Deep Learning-Based Metal Artifact Reduction in Cardiac Computed Tomography: A Preliminary Study Enabling Radiomic Analysis in Patients with Implantable Defibrillators.
Authors
Affiliations (8)
Affiliations (8)
- CardioTechLab, Centro Cardiologico Monzino IRCCS, Milan, Italy.
- Department of Electronics, Information and Bioengineering, Politecnico Di Milano Milan, Milan, Italy.
- Department of Cardiology, University Medical Center Utrecht, Utrecht, The Netherlands.
- Department of Biomedical, Surgical and Dental Sciences, University of Milan, Milan, Italy.
- Perioperative Cardiology and Cardiovascular Imaging Department, Centro Cardiologico Monzino IRCCS, Milan, Italy.
- Department of Cardiology, Cardiovascular Research Institute Maastricht (CARIM), Maastricht University Medical Center+, Maastricht, The Netherlands.
- CardioTechLab, Centro Cardiologico Monzino IRCCS, Milan, Italy. [email protected].
- Department of Electronics, Information and Bioengineering, Politecnico Di Milano Milan, Milan, Italy. [email protected].
Abstract
Idiopathic ventricular fibrillation (IVF) affects 5-10% of out-of-hospital cardiac arrest survivors, requiring implantable cardioverter-defibrillators (ICDs) for management. However, metal artifacts from ICDs may compromise cardiac computed tomography (CCT) image quality, limiting diagnostic capabilities and preventing advanced analyses like radiomics. The aim of this study is to develop a deep learning (DL)-based solution to reduce metal-induced artifacts by effectively removing both the metallic implant (primary artifact) and the associated streak artifacts (secondary artifact) in CCT scans of IVF patients with ICDs and to evaluate its effectiveness through radiomic analysis for predicting a composite clinical endpoint. CCT scans from 41 IVF patients with ICDs and 13 unaffected family members without ICDs were included. A fully convolutional neural network with 20 layers was trained using a simulated dataset combining artifact-free images from controls with real artifact masks extracted from patient scans. The network was evaluated using structural similarity index (LV-SSIM), mean absolute error (LV-MAE), and mean squared error (LV-MSE), computed on the segmented region of interest (ROI), i.e., left ventricle wall. Further, 851 radiomic features were extracted from the ROI, and machine learning models were developed to predict a composite endpoint including nonsustained ventricular tachycardia, high premature ventricular contraction burden, and ventricular arrhythmias recurrences. The DL model achieved good performance with an LV-SSIM of 0.936 ± 0.045, LV-MAE of 15.59 ± 8.48 HU, and LV-MSE of 661.89 ± 1003.80 HU on the test set. Radiomic feature analysis demonstrated that 98% of stable features were preserved after artifact removal, confirming diagnostic integrity. The predictive model achieved an F1 score of 0.85 for the composite clinical endpoint, demonstrating effective results for metal artifact reduction in CCT of IVF patients with ICDs. A promising DL solution for metal artifact reduction in CCT imaging of patients with ICD has been proposed. The demonstrated preservation of radiomic features suggests that the approach maintains diagnostic integrity, potentially expanding the utility of CCT in patients with cardiac devices who previously might have been excluded from certain studies.