classman/sam-hq-lora-dental-xray

SAM-HQ + LoRA for Individual Tooth Segmentation

Overview

This repository provides LoRA adapter weights for adapting Segment Anything Model, specifically its high-quality variant SAM-HQ, to individual tooth segmentation on dental panoramic X-ray images.

The goal of this work is to improve segmentation performance under significant domain shift (medical X-ray vs. natural images) using parameter-efficient fine-tuning, without modifying the full backbone.

This repository focuses on model artifacts only. Training code and datasets are not released.

Motivation

While SAM / SAM-HQ shows strong zero-shot performance on natural images, its performance degrades on dental panoramic X-rays due to:

• —subtle and ambiguous object boundaries,
• —high visual similarity between adjacent teeth,
• —fine-grained anatomical structures.

Full fine-tuning of SAM-HQ is impractical due to model size. LoRA enables efficient domain adaptation by training a small number of additional parameters while keeping the backbone frozen.

What This Repository Provides

• —LoRA adapter weights only
• —Four LoRA variants corresponding to different adaptation scopes
• —No training code or dataset

LoRA Adapter Variants

This repository contains four independent LoRA adapters, each trained with a different application scope:

• —Encoder + Decoder LoRA (best-performing)
• —Decoder-only LoRA
• —HQ-module-only LoRA
• —Additional ablation variant

Each adapter can be loaded independently on top of the same SAM-HQ base model, allowing controlled comparison of adaptation strategies.

Training Strategy (Summary)

• —Task formulation: individual tooth-level segmentation (each tooth treated as an independent sample)
• —Prompting strategy:
• —iterative prompt refinement following SAM (11 iterations)
• —error-driven prompt sampling from false positive / false negative regions
• —Early stopping:
• —Dice ≥ 0.85 or IoU ≥ 0.85

Implementation details of the training system are documented externally.

Dataset

• —Training set: 2,392 images
• —Test set: 269 images
• —Bad-case set: 10 images

The dataset is not publicly released due to privacy and licensing constraints.

Results

• —SAM-HQ baseline Dice: 0.748
• —SAM-HQ + LoRA Dice: 0.914

The best performance is achieved with the Encoder + Decoder LoRA variant. Iterative prompting significantly improves boundary precision in challenging cases.

Repository Scope & Reproducibility

• —This repository provides model artifacts only
• —Training code, configurations, and datasets are not released
• —Detailed system design and experimental analysis are available at:

👉 https://choihyun-1110.github.io/projects/sam-finetuning.html

Intended Use & Limitations

• —Intended for research and experimental use only
• —Not validated for clinical diagnosis
• —Performance may not generalize outside dental panoramic X-ray images

License

• —Base model license follows the original SAM / SAM-HQ license
• —This repository is intended for non-commercial research use

Authors

• —Hyun Choi — Joint First Author
• —Y. Uhm — Joint First Author

Citation

@misc{choi2025sam_hq_lora_tooth,
  author = {Choi, Hyun and Uhm, Y.},
  title = {SAM-HQ with LoRA for Individual Tooth Segmentation},
  year = {2025},
  url = {https://huggingface.co/choihyun-1110/sam-hq-lora-tooth}
}