Interpretable Multiple Instance Learning for Hematologic Diagnosis from Peripheral Blood Smears.

Area of Science:

• Hematology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate diagnosis of hematologic malignancies requires integrating cellular morphology and composition from peripheral blood smears (PBSs).
• Current automated methods focus on single-cell classification, lacking whole-slide diagnostic capabilities.
• There is a need for advanced computational tools to improve diagnostic accuracy and efficiency in hematology.

Purpose of the Study:

• To develop and evaluate a novel weakly supervised framework, CAREMIL (Cell AggRegation, Explainable, Multiple Instance Learning), for whole-slide hematologic malignancy diagnosis.
• To integrate a high-performance cell-based encoder (DeepHeme) with CAREMIL for robust feature extraction and diagnostic prediction.
• To assess the performance of CAREMIL against existing multiple instance learning (MIL) aggregation functions and image encoders.

Main Methods:

• Utilized DeepHeme, a performative cell-based encoder, for feature extraction from PBS images.
• Developed CAREMIL, an attention-based MIL framework employing weakly supervised learning for whole-slide classification.
• Evaluated various image encoders and MIL architectures, comparing CAREMIL with gated MIL for slide-level aggregation.

Main Results:

• The combination of DeepHeme and CAREMIL achieved the highest diagnostic performance across acute leukemia (AML), myelodysplastic syndromes (MDS), and hairy cell leukemia (HCL) (AUROCs 0.999, 0.891, 0.945).
• CAREMIL demonstrated superior performance as an aggregation function compared to gated MIL, especially with out-of-domain encoders like ImageNet, UNI2, and Virchow2.
• The framework successfully identified AML even in cases with minimal or absent circulating blasts, highlighting its sensitivity.
• Attention mechanisms in CAREMIL provided biological interpretability by highlighting diagnostically relevant cells and revealing disease-specific morphometric signatures.

Conclusions:

• CAREMIL, combined with DeepHeme, represents a powerful and interpretable MIL framework for hematologic slide diagnosis.
• The framework is robust to cell-level misclassifications and does not require explicit cell-level supervision.
• CAREMIL shows potential for application to other liquid biopsy specimens and supports a shift towards quantitative, morphology-informed diagnostics in hematology.