STCNet
A dual-domain spatiotemporal consistency learning framework for sparse-view 4D CBCT reconstruction. By introducing physics-driven positional encoding in the projection domain, temporal encoding in the latent space, and jointly optimizing the projection and image domains, it achieves high-fidelity, temporally coherent 4D reconstruction.
Timeline (last 90 days)
In sparse-view 4D CBCT reconstruction, the STCNet framework requires only 16 projections per respiratory phase to reconstruct high-quality images.
STCNet adopts a dual-domain collaborative strategy, introducing physics-driven positional encoding in the projection domain and temporal encoding in the latent space, jointly optimizing the projection domain and the image domain.
On the TCIA-LUNG dataset, STCNet achieves PSNR of 33.311 dB, 32.695 dB and 32.658 dB at phases 2, 5 and 8 respectively, with SSIM of 0.845, 0.823 and 0.825 respectively.
On the CIRS chest phantom dataset, STCNet achieves a maximum PSNR of 36.497 dB and a maximum SSIM of 0.877.
STCNet outperforms traditional methods and recent deep learning baselines in all three aspects: image quality, artifact suppression and structure preservation.
The PSNR at the 33 dB level and SSIM above 0.82 reported by STCNet are at the leading level in the current literature under the 16-projection condition, but these values were obtained on a specific dataset, and cross-dataset generalization still needs to be verified.
STCNet's current results are still limited to public datasets and phantoms, and complex factors such as real-patient respiratory variation, positioning errors and metal artifacts have not yet been covered.
The computational overhead of the dual-domain collaborative strategy is not discussed in detail in the paper, and it may become a bottleneck for deployment under clinical real-time requirements.
The physics-driven positional encoding introduced by STCNet provides an inductive bias that ensures the reconstruction result does not contradict the geometric relationship of the X-ray projections, and the temporal encoding constrains the motion between phases to conform to respiratory dynamics.
STCNet's design approach is in the same lineage as physics-informed neural networks (PINNs), with the core goal in both being to keep deep learning models reliable and interpretable even in data-scarce or out-of-distribution scenarios.
1 more timeline events
All Facts (11)
In sparse-view 4D CBCT reconstruction, the STCNet framework requires only 16 projections per respiratory phase to reconstruct high-quality images.
50%UnverifiedSTCNet adopts a dual-domain collaborative strategy, introducing physics-driven positional encoding in the projection domain and temporal encoding in the latent space, jointly optimizing the projection domain and the image domain.
50%UnverifiedOn the TCIA-LUNG dataset, STCNet achieves PSNR of 33.311 dB, 32.695 dB and 32.658 dB at phases 2, 5 and 8 respectively, with SSIM of 0.845, 0.823 and 0.825 respectively.
50%UnverifiedOn the CIRS chest phantom dataset, STCNet achieves a maximum PSNR of 36.497 dB and a maximum SSIM of 0.877.
50%UnverifiedSTCNet outperforms traditional methods and recent deep learning baselines in all three aspects: image quality, artifact suppression and structure preservation.
50%UnverifiedThe PSNR at the 33 dB level and SSIM above 0.82 reported by STCNet are at the leading level in the current literature under the 16-projection condition, but these values were obtained on a specific dataset, and cross-dataset generalization still needs to be verified.
50%UnverifiedSTCNet's current results are still limited to public datasets and phantoms, and complex factors such as real-patient respiratory variation, positioning errors and metal artifacts have not yet been covered.
50%UnverifiedThe computational overhead of the dual-domain collaborative strategy is not discussed in detail in the paper, and it may become a bottleneck for deployment under clinical real-time requirements.
50%UnverifiedThe physics-driven positional encoding introduced by STCNet provides an inductive bias that ensures the reconstruction result does not contradict the geometric relationship of the X-ray projections, and the temporal encoding constrains the motion between phases to conform to respiratory dynamics.
50%UnverifiedSTCNet's design approach is in the same lineage as physics-informed neural networks (PINNs), with the core goal in both being to keep deep learning models reliable and interpretable even in data-scarce or out-of-distribution scenarios.
50%UnverifiedSTCNet is a cross-disciplinary attempt at sparse-view reconstruction and dynamic sequence modeling, not aiming to replace fully sampled CT, but to provide a feasible solution for low-dose, phase-resolved 4D CBCT.
50%