Removing Artifacts from Computed Tomography Images of Heart Using Neural Network with Partial Convolution Layer

Nowadays, metal artifacts caused by the implantable device leads may impair the quality of computed tomography (CT) images of the heart. It makes processing and analysis of the images to be difficult. In this article, a method of removing metal artifacts from the CT images of the heart for patients with implantable cardiac resynchronization therapy (CRT) devices was developed.This method is based on using the UNet neural network architecture with partial convolution layers. The training dataset consisted of images with small artifacts processed by the normalized metal artifact reduction (NMAR) method and images without metal artifacts; images with large artifacts were not included in the training dataset. For each image, masks were randomly generated. These masks determine the regions where network should predict pixel values. Note, the regions were created in such a way that their shape and area were similar to a real metal artifact.Method we used in this research demonstrated better results than conventionally used NMAR. Particularly, index of peak signal to noise ratio (PSNR) and index of structure similarity (SSIM) for full size images in test dataset were 45.061 and 0.932, and the indexes for images cropped to predicted area were 36.67 and 0.92, respectively. In comparison, same indexes for NMAR processed images were much lower – 21.55 and 0.62.