ThyUS2Path: Thyroid Nodule Diagnosis from Ultrasound Images

A dataset from the Second Affiliated Hospital of Jiaxing University containing 7,992 thyroid ultrasound images from 603 patients, used to train a weakly supervised deep learning model for pathological diagnosis. The model, ThyUS2Path, was developed by Xiao-Wen Hou and published on figshare in 2026. It aims to link ultrasound phenotypes with histological reports non-invasively.

Use Cases

• Training weakly supervised deep learning models for medical image classification based on image-level annotations.
• Benchmarking attention-based neural network architectures for ultrasound analysis.
• Developing non-invasive diagnostic tools for thyroid nodule malignancy prediction.
• Studying the correlation between ultrasonography phenotypes and histopathological reports.

Strengths

• Contains 7,992 annotated ultrasound images (6,014 for training/validation, 1,978 for external testing).
• Model performance is reported with specific metrics (AUCs of 0.754 ± 0.035 and 0.735 ± 0.029).
• Includes an external test set for validation, with reported AUROCs of 0.70~0.80.

Limitations

• Row count and column-level documentation are absent; field semantics must be inferred after download.
• Data may reflect geographic bias inherent to being sourced from a single hospital in China.
• The dataset is very small in file size (15.7 KB), suggesting limited scope or that the primary data may be hosted elsewhere.

Provenance

Source

Second Affiliated Hospital of Jiaxing University

Collection Method

Retrospective analysis of patient ultrasound images followed by histopathological examination.

Freshness

Last updated 2026-06-01 04:15:08; freshness should be verified.

Geography

China