Addressing Artefacts in Anatomical MR Images: A k-space-Based Approach

Magnetic Resonance Imaging (MRI) is a valuable tool in medical diagnosis and treatment planning. Acquired data can be affected by various artefacts that can compromise image quality and result in decreased diagnostic accuracy. Artificial intelligence emerged as a robust solution in enhancing the quality of images. In the context of anatomical MRI data, we have chosen to use Deep Learning (DL) models to correct the artefacts and improve the quality of MR images.In this work we propose a library that simulates realistic primary artefacts by manipulating the k-space signal1, using standard image processing techniques. Our investigation focuses on three artefacts commonly encountered in anatomical images: Gaussian noise, blur effect and Motion Artefact (MA) with varying levels of degradation. We trained DL models to learn the artefacts features, correct them and improve the quality of the MR images. The dataset consists of in vivo 2D anatomical mouse brain images acquired on a 3T MRI system. The results of this study confirm the performance of deep learning models in the field of artefacts correction of medical images with quality improvement. The proposed method surpassed the classical denoising approach by 14.46% in terms of quality metrics and exhibited a 82.73% improvement for the deblurring task. The model performed well in the application of MA correction without any consensus image processing method has been established.