Automatic Diagnosis Of Spinal Injuries Based On Medical Images

In recent yeas,computers have been used to automatically read medical images and help doctors to make diagnosis,which attracts more and more attention.Computer-aided diagnosis systems greatly reduce the burden of doctors and alleviate the imbalance of medical resources.Recently,the rapid development of deep learning and its wide applications in computer vision provide more reference for further improving the performance of computer-aided diagnosis.Because of the characteristics of some diseases,the features of some lesions are not obvious in medical images.The visual features of patients and healthy people are very similar to each other.In addition,patients suffered from the same disease may show different features in medical images.This makes the automatic diagnosis system inevitably encounter the problem of fine-grained classification and recognition when distinguishing images between patients and healthy people.There are plenty of fine-grained classification algorithms based on natural images,but they do not take use of the nature of medical images.In this paper,we mainly focus on two common spinal diseases,spinal dislocation and vertebral collapse.We propose two kinds of networks to simulate the doctor’s decision-making process based on the expert knowledge of doctors.To capture the weak visual features of spinal dislocation,we propose a kind of data-knowledge dual driven classification model.According to the prior knowledge that doctors pay attention to the front and rear edges of spine with priority,we propose the spatial regularization to guide the attention of the model.Class activation maps are used to make the model focus on spinal edges.Thus we can use both data and knowledge to train the model.In order to backprop the gradients through class activation maps,we propose to explicitly compute activation maps by multiplying weights matrices of fully connected layers.This makes the generation of activation maps no longer rely on the backpropagation mechanism embedded in the deep learning platforms.Experiments show that the proposed method effectively improve the accuracy of classification.Furthermore,the less traning data is used,the more gains our method will bring.The highlighted regions in class activation maps are near the front and rear edges of spine,which validates the effectiveness of guiding the attention of models through ativation maps.Additionally,we also implement experiments with different backbone networks.The proposed method improves the performance of all the experimented networks.To solve the problems caused by the fuzzy classification boundary and noisy label,we propsose the vertebral collapse classification network based on the context information of adjacent vertebrae.There are three input branches with identical structures.The vertebra to be diagnosed is inputted into the middle branch and the adjacent two vertebrae are inputted into the other two branches.This simulates the decision-making process that doctors usually compare the vertebra to be diagnosed with adjacent vertebrae to make final decisions.Furthermore,the features from adjacent input branches are inputted into two discriminators with shared parameters,which makes the extracted features contain more information about the difference of adjacent vertebrae.This can help to overcome the fuzzy classification boundary to some extent.For the noisy label,we propose a simple and effective method to suppress the noise.When the model converges during training,samples with noise label causes larger loss values than clean ones.Therefore we filter noise samples by loss values and suppress backpropagation of noise samples with large loss values,which reduces the impact of noisy labels.Experiments results validate that the proposed method effectively improves the performance of the model.It is also found that the distribution of the three categories is more discriminative in t-SNE visualization maps.In addition,a series of ablation experiments are designed to verify the necessity and effectiveness of each module in our method.