Preprocess Volumes for Deep Learning

Read Volumetric Image Data

Supported file formats for volumetric image data include MAT-files, Digital Imaging and Communications in Medicine (DICOM) files, and Neuroimaging Informatics Technology Initiative (NIfTI) files.

You can read volumetric image data into an ImageDatastore. When you create the image datastore, specify the 'FileExtensions' argument as the file extensions of your data. Specify the ReadFcn property as a function handle that reads data of the file format. For more information, see Datastores for Deep Learning.

The table shows how to create an image datastore for each of the supported file formats. The filepath argument specifies the path to the files or folder containing image data.

Image File Format

Image Datastore Creation

MAT

volds = imageDatastore(filepath, ...
   'FileExtensions','.mat','ReadFcn',@(x) matRead(x));

You must create the matRead function to read data from a MAT-file. Save the function in a file called matRead.m.

function data = matRead(filename)
% data = matRead(filename) reads the image data in the MAT-file filename

inp = load(filename);
f = fields(inp);
data = inp.(f{1});
end

DICOM

volds = imageDatastore(filepath, ...
   'FileExtensions','.dcm','ReadFcn',@(x) dicomread(x));

Using DICOM files for 3-D deep learning requires the volume data be stored in a single file. You cannot volume data that is divided into multiple image files.

For more information about the DICOM file format, see dicomread.

NIfTI

volds = imageDatastore(filepath, ...
   'FileExtensions','.nii','ReadFcn',@(x) niftiread(x));

For more information about the NIfTI file format, see niftiread.

Read Volumetric Label Data for Semantic Segmentation

To perform semantic segmentation of volumetric data, you must have labels corresponding to the image data.

You can read volumetric label data into a PixelLabelDatastore. When you create the pixel label datastore, specify the 'FileExtensions' argument as the file extensions of your data. Specify the ReadFcn property as a function handle that reads data of the file format.

The table shows how to create a pixel label datastore for each of the supported file formats. The filepath argument specifies the path to the files or folder containing label data. The classNames and pixelLabelID arguments specify the mapping of voxel label values to class names.

Image File Format

Pixel Label Datastore Creation

MAT

pxds = pixelLabelDatastore(filepath,classNames,pixelLabelID, ...
    'FileExtensions','.mat','ReadFcn',@(x) matRead(x));

You must create the matRead function to read data from a MAT-file. Save the function in a file called matRead.m.

function data = matRead(filename)
% data = matRead(filename) reads the label data in the MAT-file filename

inp = load(filename);
f = fields(inp);
data = inp.(f{1});
end

DICOM

pxds = pixelLabelDatastore(filepath,classNames,pixelLabelID, ...
   'FileExtensions','.dcm','ReadFcn',@(x) dicomread(x));

Using DICOM files for 3-D deep learning requires the volume data be stored in a single file. You cannot volume data that is divided into multiple image files.

For more information about the DICOM file format, see dicomread.

NIfTI

pxds = pixelLabelDatastore(filepath,classNames,pixelLabelID, ...
   'FileExtensions','.nii','ReadFcn',@(x) niftiread(x));

For more information about the NIfTI file format, see niftiread.

Associate Image and Label Data

To associate volumetric image and label data for semantic segmentation, or two volumetric image datastores for regression, use a randomPatchExtractionDatastore. A random patch extraction datastore extracts corresponding randomly-positioned patches from two datastores. Patching is a common technique to prevent running out of memory when training with arbitrarily large volumes. Specify a patch size that matches the input size of the network and, for memory efficiency, is smaller than the full size of the volume, such as 64-by-64-by-64 voxels.

You can also use the combine function to associate two datastores. However, associating two datastores using a randomPatchExtractionDatastore has several benefits over combine.

  • randomPatchExtractionDatastore supports parallel training, multi-GPU training, and prefetch reading. Specify parallel or multi-GPU training using the 'ExecutionEnvironment' name-value pair argument of trainingOptions. Specify prefetch reading using the 'DispatchInBackground' name-value pair argument of trainingOptions. Prefetch reading requires Parallel Computing Toolbox™.

  • randomPatchExtractionDatastore inherently supports patch extraction. In contrast, to extract patches from a CombinedDatastore, you must define your own function that crops images into patches, and then use the transform function to apply the cropping operations.

  • randomPatchExtractionDatastore can generate several image patches from one test image. One-to-many patch extraction effectively increases the amount of available training data.

