Incremental diagnostic value of deep Learning-Based synthetic contrast CT: Implications for cancer treatment response assessment - A retrospective study.
Authors
Affiliations (2)
Affiliations (2)
- Machine Intelligence Laboratory, Department of Mechanical Design Engineering, Tech University of Korea, Siheung, Republic of Korea.
- Department of Radiology, College of Medicine, Ewha Womans University, Seoul, Korea.
Abstract
ObjectivePatients with malignancies often require frequent contrast-enhanced CT for treatment monitoring, but renal dysfunction or hypersensitivity reactions frequently restrict the use of iodinated contrast agents. This study aimed to evaluate whether adding deep learning (DL)-generated synthetic contrast images to precontrast CT enhances diagnostic performance in oncology patients in whom iodinated contrast is contraindicated or limited.MethodsA DL model was trained using 200 pairs of precontrast and contrast-enhanced abdominal CT scans. For clinical validation, 214 patients with various malignancies were included. A radiologist (15+ years experience in interpretation of abdominal images) evaluated two sets: precontrast CT only and precontrast CT supplemented with synthetic contrast images. A second radiologist independently re-read the cohort to assess inter-observer agreement. Diagnostic performance, including sensitivity, specificity, and area under the ROC curve (AUC), was compared. Decision curve analysis (DCA) was performed to assess clinical utility.ResultsThe addition of synthetic contrast images significantly increased sensitivity from 45.2% to 67.7% (p = 0.0391). The AUC also significantly improved from 0.718 to 0.817 (p = 0.0282). Specificity remained high at 95.6%. Qualitative assessment showed 79.4% of cases were of diagnostically acceptable quality (score 3 or 4). Inter-observer agreement was substantial to almost perfect and improved with the addition of synthetic contrast (Cohen's κ, 0.69 to 0.84). The incremental diagnostic benefit was reproduced by an independent second reader. DCA demonstrated greater net benefit for the supplemented set across all relevant threshold probabilities.ConclusionsAdding DL-based synthetic contrast images to precontrast CT was associated with improved sensitivity and diagnostic discrimination in this single-center retrospective cohort. This approach may have potential as an adjunct for assessing treatment response in oncology patients for whom iodinated contrast administration is limited or contraindicated, although external validation is required.