Semantic segmentation of plantar soft tissue histology images using several machine learning techniques. Publication under review

• lbp folder: code to complete local binary patterns
• subroutines folder: all subroutines needed to run the main scripts
• perceptual: code to extract perceptual features
• SNIC_mex: Slightly adapted SNIC method.

Running instructions

• use main_create_texture_feature_dataset.m to extract desired features from all classifier images.
• use main_select_features.m to reduce the size of the feature set
• use main_train_classifier_reduced.m to train the classifier on the training data extracted in step 1
• use one of the main_deploy_classifier*.m to apply the trained classifier to the whole slide images using desired strategy (block or superpixel)

• UNet7Channel is the caffe prototxt file describing the network architecture. Use netscope to visualize the architecture.
• MakeDeepLearningData.m file will take in images and batch crop or augment and save resulting files for input into deep neural network
• getRangAug.m is used to randomly augment the data; function called by Make*Data.m
• StitchDigitsOutput.m and AverageOverlap_Stitch.m are used to stitch the network output back into the original input size.
• caffe installation adapted from happynear

Data can be found at UW research works. There should be a zipped file containing the following folders:

• GroundTruth contains raw images correlating to ground truth label matrices
• classifierims contains the folder of single-tissue images from which texture features were extracted.
• featureSets contains extracted feature sets used to train the classifiers
• trainedClassifiers contains the trained classifiers used for whole slide iamge segmentation
• UNet_models contains 3 checkpoints of the best version of the UNet. Checkpoint 8000 was used for final segmentation comparison

• MATLAB code was run on windows and Liunux systems (Win 7, 10; Ubuntu 12)
• Python code for caffe run on Win 7.
• All other O.S. have not been tested