Viabilidade da aplicação de redes neurais para a correção de erros de segmentação em vasos sanguíneos

Abstract

Convolutional Neural Networks (CNNs) represented a significant milestone for several areas of computer vision, being widely used in tasks such as image classification, segmentation and object detection, both in research and in industry. In particular, these networks promoted important advances in the biomedical field, where they are applied to the segmentation of microscopy and X-ray images, helping to identify possible anomalies in blood vessels. However, diagnostic accuracy can be compromised by the presence of errors or discontinuities in the vascular segments extracted by a CNN. Faced with this challenge, this study seeks to evaluate the feasibility of using a second neural network to correct errors in the results of a CNN. To this end, artificial images representing the geometry of blood vessels using Bézier curves were generated, allowing the CNN's ability to restore incomplete segments to be tested. Based on this approach, CNNs were trained on different datasets, varying the number of segments and the degree of discontinuity of the branches. The results demonstrated that the network was effective in correcting the segmentations, achiving high values in quality metrics. However, as the removal of parts os vessels increased, the model had difficulty identifying clear patterns for reconstruction, mainly due to the complexity of the curve, thus compromising the continuity of the segments. In order to improve the results, two loss functions were also tested: cross-entropy and Tversky. Although the overall difference in segmentation quality was small, the Tversky function showed promising results in cases of small removals, demonstrating better recovery of lost segments.