Extending the walk-through PET with anatomical imaging capabilities : performance and applicability of CT-less and custom-CT approaches

Background and purpose The Walk-Through PET (WT-PET) is a novel long axial field-of-view (LAFOV) PET system that consists of two large vertical detector panels. The gap between the panels enables autonomous upright patient positioning, increasing throughput compared to systems with a scanner bed. Moreover, simulations have shown that the WT-PET provides similar imaging performance as a cylindrical system with the same LAFOV, despite using fewer scintillation detectors [1,2]. As such, the WT-PET shows great potential for reducing scan times and costs. An important challenge however is the integration of anatomical imaging capabilities, which is required for attenuation correction (AC) and diagnostics; whereas all commercial systems include a CT component in tandem to the PET for this purpose, their typical annular geometry does not match the WT-PET design. Instead, we propose two alternative approaches, as illustrated in Fig. 1. The first uses the WT-PET as a standalone system and leverages Deep-Learning (DL) methods to generate synthetic CT scans from the acquired non-AC PET images [3]. The second approach is to develop a custom CT, ensuring a good geometric fit. We propose a rectangular frame with spatially distributed cold-cathode X-ray sources that are activated sequentially to obtain projections along different directions. The obtained sinogram data are however incomplete, and need to be reconstructed using either (regularized) iterative methods or completed using DL-based inpainting methods. In this study, we compare approach 1 and 2a, focussing for the latter on DL-based inpainting of the incomplete sinograms before filtered back projection (FBP) reconstruction. Materials and methods Simulations were used to generate data for training and testing the DL-networks, starting from 32 high-quality FDG-PET/CT scans [4]. For approach 1, the WT-PET was simulated in GATE, using the original PET images as activity maps, and converting the CT images to tissue maps by stoichiometric conversion. The simulated coincidences were subsequently reconstructed with MLEM, and the resulting non-AC PET images served as input for a U-Net, which was trained using L1+L2 and perceptual loss. For approach 2, the rectangular CT system was modelled by forward projecting each transverse CT slice and removing the values corresponding to angles and radial coordinates that were not sampled by a source-detector pair. To fill in the gaps, a different U-Net was trained using L2 loss that combined projection-, frequency- and image-domain features. The final anatomical images were then obtained by FBP of the DL-inpainted sinograms. Results Figure 2 evaluates the two approaches in terms of anatomical image quality and AC accuracy. Visual comparison to the ground truth (GT) demonstrates that both the CT-less and rectangular-CT approaches correctly capture the overall patient anatomy. As expected, the custom CT method more reliably captures small details and challenging features such as air cavities, since these are not clearly delineated on the non-AC PET input and vary widely between patients. The difference images support these findings, and additionally reveal slight diagonal streaking artifacts in the rectangular-CT images. The AC-PET images for both approaches correspond well to the GT. Quantitative biases are relatively small, although also here the CT-less approach leads to slight deviations in e.g. air cavities. Conclusion Results suggest that the CT-less approach could be sufficient for (low-dose) AC and possibly anatomical colocalization purposes, while a custom CT component is necessary to obtain reliable anatomical information necessary for diagnostic purposes. The proposed rectangular geometry enables tight integration with the WT-PET and shows promising image quality. Results of this study will be used to decide on the design of the first physical prototype, for which a similar sinogram-inpainting pipeline will be implemented, as well as an iterative reconstruction approach. Acknowledgements and/or conflict of interest The authors would like to thank Kuangyu Shi and Christoph Clement (University of Bern) for their help with providing the PET/CT data that was used in this study. BV and FMM are supported by a personal grant from Research Foundation Flanders (FWO), with respective file numbers 1195125N and 11P0E24N. References [1] Vandenberghe S, et al. Walk-Through Flat-Panel Total-Body PET: a Patient-Centered Design for High-Throughput Imaging at Lower Cost using DOI-capable Monolithic Detectors. EJNMMI (2023). [2] Dadgar M, et al. A Simulation Study of the System Characteristics for a LAFOV PET Design based on Monolithic BGO Flat Panels compared with a pixelated LSO Cylindrical Design. EJNMMI Phys (2023). [3] Lee JS, et al. A Review of Deep-Learning-Based Approaches for Attenuation Correction in Positron Emission Tomography. IEEE TRPMS (2020). [4] Shi K, et al. “Ultra-low Dose PET Imaging Challenge 2022”. MICCAI (2022)