A deep-learning-based Signal-to-Noise Ratio (SNR) adaptive uniformity correction method

Motivation: Intensity normalization in MRI is crucial for consistent image analysis. However, naïve uniformity correction (UC) using pre-collected coil sensitivity distribution may over-compensate low-SNR regions and reduce visual quality. Goal(s): This study aims to develop an SNR-adaptive UC method to prevent over-compensation in low-SNR regions. Approach: We propose using a deep network to learn an SNR-aware uniformity correction map by suppressing over-compensation on low-SNR regions. During training, the original uniformity correction map is used to guide the weighing between consistency loss and over-compensation suppression loss. Results: The proposed method has been demonstrated to effectively suppress over-compensation on low-SNR regions in various head imaging protocols. Impact: The proposed method can improve MRI image quality by adaptively compensating for intensity variations based on the noise level in different regions. It may lead to more accurate diagnoses and better identification of subtle changes in MRI images.