Equivariance2Inverse: A Practical Self-Supervised CT Reconstruction Method Benchmarked on Real, Limited-Angle, and Blurred Data

Deep learning has improved CT reconstruction, but self-supervised methods often rely on oversimplified X-ray physics—ignoring scintillator blur, non-ideal scanning geometry, or correlated noise. As a result, they break down in real-world clinical or industrial settings. Enter Equivariance2Inverse, a new approach from researchers at CWI and Utrecht University that explicitly addresses these gaps. The method marries the robustness of Robust Equivariant Imaging (which enforces equivariance to transformations) with the inversion priors of Sparse2Inverse, creating a self-supervised pipeline that needs no ground-truth data yet stays accurate even when scans are limited in angle or suffer from sensor blur.

Benchmarked on the real 2DeteCT dataset and synthetic data with controlled blur and limited-angle geometry, Equivariance2Inverse outperformed six existing self-supervised methods. The results were telling: methods that treat noise as pixel-independent fail catastrophically under scintillator blur. More importantly, the team showed that if the objects being scanned are rotationally symmetric, exploiting that equivariance dramatically reduces artifacts from limited-angle scans. This opens the door to safer CT—fewer X-ray exposures, more robust reconstructions—without sacrificing quality.