Deep GLR: CT Reconstruction with 91K Parameters and 1000 Samples

A new paper from Veera Varuni Radhakrishnan and colleagues introduces Deep Graph Laplacian Regularization (Deep GLR), a parameter-efficient method for low-dose CT (LDCT) reconstruction. Unlike typical deep learning models that require over 500,000 parameters and 35,000 scans, Deep GLR uses just 91,848 parameters trained on 1,000 samples — a 94% reduction in data. It integrates quadratic graph regularization into a Proximal Forward-Backward Splitting optimization framework with three lightweight CNN modules, achieving 30.70 dB PSNR on the LoDoPaB-CT benchmark (6.33 dB better than filtered backprojection). This represents 5.8× better parameter efficiency and 30× better data efficiency per dB improvement compared to state-of-the-art methods.

While Deep GLR still trails top-tier models by 13 dB, its strength lies in extremely resource-limited scenarios. The learned graph bandwidth parameter (ε=1.25) converges to interpretable values, indicating the model captures meaningful image priors rather than overfitting. This makes Deep GLR particularly promising for portable CT scanners or facilities in developing regions where compute and data are scarce. The tradeoff between quality and efficiency opens a practical path for deploying AI-assisted CT reconstruction where traditional deep learning is infeasible.