how to fix this issue through running the program?

i am using Matlab for medical image classification and i get this issue:

note: i used pre-trained network (alexnet) with .dicom files dataset.

first i prepare my design network

second, i run my code.

>> deepNetworkDesigner
>> SHIVANCLASSIFY 

net = 

  SeriesNetwork with properties:

         Layers: [25×1 nnet.cnn.layer.Layer]
     InputNames: {'data'}
    OutputNames: {'output'}

Error using trainNetwork (line 170)
The training images are of size 227x227x1 but the input layer expects images of size 227x227x3.

Error in SHIVANCLASSIFY (line 36)
net = trainNetwork(augimdsTrain,layers_1,options)

net=alexnet
imds = imageDatastore('lung dataset-Labeled', ...
    'IncludeSubfolders',true, 'LabelSource','foldernames', ... % this for labeling by folder names
    'FileExtensions','.dcm','ReadFcn',@readDicomDatastoreImage); % this a function

[imdsTrain,imdsValidation] = splitEachLabel(imds,0.7);

augmenter = imageDataAugmenter( ...
    'RandRotation',[-20,20], ...
    'RandXReflection',1,...
    'RandYReflection',1,...
    'RandXTranslation',[-3 3], ...
    'RandYTranslation',[-3 3]);

%augimdsTrain = augmentedImageDatastore([224 224],imdsTrain,'DataAugmentation',augmenter);
%augimdsValidation = augmentedImageDatastore([224 224],imdsValidation,'DataAugmentation',augmenter);


augimdsTrain = augmentedImageDatastore([227 227],imdsTrain);
augimdsValidation = augmentedImageDatastore([227 227],imdsValidation);

options = trainingOptions('rmsprop', ...
    'MiniBatchSize',10, ...
    'MaxEpochs',20, ...
    'InitialLearnRate',1e-3, ...
    'Shuffle','every-epoch', ...
    'ValidationData',augimdsValidation, ...
    'ValidationFrequency',3, ...
    'Verbose',false, ...
    'Plots','training-progress');

net = trainNetwork(augimdsTrain,layers_1,options)
[YPred, probs] = classify(net,augimdsValidation);
accuracy = mean(YPred ==imdsValidation.Labels)
figure
cm=confusionchart (imdsValidation.Labels, YPred);

[Walter Roberson]

dicom images are mostly grayscale, but that code is expecting to train on RGB.

You can promote a grayscale image to RGB by IMG = repmat(IMG,[1 1 3])

[shivan artosh]

could you please show me which part of code would be changed ?

i have a big dataset so, is it possible to convert it in this way?

[Walter Roberson]

Change readDicomDatastoreImage to include a repmat()

[shivan artosh]

this is readDicomDatastoreImage function :

could you please tell me which part will be changed and how?

finaly greate thanks and that is all.

function I = readDicomDatastoreImage(filename)
    onState = warning('off', 'backtrace');
    c = onCleanup(@() warning(onState));
    I = dicomread(filename);

[Walter Roberson]

function I = readDicomDatastoreImage(filename)
    onState = warning('off', 'backtrace');
    c = onCleanup(@() warning(onState));
    I = dicomread(filename);
    I = repmat(I, 1, 1, 3);

[shivan artosh]

excuse me sir, i got this :

>> SHIVANCLASSIFY

net = 

  SeriesNetwork with properties:

         Layers: [25×1 nnet.cnn.layer.Layer]
     InputNames: {'data'}
    OutputNames: {'output'}

Error using trainNetwork (line 170)
Invalid training data. The output size (10) of the last layer does not match the number of classes (66).

Error in SHIVANCLASSIFY (line 36)
net = trainNetwork(augimdsTrain,layers_1,options)

[Walter Roberson]

Sorry, I do not know.

[shivan artosh]

thank you so much sir

Answer 1

[Khalid Babutain]

you cannot use pre-trained network unless you adjust it to your data

1. for alexnet, this pre-trained network takes 227x227x3 because it deals with RGB images

2. and that also applies to the first ConveNet which takes 3 channels because its kernels have 3 channels, in which you also have to update

3. you must update the last three classification layers to classify based on your classes

i also think that you are trying to resize your lung dataset to 227x227 in which you may lose some of its quality

this code should work for you, and if it's not clear i can clarify it for you

clear all; close all; clc;

imds = imageDatastore('lung dataset-Labeled', ...
    'IncludeSubfolders',true, 'LabelSource','foldernames', ... % this for labeling by folder names
    'FileExtensions','.dcm','ReadFcn',@readDicomDatastoreImage); % this a function

