Classification Of Knee Osteoarthritis Based On Depth Learning And Magnetic Resonance Images

Knee osteoarthritis can lead to serious consequences such as joint pain,hydrops,limited joint movement and abnormal gait.Although this disease will not lead to death,the reduction of articular cartilage is irreversible.The diagnosis of knee osteoarthritis is mainly based on the results of imaging examination.As a degenerative disease,early identification of knee osteoarthritis is very important for clinical intervention to reduce the pathological changes.At present,the research on the classification of knee osteoarthritis mainly focuses on X-ray images,but X-ray images can’t see the image details and are insensitive to the changes of early knee osteoarthritis.Magnetic Resonance(MR)images can clearly observe the early characteristics of knee osteoarthritis,so this paper proposes a classification algorithm of knee osteoarthritis based on MR images.Deep learning has been widely used in medical image classification.However,deep learning networks with excellent model performance often need very large sample data sets,and the training of small sample data sets can easily lead to over-fitting of deep learning models.In addition,for the classification task,the most important thing is the extraction of image features.The features extracted by a single deep learning model are often not sufficient,which limits the improvement of model classification performance.Therefore,this paper studies two classification algorithms based on deep transfer learning for knee MRI images.The main research work is as follows:(1)The classification method of knee MRI images based on two-stage depth fine-tuning DenseNet201 transfer learning model is studied.This method divides the training into two stages.In the first stage,all basic layers are frozen,and only the weights of all connected layers embedded in the neural network are trained.The second stage is to unfreeze some basic layers and train the weights of the unfrozen basic layer and the fully connected layer embedded in the neural network.In this step,this paper designs a block-by-block fine-tuning strategy based on Dense Block for training.The experimental results show that releasing dense blocks one by one in fine tuning can not only save time,but also improve the classification accuracy.(2)The image classification method of knee osteoarthritis based on multi-feature network fusion is studied.Aiming at the problem of insufficient feature extraction and poor robustness of single deep learning network model,a multi-feature network fusion deep learning model is further proposed based on two-stage deep fine-tuning migration convolutional neural network.The same picture can extract different features at the output end of convolution layer of different models through different deep learning models,fuse these features to get more judgmental features,and then classify the fused features through the full connection layer to get very good classification performance.The comparative experimental results show that the combination of DenseNet201 two-stage depth fine-tuning transfer learning model and Inception-ResNet-v2 network model has the best classification effect on the knee joint magnetic resonance image dataset,and the accuracy of 0.940,sensitivity of 0.975,accuracy of 0.913,specificity of 0.910,F1-Score of 0.943 and Matthews correlation coefficient of 0.884 are obtained.