Research On Meniscus Detection And Classification Technology Based On Machine Learning

Meniscal injury,a common knee disorder,may lead to a series of symptoms such as knee pain and osteoarthritis if not diagnosed or treated properly.Correct identification is an important prerequisite for clinical intervention and treatment.Magnetic Resonance Imaging(MRI)is the most commonly used imaging method for clinical diagnosis of meniscus injury,which can accurately reflect the tear position,tear type,tear grade,and so on.In traditional clinical medicine,MRI is mainly judged subjectively by doctors’ expertise or experience,in which the process takes a long time and can lead to missed or misdiagnosed cases during the diagnosis.With the continuous development of machine learning in the study of clinical imaging,machine-learning-based MRI disease prediction has become an emerging area of artificial intelligence.In the task of medical classification,machine learning can avoid the problems of complex artificial feature extraction and long time-consuming diagnosis.This paper,therefore,studies and applies the classification models and detection models in meniscus sagittal MRI using machine learning methods.The main work is as follows.(1)Dichotomous classification of meniscal injury: An automatic meniscus classification method based on improved Mobile Netv1 was proposed,which consisted of two parts: meniscal MRI preprocessing and improved classification network model.The preprocessing steps were as followed: firstly,the central deflation method was used to unify the image pixels in the data set to the same size to facilitate image cropping;secondly,the Hough Circle localization and cropping algorithm was used to determine the local region of the meniscus,which could effectively extract the local region of the meniscus in MRI;thirdly,data enhancement on the images in the original data set was performed to avoid category imbalance.Mish as the activation function of Mobile Netv1 was adopted to implement the improved network with a freeze training strategy added.The experiments showed that the method in this chapter could improve the network training accuracy performance,improve the secondary classification accuracy to97.55%,and improve the training time performance of the network.(2)Tri-classification study of meniscus detection and meniscus tears: An improved positioning and cropping method,YOLOv3,was proposed.ECANet was used as the attention mechanism module in MBConv,named MBConv+,which stacked to constitute Efficient Net-B0+ as the feature extract network of YOLOv3 for the detecting,positioning and cropping of the meniscus,and the classification experiments were performed with a support vector machine.The experimental results showed that the improved YOLOv3 network significantly reduced the training time compared with the original network under similar m AP conditions,and the frames per second increased by 1.3221 and reached 29.0691;compared with the original data set,the cropped images improved in classification performance and the accuracy increased by 7.739% to91.251%.(3)Based on PyQt5 development library,the meniscus detection and identification software was developed to implement the Hough Circle detection and the improved YOLOv3 identification and location functions in this paper.