HYPR4D Kernel Method with an Unsupervised 2.5SD+0.5TD Deep Learning Assisted Kernel Matrix

• Published in: IEEE conference record - Nuclear Science Symposium & Medical Imaging Conference. p. 1
• Main Authors: Cheng, J.-C. K., Reimers, E., Sossi, V.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.11.2023
• Subjects: Deep learning; Feature extraction; Kernel; Noise reduction; Semiconductor device measurement; Spatiotemporal phenomena; Training
• ISSN: 2577-0829
• DOI: 10.1109/NSSMICRTSD49126.2023.10338043

We describe a deep learning (DL) assisted HYPR4D kernelized reconstruction which produces nearly noise-free voxel level time-activity-curves (TACs) while preserving quantification within small structures as well as consistent spatiotemporal patterns/features within measured data. A series of iteration specific unsupervised DL networks was trained to minimize the inconsistency between the 4D kernelized OSEM subset estimates so that the network output would contain the most consistent features over the subsets for single subject training. We then construct the proposed DL HYPR4D kernel matrix based on the 4D high frequency features extracted from the network output for each reconstruction iteration. Moreover, inspired by the standard 2.5D de-noising models which use 3 slice input, we further include the corresponding slices one time frame before and after the frame of interest to provide spatiotemporal noise reduction (i.e. 2.5SD+0.5TD). Finally, we introduce a final tuning step within the reconstruction to mitigate the typical over-smoothing observed from the network output to preserve the quantification within small target structures. Contrast phantom and human [ 18 F]FDG data acquired on GE SIGNA PET/MR were used for evaluation. The proposed DL HYPR4D kernel method outperformed the standard HYPR4D kernel method as well as TOF-OSEM and TOF-BSREM (Q.Clear) in terms contrast recovery vs noise. The proposed final tuning reduced the underestimation bias due to over-smoothing within a 4mm target structure from ~15% to < 2% while maintaining nearly noise-free voxel level TACs. In addition, the proposed unsupervised DL assisted reconstruction also outperformed the supervised DL version in terms of minimizing biased patterns along the TACs.