Abstract / Summary
Abstract Background Medical-image classifiers can obtain high apparent performance from correlated slices, acquisition patterns, or anatomy outside the intended target. The downloaded CAD Cardiac MRI Dataset release contained 63,425 JPEG images but only 30 independent top-level patient folders. This study evaluated whether cardiac-region representations were more informative than exact same-slice full-image and extra-cardiac controls under patient-level validation. Methods Each Directory_* folder was treated as one patient (16 Normal, 14 Sick), as confirmed by the dataset creator. The current analysis used five patient-level cross-fitted 2.5D Attention U-Nets trained from 5,097 accepted targets (3,949 heart-present and 1,148 no-heart). Every released image received an out-of-fold mask from a model that did not train on its patient. A frozen ImageNet-pretrained EfficientNet-B0 encoded full-image, cardiac-ROI, and exact-complement representations. Slice embeddings were averaged first within folder-defined series and then across series for each patient. Nested patient-level cross-validation evaluated eleven representations using exact same-slice controls, a secondary cross-class matched cohort, and paired patient-bootstrap intervals. Results The expanded manual-target state caused all five segmentation folds and all five out-of-fold prediction parts to be recomputed. Out-of-fold prediction produced 63,107 final-valid masks among 63,425 images; positive-target patient-balanced Dice was 0.809, and 73.4% of accepted no-heart targets received empty masks. On exactly 50,735 attention-eligible slices, the automatic cardiac ROI achieved AUC 0.938 (95% CI 0.847-1.000), compared with 0.830 for the full image and 0.772 for the complement. Paired AUC differences were 0.107 (95% CI 0.009–0.235) versus the full image and 0.165 (0.029–0.344) versus the complement. In the matched cohort, ROI AUC was 0.942, versus 0.821 for the full image and 0.741 for the complement. On the manual-positive subset, automatic and manual ROIs achieved AUCs of 0.973 and 0.951, respectively; their paired difference did not exclude zero. Conclusions After expanded manual annotation and a complete rerun of every dependent stage, out-of-fold cardiac localization retained an advantage over exact same-slice and matched controls. Complement AUCs of 0.741–0.835 nevertheless show substantial residual predictive information outside the intended cardiac region. With only 30 patients, no external validation, JPEG rather than DICOM data, and outcome-informed matching in a secondary analysis, these findings support a methodological proof of concept rather than clinical use.