- Checklist-Mode PET Picture Reconstruction Utilizing Deep Picture Prior
Authors: Kibo Ote, Fumio Hashimoto, Yuya Onishi, Takashi Isobe, Yasuomi Ouchi
Summary: Checklist-mode positron emission tomography (PET) picture reconstruction is a vital instrument for PET scanners with many lines-of-response and extra info akin to time-of-flight and depth-of-interaction. Deep studying is one doable answer to reinforce the standard of PET picture reconstruction. Nonetheless, the applying of deep studying methods to list-mode PET picture reconstruction has not been progressed as a result of record knowledge is a sequence of bit codes and unsuitable for processing by convolutional neural networks (CNN). On this examine, we suggest a novel list-mode PET picture reconstruction technique utilizing an unsupervised CNN known as deep picture prior (DIP) which is the primary trial to combine list-mode PET picture reconstruction and CNN. The proposed list-mode DIP reconstruction (LM-DIPRecon) technique alternatively iterates the regularized list-mode dynamic row motion most chance algorithm (LM-DRAMA) and magnetic resonance imaging conditioned DIP (MR-DIP) utilizing an alternating course technique of multipliers. We evaluated LM-DIPRecon utilizing each simulation and scientific knowledge, and it achieved sharper photos and higher tradeoff curves between distinction and noise than the LM-DRAMA, MR-DIP and sinogram-based DIPRecon strategies. These outcomes indicated that the LM-DIPRecon is helpful for quantitative PET imaging with restricted occasions whereas maintaining correct uncooked knowledge info. As well as, as record knowledge has finer temporal info than dynamic sinograms, list-mode deep picture prior reconstruction is predicted to be helpful for 4D PET imaging and movement correction.
2. Picture Reconstruction for MRI utilizing Deep CNN Priors Skilled with out Groundtruth
Authors: Weijie Gan, Cihat Eldeniz, Jiaming Liu, Sihao Chen, Hongyu An, Ulugbek S. Kamilov
Summary: We suggest a brand new plug-and-play priors (PnP) primarily based MR picture reconstruction technique that systematically enforces knowledge consistency whereas additionally exploiting deep-learning priors. Our prior is specified by a convolutional neural community (CNN) skilled with none artifact-free floor reality to take away undersampling artifacts from MR photos. The outcomes on reconstructing free-breathing MRI knowledge into ten respiratory phases present that the strategy can type high-quality 4D photos from severely undersampled measurements akin to acquisitions of about 1 and a couple of minutes in size. The outcomes additionally spotlight the aggressive efficiency of the strategy in comparison with a number of common alternate options, together with the TGV regularization and conventional UNet3D.
3. Secure Optimization for Massive Imaginative and prescient Mannequin Primarily based Deep Picture Prior in Cone-Beam CT Reconstruction
Authors: Minghui Wu, Yangdi Xu, Yingying Xu, Guangwei Wu, Qingqing Chen, Hongxiang Lin
Summary: Massive Imaginative and prescient Mannequin (LVM) has just lately demonstrated nice potential for medical imaging duties, probably enabling picture enhancement for sparse-view Cone-Beam Computed Tomography (CBCT), regardless of requiring a considerable quantity of knowledge for coaching. In the meantime, Deep Picture Prior (DIP) successfully guides an untrained neural community to generate high-quality CBCT photos with none coaching knowledge. Nonetheless, the unique DIP technique depends on a well-defined ahead mannequin and a large-capacity spine community, which is notoriously troublesome to converge. On this paper, we suggest a secure optimization technique for the forward-model-free, LVM-based DIP mannequin for sparse-view CBCT. Our strategy consists of two principal traits: (1) multi-scale perceptual loss (MSPL) which measures the similarity of perceptual options between the reference and output photos at a number of resolutions with out the necessity for any ahead mannequin, and (2) a reweighting mechanism that stabilizes the iteration trajectory of MSPL. One shot optimization is used to concurrently and stably reweight MSPL and optimize LVM. We consider our strategy on two publicly accessible datasets: SPARE and Walnut. The outcomes present vital enhancements in each picture high quality metrics and visualization that demonstrates lowered streak artifacts. The supply code is out there upon request.
