Multimodal AI reading of genetic complexity and residual marrow composition in acute promyelocytic leukemia

Acute promyelocytic leukemia (APL) is defined by its genetics, yet interphase fluorescence in situ hybridization (FISH) for the PML::RARA fusion, the chromosomal complement seen on a metaphase spread, and the marrow's cellular composition on a Wright-Giemsa smear are read as separate, unconnected examinations. We did not identify a published model trained on two or more of these image types in the same patients with APL.

We describe two image-derived readouts for a retrospective cohort of patients with APL. A genetic complexity axis classifies the FISH fusion signal pattern and flags any additional cytogenetic abnormality (ACA) beyond the t(15;17), directly from raw, unarranged metaphase spreads. A marrow composition axis extends a cytomorphology pipeline, built on frozen DinoBloom embeddings and a per-cell classification head directly supervised on an annotated subset, from promyelocyte detection to a seven-class differential count. From this composition, we compute a residual hematopoiesis index, the model-estimated fraction of non-leukemic nucleated marrow cells, so that a higher value indicates a larger non-leukemic fraction. The index is a morphology-derived surrogate for residual non-leukemic marrow composition and carries no immune-phenotypic information. The association between the two axes is computed directly from patient-level, out-of-fold branch outputs; a small, three-node fusion module over the two genetic and one composition readout supports only a secondary, exploratory analysis, reported in the Supplementary Material. The cohort held 700 patients with smear, FISH, and karyotype linked per patient, of whom 650 had complete clinical annotation; we evaluated each branch with patient-stratified fivefold cross-validation, five repeats, pooled out-of-fold, and report bootstrap confidence intervals (CIs) over patients. Diagnostic flow cytometry was retrievable for 118 patients, and 80 slides were re-scanned with fields independently re-selected, to check the index against an external measurement and against itself.

The fusion-pattern classifier had a macro-averaged F1 of 0.82 across the five signal classes, and the karyotype ACA reader had an area under the curve (AUC) of 0.86. The residual hematopoiesis index had a median of 0.42 (interquartile range, IQR, 0.29 to 0.58) across the cohort, tracked the flow-derived non-blast fraction at a Spearman correlation of 0.80 (95% CI 0.72 to 0.86) in the 118 patients with flow data, and had an intraclass correlation of 0.91 on re-scanned slides. Genetic complexity and the residual hematopoiesis index were associated at a Spearman ρ of 0.24 (95% CI 0.17 to 0.31, p < 0.0001), the primary finding of this study, with higher genetic complexity accompanying a larger non-leukemic fraction. The association held under adjustment for metaphase yield, scorable nuclei per FISH image, age, presenting white-cell count, and accession year, with partial correlations between 0.19 and 0.24, and patients with an expert-defined additional abnormality had a higher index than those with the isolated translocation (median 0.46 against 0.41). An exploratory test of the same readouts against the Sanz risk group and DS occurrence, run on the clinically annotated subcohort, showed performance no better than chance, and an adjusted logistic model bounded any effect of an additional cytogenetic abnormality on DS at an odds ratio of 1.07 (95% CI 0.74 to 1.55), in line with two prior clinical cohorts that reported no association between cytogenetic abnormality and DS.

Image-derived genetic complexity was positively associated with the residual hematopoiesis index, and the association persisted under the adjustments and substitutions applied to it. These findings describe a cross-sectional association in a single-center, retrospective cohort, not a validated diagnostic or risk tool. The residual hematopoiesis index summarizes morphology alone, its agreement with flow cytometry is supportive rather than confirmatory, and the secondary DS analysis drew on a smaller, clinically annotated subset of the cohort rather than on every patient.