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

## TL;DR

This paper introduces CAREMIL, an interpretable machine learning framework that improves hematologic diagnosis from blood smears by combining cell-level features with whole-slide predictions.

## Contribution

CAREMIL is a novel weakly supervised attention-based MIL framework that outperforms existing methods and provides interpretable insights for hematologic diagnosis.

## Key findings

- CAREMIL paired with DeepHeme achieves state-of-the-art performance in diagnosing AML, MDS, and HCL with high AUROC scores.
- CAREMIL improves performance especially when using out-of-domain encoders like those trained on ImageNet or open-source pathology models.
- Attention values from CAREMIL highlight diagnostically relevant cells and morphometric signatures, enabling biological interpretability.

## Abstract

Accurate diagnosis of hematologic malignancies from peripheral blood smears (PBSs) requires integrating cellular morphology and composition across hundreds of white blood cells. Existing approaches primarily automate single-cell classification and do not provide whole-slide diagnostic predictions. We present a full network that utilizes a highly performative cell-based encoder (DeepHeme) for feature extraction paired with our weakly supervised framework using attention-based multiple instance learning (MIL) that we call CAREMIL (Cell AggRegation, Explainable, Multiple Instance Learning). Upon evaluating various popular image encoders and MIL architectures, the combination of DeepHeme and CAREMIL is the best performing pipeline on our disease classification task. CAREMIL proves to be a robust aggregation function that outperforms the most commonly used slide level aggregation function (gated multiple instance learning) across several encoder types. The greatest improvements in performance gain with CAREMIL is observed when using out-of-domain encoders, including an encoder trained on ImageNet and leading open-source pathology foundational models (UNI2 and Virchow2). CAREMIL plus DeepHeme achieves the highest diagnostic performance across acute leukemia (AML), myelodysplastic syndromes (MDS), and hairy cell leukemia (HCL) (AUROCs 0.999, 0.891, and 0.945, respectively), and identifies AML disease even in cases with minimal or absent circulating blasts. Attention values assigned by CAREMIL highlight diagnostically relevant cells and reveal disease-specific morphometric signatures, enabling biological interpretability and case-level insight. CAREMIL remains robust to misclassified cell types by the cell image encoder and does not require explicit cell-level supervision. These findings position CAREMIL as an effective and interpretable multiple instance learning framework for hematologic slide diagnosis, with potential to extend to bone marrow aspirates, cytology, and other liquid biopsy specimens, and to support a broader shift toward quantitative, morphology-informed diagnostics in hematology.