Search

Robotic ultrasound (US)-guided surgery is a common clinical workflow for the diagnosis and treatment of many diseases. Tissue deformation is an important factor that affects surgical accuracy. Intraoperative real-time deformable registration is therefore critical to ensure surgical precision. In this study, a real-time US image nonrigid registration method was proposed. First, the analytical derivatives and corresponding cascaded convolution operations were derived, facilitating the fast calculation of the deformation field and gradients. Second, an optimizer fusing the trust-region and limited-memory Broyden-Fletcher–Goldfarb-Shanno (L-BFGS) method was proposed, which accelerates the optimization by adopting vectorized operations instead of large-scale matrix multiplication. The proposed method achieved real-time nonrigid registration with an average registration time of 30 ms for 2-D images with a resolution of $256^{2}$ and 500 ms for 3-D images with a resolution of $112^{3}$ , respectively, on a desktop CPU computing platform. It satisfied clinical real-time requirements without prior dataset construction and training process. The proposed method was validated on public datasets, clinical cases, and robotic phantom intervention experiments. Results demonstrated that our method realized great performance in both registration speed and accuracy. This method is expected to improve the US guidance accuracy of robots, promoting the precision and automation of related surgeries.

Industrial edge computing devices typically have limited computational performance. Current high-precision grasping pose detection models are often over-parameterized, resulting in computational inefficiency and making it difficult to deploy on industrial robots. Existing efficient network design methods rely on constructing, training, and validating a large number of new networks to obtain the final target network, which also requires significant computing resources. To address this issue, the network convergence indicator (NCI) is proposed to realize the prediction of network accuracy before training. NCI can evaluate convergence based on the evasion of flat regions and saddle points theory. It is used to control the width and depth of the network model and optimize its inner layers. Based on these principles, NCIGrasp is proposed, which is a high-precision lightweight grasping network achieving outstanding performance on the Cornell and Jacquard datasets with only 10% of the parameters of previous advanced models. NCIGrasp has a small size of 0.69 MB and fast inference speed on the server (~8 ms), making it suitable for edge computing. Experimental results show that NCIGrasp has real-time grasping pose detection ability (~21 ms) for unknown objects on an ordinary laptop with limited computation capability.

Image anomaly detection is extremely challenging in industrial manufacturing processes due to unforeseen and diversified anomalies. Recently, unsupervised anomaly detection methods based on knowledge distillation have been developed and have shown remarkable potential. While most existing methods are devoted to knowledge generalization, they are inadequate for the fine-grained detection task. To address this issue, we propose a novel adaptive context-aware distillation (ACAD) paradigm that gives due consideration to distillation component dependencies and knowledge transfer optimization. Technically, a novel adaptive distillation module (ADM) is proposed for optimal context-aware knowledge transfer, which consists of contrastive decoupling distillation (CDD) and masked perceiving distillation (MPD). The proposed CDD helps to constrain the distribution of different semantic patterns and strengthen the discriminative capability. Vanilla methods treat every pixel as an equal contribution and fail to focus on critical information. To this end, the MPD is proposed to weigh different contextual knowledge adaptively. Extensive experiments with mainstream anomaly detection datasets show that ACAD outperforms the state-of-the-art competitors in accuracy and efficiency. In addition, the experimental results with a real-world inkjet printing organic electroluminescence display (OLED) panel dataset further demonstrate the effectiveness of our method.