Author: Renee Farrell, Jinkoo Kim, Xin Qian, Ziyu Shu, Zhaozheng Yin, Tiezhi Zhang 👨🔬
Affiliation: Stony Brook Medicine, Washington University in St. Louis, Stony Brook University, Stony Brook University Hospital 🌍
Purpose: Ultra-short CT scan allows fast imaging speed, dose reduction, and compact system design. We developed a deep image prior (DIP) based reconstruction method named Hybrid Prior-Enhanced Deep Image Prior (HPEDIP) for limited angle reconstruction.
Methods: DIP introduces a randomly initialized convolutional neural network as a prior for solving ill-posed inverse problems. By optimizing the network parameters solely to match the measured data, it can achieve high-accuracy reconstructions without requiring any training data. Building on this principle, we propose a limited-angle CT reconstruction method, HPEDIP, based on the DIP, which incorporates the traditional Iterative Reconstruction (IR) algorithm as an additional prior within the DIP framework. This design allows our method to inherit the strengths of both DIP and IR while addressing their respective limitations. Moreover, reference images, such as body profiles and template images, can also be utilized as priors to further improve HPEDIP’s performance.
Results: We evaluated the HPEDIP limited angle reconstruction on both phantom and real CT data. Our algorithm effectively filters out erroneous information from the reference images while leveraging their meaningful content to further enhance the algorithm's accuracy. Consequently, HPEDIP achieves accurate reconstructions for the limited-angle CT reconstruction problem. Compared with traditional IR algorithms, HPEDIP nearly perfectly eliminates the streak artifacts caused by missing projections. Moreover, compared with the standard DIP algorithm, HPEDIP suppresses network-specific artifacts and greatly improves the stability of the network.
Conclusion: HPEDIP can accurately reconstruct CT images with a restricted projection angular range, enabling ultra-short CT scans. Also, the proposed method can serve as a higher-level replacement for DIP in most existing DIP variants to further improve their performance.