Above Chance, Below the Baseline: A Majority-Anchored, Case-Grouped Ablation for Thyroid Ultrasound Segmentation and Report-Derived Suspicion Classification

Thyroid ultrasound computer-aided diagnosis is routinely reported as a segmentation Dice score and a classification F1 score, with little scrutiny of what drives either number on small, imbalanced, multi-image-per-patient benchmarks. We supply that scrutiny through two matched-protocol ablations on TN3K (segmentation) and DDTI (classification) under one Holm–Bonferroni-corrected statistical protocol. On TN3K, three ImageNet-pretrained encoder–decoder architectures (U-Net, U-Net++ and FPN, ResNet-34 encoder) exceed a from-scratch U-Net by 0.0670–0.0758 Dice under 3-fold cross-validation (adjusted p ≤ 0.0054), while the pretrained architectures separate only partially from one another: U-Net++ exceeds FPN (adjusted p = 0.024) but not U-Net (adjusted p = 0.103). A multi-axis ablation confirms pretraining (−0.0858 Dice) and augmentation (−0.0228 Dice) matter (adjusted p = 0.039 each); loss form and resolution do not reach significance. On DDTI, a deterministic majority-class predictor is a mandatory row in every table: all four backbones (ResNet50, DenseNet121, EfficientNet-B0, ViT-B/16) score below it on F1 (best 0.8896 vs. 0.9035) and accuracy (best 0.8162 vs. 0.8246), yet all exceed it on Matthews correlation coefficient (0.1541–0.3072 vs. 0.0000) and balanced accuracy (0.5660–0.6506 vs. 0.5000); no pairwise difference survives correction (best raw p = 0.040, adjusted p = 0.240). Replacing a naive image-level split with a case-grouped one moves F1 by −0.1439 but Matthews correlation coefficient by only −0.0143 and AUC by +0.0029. Re-running two identical configurations gives repeat-run spreads of 0.0494 Matthews correlation coefficient, 0.2217 F1 and 0.0420 Dice, larger than most measured effects. We further disclose a pixel-level ROC-AUC of 0.9804 against a Dice of 0.7568, and a Youden threshold saturating at 1.0000 with expected calibration error 0.1324.