Automated Segmentation and Needle Tracking in B-Mode Ultrasound Using a Novel Attention U-Net

This paper presents a deep learning-based algorithm for the segmentation of cysts or tumours of various sizes from any B-mode ultrasound image along with the needle tracking that addresses challenges such as noise, scattering, and clutter. We employ a novel attention model with an encoder and decoder that is similar to U-Net architecture trained on diverse synthetic B-mode ultrasound images of cysts, tumours, and needles. We utilize a synthetically generated dataset with the help of Field II toolbox created explicitly to model differences in cyst diameters and acoustic properties between hypoechoic and hyperechoic regions, as well as augmented to introduce diversity. The architecture has been optimized to perform in real-time, which make it suitable to robotic surgical usages. With simulated, experimental phantom test images achieved a dice score of 0.966, a structural similarity index of 0.973, and a peak signal-to-noise ratio of 29.696 dB. Our findings show high accuracy in image segmentation metrics. In future, we aim to incorporate this segmentation technique in automated surgical systems, which will further minimize operator dependency. To the best of our knowledge, the proposed method is the first fully automated algorithm that simultaneously performs the segmentation of cysts along with needle tracking.