Explainable AI identifies diagnostic cells of genetic AML subtypes.

Area of Science:

• Hematology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Explainable AI (XAI) is crucial for clinical adoption of AI decision support tools.
• Trust in AI requires understanding how models rationalize predictions.
• Acute myeloid leukemia (AML) diagnosis benefits from accurate subtype classification.

Purpose of the Study:

• To develop an inherently explainable AI model for classifying AML subtypes from blood smear images.
• To validate the model's ability to identify diagnostically relevant cells, aligning with human expert criteria.
• To provide a tool for routine diagnostics of hematopoietic neoplasms.

Main Methods:

• Trained SCEMILA, a single-cell based explainable multiple instance learning algorithm, on over 80,000 white blood cell images from 129 AML patients and 60 controls.
• Utilized a novel multi-attention module to analyze model focus.
• Developed an interactive online tool for data and prediction exploration.

Main Results:

• SCEMILA achieved perfect discrimination between AML patients and healthy controls.
• The model detected the APL subtype with an F1 score of 0.86±0.05.
• High concordance was observed between the AI's high-attention cells and expert-identified diagnostically relevant cells.

Conclusions:

• SCEMILA demonstrates high accuracy and explainability in classifying AML subtypes.
• The AI model successfully identifies diagnostically relevant cells, mirroring expert human analysis.
• This approach facilitates AI deployment in routine diagnostics for hematopoietic neoplasms.