Comparison of the whole-body MRI-guided attenuation correction approaches (segmentation-, atlas-, and deep learning-based) in the presence of outliers

Introduction: MRI-guided attenuation correction faces the challenges of the body truncation and metal-artefacts in whole-body imaging.

Purpose: The aim of this study is to compare/assess the performance of the commonly employed MRI-guided attenuation correction (AC) techniques in whole-body PET/MR imaging. These techniques include segmentation-, deep learning-, and atlas-based algorithms.

Methods: This study included 25 co-registered whole-body ¹⁸F-FDG PET/CT and PET/MR images. The in-phase Dixon images were employed to generate synthetic CT images using a residual convolutional neural network, segmentation-based approach (three-class AC map containing soft-tissue, background air, and lung), and voxel-weighting atlas-based method. Considering the CT-based AC as the reference, the bias in activity concentration (standardized uptake value (SUV)) was estimated for the different synthetic CT generation techniques. In addition to the overall performance assessment of these techniques, the primary focus of this study was on recognizing the root cause of the potential outliers such as metal-artifact, abnormal anatomy, small lesions in the lungs, and body truncation in whole-body PET/MR imaging.

Results: The deep learning approach exhibited superior performance to both segmentation- and atlas-based techniques with less than 4% SUV bias across 25 subjects compared to the atlas-based method with 9% bias in the lung and the segmentation-based approach with 20% SUV bias in the bone tissue. Overall, the deep learning-based techniques outperformed the other methods however, in case of severe metallic-artifact or body truncation in the input MRI, this technique led to a suboptimal performance compared to the atlas-based method within the affected regions. On the other hand, regarding the abnormal anatomy, notably patients presenting with small malignant lesions in the lung or pulmonary edema, the deep learning-based technique resulted in promising synthetic CT, and consequently small SUV bias in the PET images. Conclusion: The deep learning-based technique exhibited promising performance in synthetic CT generation from MRI. Nevertheless, the input MR images with severe body truncation and metal-artefacts should be specifically monitored.