Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/dataset.json

Name

Thyroid US cine-clip dataset for risk stratification (Stanford, 2017–2018)

Link

https://stanfordaimi.azurewebsites.net/datasets/a72f2b02-7b53-4c5d-963c-d7253220bfd5

Indexing

Keywords: Ultrasound, Thyroid, Thyroid nodule, Cine clips, Risk stratification, Deep learning, ACR TI-RADS

Content: US, HN

RadLex: RID10326, RID46068, RID35976, RID28950, RID49885, RID49884, RID7578

Author(s)

Rikiya Yamashita

Tara Kapoor

Minhaj Nur Alam

Alfiia Galimzianova

Saad Ali Syed

Mete Ugur Akdogan

Emel Alkim

Andrew Louis Wentland

Nikhil Madhuripan

Daniel Goff

Victoria Barbee

Natasha Diba Sheybani

Hersh Sagreiya

Daniel L. Rubin

Terry S. Desser

Organization(s)

Stanford University School of Medicine

Contact

Corresponding author email provided in article: ude.drofnats@ressedst

Funding

Supported in part by Stanford University resources; NIH awards (multiple, as listed), National Science Foundation grant; additional support as detailed in the article acknowledgments.

Ethical review

Institutional review board–approved, HIPAA-compliant retrospective study with waiver of written informed consent.

Comments

Dataset accompanying the study developing a deep learning–based risk stratification system for thyroid nodules using ultrasound cine images with biopsy-proven ground truth.

Date

Published: 2022-05-11

References

[1] Yamashita R, Kapoor T, Alam MN, et al.. "Toward Reduction in False-Positive Thyroid Nodule Biopsies with a Deep Learning–based Risk Stratification System Using US Cine-Clip Images". Radiology: Artificial Intelligence. 2022-05-11. doi:10.1148/ryai.210174. PMID: 35652118. PMCID: PMC9152684.

Dataset

Motivation

To enable development and validation of a deep learning–based risk stratification system for thyroid nodules using ultrasound cine images and to revise ACR TI-RADS recommendations.

Sampling

Retrospective cohort of biopsy-confirmed thyroid nodules from April 2017–May 2018 using ACR TI-RADS–structured reports at a single institution.

Partitioning scheme

Fivefold cross-validation with patient-level splitting; in each fold: training (3/5), validation (1/5), test (1/5); original class distribution maintained; nodules from the same patient placed in the same fold.

Missing information

No explicit dataset license, file formats, or image resolutions reported; no external validation cohort.

Relationships between instances

Multiple nodules may belong to the same patient; each nodule has a cine clip consisting of sequential frames; manual ROI segmentations provided for nodules.

Noise

Strong class imbalance (benign:malignant = 175:17).

External data

No external datasets were used for model evaluation in this study.

Confidentiality

HIPAA-compliant retrospective clinical imaging dataset with biopsy-confirmed labels.

Sensitive data

Clinical imaging and pathology-linked health data.