higher risk for leukemic transformation, shorter survival, and reduced response to immunosuppressive therapy in patients with hMDS. The current approach relies on multidisciplinary collaboration at specialized tertiary centers. We aimed to develop an artificial intelligence (AI) diagnostic system based on digitized bone marrow images that would provide objective, rapid, reproducible diagnosis across institutions.
Methods
We developed an AI algorithm to differentiate a diagnosis of hMDS and AA from patients at Moffitt Cancer Center. Patients were treated between 8/2022–8/2025, with histopathological diagnosis confirmed by an institutional expert hematopathologist. In total, 99 patients were enrolled: 88 AA and 11 hMDS. For each patient, diagnostic specimens were digitized using an Aperio AT2 scanner at 40x magnification and included hematoxylin and eosin (H&E) stained bone marrow biopsies (BMB), CD34-stained immunohistochemistry of BMB, and Wright-Giemsa (W&G) stained bone marrow aspirate (BMA) smears. The cohort was split into 60% training, 20% validation and 20% test set. Whole slide imaging (WSI) was performed by the Advanced Analytical and Digital Laboratory (AADL) under the Department of Translational Pathology, using the Leica digital pathology platform, Aperio GT450 slide scanners. Whole Slide Imaging were scanned at 40× magnification (0.26 µm/pixel) and initial QC report is generated by the Aperio GT450 and Aperio iQC Software. After QC approval, the SVS files were shared. Digitized images were preprocessed using feature extraction by pathology foundational models followed by attention-based multiple instance learning (Srisuwananukorn et al, Leukemia 2026). Given no precedent on preferred modality, one diagnostic classification model was performed on each stain modality. Benchmarking was performed to identify the optimal foundation model for feature extraction. Model performance was assessed by balanced accuracy (bACC) and area under the receiver operating characteristic curve (AUROC).
Results
Baseline clinical and morphological features were comparable between the hMDS and AA cohort, including age, sex and peripheral counts (median neutrophil count 0.86*109/L in AA vs 45% in 0.5-1*109/L range for hMDS; platelet count 19 vs 54*109/L, and reticulocyte fraction 0.038 vs 0.029). Treatment outcome differed substantially. In AA, complete response rates were 19% at 3 months and 32% at 6, with 26 relapses and a median OS (mOS) of 215 months, declining to 53 months among 10 patients who progressed to secondary MN. In hMDS, mOS was 31.9 months, declining to 18 months with progression. Given the significant outcome differences, an AI model was pursued to identify pathological discriminatory features.
The pathology foundation models benchmarked were GigaPath, Virchow2, UNI v2, and H-optimus-1. Model performance varied across stain modality, but there was no clear difference with respect to foundation model utilized. Regardless of foundation model used, performance of diagnostic classification varied upon H&E BMB with AUROC 0.4 – 0.7 and bAcc 0.4 - 0.5. In contrast, performance trended higher on CD34 BMB with bACC 0.6-1 and AUROC 0.75 – 1, or on W&G BMA with bACC 83.3% and AUROC 0.963.
Conclusion
Our AI-based digital pathology model demonstrates high diagnostic accuracy in differentiating AA from hMDS, particularly with the use of CD34 IHC and W&G BMA slides. Though likely helpful diagnostic criteria lie across all modalities, these results suggest that a multi—modal AI-based classification approach may provide optimal diagnostic accuracy, allowing clinicians to assign timely, appropriate therapies. We are currently investigating the external validation of our AI classifier on larger multi-institutional patient cohorts by integrating with clinical and molecular parameters.
DISCLOSURES
Bahga Katamesh, MD: No relevant conflicts of interest to declare.