Three-dimensional shape completion using deep convolutional neural networks: Application to truncation compensation and metal artifact reduction in PET/MRI attenuation correction

Accurate attenuation correction (AC) of PET data is a prerequisite for quantitative PET/MR imaging. However, MR images are susceptible to metal artifacts leading to void MR signal around metallic implants. Moreover, for large patients, exceeding the scanner's field-of-view (FOV), MR images of the body will be truncated, mostly in the arms, thus hampering the accurate delineation of the body contour. Both metal-induced artifacts and body truncation affect PET AC by causing segmentation errors in MRI-based attenuation map generation. In this work, a deep learning convolutional neural network-based algorithm is proposed for the completion of MR images affected by body truncation or metal-induced artifacts. The core of the network utilizes dilated convolutions and residual connections to render an end-to-end 3D shape completion. The training of the network was performed using co-registered PET, CT and MR whole-body images of 15 patients. The evaluation of the proposed method was carried out on 10 patients with severe metal-induced artifacts and truncated MR images. Body contours from the corresponding non-attenuation corrected PET (non-AC PET) or CT images were segmented to estimate the amount of truncation in MR images. The estimated truncated volumes were later used as reference to assess the efficiency of the proposed method to recover the truncated or metal artifact affected areas. Moreover, the impact of the truncation compensation and metal-induced artifact reduction was investigated in the context of segmentation-based PET/MRI attenuation correction. The activity recovery in the affected areas was estimated before and after application of the shape completion method. The body truncation affected 11.1±2.3% of the body volume and consequently the MRI segmentation-based attenuation maps of 10 patients. After shape completion using the proposed method, the amount of truncated volume dropped to 0.7±0.2%. The SUV bias in the truncated area improved from -44.5±10% to -10.5±3% considering PET-CT AC as reference. Likewise, 8.5±1.9% of the head volume was affected by the metal-induced artifact leading to SUV bias of -59.5±11%. These were reduced to 0.3±0.1% and -23.5±9%, respectively, after shape completion. It was concluded that the proposed algorithm exhibited promising results towards the completion of MRI affected by truncation and metal-induced artifacts in whole-body PET/MRI.