Equivariant Multiscale Learned Invertible Reconstruction for Cone Beam CT: From Simulated to Real Data

TL;DR

This work tackles CBCT reconstruction quality and memory constraints by introducing LIRE++—a fast, memory-efficient, rotationally equivariant multiscale invertible learned primal-dual network trained end-to-end on simulated CBCT projections. The method jointly handles scatter and primary signals through a physics-informed forward model (water–bone decomposition, poly-energetic projections) and a path-integral scatter framework with quasi-Monte Carlo sampling, enabling accurate reconstructions at 2 mm and 1 mm pitches. Across synthetic and real pelvic CBCT data, LIRE++ delivers higher PSNR/SSIM and lower HU MAE than baselines and existing state-of-the-art methods, with inference times suitable for clinical workflows. The results suggest LIRE++ can enhance online delineation and treatment planning in radiotherapy, while the multiscale, invertible, and equivariant design supports efficiency and robustness for future clinical adoption.

Abstract

Cone Beam CT (CBCT) is an important imaging modality nowadays, however lower image quality of CBCT compared to more conventional Computed Tomography (CT) remains a limiting factor in CBCT applications. Deep learning reconstruction methods are a promising alternative to classical analytical and iterative reconstruction methods, but applying such methods to CBCT is often difficult due to the lack of ground truth data, memory limitations and the need for fast inference at clinically-relevant resolutions. In this work we propose LIRE++, an end-to-end rotationally-equivariant multiscale learned invertible primal-dual scheme for fast and memory-efficient CBCT reconstruction. Memory optimizations and multiscale reconstruction allow for fast training and inference, while rotational equivariance improves parameter efficiency. LIRE++ was trained on simulated projection data from a fast quasi-Monte Carlo CBCT projection simulator that we developed as well. Evaluated on synthetic data, LIRE++ gave an average improvement of 1 dB in Peak Signal-to-Noise Ratio over alternative deep learning baselines. On real clinical data, LIRE++ improved the average Mean Absolute Error between the reconstruction and the corresponding planning CT by 10 Hounsfield Units with respect to current proprietary state-of-the-art hybrid deep-learning/iterative method.

Figures (5)

• Figure 1: Reconstruction quality metrics. Thorax in the top row, pelvic & abdominal in the bottom row.
• Figure 2: (a) Axial (top) and coronal (bottom) slices of thorax CT, HU range=(-1000, 800) and (-150, 250) for ROI, (b) U-net (c) $\partial$U-net, (d) LIRE++
• Figure 3: (a) Axial (top) and coronal (bottom) slices of abdominal CT, HU range=(-400, 400) and (-150, 250) for ROI, (b) U-net (c) $\partial$U-net, (d) LIRE++
• Figure 4: (a) Axial (top) and coronal (bottom) slices of planning pelvice CT, HU range=(-400, 400) and (-150, 250) for ROI, (b) FDK (c) TV++ (d) LIRE++
• Figure 5: (a) Axial (top) and coronal (bottom) slices of pelvic TV++ reconstruction, HU range=(-400, 400) and (-150, 250) for ROI, (b) LIRE++