Research On The Classification Of Small-sample Medical Image Based On Light-weight Convolutional Neural Network

Medical image classification is a research hotspot in the field of computer-aided diagnosis and pattern recognition.Automatic and accurate classification of human anatomical structures and diseased areas plays an important role in relieving the pressure of physicians,improving the diagnosis efficiency,and reducing the rate of misdiagnosis and missed diagnosis.Deep learning is the most widely used approach in medical image classification,and its effectiveness heavily depends on the amount and quality of training data.However,for most diseases,only limited samples of labeled data can be acquired due to such factors as the cost of data collection,physicians’ level of clinical experience,and differences in labeling habits.The limited samples will influence the performance of deep learning methods in medical image classification tasks.For small-sample medical image classification,some researchers have developed lightweight Convolutional NeuralNetwork(CNN)models to reduce the network’s demand on the amount of data.There are three types of light-weight CNN models based on classic architectures,attention and metric learning.The first type of models uses the classic architectures to ensure its performance,but its generality is poor.The second type relies on the attention module to enhance its feature extraction ability.However,the effectiveness of the attention module depends on the training process.The third type employs metric losses for training to increase the amount of information gained from the data by exploring data relationship.However,the existing metric losses cannot represent the data distribution effectively,so their advantageous effect on the network training process is limited.To solve the problems of existing light-weight CNN models,this thesis has done the following researches:(1)To reduce the difficulty of training the light-weight CNN model and improve its versatility,an unsupervised modified PCANet model has been combined with a supervised simplified DenseNet to construct a light-weight hybrid network(HybridNet)model.This model trains the kernel of the modified PCANet in an unsupervised way,and uses the output of the modified PCANet as the input to the simplified DenseNet,which is trained by the error back propagation algorithm.Thanks to the shallow features provided by the modified PCANet,the supervised sub-network is less difficult to train,and its classification accuracy is assured.Experiments on chest X-ray,mmographgy and brain magnetic resonance imaging(MRI)datasets show that the HybridNet can surpass the state-of-the-art models in the classification of different medical datasets.(2)In order to further improve the feature extraction ability of HybridNet,this thesis has designed a spiking cortical model based attention(SCMA)module.In addition,a lightweight attention model HybridNet-SCMA is generated by embedding the SCMA module into HybridNet.Experiments on chest X-ray,mmographgy,below-knee artery computed tomography(CT)images and brain MRI datasets show that the SCMA module effectively improves the classification ability of the HybridNet and the interpretability of the feature map,and HybridNet-SCMA has better performance than other compared attention networks.(3)To make full use of the distribution information of training samples and optimize the training process of light-weight CNN models,a batch similarity based triplet(BSTriplet)loss function is proposed.It analyzes all possible sample pairs in the input batch,and uses the similarity of sample pairs to help CNNs to distinguish different kinds of samples.A sample mining technology based on sample distribution to build batches is designed to ensure that the training batches have the same distribution,which helps to keep the training process stable.The BSTriplet can be combined with cross entropy(CE)loss to train CNNs and alleviate the over-fitting problem.Experiments on chest X-ray,mmographgy,belowknee artery CT images and brain MRI datasets show the joint loss of BSTriplet and CE can effectively improve the performance of various networks including HybridNet-SCMA.This thesis constructs a light-weight hybrid CNN model and improves its feature extraction ability from three different perspectives: optimizing the network structure,embedding the attention module,and improving the loss function.These improvements make the deep learning model suitable for small-sample medical image classification tasks,and provides support for the wide application of deep learning in computer-aided diagnosis.