Content-based mammographic image retrieval requires exact BIRADS categorical matching across five classes, posing far greater complexity than conventional binary classification. Existing studies remain limited by small sample sizes, improper patient-level separation, and inadequate statistical validation, restricting clinical translation. We developed a comprehensive evaluation framework systematically comparing CNN architectures (DenseNet121, ResNet50, VGG16) under advanced training strategies: fine-tuning, metric learning, and super-ensemble optimization. Rigorous patient-stratified splits (1003 patients, two images each), 602 test queries, and bootstrap confidence intervals (1000 resamples) ensured reliable assessment. Advanced fine-tuning and test-time augmentation (TTA) yielded a precision@10 of 34.71% for DenseNet121 _AdvancedFT_TTA, a 25.74% improvement over the baseline ResNet50 (27.6%). Selective super-ensemble and metric learning approaches were further benchmarked under patient-exclusive splits, confirming robust performance across architectures. Statistical analysis (bootstrap CIs, n = 1000; t-tests p 0.8) validated significant gains and reproducibility. These results establish DenseNet121_AdvancedFT_TTA as the new state-of-the-art for five-class BIRADS retrieval while reducing computational cost.