| Minimally Invasive Surgery(MIS)is highly valued by doctors and patients because it has the advantages of small surgical incisions,low risk of infection,and rapid wound healing that traditional surgery cannot provide.The key step is to make a small incision in the patient’s abdomen through which surgical instruments can enter the patient’s body to find the patient and perform surgical operations on the patient,and then dress the internal and external wounds.This greatly improves the patient’s surgical experience because it avoids the large area of trauma caused by opening the chest or abdomen during traditional surgery.Because minimally invasive surgery requires the surgeon to understand the details of the patient’s site from the camera on the instrument to guide the next step,minimally invasive surgery places higher demands on the surgeon and higher expectations on the technology used.Deep learning techniques have gained momentum in recent years due to their excellent performance and have swept through many research fields.In the field of computer vision,for example,deep learning solutions have completely overturned the status of traditional methods and have become the mainstream solution in the field of vision.Deep learning has also made breakthroughs in medical image processing,but it is not very effective to apply the generalized and mature methods in computer vision to medical image processing,because medical image datasets have many differences compared with general category datasets.First,the limited space of the working environment of surgical instruments in minimally invasive surgery makes the imaging distance of surgical instruments very short,which also leads to a dramatic change in the size of the imaging on the screen if the surgical instruments are moved slightly in front and back distance,and this volatility of scale brings challenges to the detection task.Second,there is close logical information between instruments in the field of surgical instrument detection,for example,a certain two or three surgical instruments must be used at the same time and cooperate with each other to complete some specific operations,and the acquisition and utilization of this information is essential to improve the performance of the detection algorithm.Third,during minimally invasive surgery,surgical instruments are often together at the same moment and operate together on the patient,resulting in too close proximity between instruments and mutual occlusion between instruments,which makes the use of common target detection methods may result in serious missed detection.To solve these problems above,this paper proposes a new network model,SoftANL-RCNN.for the first point mentioned above,this network introduces a feature fusion module,which brings scale invariance to the network model by fusing features at different levels and enhances the model’s ability to detect targets at different scales.To address the second point mentioned above,an improved Non-local module is introduced in the Soft-ANL-RCNN network,which enhances the feature representation by calculating the point-to-point correlation in the feature map to mine the logical relationships between surgical instruments.To address the third point above,Soft-ANLRCNN uses the Soft NMS algorithm instead of the commonly used non-maximal suppression algorithm to improve the detection capability of the model for dense targets.In this paper,we selected the m2cai-tool-localtion public dataset for experimental validation,and in order to test the robustness of Soft-ANL-RCNN,we built the surgical instrument detection dataset AJU-Set jointly with a professional medical team.All seven surgical instruments showed different degrees of improvement in detection results,especially for those surgical instruments that were used in conjunction with other instruments.In the subsequent ablation comparison experiments,the results also demonstrated that each component of Soft-ANL-RCNN plays an important role. |
PDF Full Text Request