End-to-end memory-efficient reconstruction for cone beam CT.

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Cone Beam Computed Tomography (CBCT) is crucial in medicine but suffers from lower image quality than conventional CT.
• Deep learning (DL) shows promise for CBCT reconstruction but faces challenges with high memory costs and limited generalization.

Purpose of the Study:

• To address limitations of current DL methods for CBCT reconstruction.
• To propose LIRE (Learned Invertible primal-dual REconstruction), an efficient and generalizable DL-based reconstruction scheme.

Main Methods:

• LIRE utilizes a learned invertible primal-dual iterative scheme with U-Net and residual CNN architectures.
• Memory efficiency is achieved through invertible blocks and patch-wise computations, enabling training on high-resolution data with limited VRAM.
• The method is trained and validated on thorax CT scans and tested on both thorax and head/neck CT datasets.

Main Results:

• LIRE outperformed classical and baseline DL methods on both thorax and head/neck CT datasets.
• Achieved superior Peak Signal-to-Noise Ratio (PSNR) values compared to U-Net baseline for both small and large field-of-view settings.
• Demonstrated successful fine-tuning for high-resolution (1 mm voxel spacing) CBCT reconstruction, outperforming the baseline.

Conclusions:

• Learned invertible primal-dual schemes with memory optimizations can effectively reconstruct CBCT volumes.
• LIRE offers improved reconstruction quality and better generalization than traditional DL approaches for CBCT.