Help Understanding groundTruth for CNN segmentation.

I am attempting binary xray segementation using convolutional neural networks in matlab. I have a folder of the preoprocessed images, and a folder of binary segementations which match those images.

The segmentaions are binary so they have two class outputs denoted by a 0 and 1 respectively, "Background", "Cervical_Masks".

My problem stems from using the groundTruth function in matlab.

imds=imageDatastore('Stringtofiles', 'FileExtensions', '.tif', 'IncludeSubfolders', 0, 'LabelSource', 'foldernames');
imds_Masks=imageDatastore('/Stringtofiles', 'FileExtensions', '.tif', 'IncludeSubfolders', 0, 'LabelSource', 'foldernames');
groundT=groundTruthDataSource(imds);

ldc = labelDefinitionCreator();
addLabel(ldc,'Cervical_Masks',labelType.PixelLabel);
addLabel(ldc,'Background',labelType.PixelLabel, PixeWhenlLabelID.0);
labelDefs = create(ldc);
labelData=table(imds_Masks.Files, 'VariableNames', {'PixelLabelData'});

gTruth = groundTruth(groundT,labelDefs,labelData);

[imds,pxds] = pixelLabelTrainingData(gTruth);

After this point my data is wrong.

Simply checking it for any subject reveals that the pxds file which should represent the mask is incorrect.

subject=22
I = readimage(imds,subject);
C = readimage(pxds,subject);
imshowpair(I,uint8(C),'montage')

I think this is because I am not using the labeling correct.

Could someone please help?

Thanks!

Answer 1

[Anshika Chaurasia]

Hi,

On looking the provided code, the possible reason for incorrect label could be in following line:

addLabel(ldc,'Background',labelType.PixelLabel, PixeWhenlLabelID.0);

I am not sure about "PixeWhenlLabelID.0".

It would be better if you can share the images or some test code. It would be easy for us to reproduce your issue at our end.

Refer to the following blog:

Semantic Segmentation Using Deep Learning » Deep Learning - MATLAB & Simulink (mathworks.com)

Answer 2

[yanqi liu]

sir，please check the follow code to get some information

for more information,you can read the book 《计算机视觉与深度学习实战》，Thank you！

clc; clear all; close all;
% image file
data = load('stopSignsAndCars.mat');
imageFilenames = data.stopSignsAndCars.imageFilename(1:2)

imageFilenames = 2×1 cell array
    {'stopSignImages/image001.jpg'}
    {'stopSignImages/image002.jpg'}

imageFilenames = fullfile(toolboxdir('vision'),'visiondata',imageFilenames);
dataSource = groundTruthDataSource(imageFilenames);
% label
ldc = labelDefinitionCreator();
addLabel(ldc,'stopSign',labelType.Rectangle);
addLabel(ldc,'carRear',labelType.Rectangle);
labelDefs = create(ldc)

labelDefs = 2×5 table
        Name          Type       LabelColor     Group      Description
    ____________    _________    __________    ________    ___________

    {'stopSign'}    Rectangle    {0×0 char}    {'None'}       {' '}   
    {'carRear' }    Rectangle    {0×0 char}    {'None'}       {' '}   

% add target
stopSignTruth = {[856   318    39    41];[445   523    52    54]};
carRearTruth = {[398   378   315   210];[332   633   691   287]};
% make label
labelNames = {'stopSign';'carRear'};
labelData = table(stopSignTruth,carRearTruth,'VariableNames',labelNames)

labelData = 2×2 table
        stopSign               carRear      
    _________________    ___________________

    {[856 318 39 41]}    {[398 378 315 210]}
    {[445 523 52 54]}    {[332 633 691 287]}

% display
gTruth = groundTruth(dataSource,labelDefs,labelData)

gTruth = 
  groundTruth with properties:

          DataSource: [1×1 groundTruthDataSource]
    LabelDefinitions: [2×5 table]
           LabelData: [2×2 table]

figure; imshow(gTruth.DataSource.Source{1});
hold on;
rectangle('Position', gTruth.LabelData.stopSign{1}, 'EdgeColor', 'g', 'LineWidth', 2);
rectangle('Position', gTruth.LabelData.carRear{1}, 'EdgeColor', 'c', 'LineWidth', 2);