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

Abstract

The anatomical variations of the cerebrovascular system of infants may have an impact on the neurological development of the infants and may even lead to death. Intracranial vascular abnormalities include multiple diseases such as arteriovenous malformations and moyamoya. For evaluating this kind of diseases, there are non-invasive imaging techniques such as TRANCE which is a type of MRI. It allows the visualization of the vasculature isolated from the white matter. In the process of analyzing the images, they go through manual techniques for processing the images and segmenting the vessels depending on the patient. Besides the images are challenging to analyze due to the nature of the images they present noise related to their acquisition. This is difficult not only for the analysis and the diagnosis of the specialists but also for the performance of AI techniques. This thesis focuses on the problem of denoising these images and evaluating them using unsupervised methods due to the lack of clean images available and in the automation of segmentation. The purpose is to enhance the images for a better visual analysis of the specialist and segmentation for improving the quantification by doing feature extraction of the vessels. For doing so a pipeline was proposed. It consists of using a combination of traditional methods and deep learning-based unsupervised methods for denoising the images. The results were evaluated quantitatively using non-reference image quality evaluators and qualitatively by specialists. For the segmentation, a model was trained using noisy images. Then it was tested using the noisy images and the denoised ones. In total there were 6 comparisons of denoising techniques. The use of unsupervised denoising models utilizing noise2void and probabilistic Noise2Void, in contrast to the application of traditional approaches, as well as the combination of both was compared. The non-reference image quality evaluators, the NIQE and PIQE scores, were used to evaluate the results qualitatively and quantitatively. 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 of noise in the central and most dense areas, according to the qualitative results. The segmentation was done with a UNet model. The two approaches were compared. The results showed that training the model with noisy images and obtaining the segmentation using the denoised images showed an improvement of 9.4\% of the dice score and almost 16\% in the Hausdorff distance.