Abstract:Nowadays, the ultrasound image plays an important role in diagnosis and treatment. During the peripheral nerve block (PNB), doctors firstly need to locate the nerve structure and its surrounding structures (muscle, veins, and arteries). However, how to automatically identify these objects from the brachial plexus ultrasound image is still a challenging task. This paper proposed a novel Multi-Object Iterative Detection Network, Mul-IDNet, to realize the multi-object detection from the ultrasound image, which is designed by following the Mask RCNN. The proposed network mainly consists of four parts: the feature extraction network, the FCFPN (Full Connected Feature pyramid network), the MIRPN(Mean Iterative Region Proposal Network) and the head architecture. In detailed, the FCFPN and the MIRPN are proposed to replace the original FPN and RPN module in the Mask RCNN. And the FCFPN is mainly used to fully utilize the multi-scale feature information, while the MIRPN is proposed to improve the ability of proposals predicting by further utilizing the boundary features of the target object. By evaluating the proposed Mul-IDNet on the UBPD dataset,the extensive experiments show that the proposed method achieve a better detection performance compared to other counterpart methods.

Mul-IDNet: Multi-Objects Detection from the Brachial Plexus Ultrasound Image