JMIR Publications

ABSTRACT

Background:

Thyroid nodule segmentation in ultrasound imaging is clinically important but typically requires large volumes of expert annotations. Foundation models offer a promising approach to reducing labeled data requirements, yet the relationship between annotation budget and segmentation performance across adaptation strategies remains insufficiently characterized.

Objective:

Thyroid nodule segmentation in ultrasound imaging is clinically important but typically requires large volumes of expert annotations. Foundation models offer a promising approach to reducing labeled data requirements, yet the relationship between annotation budget and segmentation performance across adaptation strategies remains insufficiently characterized.

Methods:

We evaluated four SAM3-based adaptation strategies on TN5000 (5,000 biopsy-confirmed thyroid ultrasound images): full fine-tuning (MedSAM3), zero-shot inference (SAM3-Zero), and Low-Rank Adaptation at ranks 8 and 16 (SAM3-LoRA8, SAM3-LoRA16). All models were initialized via Masked Autoencoder pretraining on the unlabeled corpus and evaluated under four annotation regimes (10%, 25%, 50%, and 100% of labeled data) using Dice, IoU, AUC, sensitivity, specificity, and F1-score, with 95% bootstrap confidence intervals and Bonferroni-corrected significance testing. External validation was conducted on the independent DDTI dataset.

Results:

At full annotation budget, MedSAM3 achieved the highest internal performance (Dice = 0.815, AUC = 0.989), significantly outperforming the zero-shot baseline (Dice = 0.097, p < 0.001 across Wilcoxon, DeLong, and McNemar tests). At 25% label availability, SAM3-LoRA8 matched or exceeded MedSAM3 (Dice = 0.736 vs 0.726) while updating only 1.85% of model parameters, identifying 25% as the annotation-efficiency crossover threshold. No statistically significant difference was observed between LoRA ranks 8 and 16 on per-image Dice (Wilcoxon p = 0.704), indicating that increasing rank beyond 8 yields no measurable performance benefit for this task. External validation on the DDTI dataset preserved model ranking (MedSAM3 AUC = 0.911); a dissociation between AUC and Dice under domain shift was observed across all fine-tuned models, with implications for cross-site deployment.

Conclusions:

Full fine-tuning with domain-adapted initialization provides the best segmentation performance at large annotation budgets. LoRA adaptation with rank 8 offers a more annotation-efficient and parameter-efficient alternative, matching full fine-tuning at 25% label availability and requiring approximately half the training time. These findings provide a concrete decision framework for selecting adaptation strategies in annotation-constrained clinical deployments of thyroid ultrasound segmentation.