A deep learning approach to generate contrast-enhanced computerised tomography angiograms without the use of intravenous contrast agents

Abstract Introduction Contrast-enhanced computerised tomographic (CT) angiograms are widely used in cardiovascular imaging to obtain a non-invasive view of arterial structures. In aortic aneurysmal disease (AAA), CT angiograms are required prior to surgical intervention to differentiate between blood and the intra-luminal thrombus, which is present in 95% of cases. However, contrast agents are associated with complications at the injection site as well as renal toxicity leading to contrast-induced nephropathy (CIN) and renal failure. Purpose We hypothesised that the raw data acquired from a non-contrast CT contains sufficient information to differentiate blood and other soft tissue components. Therefore, we utilised deep learning methods to define the subtleties between the various components of soft tissue in order to simulate contrast enhanced CT images without the need of contrast agents. Methods Twenty-six AAA patients with paired non-contrast and contrast-enhanced CT images were randomly selected from an ethically approved ongoing study (Ethics Ref 13/SC/0250) and used for model training and evaluation (13/13). Non-contrast axial slices within the aneurysmal region from 10 patients (n=100) were sampled for the underlying Hounsfield unit (HU) distribution at the lumen, intra-luminal thrombus and interface locations, identified from their paired contrast axial slices. Subsequently, paired axial slices within the training cohort were augmented in a ratio of 10:1 to produce a total of 23,551 2-D images. We trained a 2-D Cycle Generative Adversarial Network (cycleGAN) for this non-contrast to contrast transformation task. Model output was assessed by comparison to the contrast image, which serves as a gold standard, using image similarity metrics (ex. SSIM Index). Results Sampling HUs within the non-contrast CT scan across multiple axial slices (Figure 1A) revealed significant differences between the blood flow lumen (yellow), blood/thrombus interface (red), and thrombus (blue) regions (p<0.001 for all comparisons). This highlighted the intrinsic differences between the regions and established the foundation for subsequent deep learning methods. The Non-Contrast-to-Contrast (NC2C)-cycleGAN was trained with a learning rate of 0.0002 for 200 epochs on 256 x 256 images centred around the aorta. Figure 1B depicts “contrast-enhanced” images generated from non-contrast CT images across the aortic length from the testing cohort. This preliminary model is able to differentiate between the lumen and intra-luminal thrombus of aneurysmal sections with reasonable resemblance to the ground truth. Conclusion This study describes, for the first time, the ability to differentiate between visually incoherent soft tissue regions in non-contrast CT images using deep learning methods. Ultimately, refinement of this methodology may negate the use of intravenous contrast and prevent related complications.