Artificial Intelligence Benchmarking of Sparse-View Cranial Computed Tomography Reconstruction Under Unmeasured X-Ray Tube Current Drift: A Healthcare Diagnostic-Confidence Evaluation

Sparse-view cranial computed tomography reconstruction is benchmarked almost exclusively on image-similarity metrics computed against a fully sampled reference, an evaluation design that presumes the reference is itself stable. In clinical practice the X-ray tube current drifts between calibration and acquisition, and the drift is unmeasured, so the reference against which a reconstruction is scored carries an unknown and correlated error. the present analysis treats that evaluation problem as a formal one. It defines a benchmark estimator for reconstruction fidelity under unmeasured tube-current drift, derives its bias as a function of the drift-to-signal ratio and the angular sampling fraction, and shows that the bias is not removed by increasing the number of test volumes because the drift is shared across the reconstruction and its reference. A design-based operating-characteristics analysis is then reported: a before-after-control-impact estimator was evaluated across 2000 replications per design cell, with cluster counts of 8, 16, 24 and 32, effect sizes of 3, 5 and 8 percentage points, a between-cluster standard deviation of 6.0, a residual standard deviation of 4.5 and an intraclass correlation of 0.64. Fifty-four clusters were required for 80% power at a 5-percentage-point effect. The contribution is a diagnostic-confidence framework that separates reconstruction error from reference error and reports the drift-conditional confidence interval a benchmark can legitimately support.