Purpose: This dissertation aims to evaluate the clinical utility of deep learning-based synthetic contrast-enhanced computed tomography (DL-SynCCT) in abdominal imaging, particularly in scenarios where the use of iodinated contrast media is contraindi...
Purpose: This dissertation aims to evaluate the clinical utility of deep learning-based synthetic contrast-enhanced computed tomography (DL-SynCCT) in abdominal imaging, particularly in scenarios where the use of iodinated contrast media is contraindicated. The study focuses on two distinct patient populations: those undergoing non-enhanced CT (NECT) due to medical limitations and patients in oncologic or post-surgical settings who require frequent follow-up imaging. The primary objective is to assess whether DL-SynCCT enhances lesion detectability, diagnostic confidence, and overall image quality in these high-risk groups.
Materials and Methods: A weakly supervised deep learning algorithm was developed using a two-stage training process involving virtual non-contrast (VNC) and true contrast-enhanced CT datasets. The algorithm was validated on two independent cohorts: 398 general NECT patients and 300 patients in specialized clinical settings (241 with malignancies and 59 post-surgical or interventional cases). Three board-certified radiologists conducted blinded, paired readings (NECT-only vs. NECT with DL-SynCCT) for each case, separated by a two-week washout period. Sensitivity, specificity, diagnostic confidence, and image quality were quantitatively assessed, with statistical analysis performed using McNemar’s test, Wilcoxon signed rank test, and Generalized Estimating Equations (GEE).
Results: DL-SynCCT consistently improved diagnostic performance across both patient populations. In general NECT cases, lesion sensitivity increased from 72.0% to 76.4% (p < 0.001), while diagnostic confidence rose significantly without compromising specificity. In the oncology cohort, pooled sensitivity improved from 48.5% to 55.2% (p < 0.001), particularly in gastrointestinal malignancies, where sensitivity increased from 34.3% to 48.1%. Among post-surgical patients, DL-SynCCT enhanced sensitivity for detecting complications from 68.4% to 74.4% (p = 0.01) and specificity from 83.3% to 93.3% (p = 0.03). Radiologists reported improved confidence and fewer missed findings in cases with subtle or ambiguous findings.
Conclusions: DL-SynCCT offers a safe and effective alternative to traditional contrast-enhanced CT, particularly in patients for whom contrast administration is not feasible. Its implementation results in higher lesion detection sensitivity and diagnostic confidence, supporting its potential role in routine clinical practice. The algorithm's robust technical performance and ease of integration into existing workflows further support its clinical viability. These findings highlight DL-SynCCT as a promising tool for improving the diagnostic value of NECT, with implications for oncologic surveillance, postoperative care, and broader radiologic applications.