【CNN回归预测】卷积神经网络CNN数据回归预测【含Matlab源码 2003期】.zip

clc, clearvars, close all %------------------------------------------------------------------------------------------------ %%%%% Deep learning convolutional neural network regression v1.1 %%%%% % With network parameter gridsearch, input normalization and geometric image augmenation: % Network is designed to learn to predict a numerical value from images % Provided example input (X)data is 3 channel (R2*, QSM, GRE mag) MR image slices of ex vivo stroke blood clots % Provided example output (Y)data is RBC content of clots determined through histological analysis % Built to iterate cross-validation experiments over a set of specified network parameters (gridsearch) % Can customize which parameters to iterate over by changing the for loops % Network settings which produce the highest accuracy will be used to form predictions on a separate test set % You must add iterated variables in for loops (line 87) and results table pull (line 275) for this to work % Network can normalize training data input based on distribution of the output variable using random oversampling (ROS) % Network can apply random geometric augmentation operations to training data % And duplicate training dataset to increase size, prior to augmenation (duplication factor) % Handles input predictor (x) data as [X, Y, nchannel, nslice] where nchannel = 1 or 3 % Handles input prediction (y) data as [y(1):y(nslice)] % v1.0.1 code written and posted by Spencer D. Christiansen, October 2021. Tested on Matlab 2020b % v1.1 updated March 2022: % -fixed overwrite_crossval_results_table glitch that created new table for every iteration % -fixed nonexistent aug_params matrix when augment_training_data set to 0 % -minor changes to comments and variable names for clarity %------------------------------------------------------------------------------------------------ load([ 'test_input_xdata.mat']) load([ 'test_input_ydata.mat']) save_directory = ''; summary_file_name = 'Crossval_results_summary_table'; %Network settings: augment_training_data = 1; %Randomly augment training dataset normalize_training_input = 1; %Normalize training data based on predictor distribution overwrite_crossval_results_table = 1; %1: create new cross validation results table, 0: add lines to existing table %Input data pre-processing: xdata = permute(xdata,[1,2,4,3]); %input needs to be [X, Y, nchannel, nslice] n_images = size(ydata,1); %For random splitting of input into training/testing groups img_number = 1:n_images; rng(20211013); idx = randperm(n_images); %Training/testing data split: n_test_images = 28; %Number of images to kept for independent test set n_crossval_folds = 8; %Number of training cross-validation folds n_train_images = n_images-n_test_images; idx_test = find(ismember(img_number, idx(end-(n_test_images-1):end))); XTest = xdata(:,:,:,idx_test); YTest = ydata(idx_test); xdata(:,:,:,idx_test) = []; ydata(idx_test) = []; img_number(idx_test) = []; idx(end-(n_test_images-1):end) = []; %Network parameters (can iterate over by moving into for loop): %params.optimizer = {'sgdm','adam'}; %'sgdm' | 'adam' params.batch_size = 8; %params.max_epochs = [4,6,8]; params.learn_rate = 0.001; params.learn_rate_drop_factor = 0.1; params.learn_rate_drop_period = 20; params.learn_rate_schedule = 'none'; %'none', 'piecewise' params.shuffle = 'every-epoch'; params.momentum = 0.9; %for sgdm optimizer params.L2_reg = 0.01; params.conv_features = [16, 16, 32]; %Number of feature channels in convolutional layers params.conv_filter_size = 3; params.conv_padding = 'same'; params.pooling_size = 2; params.pooling_stride = 1; params.dropout_factor = 0.2; params.duplication_factor = 3; %Duplicate training set by N times show_plots = 0; %1: show plots of training progress tic iter = 1; disp('Performing cross validation evaluation over all network iterations:') for var1 = {'sgdm','adam'} params.optimizer = var1; for var2 = [4,6,8] params.max_epochs = var2; %for var3 = X:Y %params.example = var3; %etc. %Splitting training data into k-folds for k = 1:n_crossval_folds images_per_fold = floor(n_train_images/n_crossval_folds); idx_val = find(ismember(img_number, idx(1+(k-1)*images_per_fold:images_per_fold+(k-1)*images_per_fold))); YValidation = ydata(idx_val); XValidation = xdata(:,:,:,idx_val); XTrain = xdata; XTrain(:,:,:,idx_val) = []; YTrain = ydata; YTrain(idx_val) = []; %ROS input normalization: if normalize_training_input == 1 [XTrain, YTrain] = ROS(XTrain, YTrain, params.duplication_factor); else XTrain = repmat(XTrain,1,1,1,params.duplication_factor); YTrain = repmat(YTrain,params.duplication_factor,1); end %Random geometric image augmenation: %Augmentation parameters aug_params.rot = [-90,90]; %Image rotation range aug_params.trans_x = [-5 5]; %Image translation in X direction range aug_params.trans_y = [-5 5]; %Image translation in Y direction range aug_params.refl_x = 1; %Image reflection across X axis aug_params.refl_y = 1; %Image reflection across Y axis aug_params.scale = [0.7,1.3]; %Imaging scaling range aug_params.shear_x = [-30,50]; %Image shearing in X direction range aug_params.shear_y = [-30,50]; %Image shearing in Y direction range aug_params.add_gauss_noise = 0; %Add Gaussian noise aug_params.gauss_noise_var = 0.0005; %Gaussian noise variance if augment_training_data == 1 XTrain = image_augmentation(XTrain,aug_params); else aug_params = structfun(@(x) [], aug_params, 'UniformOutput', false); end %Network structure: layers = [ imageInputLayer([size(XTrain,1),size(XTrain,2),size(XTrain,3)]) convolution2dLayer(params.conv_filter_size,params.conv_features(1),'Padding',params.conv_padding) %batchNormalizationLayer reluLayer averagePooling2dLayer(params.pooling_size,'Stride',params.pooling_stride) convolution2dLayer(params.conv_filter_size,params.conv_features(2),'Padding',params.conv_padding) %batchNormalizationLayer reluLayer averagePooling2dLayer(params.pooling_size,'Stride',params.pooling_stride) convolution2dLayer(params.conv_filter_size,params.conv_features(3),'Padding',params.conv_padding) %batchNormalizationLayer reluLayer dropoutLayer(params.dropout_factor) fullyConnectedLayer(1) regressionLayer]; params.validationFrequency = floor(numel(YTrain)/params.batch_size); options = network_options(params,XValidation,YValidation,show_plots); net = trainNetwork(XTrain,YTrain,layers,options); %Network results: accuracy_threshold = 0.1; %Predictions within 10% will be considered 'accurate' predicted_train = predict(net,XTrain); predictionError_train = YTrain - predicted_train; numCorrect_train = sum(abs(predictionError_train) < accuracy_threshold); accuracy_train(k) = numCorrect_train/numel(YTrain); error_abs_train(k) = mean(abs(predictionError_train)); rmse_train(k) = sqrt(mean(predictionError_train.^2)); predicted_val = predict(net,XValidation); predictionError_val = YValidation - predicted_val; numCorrect_val = sum(abs(predictionError_val) < accuracy_threshold); accuracy_val(k) = numCorrect_val/numel(YValidation); error_abs_val(k) = mean(abs(predictionError_val)); rmse_val(k) = sqrt(mean(predictionError_val.^2)); if k == 1 predicted_val_table(1:numel(YValidation),1) = predict(net,XValidation); if iter == 1 YValidation_table(1:numel(YValidation),1) = YValidation; end else predicted_val_table(end+1:end+numel(YValidation),1) = predict(net,XValidation); if