Author: Justus Adamson, Mu Chen, Ke Lu, Zhenyu Yang, Fang-Fang Yin, Rihui Zhang, Yaogong Zhang, Haipeng Zhao, Haiming Zhu, Yuchun Zhu 👨🔬
Affiliation: Shanghai Dacheng Medical Technology, Duke University, Medical Physics Graduate Program, Duke Kunshan University, Duke Kunshan University, The First People's Hospital of Kunshan 🌍
Purpose: In filtered back-projection (FBP) reconstruction, conventional filters often reduce noise at the expense of high-frequency details, leading to structural details loss. To address this limitation, we propose a deep-learning based “deep-filter” to replace the conventional FBP filter, aiming for an improved balance between noise suppression and detail preservation.
Methods: We developed an end-to-end reconstruction framework that maps sinograms to reconstructed images by combining learnable deep-filter with fixed back-projection operator. The deep-filter exploits the powerful feature extraction and nonlinear representation of neural network, while we preserve the physical back-projection mechanism by keeping the back-projection operator non-trainable, allowing the deep-filter to discover the optimal filter during training. The large convolution kernels were employed in the projection domain to capture a wide range of contextual features, and the small convolution kernels were subsequently used with increasing depth to achieve hierarchical feature encoding and decoding. The filtered sinogram was then mapped back to the image domain through the back-projection operator to generate the reconstructed CT image. Poisson noise was additionally added to sinogram to simulate the clinical image acquition situation. The generated noisy sinograms were used for training and evaluation to ensure robustness and generality.
Results: The results show that the proposed deep-filter outperforms all tested conventional filters Shepp-Logan, Hamming, Hann, and Cosine, achieving a lower RMSE of 0.022 (conventional filters range from 0.023–0.028) and a higher SSIM of 0.895 and PSNR of 32.989dB.
Conclusion: This study verifies the feasibility of replacing conventional FBP filter with deep-learning based "Deep-filter” and achieves better noise-signal balance and preserves fine structural details through a reconstruction framework that integrates neural network and physics-driven back-projection operator. It explores the feasibility of combining data-driven deep learning techniques with physical constraints in the field of reconstruction.