Hi, in MATLAB, when dealing with 4D images where the fourth dimension represents frequency, there are two key implementation approaches depending on how the frequency information is handled:
1.Treating Frequency as Additional Channels - Reshape the dataset to merge the frequency dimension into the channel dimension: Xtrain:(H,W,C,F,N)→(H,W,C×F,N) . The CNN treats different frequency bands as additional feature channels, similar to how colour channels (RGB) work.
2.Treating Frequency as an Additional Input Dimension - Instead of using “imageInputLayer”, use custom 3D convolutions (convolution3dLayer) to process the frequency dimension separately: Xtrain:(H,W,C,F,N)
While the first approach has simple implementation leveraging built in MATLAB deep learning tools it loses explicit frequency structure.
MATLAB Implementation of a 3D CNN for 4D Images on random data
% Define Dummy Data
imageSize = [32, 32, 3, 10]; % (Height, Width, Channels, Frequency)
numSamples = 200; % More samples improve learning
numClasses = 5; % Classification task
% Generate Random 4D Data (height, width, channels, frequency, samples)
XTrain = rand(imageSize(1), imageSize(2), imageSize(3), imageSize(4), numSamples);
YTrain = categorical(randi(numClasses, [numSamples, 1])); % Random labels
% Define a deeper 3D CNN
layers = [
image3dInputLayer(imageSize, 'Normalization', 'none')
% First convolutional block
convolution3dLayer([3 3 3], 32, 'Padding', 'same')
batchNormalizationLayer
reluLayer
maxPooling3dLayer([2 2 1], 'Stride', [2 2 1])
% Second convolutional block
convolution3dLayer([3 3 3], 64, 'Padding', 'same')
batchNormalizationLayer
reluLayer
maxPooling3dLayer([2 2 1], 'Stride', [2 2 1])
% Third convolutional block
convolution3dLayer([3 3 3], 128, 'Padding', 'same')
batchNormalizationLayer
reluLayer
% Fourth convolutional block
convolution3dLayer([3 3 3], 256, 'Padding', 'same')
batchNormalizationLayer
reluLayer
maxPooling3dLayer([2 2 2], 'Stride', [2 2 2])
% Fully connected layers
fullyConnectedLayer(512)
reluLayer
dropoutLayer(0.5) % Reduces overfitting
fullyConnectedLayer(numClasses)
softmaxLayer
classificationLayer];
% Training Options with Adam Optimizer
options = trainingOptions('adam', ... % Faster convergence than SGDM
'InitialLearnRate', 0.001, ...
'MaxEpochs', 20, ... % More epochs for better learning
'MiniBatchSize', 16, ...
'Verbose', true, ... % Ensures loss printing
'Plots', 'training-progress'); % Visualize training
% Train the Network
net = trainNetwork(XTrain, YTrain, layers, options);
You can also refer this paper for information about 4-d convlutions
Please refer this MATLAB documentation for “convn” function in MATLAB
