Kernel-weighted contribution: a method of visual attribution for 3D deep learning segmentation in medical imaging

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models are increasingly used in medical image segmentation.
• Explaining these models' decisions is crucial for trust and adoption in healthcare.
• Current attribution methods may not be optimal for medical image segmentation tasks.

Purpose of the Study:

• Introduce kernel-weighted contribution, a novel visual explanation method for 3D medical image segmentation.
• Develop accurate and interpretable explanations for deep learning model decisions.
• Assess feature importance based on activation map contributions.

Main Methods:

• Evaluate the kernel-weighted contribution method on a synthetic dataset with known ground truth.
• Compare the method against three other attribution techniques across five model layer combinations.
• Analyze explanations for 100 test samples using a comprehensive quality metric.

Main Results:

• Kernel-weighted contribution yielded superior explanations when applied to both encoder and decoder sections.
• The method outperformed others in four of five experiments when using the same model layers.
• Equivalent superior performance was observed compared to GradCAM++ on specific decoder layers.

Conclusions:

• The kernel-weighted contribution method provides high-quality explanations tailored for medical image segmentation models.
• This approach effectively leverages architectural specifics of segmentation models.
• The synthetic dataset and method implementation are publicly available for research.