Do AutoReplace and AutoMove data augmentation methods improve U-Net model segmentation performance of ablation catheter tips in transesophageal echocardiography compared to conventional methods?
Novel data augmentation methods (AutoReplace and AutoMove) improve deep learning-based segmentation of ablation catheters in transesophageal echocardiography, potentially enhancing the safety and dependability of cardiac interventions.
Interventional cardiology is a minimally invasive surgery that focuses on treating cardiovascular disorders with catheters, such as coronary artery disease, strokes, peripheral arterial disease, and aortic disease. Ultrasound imaging, also known as echocardiography, is a common imaging technology used to keep track of catheter punctures. Precisely segmenting a medical device during cardiac interventions can increase the procedure's safety and dependability. Deep learning has been widely applied to medical image segmentation tasks. It needs an abundant and representative training dataset to perform well. However, in many cases, the process of collecting and labelling data is costly and time-consuming. In this article, we propose two data augmentation methods, AutoReplace and AutoMove, to solve the problem of data efficiency. We used these methods to train a U-Net model to segment the tiny ablation catheter tip in transesophageal echocardiography into the standard five-chamber views and compared the results with conventional data augmentation methods. The comparison showed that our proposed methods have a better effect on the segmentation performance in this task. The principles of these two methods are simple, and they can be easily applied to other medical images with similar annotations.
Jia et al. (Sun,) studied this question.