Segmentação do lúmen e identificação de região de bifurcações em imagens de tomografia por coerência óptica intravascular utilizando redes neurais convolutivas

Abstract

Intravascular Optical Coherence Tomography (IVOCT) technology enables the experts to analyze coronary lesions from high-resolution images. Some level of automation could benefit experts since visual analysis of pullback frames is a laborious and time-consuming task. Even with the growing popularity of Convolutional Neural Networks (CNN) in the medical area, there are few works in the literature applying them to lumen segmentation and classification of bifurcation regions tasks. In this work, we evaluated three CNN architectures for the lumen segmentation task, and four architectures for bifurcation region classification, using an IVOCT image set of nine pullbacks from nine different patients. Regarding lumen segmentation task, direct networks and direct acyclic graph (DAG) networks were evaluated in different datasets, varying spatial resolution, coordinate systems, and color space. Regarding bifurcation region classification, besides variations in the coordinate systems and color space of the datasets, data augmentation techniques were used to balance them, in order to compensate for the smaller number of bifurcation images, besides using transfer learning in some of the evaluated networks, applying knowledge acquired from one of the segmentation networks. Our results are comparable to the other works found in the literature, presenting, for segmentation, best results in accuracy, Dice coefficient, and Jaccard over 99%, 98%, and 97%, respectively. In classification, better results were presented in F1 score (99,68%), and AUC (99,72%) obtained by a CNN with transferred knowledge.