Disentangling weak spatial priors from generic local context in radiographic abnormality detection: a controlled mechanism audit

Background: Improvements from spatially guided crops may reflect generic local context, optimization or operating-point selection rather than informative spatial guidance. Purpose: To assess whether a coarse foreground guidance map improves radiographic abnormality discrimination beyond non-prior local controls. Methods: We audited original frozen DenseNet121 pipelines and conducted a revision-stage matched experiment using five seeds, patient-disjoint development partitions, and prior-guided, centred and random-position crops. Crops retained 90% of image width and height. The primary revised comparison used a common eight-epoch local budget and study ROC-AUC. Validation-selected checkpoints and fixed 50:50 global–local fusions were secondary. Evaluation used 1,199 studies from 783 patients, with paired patient-cluster bootstrap intervals. This cohort had already been evaluated in the original study. Results: At the common final epoch, mean ROC-AUC was 0.892269 for prior-guided, 0.891573 for centred and 0.892221 for random-position crops. Mean paired AUC differences were +0.000696 (seed SD 0.002757) versus centred and +0.000048 (SD 0.002693) versus random. All ten primary per-seed 95% intervals included zero. Fallback exclusion and alternative aggregation yielded small descriptive differences. Original fusion pipelines had similar AUC but different operating profiles at their independently frozen thresholds. Conclusion: Within MURA and this conservative DenseNet121 crop procedure, spatial guidance showed no consistent ranking advantage over the tested controls. The findings do not establish equivalence or exclude benefits from stronger or different priors. Matching training exposure and reporting discrimination, calibration and operating-point behaviour separately strengthen the audit without establishing clinical trustworthiness.