Merge pull request #26 from MATLAB-Community-Toolboxes-at-INCF/fmri

fMRI denoising example + helper function

Author: stevevanhooser

examples/fmriDenoiserNetwork.m

@@ -0,0 +1,65 @@
+function net = fmriDenoiserNetwork(modelPath, pretrainedNetwork)
+%fmriDenoiserNetwork - generate MATLAB Version of the DeepInterpolation fMRI-Denoiser
+
+lgraph = layerGraph();
+
+tempLayers = [
+    image3dInputLayer([7 7 7 5],"Name","image3dinput",'Normalization','none');
+    convolution3dLayer([3 3 3],8,"Name","conv3d","Padding","same","Weights",mywini(1),"Bias",mybini(1));
+    reluLayer("Name","relu1")];
+lgraph = addLayers(lgraph,tempLayers);
+
+tempLayers = [
+    maxPooling3dLayer([3 3 3],"Name","pool1","Padding","same")
+    convolution3dLayer([3 3 3],16,"Name","conv3d_1","Padding","same","Weights",mywini(2),"Bias",mybini(2))
+    reluLayer("Name","relu2")];
+lgraph = addLayers(lgraph,tempLayers);
+
+tempLayers = [
+    maxPooling3dLayer([3 3 3],"Name","pool2","Padding","same")
+    convolution3dLayer([3 3 3],32,"Name","conv3d_2","Padding","same","Weights",mywini(3),"Bias",mybini(3))
+    reluLayer("Name","relu3")];
+lgraph = addLayers(lgraph,tempLayers);
+
+tempLayers = [
+    depthConcatenationLayer(2,"Name","conc_up_1")
+    convolution3dLayer([3 3 3],16,"Name","conv3d_3","Padding","same","Weights",mywini(4),"Bias",mybini(4))
+    reluLayer("Name","relu4")];
+lgraph = addLayers(lgraph,tempLayers);
+
+tempLayers = [
+    depthConcatenationLayer(2,"Name","conc_up_2")
+    convolution3dLayer([3 3 3],8,"Name","conv3d_4","Padding","same","Weights",mywini(5),"Bias",mybini(5))
+    reluLayer("Name","relu5")
+    convolution3dLayer([1 1 1],1,"Name","conv3d_5","Padding","same","Weights",mywini(6),"Bias",mybini(6))
+    regressionLayer("Name","out_r")];
+lgraph = addLayers(lgraph,tempLayers);
+
+clear tempLayers;
+
+lgraph = connectLayers(lgraph,"relu1","pool1");
+lgraph = connectLayers(lgraph,"relu1","conc_up_2/in2");
+lgraph = connectLayers(lgraph,"relu2","pool2");
+lgraph = connectLayers(lgraph,"relu2","conc_up_1/in2");
+lgraph = connectLayers(lgraph,"relu3","conc_up_1/in1");
+lgraph = connectLayers(lgraph,"relu4","conc_up_2/in1");
+
+net = assembleNetwork(lgraph);
+save(modelPath,"net")
+
+function w = mywini(ilayer)
+lwlnames = {'conv3d','conv3d_1','conv3d_2','conv3d_3','conv3d_4','conv3d_5'}; %layers_with_learnables
+lname = lwlnames{ilayer};
+thisweights = h5read(pretrainedNetwork,strcat('/model_weights/',lname,'/',lname,'/kernel:0'));
+w = permute(thisweights,[5,4,3,2,1]);
+end
+
+function b = mybini(ilayer)
+lwlnames = {'conv3d','conv3d_1','conv3d_2','conv3d_3','conv3d_4','conv3d_5'}; %layers_with_learnables
+lwldims = [8 16 32 16 8 1];
+lname = lwlnames{ilayer};
+b = h5read(pretrainedNetwork,strcat('/model_weights/',lname,'/',lname,'/bias:0'));
+b = reshape(b,[1 1 1 lwldims(ilayer)]);
+end
+
+end