The EfficientNet-based MRI Brain Imaging Data Classification For Mild Cognitive Impairment

In recent years,with the continuous improvement of people’s living standards,the average life expectancy is also getting longer and longer,followed by various elderly diseases,such as Alzheimer’s disease（AD）.The disease has brought great losses and injuries to the country,society,individuals,and patients’families.However,there is still a lack of effective treatment for moderate and severe AD.The current research generally focus on the prevention and treatment of mild AD.As a medical aid,this paper uses the deep learning method to classify pMCI（Progressive Middle Cognitive Improvement）and sMCI（Stable Middle Cognitive Improvement）on the open source dataset ADNI^[4]（Alzheimer’s Disease Neuroimaging Initiative）.The details are as follows:（1）In view of the poor classification effect caused by the class imbalance and sample difficulty imbalance of the training data set and the reduction of the output of the fully connected classifier from ultra-high dimension to two dimensions,this paper introduces the prediction error of the focus loss calculation model to optimize the whole model.The experimental results show that the output of the optimized classifier is more balanced and the discrimination is more obvious.The specific work is as follows:the image data of white matter,gray matter and skull removal are fused,and then input into Efficientnet-V2-S feature extraction network.Finally,the loss value is predicted by the model calculated by focus loss,and the back-propagation training model is used.（2）Because of the dimensions of whole brain MRI images are too high and difficult to distinguish,this paper constructs a multimodal feature fusion module and a feature extraction module.The specific work is as follows:The original images,Sobel images and discrete Cosine transform images corresponding to the five parts of the left hippocampus,right hippocampus,left temporal lobe,right temporal lobe and spastic fissure of the brain region are combined respectively,and then longitudinal stitching is carried out after multi-modal feature fusion and feature extraction,so as to achieve the purpose of dimensionality reduction and feature extraction.（3）In view of the insufficient training samples of MRI images of Alzheimer’s disease and the over fitting problem caused by small sample training,a general framework for deep learning of small sample images is proposed based on twin network and Contrast loss.The model proposed in this paper can be flexibly replaced and the model weight parameters can be tuned and migrated.For example,the model trained in Chapter 4 is tuned and migrated in Chapter 5.The specific work is as follows:Firstly,the network receives different inputs from different brain regions of a pair of test samples,and sends them to different branch networks for feature extraction according to different brain regions,so as to obtain the corresponding image features of the brain region;Then,the image features corresponding to different brain regions are fused through the feature fusion module to obtain the image features corresponding to the sample;Finally,the extracted image features are sent to the contrast loss to obtain the similarity between sample pairs,and the extracted image features are sent to the sub classifier to obtain the classification probability results.The experimental results show that the proposed model can better classify pMCI and sMCI image data.