High throughput image labeling on chest computed tomography by deep learning.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Radiology Informatics

Background:

• Mining large medical image datasets from Picture Archiving and Communication Systems (PACS) or clinical trials requires identifying relevant scans.
• Manual review of individual image series is impractical for big data processing.
• Digital Imaging and Communications in Medicine (DICOM) headers often lack complete or accurate labeling information.

Purpose of the Study:

• To develop an automated, high-throughput image labeling pipeline using deep learning.
• To identify key scan characteristics including scan direction, patient posture, lung coverage, contrast agent usage, and breath-hold status.
• To enable practical big data mining by overcoming limitations of manual review and incomplete DICOM headers.

Main Methods:

• An image-based deep learning pipeline was developed for automated labeling.
• Specific tasks were framed as classification problems, with some requiring segmentation and anatomical landmark identification.
• Models utilized images from different view planes tailored to each classification task.
• A research database from multicenter clinical trials was used for model training and testing.

Main Results:

• The proposed deep learning models achieved high accuracy on the test set across various labeling tasks.
• The pipeline successfully identified scan direction, posture, lung coverage, contrast usage, and breath-hold types.
• The system demonstrated effectiveness on a diverse dataset from multicenter clinical trials.

Conclusions:

• The image-based deep learning pipeline provides an effective solution for high-throughput labeling of medical imaging data.
• Automated labeling significantly enhances the practicality of big data mining in medical research.
• The developed models show promising performance for quality control and data organization in large-scale clinical trial datasets.