Frequency Separation and Aggregation-induced Contrastive Learning for Ultrasound Thyroid Nodule Segmentation

Data

The DDTI, TN3K, and TUCC are publicly available datasets, they differ considerably in acquisition devices, imaging protocols, and patient populations. Therefore, we integrate them under a unified evaluation framework to establish a benchmark protocol for thyroid ultrasound segmentation. This benchmark enables comprehensive assessment of both segmentation accuracy and cross-dataset generalization. We publicly disclose the organization of our data and the list of image file partitions.

Downloading data

• Download the ultrasound datasets DDTI, TN3K, and TUCC.

Dataset Organization

TNS/
├── TrainDataset/       # Training set
│   └── file_list.txt/  # File name list
├── ValidDataset/       # Validation set
│   ├── Imgs/          
│   └── file_list.txt/           # Validation masks
└── TestDataset/        # Test set
    ├── DDTI/           # DDTI test set
    │   ├── Imgs/
    │   └── file_list.txt/
    ├── TN3K/           # TN3K test set
    │   ├── Imgs/
    │   └── file_list.txt/
    └── TUCC/           # TUCC test set
        ├── Imgs/
        └── file_list.txt/

Method and Training

The method and training code have been uploaded, and we will continue to optimize them.

Prediction maps for all models can be found from Google Drive