Work

Fast Prediction of CT Spatial Resolution from Acquisition Parameters and Image Patches Using Physics-Simulated Data and Machine Learning

Year: 2026
Type: preprint
Abstract: The spatial resolution of a computed tomography (CT) scan, summarised by the frequency at which the modulation transfer function falls to one half (MTF50), depends jointly on focal-spot size, geometric magnification, detector blur and reconstruction voxel size. We built a simulation pipeline that scans an analytic slanted-edge and bar-pattern phantom at 400 acquisition settings and measures the ground-truth MTF with the edge-spread-function method and used it to train two kinds of fast estimators. A Gaussian-process surrogate predicts MTF50 from the four settings with a mean absolute error (MAE) of 0.106 lp/mm and R 2 = 0.983 on held-out settings, compared with 1.285 lp/mm for the classical quadrature blur formula and 0.293 lp/mm for a calibrated version of it, about 2,000 times faster than a single simulation. Adding physics-derived blur features improved every model and kept errors low in several extrapolation scenarios in which raw-feature models failed. A small residual convolutional network that sees only a 64 × 64 image patch reached 0.121 lp/mm per setting, close to the surrogate, but its error rose to 0.60 lp/mm on unseen generic phantoms, exposing a clear out-of-distribution limit. Within the studied ranges, magnification and detector blur were the most influential settings. Code, data and trained models are released.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7225918680
Cites: 10
Cited by:
Related to: 10