Research On Classification And Typing Method Of Femoral Head Image Based On Deep Learning Technology

In today’s society,alcohol consumption and the use of hormone drugs are becoming more and more common,leading to an increasing incidence of osteonecrosis of the femoral head(ONFH)every year,and showing a trend towards younger age.ONFH has difficulty in diagnosis such as the symptoms being overlooked in the early stages and the magnetic resonance imaging(Magnetic Resonance Imaging,MRI)images of the femoral head showing small areas,which can result in misdiagnosis.To solve these problems,this paper studies the femoral head image classification problem and the ONFH image typing problem based on deep learning technology,proposing a two-stage classification method combining image semantic segmentation,attention mechanism,and image classification.For the femoral head image classification problem,a two-stage classification method Classifier-FHI is designed based on the Res UNet semantic segmentation model and the Swin-Transformer image classification network model.For the ONFH image typing problem,a Dense Net network model with Dual Attention attention module is proposed as the classifier for the ONFH typing,and a typing method Typing-ONFHI is designed based on the Res UNet semantic segmentation.Experimental results show that the proposed models and methods are better than the currently available methods in femoral head image classification and typing problems.They can reach the diagnostic level of medical experts and the diagnostic efficiency is dozens of times higher than that of medical experts.The research results in this paper provide an efficient and automatic technical method for the diagnosis of ONFH,specifically completing the following tasks:(1)A femoral head MRI image dataset JLU-ONFH is constructed and publicly released for the first time.This dataset includes classification and typing datasets,with2286 positive samples and 8782 negative samples in the classification dataset,both original and semantically segmented images are included,without data augmentation.The typing dataset includes 1448 middle-sectional images of ONFH and the images are preprocessed by central cropping.This dataset can be used for the research of femoral head image classification and typing.JLU-ONFH provides the first benchmark dataset for the research of femoral head image classification and typing based on deep learning technology.(2)The proposed femoral head image classification method and ONFH image typing method are both two-stage methods.The first stage task is to perform image semantic segmentation on the femoral head MRI image,segmenting the femoral head and the acetabulum,and using the segmentation results as input to the second stage,which can effectively improve the accuracy of the subsequent classification task.In this paper,Res UNet is used as the network model for semantic segmentation,and the m IOU and Dice coefficient of the segmentation task are 88.6% and 0.909.(3)A femoral head image classification method Classifier-FHI is proposed in this paper,using Swin-Transformer as the image classifier.The experimental results show that the accuracy of the method directly using Swin-Transformer and the Classifier-FHI method for classification on JLU-ONFH dataset are 91.75% and 98.09%.Moreover,the Classifier-FHI method proposed in this paper has an accuracy rate 2% higher than the SOTA model classification method.(4)A femoral head ONFH image typing method Typing-ONFHI is proposed in this paper.In this method,a DA-Dense Net network model combining Dense Net and Dual Attention attention mechanism is proposed as the classifier for typing,and Focal loss is used as the loss function.Experimental results show that the accuracy of the DA-Dense Net model is 2.7% higher than that of the baseline model Dense Net,and the final typing experimental results have an accuracy rate of 88%,with macro-average and micro-average AUC areas of 0.9 and 0.92.