A Machine Learning-Based Approach To Classify MRI-Negative Epilepsy

Magnetic Resonance Imaging(MRI)is widely used in the detection of brain diseases.With the rise of machine learning techniques,MRI classification methods incorporating machine learning have been proposed,and these methods are widely used to assist physicians in aiding disease diagnosis and for further medical research.When diagnosing disease,patients can be classified as negative or positive.For the use of MRI as an aid to diagnosis,a negative patient is one who has actual disease but no visible lesions on MRI,i.e.,the physician cannot determine whether the patient has disease by looking at the MRI with the naked eye.Addressing the above issues,feature extraction and classification models based on machine learning designing for three classification tasks for negative frontal lobe epilepsy,negative temporal lobe epilepsy,and negative patients were investigated in thesis,with the main research elements including:1.MRI classification algorithm based on traditional machine learning: firstly,the negative epilepsy MRI dataset(total 270 cases)was built and processed using Free Surfer medical image processing software dataset;the MRI was denoised,normalized,and brain region segmented,and then the medical features were quantified according to the medical atlas.Comparing the effectiveness of multiple traditional machine learning classifiers for negative epilepsy MRI classification,the highest classification accuracy was achieved by SVM(linear kernel).Meanwhile,based on the SVM classifier,the best screening model is RFE(recursive feature elimination)when comparing the effect of multiple feature screening models on negative epilepsy MRI feature screening.Meanwhile,the feature screening models were combined to obtain the RFE-PCA-RFE feature screening model with better performance,and the accuracy of this model was 66.32%,the sensitivity was 79.09%.2.MRI classification algorithm based on deep learning: MRI is preprocessed and expanded into 2D MRI sequences by coordinate axes as model input,CNN is used to extract negative epilepsy MRI features and test and compare the effect of multiple CNN networks;meanwhile,the neural network visualization model Grad-CAM++(Gradientweighted Class Activation Mapping++)was used to map the features of the neural network by generating a heat map,and the visualization effect and model accuracy were combined to obtain a y-axis segmentation standard MRI negative epilepsy classification model based on the Inception V3 network,and the accuracy of this model was 52.60%.The Inception V3-LSTM classification model was obtained by inputting this Inception V3 network mix10 layer feature extraction into LSTM,and the accuracy of this model was 64.28%,the sensitivity was 71.43%.Meanwhile,the study conducted comparison experiments on MRI de-contextualized input 3D convolutional neural network and obtained the optimal 3D convolutional neural network as Shuffle Net with an accuracy of 50.00% and a sensitivity of 57.14%.