Measurement-derived residual-risk verification through four-way k-space partitioning in self-supervised accelerated brain MRI

Self-supervised accelerated MRI reconstruction reduces dependence on fully sampled training targets, but it does not inherently provide a separate basis for verifying whether reconstructed image content remains locally consistent with acquired measurements. We developed a measurement-derived verification framework that partitions acquired k-space into four roles: reconstruction input, self-supervised reconstruction-loss evaluation, residual-risk learning, and reserved verification. Residuals from the reserved measurements were transformed into spatial residual-risk targets using density-compensated adjoint projection, probe-based gain normalization, anatomical support weighting, percentile scaling, and logarithmic compression. Using fastMRI brain multicoil data, we evaluated 201 training, 40 calibration, and 40 AXT2 test volumes, comprising 636 test slices and 999,792 sampled pixels. The primary Full-CNN achieved AUPRC 0.8348, mean absolute error 0.0834, Spearman correlation 0.9389, and AUROC 0.9818. The secondary A4-CNN achieved AUPRC 0.8452 with similar scale and rank agreement. MRI-volume intervals achieved at least 90% sampled-location coverage in 39 of 40 test volumes for both predictors. AUPRC declined and MAE increased with fewer reserved measurements and under AXFLAIR sequence shift. Residual-risk maps showed moderate correspondence with fully sampled reference error and differed from the tested predictive-variance measures. The framework provides a measurement-derived verification layer for assessing local residual inconsistency in self-supervised accelerated MRI without requiring fully sampled development targets.