Research On Auxiliary Diagnosis Of Femoral Head Necrosis Based On Convolutional Neural Network

Femoral head necrosis is a disease that affect human health seriously.Every year,a large number of people suffer from this disease.At present,the diagnosis of femoral head necrosis is mainly obtained by doctors’ analysis of femoral heads DR images taken by patients.The DR image of the femoral head will be interfered by the equipment and the external environment during the filming process.In addition,the large patient population brings great work pressure to doctors,which eventually leads to misdiagnosis.Aiming at the above problems,image enhancement technology and convolutional neural network method are used in thesis to study the auxiliary diagnosis of femoral head necrosis.First,aiming at the problems of poor quality and excessive noise in DR images of femoral heads,a new image enhancement method was proposed.The generalization of traditional image enhancement methods is not high,and it is not possible to obtain good enhancement results on complex and diverse femoral head DR images.Therefore,this thesis proposed a DR image enhancement method combining adaptive fuzzy values and fractional differential.By extracting different frequency bands of the original DR image,proposed an adaptive fuzzy set method to enhance the extracted low frequency band and using a fractional differential method to enhance the high frequency band.The comparison of different enhancement methods proves the effectiveness of this method.Secondly,in order to solve the problems of over-reliance on computer resources in traditional network pre-training methods,which lead to time-consuming network pre-training and waste of computing resources,a new network pre-training method based on the outlier detection is proposed.Refining the pre-training data at the feature level can better deduplicate similar features in the pre-training data set.Compared with the traditional pre-training method,the proposed method can make the model have better initialization parameters and higher training accuracy in a lower training time.Finally,in the analysis of femoral head necrosis disease not only to consider the overall characteristics of the feature area,but also the texture details on the image.Because the traditional image processing method dependent on a certain feature and cannot be applied to the complex and diverse DR image data of the femoral head.Therefore,a convolutional neural network combined with batch normalization is proposed to analyze the DR image data of the femoral head.The convolution stack is used to reduce the training parameters of the network and the batch normalization layer is used to speed up the network training.Experiments performed on the network using the enhanced DR image data shows that the network has better classification on complex femoral head DR images and reduces the problem of training costs.