run.m

%-------------------------------------------------------------------------%
%-------------------------------------------------------------------------%
%-------------------------------------------------------------------------%
% ------ Two-Tier Tissue Decomposition for Histopathological Image ------ %
% ------------------ Representation and Classification ------------------ %
%-------------------------------------------------------------------------%
% 
% In this work, aim is to design a classification system for histopatholo-
% gical images. Towards this end, we present a new model for effective 
% representation of these images that will be used by the classification 
% system. The contributions of this model are twofold. First, it introduces 
% a new two-tier tissue decomposition method for defining a set of multi-
% typed objects in an image. Different than the previous studies, these 
% objects are defined combining texture, shape, and size information and 
% they may correspond to individual histological tissue components as well 
% as local tissue subregions of different characteristics. As its second 
% contribution, it defines a new metric, which we call dominant blob scale, 
% to characterize the shape and size of an object with a single scalar 
% value. Our experiments on colon tissue images reveal that this new object 
% definition and characterization provides distinguishing representation of 
% normal and cancerous histopathological images, which is effective to 
% obtain more accurate classification results compared to its counterparts.
% 
% NOTE: The following source codes are provided for research purposes only.
% The authors have no responsibility for any consequences of use of these 
% source codes. If you use any part of the codes, please cite the following 
% paper.
% 
% T. Gultekin, C. Koyuncu, C. Sokmensuer, and C. Gunduz-Demir, "Two-tier 
% tissue decomposition for histopathological image representation and 
% classification," IEEE Trans. Med. Imag., vol.34, no.1, pp.275–283, Jan. 
% 2015.
% 
% Model parameters to be adjusted ():
% K        : cluster number
% sizeThr  : area threshold
% cPercent : covered pixel percentage
% edgeThr  : edge threshold
% C        : SVM optimization parameter
% 
% Before run this program, create two txt files containing filenames of 
% images together with their class labels, for the training and test sets. 
% After that, assign the name of the created filenames to the variables 
% trainFilename and testFilename in lines 59 and 60.
% 
% To improve the efficiency, the "for" loop in line 5 of "src/getDataset.m"
% can be executed in parallel using "parfor".
% 
% Each line should have the following format:
% [image_file_name_with_its_path] [class_label]
% 
% For further questions feel free to email me at canfkoyuncu@gmail.com
%-------------------------------------------------------------------------%
%-------------------------------------------------------------------------%
%-------------------------------------------------------------------------%
K           = 6;
cPercent    = 0.05;
sizeThr     = 40;
edgeThr     = 20;
C           = 0.01;

trainFilename = 'train_imgs.txt';
testFilename  = 'test_imgs.txt';

if exist(trainFilename, 'file') ~= 2
    error(['Couldn''t locate the file : ' trainFilename]);
end
if exist(testFilename, 'file') ~= 2
    error(['Couldn''t locate the file : ' testFilename]);
end 

addpath(genpath('libsvm_3.22/'));
addpath('src/');
[trainImgFilenames, trainLabels] = processFile(trainFilename);
[testImgFilenames, testLabels]   = processFile(testFilename);

uTriIndices = triu(true(K*3, K*3));
textFilters = makeSfilters;
diskFilters = getDiskFilters;
%% creatining texton vocabularies
vocabulary = createTextonVocabulary(trainImgFilenames, trainLabels, K, textFilters);
save('__vocabulary.mat', 'vocabulary');
%% obtaining datasets
load __vocabulary; 
train   = getDataset(trainImgFilenames, vocabulary, sizeThr, K, cPercent, uTriIndices, textFilters, diskFilters);
test    = getDataset(testImgFilenames, vocabulary, sizeThr, K, cPercent, uTriIndices, textFilters, diskFilters);
save('__rawData.mat', 'train', 'trainLabels', 'test', 'testLabels');
%% eliminating less frequently occured edges
load '__rawData.mat';
elimColIds = max(train,[],1) <= edgeThr;
train(:, elimColIds) = [];
test(:, elimColIds)  = [];
save('__elimData.mat', 'train', 'trainLabels', 'test', 'testLabels', 'elimColIds');
%% normalizing datasets
load '__elimData';
trMean = mean(train);
trStd  = std(train);
train  = (train - repmat(trMean, size(train,1), 1))./ repmat(trStd, size(train,1), 1);
test   = (test - repmat(trMean, size(test,1), 1)) ./ repmat(trStd, size(test,1), 1);
save('__normData.mat', 'train', 'trainLabels', 'test', 'testLabels', 'trMean', 'trStd');
%% classifying test data
load '__normData';
[train, trainLabels] = balanceSet(train, trainLabels);
svmModel = svmtrain(trainLabels, train, [' -t 0 -c ' num2str(C) ' -b 1']);
[predicted_label, accuracy, prob_estimates] = svmpredict(testLabels, test, svmModel,' -b 1');