Hybrid Framework Combining Pre-Trained Supervised Image Priors with Neural Representation for Extremely Sparse-View CT Reconstruction (HYPER)
Abstract
Purpose
Extremely sparse-view CT benefits for reducing radiation dose while causing streak artifact when using the traditional filtered-back projection (FBP). We propose a new learning-based reconstruction method, named HYPER (HYbrid framework combining pre-trained supervised image Priors with Neural Representation for Extremely sparse-view CT Reconstruction).
Methods
The HYPER framework consists of three features: (1) deep learning-based pre-trained networks, (2) streak artifact removal applied to the generated images, (3) self-supervised Implicit Neural Representation(INR) reconstruction scheme that enforces data fidelity guided by image priors. ResNet model trained on paired FBP-reconstructed and ground-truth images generated high-quality CT images. These images were forwardly projected to generate full- and sparse-projection views, therefore producing realistic, streak artifact-free images. Finally, the INR was iteratively updated with the image priors generated by pre-trained ResNet and artifact removal process. We validated on the AAPM 2016 Low-Dose CT dataset using 9 subjects for training/validation and 1 unseen subject for testing with CT images resized to 256 × 256. We simulated parallel-beam projections with 10 views sampled from a total of 720 views. The performance of the HYPER framework was compared to previous learning-based reconstruction methods (FBPConvNet, NeRP, MCG, and DiffusionMBIR) in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).
Results
The baseline FBP reconstruction achieved PSNR(dB)/SSIM of 22.08/0.3536. FBPConvNet, NeRP, MCG, and DiffusionMBIR yielded 30.29/0.8275, 29.92/0.7746, 33.98/0.8769, and 34.71/0.9139, respectively. In contrast, the proposed HYPER framework further improved reconstruction accuracy, achieving a PSNR of 39.22 dB and an SSIM of 0.9531 under the 10-view extremely sparse-view condition.
Conclusion
The proposed HYPER framework for extremely sparse-view CT reconstruction combined pre-trained supervised image priors with self-supervised implicit neural representation for extremely sparse-view CT reconstruction. This hybrid approach successfully produced highly qualified CT images reconstructed from only 10 views, outperforming state-of-the-art reconstruction methods.