ThyroidDataSet: A Collection of Thyroid Ultrasound Datasets

This repository hosts a curated collection of thyroid ultrasound datasets for research on segmentation, classification, detection, and report generation.
All datasets are provided as ZIP archives under the data/ directory in this Hugging Face dataset:

https://huggingface.co/datasets/Doul0414/ThyroidDataSet

The collection currently includes the following datasets:

• TGVideo
• DDTI
• TN3K
• 24-TMI
• TN5K
• ThyUS2Path
• Cine-clip
• AHU
• ThyroidXL
• PKTN
• FinalData (an additional in-house dataset, 1,854 images)
• TN5K_MASK_SAM2_L (new, in-house processed masks for TN5K)
• VIDEO_ImageAndMask (new, in-house processed video frames and masks)
• VIDEO_DATASET (new, in-house processed video, Including ultrasound imaging video, masked video, and visualized masked ultrasound imaging video)

Please always refer to the original dataset papers and licenses when using these data. This repository only provides a consolidated entry point for research and comparison.

---

1. Dataset Summary

1.1 Public/open datasets

Dataset	Images / Cases	Train	Test	File format	Task(s)	Location	Ultrasonic imaging device(s)
TGVideo	15,186 images (16 cases)	15,186	N/A	Image: DICOM; Mask: DICOM	Segmentation	Germany	GE Logiq E9
DDTI	637 images	N/A	637	Image: PNG; Mask: PNG; Label: CSV	Segmentation, Classification	Colombia	TOSHIBA Nemio 30, TOSHIBA Nemio MX
TN3K	3,493 images	2,879	614	Image: JPG; Mask: JPG; Label: CSV	Segmentation, Classification	Guangzhou, China	GE Logiq E9, ARIETTA 850, RESONA 70B
24-TMI	2,460 cases (4,920 images)	3,934	986	JPEG	Report generation	Beijing, China	N/A
TN5K	5,000 images	4,000	1,000	Image: JPG; Label: XML	Detection, Classification	Beijing, China	GE Logiq E9, GE S7
ThyUS2Path	8,508 images	5,457	3,051	Image: JPG; Label: CSV	Classification	Zhejiang, China	Esaote MyLab (Portable)
Cine-clip	17,412 frames (192 cases, avg 90 frames per case)	N/A	N/A	Image: HDF5; Label: CSV	Segmentation, Classification	California, USA	N/A
AHU	1,833 cases (125,896 images)	N/A	N/A	Image: JPG; Label: folder-level	Classification	China (web scraping)	Heterogeneous
ThyroidXL	11,635 images	9,541	2,094	Image: PNG; Mask: PNG; Label: TXT	Segmentation, Classification, Detection	Vietnam	Hitachi Aloka Arietta V70
PKTN	1,005 images	N/A	N/A	Image: JPG; Mask: JPG	Segmentation	Beijing, China	N/A

The ZIP files in data/ follow the original dataset names as closely as possible (e.g. TGVideo.zip, TN3K.zip, ThyroidXL.zip, PKTN.zip, Cine-clip.zip, 24TMI.zip, AHU-ProcessedData.zip, etc.).

1.2 FinalData (in-house dataset)

• Archive name: finall data.zip
• Images: 1,854 ultrasound images
• Tasks: Segmentation, Classification
• Location: China
• Ultrasonic imaging devices: Unknown / heterogeneous

---

1.3 New in-house processed datasets

1.3.1 TN5K_MASK_SAM2_L

• Archive name: TN5K_MASK_SAM2_L.zip
• Source dataset: TN5K
• Content: Automatically generated lesion segmentation masks for the TN5K images, produced using the SAM2-based medical segmentation pipeline.

---

1.3.2 VIDEO_DATASET AND VIDEO_ImageAndMask

• Archive name: VIDEO_DATASET.zip、VIDEO_ImageAndMask.zip
• Source data: In-house thyroid ultrasound video.
• VIDEO_DATASET Content: Including ultrasound imaging video, masked video, and visualized masked ultrasound imaging video
• VIDEO_ImageAndMask Content: Source video frames and masks.

---

2. Repository Structure

On Hugging Face, the dataset files are organized as:

ThyroidDataSet
└── data/
    ├── 24TMI.zip
    ├── AHU-ProcessedData.zip
    ├── Cine-clip.zip
    ├── DDTI dataset.zip
    ├── PKTN.zip
    ├── TGVideo.zip
    ├── ThyUS2Path.zip
    ├── ThyroidXL.zip
    ├── TN3K.zip
    ├── TN5000.zip
    ├── tg3k(TGVideoSubset).zip
    ├── finall data.zip              # FinalData (in-house)
    ├── TN5K_MASK_SAM2_L.zip         # TN5K SAM2-based lesion masks
    ├── VIDEO_DATASET.zip            # Internal video data
    └── VIDEO_ImageAndMask.zip       # Internal raw video frames and mask images