-Author name in bold denotes the presenting author
-Asterisk * with author name denotes a Non-ASH member
Clinically Relevant Abstract denotes an abstract that is clinically relevant.

PhD Trainee denotes that this is a recommended PHD Trainee Session.

Ticketed Session denotes that this is a ticketed session.

2227 AI-Driven Quantitation of Blast Cells and Vessel Density in Bone Marrow Biopsies Highlights Limitations of Manual Assessment

Program: Oral and Poster Abstracts
Session: 803. Emerging Tools, Techniques, and Artificial Intelligence in Hematology: Poster I
Hematology Disease Topics & Pathways:
Artificial intelligence (AI), Emerging technologies, Technology and Procedures, Machine learning, Pathology
Saturday, December 7, 2024, 5:30 PM-7:30 PM

Xuezi Hu1*, Korsuk Sirinukunwattana1*, Ka Ho Tam1*, Alan Aberdeen1*, Daniel Royston, DPhil, FRCPath, MBBChir2 and Carlo Pescia3*

1Ground Truth Labs Ltd., Oxford, United Kingdom
2Nuffield Division of Clinical Laboratory Sciences, University of Oxford, Oxford, United Kingdom
3Università degli Studi di Milano, MILANO, Italy

Introduction:

Accurate assessment of blast cells using CD34 immunohistochemistry (IHC) in bone marrow trephines (BMTs) is essential for diagnosing myeloid neoplasms, monitoring disease progression, and guiding treatment. Accurate blast quantitation is particularly important for classifying acute myeloid leukemia (AML) and myelodysplastic neoplasms (MDS), particularly with the recent introduction of the MDS/AML category in the International Consensus Classification (ICC), which lowers the threshold for AML-like therapies to a 10% blast cutoff. In myeloproliferative neoplasms (MPN), blast quantitation defines accelerated phase as >10% blasts. Manual quantitation of CD34+ blasts by IHC is subjective and therefore prone to inter-observer variability. Immunostaining for CD34 has also been used to evaluate microvessel density in MPNs, with increased MVD described as characteristic of myelofibrosis. We developed an AI strategy to improve the detection and quantitation of CD34+ blasts and vessels using routine diagonostic samples.

Method

CD34-stained BMT slides of 38 CALR+ MPN (14 ET, 8 pre-PMF, 10 overt PMF, 5 post-ET MF), 10 AML, and 30 MDS (15 low-risk MDS, 6 MDS-EB, and 9 MDS/AML) from 2022 to 2024 were identified for this study. All samples were of diagnostic quality. Two hematopathologists performed a visual estimation of CD34+ blasts per whole slide image (WSI) and MVD estimation through the “hotspot” method, and the mean value of both was calculated. A YOLOv7 model was trained using expert annotations on WSIs to identify blasts. Vessel quantitation was obtained by combining the CD34-positive mask and the blast model prediction, generating a vessel mask.

Results

The correlation between two hematopathologists’ blast counts was 0.56 (95% CI: 0.318, 0.743),with both consistently overestimating blast counts compared to the AI-defined exact count, with a median difference (MD) of 4.99 and limits of agreement (LoA) of (16.124, -6.138). Notably. the total CD34+ blast cell area/total cell area ratio from the model showed an MD of 0.198 (LoA: 9.160, -8.763) compared to hematopathologists’ average estimation, indicating that manual estimation is likely influenced by cell area rather than true cell enumeration. In MDS and AML cases, AI blast quantitation demonstrated a strong correlation with bone marrow (BM) aspirate (ρ = 0.705, MD = 3.962, LoA: 19.394 to -11.469) and a moderate correlation with flow cytometry (ρ = 0.577), but with a lower MD of 1.920 (LoA: 13.634 to -9.794). This suggests that AI quantitation is closer to flow cytometry based blast cell quantitation, challenging the convention that flow cytometry may systematically underestimate blast cell proportions.

Importantly, AI-based quantitation reshaped MDS classification, downgrading 8 MDS/AML and 5 MDS-EB to low-risk MDS. The area estimation method downgraded only 2 MDS/AML, but interestingly upgraded 2 low-risk MDS to MDS-EB, 1 MDS-EB to MDS/AML, and even 1 MDS/AML to AML. AI quantitation also downgraded 8 AML cases (2 when considering area).

As expected, both manual and AI-driven MVD estimations were higher in MPN (mean values of 19.08 and 204.3, respectively) compared to MDS (10.77 and 138.9) and AML (10.25 and 182.7). The correlation between manual MVD assessment and the vessel mask was moderate (ρ = 0.6134), with a notable mean difference of -77.37 (LoA:43.38, -198).

Conclusions

We present a systematic AI-driven quantitation of CD34+ blasts and vessels in BMTs, automating assessment with high accuracy in myeloid neoplasms. Our findings suggest that conventional manual assessment is inaccurate and consistently overestimates CD34+ blast proportions in several important myeloid disorders, possibly influenced by the inadvertent evaluation of cell area. This discrepancy has potential to significantly impact classification and prognostication in BMT assessment, and clearly warrants further validation studies. The successful AI-based detection of bone marrow CD34+ vascular structures also has potential to improve the evaluation of MVD in myeloid disorders, and further demonstrates the potential of these technologies in future translational research studies.

Disclosures: Hu: Ground Truth Labs Ltd.: Current Employment. Sirinukunwattana: Ground Truth Labs Ltd.: Current Employment. Tam: Ground Truth Labs Ltd.: Current Employment. Aberdeen: Ground Truth Labs Ltd.: Current Employment. Royston: Johnson & Johnson: Consultancy; Ground Truth Labs Ltd.: Consultancy.

*signifies non-member of ASH