SaRF: Saliency regularized feature learning improves MRI sequence classification.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Radiology
• Neuroimaging

Background:

• Deep learning for medical image analysis offers potential workflow improvements for neuro-radiologists using multi-sequence MRI.
• Accurate assignment of MRI sequence types is crucial for deep learning systems but is prone to errors in clinical practice.
• Current deep learning models for image-based sequence classification face robustness and reliability challenges.

Purpose of the Study:

• To develop a novel method for robust and reliable magnetic resonance imaging (MRI) sequence classification using deep learning.
• To enhance the accuracy and interpretability of deep learning models in neuroimaging by leveraging saliency information.

Main Methods:

• A novel method employing saliency information to guide feature learning for sequence classification.
• Utilized two self-supervised loss terms to improve distinctiveness of class-specific saliency maps and promote similarity with deep features.

Main Results:

• Achieved a mean accuracy improvement of 4.4% (0.935 to 0.976) on a cohort of 2100 patient cases.
• Demonstrated improvements in mean AUC (1.2%) and mean F1 score (20.5%), alongside reduced expected calibration error (30.8%).
• Expert feedback confirmed enhanced model interpretability and calibration.

Conclusions:

• The proposed method significantly improves accuracy, AUC, and F1 score for MRI sequence classification.
• The approach enhances model calibration and interpretability of saliency maps in deep learning for neuroimaging.