By retraining a segmentation network using scans of patients with defibrillators, researchers have improved the delineation of 20 substructures of the heart despite metallic artefacts.
Stereotactic radiotherapy is increasingly being used to treat ventricular tachycardia, a serious arrhythmia. However, these patients often wear an implantable cardioverter defibrillator, whose metal blurs CT images and degrades the automatic delineation of cardiac structures, which is necessary to precisely dose the rays.
Images with implants in training
The authors took an already validated deep learning model and retrained it on a carefully annotated cohort, supplemented with images of patients with implants, in order to predict 20 substructures of the heart. Tested on 10 patients treated for ventricular tachycardia, it achieved an average Dice coefficient of 0.71, significantly better than the trained version without implant images. In cancer patients with defibrillators, the score was 0.70; in those without implants, performance remained virtually unchanged (0.74 versus 0.75).
Dosimetric analysis shows that, during these radioablations, the ventricles, the tricuspid valve, the atrioventricular node and the left coronary arteries receive a higher relative dose than the other substructures, depending on the location of the target.
Read the full article on the original source
Go to the source →