Using Machine Learning to Identify Intravenous Contrast Phases on Computed Tomography.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate identification of intravenous (IV) contrast phase in Computed Tomography (CT) is crucial for diagnosis.
• Manual assessment can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To apply machine learning (ML) techniques for automatic identification of IV contrast presence and physiologic phase in CT scans.
• To evaluate the performance of ML models in classifying contrast phases and predicting timing.

Main Methods:

• Utilized a large dataset of 82,690 chest and abdomen CT examinations from 17 institutions.
• Employed semi-supervised learning with weak labels derived from DICOM metadata.
• Developed a 12-layer CNN and ResNet18 for classification and regression tasks, incorporating transfer learning for CT Chest analysis.

Main Results:

• Achieved 93.3% test accuracy in predicting contrast phase using ML models trained on DICOM metadata.
• Regression analysis successfully delineated arterial and nephrogenic phases.
• Transfer learning on CT Chest yielded an AUROC of 0.999 with fine-tuning.

Conclusions:

• Machine learning algorithms can accurately and automatically ascertain contrast presence and phase in CT scans.
• Transfer learning enables high-precision contrast phase prediction on CT Chest with minimal labeled data.