Open-Access Fully Automated Intravenous Contrast Detection and Body Part Classification for Computed Tomography Scans: The FALCON Model.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiology, LMU Munich, Munich, Germany. [email protected].
- Department of Radiology, Division of Thoracic Imaging and Intervention, Massachusetts General Hospital, Boston, MA, USA. [email protected].
- Department of Radiology, Division of Thoracic Imaging and Intervention, Massachusetts General Hospital, Boston, MA, USA.
- Chair for AI in Healthcare and Medicine, Technical University of Munich (TUM), and TUM University Hospital, Munich, Germany.
- Department of Diagnostic and Interventional Radiology, University Medical Center of the Johannes Gutenberg University Mainz, Mainz, Germany.
- Department of Radiology, LMU Munich, Munich, Germany.
- Harvard Medical School, 25 Shattuck Street, Boston, MA, 02115, USA.
Abstract
Presence of intravenous contrast on computed tomography (CT) scans is often unreliably documented, especially in large research datasets. FALCON is an open-access fully automated deep learning model enabling large-scale intravenous contrast detection and body part classification for CT scans of the head and neck (HN), chest, abdomen, and pelvis (AP). This study used six independent datasets consisting of 3138 CT scans of the HN, chest, and AP of 3126 patients from five institutions between 1996 and 2023 to train and validate four CNN models for intravenous contrast detection and body part classification. The ground truth of intravenous contrast presence was verified by a radiologist. We used ResNet9 network architecture and integrated the four models into a graphical user interface. We assessed FALCON's performance with F1 scores and compared FALCON's annotation time to manual annotation by human experts. In the external test set containing 1348 scans, the F1 score for intravenous contrast detection was 99.4% (95%CI: 98.8, 99.9) for HN CT, 98.3% (95%CI: 96.9, 99.5) for chest CT, and 98.1% (95%CI: 96.9, 99.1) for AP CT. The F1 score for body part classification alone on unseen data was 100% for HN, chest, and AP CT. Compared to human experts, annotation of a single scan with FALCON required 1.3 s vs. 21 s for HN CT, 1.8 s vs. 33 s for chest CT, and 3.7 s vs. 1.6 s for AP CT. The open-access FALCON model ( https://github.com/FintelmannLabDevelopmentTeam/Falcon ) quickly and reliably detects intravenous contrast and classifies body part on CT scans.