PET-QA-Net: Towards Routine PET Image Artifact Detection and Correction using Deep Convolutional Neural Networks

Abstract

Nowadays PET imaging is routinely coupled with anatomical imaging in the form of PET/CT and PET/MRI. CT or MR images are commonly used to correct for attenuation and scatter during the image reconstruction process to enable quantitative PET imaging. PET image artifacts arise from three main domains, (i) PET images (halo-artifact), (ii) interface between PET and CT/MR images (mismatch, misregistration and motion artifacts in the upper abdomen or thorax regions), and (iii) CT/MRI artifacts (e.g. metal, contrast, agent, truncation) propagate to PET images. We propose an emission-based attenuation and scatter correction method that would be immune to these artifacts. Our aim is to embed this framework within routine clinical imaging as a fast and efficient quality assurance (QA) tool that can be used to detect and correct artifacts for improved PET image quality and quantification. In this study, we analyzed 1650 clinical PET/CT images. All PET images were corrected using CT images for attenuation and scatter during PET image reconstruction. The dataset was randomly split into train (1200), validation (250) and test (200) sets. Uncorrected (for ttenuation and scatter) PET images were used as input to our proposed deep neural network (PET-QA-Net) to generate attenuation corrected (CT-based) PET images. We implemented a modified Res-U-Net architecture composed of encoder and decoder parts, trained in a 2D manner. We used 6 blocks for the encoder and 6 blocks for the decoder part which each 2 block were connected using residual connection. For model evaluation, voxel-wise mean error (ME), mean absolute error (MAE), relative error (RE%), absolute relative error (ARE%), and structural similarity index (SSIM) were calculated. We achieved a ME of 0.02±0.05, MAE of 0.20±0.07, RE of 1.52±3.5%, ARE of 10.0±4.5% and SSIM of 0. 97±0.02 in the test set. We considered MAE>0.4 as an outlier and inspected the images visually.We reported different artifacts that were detected and corrected using the proposed framework. Overall, we have built a highly effective and fast quality control tool that can be used routinely to detect and correct PET image artifacts in clinical setting.