Leukocyte deep learning classification assessment using Shapley additive explanations algorithm

Area of Science:

• Hematology
• Artificial Intelligence
• Medical Imaging

Background:

• Peripheral blood smear analysis is crucial for diagnosing hematological diseases but is labor-intensive and requires expertise.
• Digital morphology analyzers, including those using deep learning, offer automation but often lack transparency.
• Interpreting deep learning models in medical diagnostics remains a challenge for scientists.

Purpose of the Study:

• To enhance the interpretability of deep learning models for leukocyte classification in peripheral blood smears.
• To integrate explanatory factors into deep learning workflows for better understanding of classification decisions.
• To validate the utility of Shapley Additive Explanations (SHAP) for visualizing classification drivers.

Main Methods:

• Utilized 10,297 leukocyte images from peripheral blood smears.
• Employed VGG16 and VGG19 deep learning models, both pre-trained and fully trained, for leukocyte classification.
• Applied SHAP DeepExplainer to generate visual explanations, highlighting significant cellular regions for classification.

Main Results:

• Achieved high classification accuracy: 99.81% for VGG16 and 99.79% for VGG19 with fully trained models.
• Fully trained models slightly outperformed partially trained models (98.67% VGG16, 98.33% VGG19).
• SHAP explanations revealed that cell and nucleus contours were important for pre-trained models, while the cytoplasm was key for fully trained models.

Conclusions:

• SHAP DeepExplainer provides valuable insights into leukocyte classification by deep learning models.
• The method aids in verifying automated classification results in peripheral blood smear analysis.
• Despite variations in explanations between model training states, SHAP enhances the transparency of automated hematology diagnostics.