Photodiagnosis Photodyn Ther. 2026 Nov 04. pii: S1572-1000(26)00371-6. [Epub ahead of print]
105706
BACKGROUND: Acute myeloid leukemia (AML) and acute lymphoblastic leukemia (ALL) may show overlapping morphology on bone-marrow trephine sections, while definitive ancillary testing is not always immediately available. We evaluated whether microscopic hyperspectral imaging (HSI) combined with patient-level spectral-spatial deep learning could support differentiation of AML, ALL, and normal marrow.
METHODS: This retrospective single-center study included 78 patients (26 AML, 26 ALL, and 26 normal marrow). H&E-stained bone-marrow trephine sections were imaged across 400-1000 nm. Patient-level nested stratified five-fold cross-validation was defined before region-of-interest (ROI) sampling to prevent information leakage. We developed HemaSpec-MIL, a new hierarchical spectral-spatial multiple-instance learning algorithm and a major methodological output of this study. The framework combines multi-scale spectral encoding, morphology encoding, bidirectional cross-attention, and gated multiple-instance aggregation to generate patient-level probabilities. The primary endpoint was pooled out-of-fold macro-AUC.
RESULTS: HemaSpec-MIL achieved a pooled out-of-fold accuracy of 0.897, macro-F1 of 0.897, and macro-AUC of 0.957. One-vs-rest AUCs were 0.964 for AML, 0.945 for ALL, and 0.962 for normal marrow. Performance exceeded spectral-only, morphology-only, 3D-CNN, Vision Transformer, and spectral Transformer baselines. Ablation analysis supported complementary contributions from the spectral branch, morphology branch, cross-attention fusion, gated patient-level aggregation, and spectral augmentation.
CONCLUSIONS: Microscopic HSI combined with the newly developed HemaSpec-MIL framework enabled accurate internal differentiation of AML, ALL, and normal marrow while preserving wavelength- and tissue-level interpretability. Independent multicenter validation and acquisition harmonization are required before clinical translation.
Keywords: acute lymphoblastic leukemia; acute myeloid leukemia; bone marrow pathology; deep learning; hyperspectral imaging; multiple-instance learning