Abstract
242464
Introduction: PET quantification methods require high-resolution 3D MRI which may not be available at the time of analysis. Utilizing CT from PET/CT imaging for the segmentation of cortical regions can be a valuable alternative with many potential advantages. Therefore, this study aims to develop a deep learning (DL) model for parcellating brain CT from PET/CT into 46 volumes of interest (VOIs) for accurate amyloid quantification in 18F-florbetaben (FBB) PET/CT imaging, without the need for manual labeling.
Methods: The pairs of brain axial FBB PET/CT and T1-weighted MRI in 226 patients presented in dementia clinic from February 2016 to November 2022 were included. For the training and evaluation of the DL model performing whole brain CT parcellation, the dataset was divided into training (54%), validation (6%), and independent test (40%) sets. Utilizing automatically generated pseudo segmentation labels, three UNets were independently trained for multi-planar whole brain parcellation on CT scans, and subsequently ensembled.
The proposed CT-based DL model was compared to MRI-based (FreeSurfer) method by assessing the agreement of standardized uptake value ratio (SUVR) for FBB using linear regression analysis, and intraclass correlation coefficient (ICC). Through the same approach, the consistency between CT-based SUVR and MRI-based SUVR for different clinical dementia rating (CDR) score groups was evaluated within the independent test set, composed of 30 patients each with CDR scores of 0, 0.5, and 1.
Results: The training and validation set, as well as the independent test set, consisted of patients who underwent both FBB PET/CT and MRI, totaling 136 (mean age 71.8 ± 7.7) and 90 (71.9 ± 8.0) patients respectively. Trained DL-based CT parcellation model showed a mean dice similarity coefficient of 0.80 for 46 total VOIs, and 0.72 for 16 cortical and limbic VOIs. For global SUVR, the linear regression analysis yielded a slope of 1.013, an y-intercept of -0.017, and an R^2 of 0.997 (P < 0.001), and the ICC was 0.999 (P < 0.001). Additionally, in each CDR group, both global SUVRs showed linear regression analysis results with a slope within 1 ± 0.020, a y-intercept within 0 ± 0.026, R^2 of at least 0.993, and an ICC of at least 0.996.
Conclusions: The deep learning-based quantification using CT from PET/CT demonstrated strong agreement in assessing SUVR in amyloid PET scans compared to quantification based on MRI. The application of deep learning to CT data has the potential to enhance the usefulness of amyloid PET quantification at the individual subject level.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.