Thyroid nodule segment SAM model
This repository provides the implementation of MedSAM, a medical image segmentation model, applied to the TN3K public dataset for thyroid nodule segmentation.
We fine-tuned MedSAM for precise segmentation of thyroid nodules in ultrasound images. Both quantitative and qualitative evaluations show that MedSAM achieves high performance across standard segmentation metrics.
| Metric | Value |
|---|---|
| Dice (DSC) | 0.8861 |
| IoU | 0.8075 |
| Precision | 0.9440 |
| Recall | 0.8505 |
| Specificity | 0.9926 |
| Accuracy | 0.9805 |
The following figure shows visual examples of thyroid nodule segmentation results on the TN3K test set.
Left: Input Ultrasound Image | Middle: Ground Truth Mask | Right: MedSAM Prediction
To view more examples, see the tn3k_infer_results/ directory.
We used the TN3K dataset for thyroid ultrasound segmentation tasks.
If you need access to the original dataset, please contact me at 424umar@gmail.com.
- Clone this repo
- Prepare the TN3K dataset in the expected format
- Run the training script using our MedSAM configuration
- Visual and numerical results will be saved under the
results/directory
For questions or collaborations, please get in touch with 424umar@gmail.com & umarfarooq@hanyang.ac.kr or open an issue in this repository.