ShapeAXI: Shape Analysis Explainability and Interpretability.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Machine learning for healthcare

Background:

• 3D shape analysis is crucial for medical diagnostics.
• Current methods may lack interpretability and efficiency.
• Need for advanced tools in orthopedic and craniofacial research.

Purpose of the Study:

• Introduce ShapeAXI, a novel framework for 3D shape analysis.
• Demonstrate its utility in medical classification tasks.
• Enhance the interpretability of shape analysis through explainability heatmaps.

Main Methods:

• Utilizes a multi-view approach to capture 3D objects.
• Applies 2D Convolutional Neural Networks (CNNs) for analysis.
• Implements automatic N-fold cross-validation and result aggregation.

Main Results:

• Successfully classified condyles into healthy and degenerative states.
• Efficiently categorized cleft patient shapes from CBCT scans into four severity classes.
• Generated insightful, class-specific explainability heatmaps for enhanced interpretability.

Conclusions:

• ShapeAXI offers a versatile and interpretable approach to 3D object classification.
• The framework shows promise for advancing condyle assessment and cleft patient analysis.
• ShapeAXI provides a new benchmark for 3D interpretation with potential broad applications.