← Back to Home

Beyond Blast Counts: AI Identifies Blood Cancer Activity From the Spatial Organization of Bone Marrow

A research team condensed the locations, morphology, and neighborhood relationships of more than 5 million cells in bone marrow sections into a single score that preliminarily distinguished active myelodysplastic neoplasms from remission; however, the method still requires validation across hospitals and in large populations and cannot yet replace current diagnostic approaches.

By SURL BioNews

Bone marrow sections preserve more than cell counts; they also provide a map of the hematopoietic ecosystem. Whether stem cells remain close to blood vessels, whether red blood cell precursors can maintain their communities, and how abnormal cells cluster may all reveal whether a blood cancer is receding or returning. A team at Weill Cornell Medicine used image analysis and artificial intelligence to quantify these spatial changes, hoping to add a layer of information that conventional pathological assessment has difficulty capturing.

The method, called the Myelodysplastic Neoplasm Microarchitecture Perturbation Score (MDS-MAPS), was designed for myelodysplastic neoplasms (MDS). MDS impairs normal blood formation, and some patients may progress to acute myeloid leukemia. Clinicians currently often assess disease status based on blast percentage, cell morphology, flow cytometry, and genetic testing, but gray areas remain.

Published in *Leukemia*, the study included 36 diagnostic sections, 29 longitudinal samples collected during treatment, 13 samples representing precancerous conditions, and 21 normal controls. Using multiplex immunofluorescence and whole-slide imaging, the team identified more than 5 million cells with spatial coordinates, then developed a composite score from 82 features covering cellular composition, morphology, and arrangement. A higher score indicates that the bone marrow structure deviates further from a healthy state.

In a leave-one-patient-out validation analysis, MDS-MAPS distinguished active disease from remission with an area under the curve (AUC) of 0.883, higher than the 0.660 achieved using blast percentage alone. It also distinguished low-blast MDS from clonal cytopenia of undetermined significance, with an AUC of 0.815. These figures indicate that the model had better classification ability in the existing data, but they do not demonstrate that it can improve treatment decisions or patient outcomes.

The spatial maps also outlined the disease’s biological profile. Erythroblastic islands in MDS bone marrow were disrupted, while hematopoietic stem and progenitor cells gradually moved away from supportive perivascular niches. Some changes showed genotype-specific features, particularly pronounced in cases with TP53 mutations. Longitudinal samples showed that bone marrow architecture moved closer to normal during remission and became disorganized again at relapse. The decline in the score was also not fully explained statistically by a reduction in blasts.

However, this remains a retrospective study conducted at a single academic medical center, with limited numbers of patients and longitudinal samples, and the model’s performance was derived primarily from internal cross-validation. Differences among hospitals in section processing, staining, scanning equipment, and patient composition could all alter the results. The next step is to confirm reproducibility in larger, independent, multicenter populations, establish thresholds suitable for clinical use, and demonstrate through prospective studies that the method provides practical benefits beyond current tests. Until then, MDS-MAPS is better regarded as a candidate adjunctive marker rather than a standalone diagnostic tool.

References

  1. Weill Cornell Medicine
  2. Leukemia