Improving Interpretability of Leucocyte Classification with Multimodal Network.

Area of Science:

• Hematology
• Computational Biology
• Medical Imaging

Background:

• Accurate white blood cell classification is crucial for diagnosing hematologic diseases.
• Current methods rely on image-based or feature-based classification, each with limitations.
• Image-based classification offers high performance, while feature-based classification provides better interpretability.

Purpose of the Study:

• To develop and evaluate a Multimodal neural network for white blood cell classification.
• To compare the performance and interpretability of the Multimodal approach against image-only and feature-only models.
• To identify key morphological features contributing to cell characterization.

Main Methods:

• Utilized a Multimodal neural network integrating both cell images and morphological features.
• Trained and compared three models: image-only, feature-only, and Multimodal.
• Employed SHAP (SHapley Additive exPlanations) values for model interpretability.

Main Results:

• Image-only training achieved the highest classification performance.
• The Multimodal model demonstrated enhanced interpretability compared to image-only methods.
• Crucial morphological features for cell characterization were identified by the Multimodal model.

Conclusions:

• Multimodal neural networks offer a balance between performance and interpretability in white blood cell classification.
• SHAP values can be effectively used to interpret complex models in hematology.
• The study highlights the importance of integrating diverse data types for robust and understandable diagnostic tools.