CT-free Total-body PET segmentation

Introduction: Low-dose positron emission tomography (PET) imaging is made possible with the use of high sensitivity long axial field of view (FOV) PET/computed tomography (CT) scanners. However, the use of CT in this process introduces a considerable radiation burden. To address this issue, deep learning (DL)-based methods have been proposed as a substitute for CT-based PET attenuation and scatter correction. However, the anatomical localization of CT remains unrealized. With more anatomic information provided by extra-high sensitivity total-body PET, we aim to achieve total-body PET multi-organ segmentation on non-corrected PET imaging using a DL approach as a step towards true CT-free PET imaging.

Methods: In this study, we developed a DL-based multi-organ segmentation pipeline for total-body ¹⁸F-FDG PET imaging using a dataset of 114 patients scanned with a Siemens Biograph Vision Quadra. The ground-truth multi-organ segmentation labels were generated using the CT images as input to the Multi-Organ Objective Segmentation (MOOSE) software. A 3D U-Net-like architecture was trained on the non-attenuation and non-scatter corrected PET images. The segmentation pipeline was implemented using the nnU-Net software. Three nuclear medicine physicians independently assessed the utility of the results in a clinical setting.

Results: The trained model achieved an average Dice similarity coefficient (DSC) of 0.82 in the test dataset. The preliminary results show an accurate overlap between the MOOSE-generated labels and our predicted organ segmentations: 70% of targeted organs achieved DSCs of more than 0.80, whereas a few organs exhibited lower scores (e.g., bladder [0.70], thyroid [0.69] and pancreases [0.59]). Visual readings of three nuclear medicine physicians confirmed the accountability of generated segmentations.

Conclusions: Our study explored the possibility of total-body PET multi-organ segmentation using a deep learning-based method that does not require the anatomical information from CT, which represents an important step towards true CT-free PET imaging and has the potential to improve the efficiency and safety of PET scans. Future work will develop models based on manually delineated segmentation labels and expanding the dataset to include data from different types of scanners.

• Download figure
• Open in new tab
• Download powerpoint