Sistema leve colaborativo com redes neurais convolucionais e vision transformers para segmentação de imagens médicas

Abstract

Medical image segmentation is fundamental for providing clear visual information about anatomical structures and anomalies. Currently, tasks related to the extraction of regions of interest increasingly rely on deep learning approaches, which overcome the limitations of traditional methods. In recent years, Convolutional Neural Networks (CNNs) have been widely adopted in numerous studies due to their ability to capture local patterns present in images. However, CNNs exhibit limitations when the objective is to model global relationships among different image regions. To address this issue, a new research direction based on Vision Transformers (ViTs) has emerged, achieving strong performance in various computer vision tasks. In this work, I propose to combine the complementary strengths of CNNs and ViTs in a promising approach for the segmentation of chest computed tomography (CT) images with pulmonary lesions associated with COVID-19. The proposed method integrates CNNs and Vision Transformers within a collaborative system composed of three specialized agents. Each agent operates in a complementary manner, focusing on distinct image characteristics to improve segmentation quality. The system is designed to be lightweight, making it suitable for healthcare systems in developing countries and clinical environments with limited computational resources. The effectiveness of the proposed system was validated on a public lung CT image segmentation dataset using evaluation metrics widely adopted in the literature, including the Dice Coefficient, Jaccard Index, Hausdorff Distance, ASD, and bAHD. The framework achieved consistent performance, reaching Dice Coefficient = 0.8260, Jaccard Index = 0.7330, Hausdorff Distance = 9.3301, ASD = 1.3401, and bAHD = 5.7430. These results demonstrate not only the feasibility of the proposed lightweight collaborative approach but also superior performance when compared to a reference model based on a traditional CNN, implemented exclusively for comparative purposes.