A Deep Learning Based Technique For Truncation Completion In Pet/Mr
Haojun Yu,Shuguang Chen,Hui Liu,Tuoyu Cao,Lingzhi Hu,Hongcheng Shi
2019-01-01
Journal of Nuclear Medicine
Abstract:2028 Objectives: Arm truncation artifact was the common artifact in PET/MR, and may lead to misidentification of smaller lesions or quantification error. Conventionally the contour extracted from 18-F-FDG-PET images was combined with arm-truncated-MR water / fat images for 5-compartment attenuation correction (AC) map generation. However, the PET contour ignored the tissue context, and may lead to incorrect estimation of AC map. The cycle generative adversarial networks (cycle-GAN) trained with paired or un-paired A-B datasets was able to synthesize B from A. In this study, we modified generator part of the cycle-GAN with fast convergence to map domain A (PET + truncated MR water and fat images) and B (none-truncated MR water and fat images) for PET/MR AC map generation. Methods: 10 patients’ whole body PET/MR images with partial-arm covered (acquired at uPMR-790, United-Imaging Healthcare, Shanghai, China) were respectively included into this study. 8 patients’ images were used for training and 2 patients’ images were for test purpose. The input shape to the cycle-GAN A was 256x256x3 representing PET + MR water + MR fat images (matrix size: 256x256), and the input shape to cycle-GAN B was 256x256x2 designed for MR water + MR fat images. An open dataset for style transferring (horse vs. zebra, 1067 images) was processed to accommodate the inputs and used as the base for transfer-learning. In the training phase, the preparation of inputs to A and B includes 1) cropping both A and B in the left-right (LR) direction to 80%, 2) cropping MR images (water and fat images) in A by 80% in order to have B as reference. In the testing phase, the generator for A-to-B were used to take the original PET + MR water + MR fat images as input. 20 coronal, transversal slices of corresponding contrast (PET, MR water,MR fat) images were selected from each patient’s PET/MR whole body scan. Flip, mirror and scale operations were performed to augment datasets, and total 480 slices were used for training, where 8 coronal slices were used for testing. All the images were normalized and scaled to [0, 1] before the input to network. Tissue compartment consistency was evaluated by comparing the consistency between synthesized tissue compartment with common knowledge (fat, muscle, bone), where consistency in one compartment distribution contributes to score by adding 1 point. The shape consistency was scored visually in range of 0 - 3, with 1 point representing roughly 30% consistency. Results: The synthesized MR water images were used for performance evaluation. The mean absolute errors of normalized images were 8.9, 9.6 for train, test image sets respectively. The shape tissue compartment and shape consistency were visually evaluated in 10 slices for each case. The shape consistency score was 2.5, where tissue compartment consistency score was 2.1 (Figure 3). Conclusions: We demonstrated the feasibility of synthesis none-arm-truncated MR water and fat images from PET, truncated MR water and fat images. It provides new opportunity for PET/MR attenuation correction.