[imdsTrain,imdsValidation] = splitEachLabel(imds,0.7);


net = alexnet(); % analyzeNetwork(lgraph)
numClasses = numel(categories(imdsTrain.Labels)); % number of classes = number of folders
imageSize = [227 227]; % you can use here the original dataset size

global GinputSize
GinputSize = imageSize;

lgraph = layerGraph(net.Layers);

lgraph = removeLayers(lgraph, 'fc8');
lgraph = removeLayers(lgraph, 'prob');
lgraph = removeLayers(lgraph, 'output');

% create and add layers
inputLayer = imageInputLayer([imageSize 1], 'Name', net.Layers(1).Name,...
    'DataAugmentation', net.Layers(1).DataAugmentation, ...
    'Normalization', net.Layers(1).Normalization);
lgraph = replaceLayer(lgraph,net.Layers(1).Name,inputLayer);

newConv1_Weights = net.Layers(2).Weights;
newConv1_Weights = mean(newConv1_Weights(:,:,1:3,:), 3); % taking the mean of kernal channels

newConv1 = convolution2dLayer(net.Layers(2).FilterSize(1), net.Layers(2).NumFilters,...
    'Name', net.Layers(2).Name,...
    'NumChannels', inputLayer.InputSize(3),...
    'Stride', net.Layers(2).Stride,...
    'DilationFactor', net.Layers(2).DilationFactor,...
    'Padding', net.Layers(2).PaddingSize,...
    'Weights', newConv1_Weights,...BiasLearnRateFactor 
    'Bias', net.Layers(2).Bias,...
    'BiasLearnRateFactor', net.Layers(2).BiasLearnRateFactor);
lgraph = replaceLayer(lgraph,net.Layers(2).Name,newConv1);

lgraph = addLayers(lgraph, fullyConnectedLayer(numClasses,'Name', 'fc2'));
lgraph = addLayers(lgraph, softmaxLayer('Name', 'softmax'));
lgraph = addLayers(lgraph, classificationLayer('Name','output'));

lgraph = connectLayers(lgraph, 'drop7', 'fc2');
lgraph = connectLayers(lgraph, 'fc2', 'softmax');
lgraph = connectLayers(lgraph, 'softmax', 'output');
% -------------------------------------------------------------------------

augmenter = imageDataAugmenter( ...
    'RandRotation',[-20,20], ...
    'RandXReflection',1,...
    'RandYReflection',1,...
    'RandXTranslation',[-3 3], ...
    'RandYTranslation',[-3 3]);

%augimdsTrain = augmentedImageDatastore([224 224],imdsTrain,'DataAugmentation',augmenter);
%augimdsValidation = augmentedImageDatastore([224 224],imdsValidation,'DataAugmentation',augmenter);

augimdsTrain = augmentedImageDatastore(imageSize,imdsTrain);
augimdsValidation = augmentedImageDatastore(imageSize,imdsValidation);
options = trainingOptions('rmsprop', ...
    'MiniBatchSize',10, ...
    'MaxEpochs',20, ...
    'InitialLearnRate',1e-3, ...
    'Shuffle','every-epoch', ...
    'ValidationData',augimdsValidation, ...
    'ValidationFrequency',3, ...
    'Verbose',false, ...
    'Plots','training-progress');

net = trainNetwork(augimdsTrain,lgraph,options)
[YPred, probs] = classify(net,augimdsValidation);
accuracy = mean(YPred ==imdsValidation.Labels)
figure
cm=confusionchart (imdsValidation.Labels, YPred);

you also need to update the readDicomDatastoreImage function to resize every image you read with specified size:

function I = readDicomDatastoreImage(filename)

    onState = warning('off', 'backtrace');
    c = onCleanup(@() warning(onState));
    I = dicomread(filename);
    global GinputSize;
    I = imresize(I,GinputSize(1:2));

[shivan artosh]

thank you so much sir, it has worked for me.

* i am beginner in DL and Matlab, could you please clearify the code by writting a comment in it (code).

• how can we get high accuracy? may i use another network(googlenet , vgg, resenet and how?

[shivan artosh]

hello sir

could you please tell me what is that mechanism you have done form me in coding? i mean what is this model name ?

[Walter Roberson]

I doubt that there is a name for this model.

[shivan artosh]

so, what do you think of it?

i use this for classification on medical imgase dataset and obtain a good accuracy.

how can i substituted with CNN ? i want to use CNN model. could you please help me?