Magnetic resonance image enhancement and segmentation using conventional and deep learning denoising techniques for...

The study of brain vascular dynamic patterns in infants, through dynamic angio MRI (TRANCE-MRI) images, is relevant to identify pathologies associated with brain flow and perfusion. However, several drawbacks arise while using these types of images for diagnosis, such as noisy images and difficulties in the quantification of the vessel patterns. Depending on the patient, specialists can use manual procedures to analyze the images and segment the veins during the image analysis. Image acquisition in infants is often affected by motion artifacts, variable contrast due to short acquisition times, and scanner hardware limitations, which together increase noise and reduce vessel visibility. Furthermore, this fact poses serious challenges for both the use of AI tools as well as the analysis and diagnosis of professionals. The goal of this research is to assess automatic denoising pipelines for enhancing image quality to aid visual analysis and automated vessel segmentation for improved quantification through vessel feature extraction. As a result of this research, an entire pipeline is presented as a solution. For denoising the images, we have explored the combination of conventional techniques with unsupervised techniques based on deep learning. The outcomes were subjectively assessed by experts and quantitatively by non-reference image quality evaluators. Using Noise2Void and PPN2V GMM produced the best outcomes, according to the scores. However, employing a combination of traditional methods and deep learning-based methods, the vessels showed a reduction in noise in the central and most dense areas, according to qualitative results. A model was trained using noisy images for segmentation. Then it was put to the test using both denoised and noisy images. The findings demonstrated an improvement of 9.4% in the dice score and nearly 16% in the Hausdorff distance when the model was trained using noisy images and segmentation was obtained using denoised images.

The online version contains supplementary material available at 10.1007/s13755-025-00406-x.