Related Experiment Videos
Aligning deep learning and interpretable models for blood cell classification: A dual-model framework for
Magdalene Yeok Yu Cheong1,2, Xinran Xu2, Li Rong Wang1,3
1College of Computing and Data Science, Nanyang Technological University, Singapore, Singapore.
PLOS Digital Health
|July 30, 2026
Summary
This study introduces a dual-model AI framework for transparent blood cell classification. It combines deep learning with interpretable methods, enhancing diagnostic accuracy and clinical insight in hematology.
Area of Science:
- Hematology
- Artificial Intelligence
- Medical Diagnostics
Background:
- Blood smear examination is crucial but labor-intensive and subjective.
- Current deep learning models for blood cell classification lack clinical transparency.
Purpose of the Study:
- To develop a transparent dual-model framework for blood cell classification.
- To enhance the interpretability and clinical utility of AI in hematology.
Main Methods:
- A deep YOLO (You Only Look Once) classifier was paired with a shallow, interpretable explainer.
- Clinically informed features were extracted from segmented images to train the explainer.
- SHapley Additive exPlanations (SHAP) were used to quantify feature importance.
Main Results:
- YOLO achieved high AUCs (0.997, 0.994, 1.000) on multiple datasets.
- The explainer demonstrated strong alignment with YOLO predictions.
- User studies showed 96.9% concurrence with expert predictions.
Conclusions:
- Coupling deep learning with interpretable models improves prediction confidence.
- The framework provides clinically meaningful insights, enhancing AI transparency in hematology.