A Cross-Domain Benchmark of Intrinsic and Post Hoc Explainability for 3D Deep Learning Models.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Medical Imaging

Background:

• Deep learning for 3D data is prevalent in medical imaging, object recognition, and robotics.
• The increasing use of AI necessitates explainability due to the black-box nature of models.
• Lack of standardized benchmarks hinders reliable comparison of explainable AI (XAI) methods for 3D data.

Purpose of the Study:

• To present a unified benchmarking framework for evaluating intrinsic and post hoc XAI methods on diverse 3D datasets.
• To quantitatively assess the performance of various XAI techniques across medical CT scans, CAD models, and point clouds.
• To provide comparative insights for selecting appropriate XAI methods in different 3D data domains.

Main Methods:

• Developed a unified benchmarking framework for 3D explainability.
• Evaluated XAI methods (Grad-CAM, Integrated Gradients, Saliency, Occlusion, ResAttNet-3D) on CT scans (MosMed), CAD models (ModelNet40), and point clouds (ScanObjectNN).
• Assessed explanation quality using Correctness (AOPC), Completeness (AUPC), and Compactness metrics.

Main Results:

• Explanation quality varied significantly across methods and domains.
• Grad-CAM and intrinsic attention performed best on medical CT scans.
• Gradient-based methods showed superior performance on voxelized and point-based data.
• Statistical tests confirmed significant performance differences among methods.

Conclusions:

• No single XAI method is universally superior across all 3D data domains.
• Domain-specific evaluation and multi-metric assessment are crucial for selecting effective XAI techniques.
• The proposed framework enables reproducible and standardized assessment of 3D explainability.