What Actually Matters in Deep Learning for Breast Ultrasound Classification: A Systematic Ablation Study With Wilcoxon Signed-Rank and Holm–Bonferroni Analysis

Deep learning pipelines for breast ultrasound classification often combine multiple training-stabilization components, yet the contribution of each remains uncertain. Using a ResNet34 backbone, we evaluated twelve configurations in a top-down ablation sequence from a fully equipped reference. Ten transitions removed a training component or predefined coupled block, and one disabled a monitoring-only calculation. Each configuration used 10-fold cross-validation on a balanced set of 5,000 training images and a separate 200-image hold-out test set. Fold differences were assessed using the Wilcoxon signed-rank test with Holm–Bonferroni correction and rank-biserial effect sizes. Only removal of pretrained initialization with its coupled differential learning rate and removal of geometric augmentation produced significant sequential accuracy degradation. None of the seven directly interpretable intermediate fold-level removals showed a significant sequential effect. SWA was used only for the final hold-out model and was therefore not directly evaluated by the fold-level Wilcoxon comparison. Anchor comparisons showed collective superiority of the full pipeline but did not isolate individual component effects within the pretrained reference. Non-significant findings indicate no detectable marginal effect, not equivalence. Conclusions are restricted to the dataset, ResNet34 architecture, training budget, random-seed scheme, image-level split, and sequential ablation order studied.