Research On Semi-Supervised Patella Misalignment Detection Algorithm

Patella misalignment,which refers to the horizontal misalignment between the patella and the center of the knee,is one of the key factors affecting both lower limbs’ radiographic quality.Reducing the imaging quality problems caused by the misalignment of the patella can reduce the misdiagnosis rate of subsequent orthopedic diseases(such as knee varus,knee valgus,etc.).However,current radiography process mainly relies on the radiologist’s subjective assessment of the patient’s posture.The process is time-consuming and exists inter-operator differences.Therefore,in order to solve the above problems,this thesis proposes a patella misalignment detection algorithm based on convolutional neural networks(CNNs)to automatically detect patella misalignment,which could reduce the impact of imaging quality problems on the subsequent diagnosis.Besides,to take full advantage of limited labeled data and tremendous unlabeled data in the medical scene and improve the robustness of the model,this thesis takes deep research on semi-supervised learning methods.The task of detecting patella misalignment is divided into two subtasks: the lower limb mechanical axis key points detection and the patella detection.And this thesis established the lower limb mechanical axis key point detection dataset and the patella detection dataset.For the research on the detection algorithm of the key points of the lower limb mechanical axis,this thesis first analyzes the quantization error problem of the existing algorithm in this task,then introduces an offset branch to predict the offset value to compensate quantization error.This method improves the average accuracy from 66.0% to75.2%.Besides,for the complex and dense bone structure,a stacked hourglass network with deformable convolution is proposed to explicitly encode the bone structure,further increasing the average accuracy from 75.2 % to 83.0%.The key points of the lower limb mechanical axis are not only used to provide the position of the knee center point,but also form the lower limb mechanical axis.The detected limb mechanical axis can be used for auxiliary diagnosis of knee varus and knee valgus,which greatly reduces the workload of orthopedic surgeons.For the design of the patella detection algorithm,in order to make full use of the limited labeled data and tremendous unlabeled data,this thesis designs a teacher-student mutual learning framework based on the classic two-stage object detection network to complete the semi-supervised patella detection task.Compared with the fully supervised training method which only uses the labeled data,the proposed semi-supervised object detection algorithm improves AP from 64.72% to 66.45%,while AP@0.75 increased from 79.84% to 86.64%,The results prove the effectiveness of the semi-supervised object detection framework proposed in this thesis.