Interpretable multiple instance learning for hematologic diagnosis from peripheral blood smears.

Area of Science:

• Computational pathology
• Hematologic diagnostics
• Machine learning in medicine

Background:

• Accurate diagnosis of hematologic malignancies from peripheral blood smears (PBSs) requires integrating cellular morphology and composition.
• Current computational methods often focus on single-cell classification, lacking holistic, slide-level diagnostic capabilities.

Purpose of the Study:

• To develop and validate a novel computational framework for accurate, interpretable, slide-level diagnosis of hematologic malignancies.
• To integrate high-performance cell-based feature extraction with a weakly supervised, attention-based multiple instance learning (MIL) model.

Main Methods:

• A framework combining DeepHeme (cell-based encoder) and CAREMIL (weakly supervised, attention-based MIL) was developed.
• Evaluated performance against leading image encoders and MIL architectures, including out-of-domain models.
• CAREMIL demonstrated robust cell aggregation and attention mechanisms for slide-level analysis.

Main Results:

• CAREMIL with DeepHeme achieved high diagnostic accuracy for acute myeloid leukemia (AML), myelodysplastic syndromes (MDS), and hairy cell leukemia (HCL) (AUROCs: 0.999, 0.891, 0.945).
• The model successfully identified AML even with minimal circulating blasts and provided interpretable insights via attention values highlighting diagnostically relevant cells.
• The framework showed resilience to individual cell misclassifications and did not require cell-level supervision.

Conclusions:

• CAREMIL is an effective and interpretable MIL framework for hematologic slide diagnosis.
• The framework is extendable to other specimen types like bone marrow aspirates and cytology.
• Supports a move towards quantitative, morphology-informed hematologic diagnostics.