Avaliação de modelos de segmentação para remoção de ruídos em tomografias computadorizadas

Abstract

Low-dose computed tomography (LDCT) reduces exposure to ionizing radiation, but ge- nerates noisy images with poor quality. This work analyzes the use of convolutional neural networks for segmentation to remove noise from LDCT images, aiming to preserve structural details and relevant information. Four architectures were evaluated: U-Net, LinkNet, PSPNet, and MA-Net. These networks were combined with different backbones (ResNet50, VGG16, DenseNet121, Inceptionv4, MobileNetV2, and EfficientNet-B2). The AAPM 2016 dataset from the Mayo Clinic was used. The results demonstrate that the combination of U-Net + Incepti- onv4 obtained the best performance, with a PSNR of 48.797 and SSIM of 0.977, outperforming traditional methods. The analyses show that the use of more advanced attention mechanisms and backbones can improve detail recovery in noisy images.