A Weakly Supervised Gradient Attribution Constraint for Interpretable Classification and Anomaly Detection.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Pathology

Background:

• Deep learning interpretability is crucial for medical applications, yet current methods are often post-hoc and not integrated into training.
• Lack of transparency hinders trust and adoption in critical fields like medicine, where decisions must be understandable.

Purpose of the Study:

• To develop a weakly supervised method for interpretable healthy vs. pathological classification and anomaly detection in medical images.
• To improve the transparency and reliability of deep learning models in critical medical diagnostics.

Main Methods:

• Introduced a novel loss function to constrain deep learning models during training using gradient-based attributions.
• Constrained healthy image voxels to drive network decisions towards the healthy class, enabling unsupervised segmentation of pathologies.
• Evaluated the efficacy of constrained training with simple Gradient attribution versus Expected Gradient, and proposed combining attributions for robustness.

Main Results:

• The proposed method achieved more relevant, pathology-driven classification for brain tumors and multiple sclerosis.
• Outperformed state-of-the-art methods in anomaly detection, particularly in segmenting multiple sclerosis lesions, with a 15-point Dice improvement.
• Demonstrated that constrained training with Gradient attribution is computationally efficient and comparable to Expected Gradient.

Conclusions:

• The new weakly supervised constraint enhances interpretability and classification accuracy in medical deep learning.
• The method effectively detects and segments pathologies, offering a more transparent and reliable alternative to existing approaches.
• This approach reduces computational costs while maintaining high performance, making interpretable deep learning more accessible for medical use.