Improved bias field correction by combining Deep Learning and pre-scan information

Motivation: MR images can be affected by technically based signal intensity inhomogeneities (bias fields), which may complicate reading or even lead to clinical misinterpretations. Goal(s): To develop a generic method to detect and correct such signal intensity inhomogeneities across a broad range of body regions and magnetic field strengths. Approach: A deep learning algorithm was trained in a supervised fashion to correct intensity biases. Performance of multiple models trained with different combinations of input data were compared against common correction techniques. Results: The model combining conventionally corrected images and additional pre-scan information as input achieved the best performance of all compared approaches. Impact: This research addresses MRI intensity inhomogeneity issues by combining deep learning models and additional context information. The approach improves image quality and holds the promise to enhance bias field correction and potentially reduce diagnostic errors with minimal processing overhead.