NIH CT Scans with Lesion Bounding Boxes and Measurements

32,120 axial computed tomography slices from 10,594 CT scans of 4,427 patients. The dataset includes 1-3 lesions per image with bounding boxes and size measurements, totaling 32,735 annotated lesions. It was created by farrell236 from NIH PACS data and last updated on Hugging Face in June 2024.

Use Cases

• Train object detection models to locate lesions using the provided bounding box annotations.
• Develop regression models to predict lesion size from CT image features and the provided measurement data.
• Benchmark multi-lesion segmentation algorithms on images containing 1-3 lesions per slice.
• Study lesion characteristics and distributions across a large cohort of 4,427 patients.

Strengths

• Contains 32,120 annotated CT image slices.
• Includes 32,735 lesion instances with bounding boxes and size measurements.
• Derived from a substantial patient cohort of 4,427 individuals.

Limitations

• Limited to axial CT slices; other imaging planes or modalities are not included.
• Lesion annotations were mined from clinical PACS, which may introduce label noise compared to expert review.
• The dataset focuses on lesions present in CT scans, which may not represent the full spectrum of radiological findings.

Provenance

Source

National Institutes of Health (NIH) picture archiving and communication system (PACS).

Collection Method

Lesion annotations were mined from clinical radiology reports in the PACS.

Freshness

Last updated on the platform in June 2024.

Geography

United States (implied by NIH source).