Segmentação de vasos sanguíneos utilizando redes neurais convolucionais: investigação da prevalência de descontinuidades e desenvolvimento de técnicas para mitigá-las

Abstract

With the current advancements in technology and medical techniques, the pursuit of improving diagnostic methods and scientific analyses brings forth a significant challenge: the efficiency and effectiveness in processing clinical data. In the field of medical image processing, there is a crucial phase that can influence all subsequent steps and even the final diagnosis, which is the segmentation phase. Particularly, segmentation is vital in examinations involving images of blood vessels, as these structures pose a great difficulty in analysis due to their complex and thin nature. Seeking improvements in segmentation techniques is of paramount importance, considering it is a highly sensitive phase of analysis, and a simple change of lens or imaging acquisition device can compromise the quality of the samples. Literature studies focus on quantifying the quality of segmentation methods using global metrics such as accuracy, precision, and recall, but often this focus may lead to problems in vessel geometry, where important information such as bifurcations, continuity, and diameter are lost, thereby causing various diagnostic issues. This work focuses on the analysis of continuity problems in vessel segmentation, i.e., cases in which parts of vessels are not correctly detected. An accuracy metric is defined to specifically quantify the segmentation quality in regions of vessels that are difficult to segment. It is demonstrated that this metric enables more precise quantification of the quality of low-saliency vessel detection in images than traditional metrics. Additionally, a data augmentation technique is defined for training neural networks, enabling improved segmentation quality of low-saliency vessels. The technique involves the creation of regions with drops in intensity and vessel discontinuities. Based on the analyses conducted, it is expected that the developed techniques can assist in improving diagnoses and future research in biology, creating new possibilities for addressing segmentation problems.