3658 - Translate Contrast-Enhanced CT from Non-Contrast CT via Subtraction Image Prediction

Purpose/Objective(s):

Automated generation of contrast-enhanced CT (CECT) from non-contrast CT (NCCT) holds significant clinical potential by avoiding the risks associated with contrast agents. Existing methods rely on direct end-to-end mapping between NCCT and CECT. However, in clinical practice, radiologists often use subtraction imaging, obtained by subtracting NCCT from CECT, to highlight contrast agent-enhanced regions for diagnosis. In this study, we propose a novel approach that directly predicts the subtraction image.

Materials/Methods:

We develop a subtraction image prediction model that accurately generates contrast-enhanced regions while preserving anatomical structures. To achieve this, we employ a UNet-based architecture, where the input is a non-contrast CT image, and the output is the predicted subtraction image. To further improve the structural consistency between the predicted subtraction image and the ground truth subtraction image, we introduce a structure-invariant loss. This loss function emphasizes structural differences rather than pixel-wise intensity mismatches, ensuring that the network learns to predict subtraction images that retain both local contrast variations and anatomical integrity. For training and evaluation, we curated a large-scale abdominal CT dataset comprising 558 patient cases. The dataset is split into 412 cases for training and 146 cases for testing, ensuring robust generalization. Additionally, we evaluated our model on two publicly available abdominal CT datasets, containing 100 cases (75/25) and 92 cases (70/22), respectively. The performance of our model is assessed using PSNR and SSIM, demonstrating its effectiveness in generating high-fidelity subtraction images with enhanced contrast and preserved anatomical structures. Furthermore, our model is trained separately to generate venous-phase or arterial-phase contrast-enhanced CT images, allowing for phase-specific contrast recovery and improved diagnostic utility in different clinical scenarios.

Results:

Experimental results demonstrate that our subtraction-based approach significantly improves image quality. Compared to baseline methods, our approach achieves a 1.3% increase in PSNR and a 3.1% improvement in SSIM on average. The predicted subtraction images effectively highlight contrast-induced changes and preserve anatomical structures.

Conclusion:

By leveraging subtraction image prediction, our method improves anatomical fidelity in contrast enhancement translation while reducing reliance on direct end-to-end mapping. This approach has strong clinical implications, as it better aligns with radiological workflows and can enhance diagnostic accuracy. Future work will explore broader applications, including multi-organ datasets and real-time clinical integration.