Hugging Face's logo

Datasets:
MedXAgent
/
OpenThyroidDB

Tasks:
Image Segmentation
Image Classification
Object Detection
Modalities:
Text
Formats:
json
Languages:
English
Chinese
Size:
< 1K
Tags:
medical-imaging
ultrasound
thyroid
multicenter
report-generation
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

OpenThyroidDB

A multicenter, multitask thyroid ultrasound data resource supporting the ThyroidXAgent evidence workflow.

OpenThyroidDB integrates curated public thyroid ultrasound resources and institutionally governed clinical cohorts into a full-spectrum resource for segmentation, benign–malignant classification, report generation, and malignant-lesion stratification across diverse scanners and acquisition settings.

This Hugging Face repository (MedXAgent/OpenThyroidDB) is the data-distribution layer of OpenThyroidDB. It separates redistributed or processed external resources from MedXAgent-governed cohorts, while preserving dataset-level provenance and archive boundaries.

Scope of the resource

The task organization follows Figure 1 of the ThyroidXAgent study:

Task Evidence represented in the current repository Primary use
Segmentation CineClip, TGVideo, ThyroidXL, TN3K, PKTN, DDTI, TN5000 masks, TNVideo Pixel-/region-level delineation of thyroid structures and lesions
Classification AHU, ThyroidXL, ThyUS2Path, TN5000/TN5K, TN3K, DDTI, TNVideo Thyroid nodule diagnosis and categorization
Report generation KMVE, SMU-HMC, ZJH-TS, TNVideo Standardized report generation from ultrasound images or videos
Malignant-lesion stratification Supporting thyroid image resources and task-specific ThyroidXAgent cohorts FTC/PTC prediction and lymph-node-metastasis subtype classification

Figure 1 reports task-specific curated subsets rather than raw archive totals. Examples include:

  • Public segmentation resources: CineClip (17,417), TGVideo (15,186), ThyroidXL (11,631), TN3K (5,347), PKTN (1,003), and DDTI (637 images).
  • Public classification resources: AHU (125,896), ThyroidXL (11,631), ThyUS2Path (8,508), TN5K (5,000), TN3K (5,347), and DDTI (349 images).
  • Report-generation resources: KMVE (2,474 reports / 15,921 images), SMU-HMC (23,955 reports / 248,194 images), ZJH-TS (353 reports / 4,471 images), and a TNVideo report-generation subset (145 case videos).
  • TNVideo task subsets: 148 cases for segmentation/classification and 145 case videos for report generation in the paper workflow.

Scanner and acquisition diversity

OpenThyroidDB spans still images, cine clips, videos, masks, bounding boxes, categorical labels, and free-text reports collected across heterogeneous acquisition environments. Scanner families represented in the study include, among others:

  • GE LOGIQ E9 and GE S7
  • Fujifilm ARIETTA 850 and Hitachi Aloka ARIETTA V70
  • Mindray RESONA 70B
  • Toshiba Nemio 30 / Nemio MX
  • Esaote MyLab
  • Samsung Medison RS80
  • SuperSonic Imagine Aixplorer Ultimate

Scanner availability and metadata completeness vary by source dataset.

Current repository structure

The layout below reflects the current Hugging Face release:

OpenThyroidDB/
├── ExternalDatasets/
│   ├── AHU/
│   │   └── AHU-ProcessedData.zip
│   ├── CineClip/
│   │   └── Cine-clip.zip
│   ├── DDTI/
│   │   ├── DDTI_SourceData.zip
│   │   └── DDTI_processed.zip
│   ├── KMVE/
│   │   └── KMVE.zip
│   ├── PKTN/
│   │   └── PKTN.zip
│   ├── TG3K/
│   │   └── tg3k.zip
│   ├── TGVideo/
│   │   └── TGVideo.zip
│   ├── TN3K/
│   │   └── TN3K.zip
│   ├── TN5000/
│   │   ├── TN5000.zip
│   │   └── TN5K_MASK_SAM2_L.zip
│   ├── ThyUS2Path/
│   │   └── ThyUS2Path.zip
│   ├── ThyroidXL/
│   │   └── ThyroidXL.zip
│   └── ZJH-2K/
│       └── ZJH-2K.zip
└── MedXAgentDatasets/
    ├── SMU-HMC/
    │   ├── Batch1.zip ... Batch15.zip
    │   ├── Batch_Multiple_Cases.zip
    │   └── test_set_single.json
    ├── TNVideo/
    │   ├── SourceVideoAndAnno.zip
    │   ├── VIDEO_DATASET.zip
    │   └── VIDEO_ImageAndMask.zip
    └── ZJH-TS/
        ├── ZJH-TS.zip
        └── test_150cases_manifest.csv

