2024 · December
AI Blast Quantitation Reveals Systematic Overestimation in Manual Assessment
Xuezi Hu
Machine Learning Scientist
At ASH 2024, we presented findings showing that manual assessment of blast cells in bone marrow consistently overestimates blast percentages compared to AI-driven quantitation. This has direct implications for patient classification—our analysis suggests that a substantial number of cases may be assigned to higher-risk categories than warranted.
Why blast counts matter
Blast cell percentage determines how patients with myeloid neoplasms are classified and treated. The threshold matters: the International Consensus Classification now defines MDS/AML at 10% blasts, lowering the bar for AML-like therapies. In myeloproliferative neoplasms, exceeding 10% blasts signals accelerated phase disease.
Getting this number wrong has consequences. Overestimation could mean patients receive more aggressive treatment than necessary. Underestimation could delay appropriate intervention.
What we found
We developed an AI model to quantify CD34+ blast cells in bone marrow trephine biopsies and compared it against manual assessment by two expert hematopathologists across 78 diagnostic samples.
The inter-observer correlation between pathologists was 0.56—moderate at best. Both consistently overestimated blast counts compared to the AI, with a median difference of approximately 5 percentage points. This overestimation appears driven by cell area rather than true cell count: pathologists may be unconsciously influenced by how much visual space blasts occupy rather than their actual number.
Notably, AI quantitation correlated more closely with flow cytometry (ρ = 0.577) than manual assessment did. This challenges the assumption that flow cytometry systematically underestimates blasts—it may be that manual counting systematically overestimates them.
Classification impact
The implications are significant. Using AI-based counts instead of manual assessment would have reclassified:
- 8 MDS/AML cases → low-risk MDS
- 5 MDS-EB cases → low-risk MDS
- 8 AML cases → lower categories
That's 21 potential reclassifications across 40 MDS and AML cases—over half the cohort.
Microvessel density
We also demonstrated AI quantitation of microvessel density using the same CD34 staining. As expected, MVD was higher in MPN compared to MDS and AML, confirming the utility of automated vessel detection for characterizing the bone marrow microenvironment.
What's next
These findings warrant further validation. If confirmed in larger cohorts, AI-assisted blast quantitation could help standardize assessment and reduce the variability that currently affects patient classification and treatment decisions.
Read the full abstract: DOI: 10.1182/blood-2024-203118