Channel and Spatial Attention Mechanism-based Yolo Network for Target Detection of the Lung Ultrasound Scanning Robot

Ultrasonography is the preferred method for detecting lung lesions, which has been widely recognized and adopted. In order to reduce the work stress and risk of infection for health care workers, lung ultrasound (LUS) scanning robots can be used to perform this task instead of doctors. For the LUS scanning robot, the first critical step is the detection of the scanned areas of the patient's lungs. The algorithms using computer vision and deep learning have made significant progress in the field of target detection, which are being applied more and more frequently to robotic autonomous decision-making. However, some of the performance of the traditional convolutional neural networks (CNN) still needs to be improved due to the more stringent requirements for real-time, accuracy and hardware cost of LUS scanning robots. The effectiveness of CNN combined with computer vision for detecting scanned sites in LUS scanning robots is explored and a lightweight Yolo network based on the attention mechanism is proposed. We use Yolo V4- Tiny model as the backbone network. Then the depth feature information of the feature layer output from the backbone network are extracted by adding space and channel attention mechanism, and the prediction results are output by using Yolo head model. Meanwhile, we also optimize the loss function of the network to further improve the performance of the network. Compared to several classical algorithms, the proposed method improves the target detection performance of the LUS scanning robot, which achieves 98.83% average precision, 96.87% precision, and 91.00% F1 score. We apply the proposed algorithm to the designed LUS scanning robot system and conduct clinical experiments. The results have shown that the proposed method has a 3D localization accuracy of 7.53 ± 0.37mm for the lung scanned sites, which has the potential to be applied to the target detection of the LUS scanning robot.