Abstract 11285: Synthetic-Contrast-Enhanced Computed Tomography Creating by Machine Learning Without Contrast Media

Abstract

Introduction: Contrast-enhanced computed tomography (CECT) is the gold standard for the diagnosis of aortic dissection (AD), at the same time, CECT has risks of nephropathy and allergic reactions caused by contrast media. Hypothesis: Synthetic-contrast-enhanced computed tomography (SECT) created from non-contrast-enhanced computed tomography (NECT) by machine learning without contrast agents could be an alternative to CECT in the diagnosis of AD. Methods: 50 patients with type A AD and 200 patients without AD who underwent NECT and CECT in pairs at Kobe university hospital were randomly selected and used for model training and testing. 3120 CT slices were used for training and 379 slices were used for testing. AD detection algorithms for NECT or CECT were developed by convolutional neural networks using NECT datasets or CECT datasets respectively. A SECT creating algorithm was developed by u-net using the pairs of CT datasets. To provide visual explanations of diagnosis, the gradient-weighted class activation mapping (Grad-CAM) method was adapted. Results: The provided sensitivities of AD detection algorithms for NECT or SECT were 90 % and 97 %, and the specificities were 100 % and 100 % respectively. When the developed AD detection algorithms for CECT were adapted to SECT, the provided sensitivity of 100 %, and sensitivity of 95 %. Most slices SECT of patients with AD demonstrated intimal flap and/or thrombosed false lumen as the almost same as CECT. The Grad-CAM visualization revealed that the AD detection algorithm using SECT images identified intimal flap or thrombosed false lumen in images more clearly than the AD detection algorithm using NECT images. Conclusions: The SECT creating algorithm improved diagnostic sensitivity for detecting AD using NECT datasets, which suggests the potential of the proposed method to support the clinical practice by reducing complications of contrast agents.