Concept attribution: Explaining CNN decisions to physicians.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Gradient-based methods for deep learning explainability often focus on input pixels, which can be challenging to correlate with meaningful image features in medical diagnostics.
• Existing methods may not directly align with expert knowledge or diagnostic measures crucial for clinical trust.

Purpose of the Study:

• To develop a post-hoc framework for deep learning explainability that shifts attribution from pixels to user-defined concepts.
• To enable experts to verify the presence of specific diagnostic measures within network representations for improved interpretability and trust.
• To offer a method that integrates seamlessly with existing convolutional neural networks without altering their training.

Main Methods:

• A novel framework for concept-based attribution in deep learning models.
• Introduction of a spatial pooling operation on feature maps for enhanced interpretability.
• Analysis of regularized regression to address overfitting in high-dimensional latent spaces.

Main Results:

• Demonstrated versatility across medical applications (histopathology, retinopathy) and non-medical tasks (handwritten digit classification).
• Generated explanations that align with clinical guidelines.
• Provided explanations complementary to traditional saliency maps.

Conclusions:

• The proposed concept-attribution framework offers a robust and interpretable approach to deep learning explainability, particularly valuable in medical imaging.
• The method enhances trust by allowing experts to validate network reasoning against domain-specific concepts.
• This technique is easily integrated into current deep learning architectures, offering practical benefits for AI deployment.