Generative Adversarial Networks for Peripheral Cell Blood Images Standardization

Bachelor's Thesis in Biomedical Engineering (Politechnic University of Catalonia)

Abstract

Our multidisciplinary research group has been developing models based on Convolutional Neural Networks for the automatic morphological recognition of normal, reactive and abnormal blood cells circulating in blood, lymphoma and acute leukaemia. The problem addressed in this study is related to the effect of different staining protocols used to prepare the blood smear from which microscope images are analyzed. Specifically, the automatic blood cell classification with models trained using an image database collected in a single hospital may be affected when images are coming from another hospital. Thus resulting in a certain number of misclassified images. Such accuracy decrease is because some color, staining and texture characteristics may suffer variations in the images depending on the staining protocol. Generative Adversarial Networks (GANs) were leveraged to transform the images of the X domain (Hospital Germans Trias i Pujol) into the Y domain (Hospital Clínic). Characteristics of color, staining and texture of the latter were emulated. After this transformation, the classification accuracy of the images taken at Hospital Germans Trias i Pujol improved. CycleGAN and ColorizationGAN architectures were used for the transformations. ResNet 18, ResNet 34, ResNet 18 with Focal Loss and ResNet 34 with Focal Loss were used for the classifications. The classification accuracy improved from 34% to 81%.

Structure of the Project

└───Notebooks_Results_TFG
    ├───.ipynb_checkpoints
    ├───Experiments_3_Classes
    │   ├───CNN_CanRuti_Fake_Colorization
    │   ├───CNN_CanRuti_Original 
    │   ├───CNN_Clinic_Fine_Tuning
    │   ├───CNN_No_Separate_Dataset_CycleGAN_Classification
    │   │   ├───CNN_CanRuti_Fake_Cyc_15_Epoch 
    │   │   ├───CNN_CanRuti_Fake_Cyc_200_Epoch
    │   │   ├───CNN_CanRuti_Fake_Cyc_20_Epoch  
    │   │   └───Experiments_Bad_Epochs
    │   │       ├───CNN_CanRuti_Fake_Cyc_10_Epoch 
    │   │       └───CNN_CanRuti_Fake_Cyc_25_Epoch      
    │   └───CNN_Separate_Dataset_CycleGAN_Classification
    │       ├───CNN_CanRuti_Fake_Cyc_Separate_200_Epoch
    │       │  
    │       └───CNN_CanRuti_Fake_Cyc_Separate_Optim_Epoch   
    └───GAN_Transformation_NBS
        ├───Colorization_NBS
        ├───CycleGAN_Transfo_CanRuti
        └───CycleGAN_Transfo_Clinic