PatchChestCT: A patch-level spatial annotation dataset for nine abnormalities in chest CT

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Developing generalist artificial intelligence (AI) for radiology requires large-scale 3D imaging datasets with precise spatial annotations.
• Existing chest computed tomography (CT) datasets often lack detailed localization information, hindering AI model training for accurate finding identification.

Purpose of the Study:

• To introduce PatchChestCT, a novel, large-scale, publicly available dataset for multi-abnormality localization in non-contrast chest CT scans.
• To provide 3D patch-level annotations for nine significant abnormalities across 2,201 physician-reviewed CT studies.

Main Methods:

• Utilized the CT-RATE cohort as the source for the dataset.
• Developed a token-aligned annotation scheme for efficient and scalable integration with deep learning architectures like Transformers.
• Annotated one reconstructed 3D volume per study, focusing on patch-level details for nine distinct abnormalities.

Main Results:

• Models trained with PatchChestCT's patch-level labels demonstrated superior localization performance compared to models trained solely on image-level labels.
• Validation across multiple deep learning architectures confirmed the effectiveness of the patch-level annotations for localization tasks.

Conclusions:

• PatchChestCT serves as a valuable public resource for advancing the development of localization-aware AI models in chest CT analysis.
• The dataset facilitates training more accurate AI systems capable of precisely identifying abnormalities in medical imaging.