Jigsaw training-based background reverse attention transformer network for guidewire segmentation.

Abstract

Guidewire segmentation plays a crucial role in percutaneous coronary intervention. However, it is a challenging task due to the low signal-to-noise ratio of X-ray sequences and the great imbalance between the number of foreground and background pixels. Besides, most existing guidewire segmentation methods are designed for single guidewire segmentation. This paper aims to solve the task of single and dual guidewire segmentation in X-ray fluoroscopy sequences. A jigsaw training-based background reverse attention (BRA) transformer network is proposed. A jigsaw training strategy is used to train the guidewire segmentation network. A BRA module is also designed to reduce the influence of background information. First, robust principal component is conducted to generate background maps for guidewire sequences. Then, BRA is computed on the basis of the background features. The experimental results on the dataset collected from three hospitals show that the proposed method can achieve single and dual guidewire segmentation in X-ray fluoroscopy sequences. Higher F1 score and precision than state-of-the-art guidewire segmentation methods can be obtained in most cases. The jigsaw training strategy helps reduce the need for dual guidewire data and improve the performance of the network. Our BRA module helps reduce the influence of background information and distinguish the guidewire. The proposed methods can obtain higher performance than state-of-the-art guidewire segmentation methods.