Preprocess Volumetric Data

Deep learning frequently requires the data to be preprocessed and augmented. For example, you may want to normalize image intensities, enhance image contrast, or add randomized affine transformations to prevent overfitting.

To preprocess volumetric data, use the transform function. transform creates an altered form of a datastore, called an underlying datastore, by transforming the data read by the underlying datastore according to the set of operations you define in a custom function. Image Processing Toolbox™ provides several functions that accept volumetric input. For a full list of functions, see 3-D Volumetric Image Processing (Image Processing Toolbox). You can also preprocess volumetric images using functions in MATLAB® that work on multidimensional arrays.

The custom transformation function must accept data in the format returned by the read function of the underlying datastore.

Underlying Datastore

Format of Input to Custom Transformation Function

ImageDatastore

The input to the custom transformation function depends on the ReadSize property.

  • When ReadSize is 1, the transformation function must accept an integer array. The size of the array is consistent with the type of images in the ImageDatastore. For example, a grayscale image has size m-by-n, a truecolor image has size m-by-n-by-3, and a multispectral image with c channels has size m-by-n-by-c.

  • When ReadSize is greater than 1, the transformation function must accept a cell array of image data corresponding to each image in the batch.

For more information, see the read function of ImageDatastore.

PixelLabelDatastore

The input to the custom transformation function depends on the ReadSize property.

  • When ReadSize is 1, the transformation function must accept a categorical matrix.

  • When ReadSize is greater than 1, the transformation function must accept a cell array of categorical matrices.

For more information, see the read function of PixelLabelDatastore.

randomPatchExtractionDatastore

The input to the custom transformation function must be a table with two columns.

For more information, see the read function of randomPatchExtractionDatastore.

RandomPatchExtractionDatastore does not support the DataAugmentation property for volumetric data. To apply random affine transformations to volumetric data, you must use transform.

The transform function must return data that matches the input size of the network. The transform function does not support one-to-many observation mappings.

Example: Transform Volumetric Data in Image Datastore

This sample code shows how to transform volumetric data in image datastore volds using an arbitrary preprocessing pipeline defined in the function preprocessVolumetricIMDS. The example assumes that the ReadSize of volds is greater than 1.

dsTrain = transform(volds,@(x) preprocessVolumetricIMDS(x,inputSize));

Define the preprocessVolumetricIMDS function that performs the desired transformations of data read from the underlying datastore. The function must accept a cell array of image data. The function loops through each read image and transforms the data according to this preprocessing pipeline:

  • Randomly rotate the image about the z-axis.

  • Resize the volume to the size expected by the network.

  • Create a noisy version of the image with Gaussian noise.

  • Return the image in a cell array.

function dataOut = preprocessVolumetricIMDS(data,inputSize)
 
numRows = size(data,1);
dataOut = cell(numRows,1);
 
for idx = 1:numRows
    
    % Perform randomized 90 degree rotation about the z-axis
    data = imrotate3(data{idx,1},90*(randi(4)-1),[0 0 1]);

    % Resize the volume to the size expected by the network
    dataClean = imresize(data,inputSize);
    
    % Add zero-mean Gaussian noise with a normalized variance of 0.01
    dataNoisy = imnoise(dataClean,'gaussian',0.01);

    % Return the preprocessed data
    dataOut(idx) = dataNoisy;
    
end
end

Example: Transform Volumetric Data in Random Patch Extraction Datastore

This sample code shows how to transform volumetric data in random patch extraction datastore volds using an arbitrary preprocessing pipeline defined in the function preprocessVolumetricPatchDS. The example assumes that the ReadSize of volds is 1.

dsTrain = transform(volds,@preprocessVolumetricPatchDS);

Define the preprocessVolumetricPatchDS function that performs the desired transformations of data read from the underlying datastore. The function must accept a table. The function transforms the data according to this preprocessing pipeline:

  • Randomly select one of five augmentations.

  • Apply the same augmentation to the data in both columns of the table.

  • Return the augmented image pair in a table.

function dataOut = preprocessVolumetricPatchDS(data)

img = data(1);
resp = data(2);

% 5 augmentations: nil,rot90,fliplr,flipud,rot90(fliplr)
augType = {@(x) x,@rot90,@fliplr,@flipud,@(x) rot90(fliplr(x))};

rndIdx = randi(5,1);
imgOut = augType{rndIdx}(img);
respOut = augType{rndIdx}(resp);

% Return the preprocessed data
dataOut = table(imgOut,respOut};

end

See Also

| | | |

Related Examples

More About