Released components

ExternalDatasets

ExternalDatasets/ contains source-aligned or processed archives derived from publicly described datasets. Each subdirectory is kept separate to make provenance, licensing, and citation easier to audit.

Directory Current archive(s) Main task relevance
AHU/ AHU-ProcessedData.zip Classification
CineClip/ Cine-clip.zip Segmentation, classification
DDTI/ DDTI_SourceData.zip, DDTI_processed.zip Segmentation, classification
KMVE/ KMVE.zip Report generation
PKTN/ PKTN.zip Segmentation
TG3K/ tg3k.zip Thyroid ultrasound image analysis
TGVideo/ TGVideo.zip Video/frame segmentation
TN3K/ TN3K.zip Segmentation, classification
TN5000/ TN5000.zip, TN5K_MASK_SAM2_L.zip Detection/classification and derived SAM2 masks
ThyUS2Path/ ThyUS2Path.zip Classification
ThyroidXL/ ThyroidXL.zip Segmentation, classification, detection
ZJH-2K/ ZJH-2K.zip Thyroid ultrasound image analysis

Users must consult and cite the original publication and license for every external dataset they use. Inclusion in this repository does not relicense an external resource.

MedXAgentDatasets

MedXAgentDatasets/ contains institutionally governed or MedXAgent-curated releases used by the ThyroidXAgent workflow.

SMU-HMC

  • Multicenter image–report cohort for thyroid ultrasound report generation.
  • The paper cohort contains 23,955 reports and 248,194 images.
  • The current release is distributed as 15 batch archives, an additional multiple-case archive, and test_set_single.json.

ZJH-TS

  • De-identified thyroid ultrasound image–report cohort.
  • Full archive: 353 reports and 4,471 images.
  • test_150cases_manifest.csv provides the final selected evaluation subset of 150 reports and 2,034 images.
  • Manifest columns are index, report_id, data_dir, n_images, findings, impression, and time.

TNVideo

  • Thyroid ultrasound video resource supporting segmentation, classification, and video-conditioned report generation.
  • SourceVideoAndAnno.zip: source videos and sparse frame-level annotations.
  • VIDEO_ImageAndMask.zip: extracted images and masks.
  • VIDEO_DATASET.zip: processed video-oriented release.
  • Frame annotations in the source archive are sparse Pascal VOC bounding boxes rather than dense per-frame segmentation masks.

Access, governance, and responsible use

By requesting or using this repository, users agree to:

  1. Use the data only for legitimate research, education, and method development permitted by the applicable source terms.
  2. Respect the original licenses, citations, and redistribution restrictions of all external datasets.
  3. Follow the approved access conditions for institutionally governed cohorts.
  4. Not attempt to identify or contact any individual represented in the data.
  5. Not use the repository as a substitute for clinical judgment or as a clinically validated diagnostic product.
  6. Apply appropriate security, ethics, and institutional review procedures when handling downloaded data.

Download

Authenticate and request access on the dataset page before downloading gated files:

hf auth login

Download a specific component:

hf download MedXAgent/OpenThyroidDB \
  "MedXAgentDatasets/ZJH-TS/ZJH-TS.zip" \
  --repo-type dataset

Download a directory with Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="MedXAgent/OpenThyroidDB",
    repo_type="dataset",
    allow_patterns=["MedXAgentDatasets/ZJH-TS/*"],
    local_dir="OpenThyroidDB",
)
Downloads last month
-
Total file size:
62.3 GB