Segmentation and classification of medical wound images with deep learning approaches

In the field of medicine, although doctors have high skills in the specialty they work in, different decisions and treatment recommendations can be made for the same disease at different times. This may lead to differences among observers in decision-making processes. In order to reduce and gradually eliminate such subjectivity, efforts to develop diagnostic systems based on quantitative criteria that can help doctors all over the world are increasing day by day. In this thesis, different methods have been proposed to provide segmentation and classification of medical wound images with deep learning convolutional neural network methods.The scope of the thesis work is as follows: (i) Segmentation of wound images from the existing image by semantic segmentation methods, (ii) Classification of the segmented images with popular deep learning architectures in the literature, (iii) Parameter optimization for the architecture that gives the best results in the classification phase, (iv) Designing new convolutional neural network models based on the most successful method. First, 20 different encoder-decoder-based methods are discussed for segmenting medical wound images. The study discusses how deep learning performs in this field by examining the effects of these approaches on pixel-level (pixel-based) classification when used together with different featureextracting CNN architectures in the segmentation of wound images. Thus, the study also examines the segmentation success of the base (backbone) model. In the segmentation process, in addition to the pre-trained deep learning architectures, a 5-layer vanilla CNN network was designed and used as the base model in the segmentation. To our knowledge of the literature, this is the first attempt to segment medical wound images with encoder-decoder semantic segmentation methods undertaken within the scope of this thesis. Secondly, classification performances were investigated to decide whether the wound images belonged to granule, necrotic, and slough classes. For this, it is aimed to measure the success of 19 CNN architectures with constantly initialized parameters on the data set. The next stage discusses the effect of parameter values used in CNN architectures on the results. At this stage, finding the epoch number, batch size, and learning rate parameters that find the best success in the classification of wound images are also discussed. Thirdly, AlexNet architecture, which is the most successful method among 19 CNN architectures, is discussed. The effect of training parameters used in the classification of pressure and diabetic foot wound images on success was examined and investigated which parameter/s were effective in finding the optimum results. In the parameter optimization experiments, when the changes made to the evaluation metrics were examined, it was observed that the learning rate parameter produced the optimum values at the value of 1e-4. Finally, by using the parameters with which optimum results are obtained, the AlexNet architecture, in which the best results are obtained, has been modified, different versions have been proposed and the classification successes have been examined. The performances of the models with different number of convolution layers were measured and their performances in defining the wound images were compared. In addition, comparative experiments were carried out by using the SVM classifier instead of the Softmax Classifier in the classification layer of the AlexNet architecture. A unique data set consisting of 2100 images of wound images was created to be used in the training and testing stages of these models. The success of the MobileNet-UNet model, which achieved the highest accuracy among the encoder-decoder-based models designed for segmentation, was achieved at 99.67%. The success of the AlexNet architecture, which is the CNN model with the highest accuracy for the classification process, has been achieved at 95.83%. In addition, with the development of the AlexNet architecture, the classification accuracy was increased to 98.85% with the 6Conv_SVM model, which has six convolution layers, among the proposed models. In summary, in this thesis, new methods have been proposed for the segmentation and classification of medical wound images with deep learning CNN methods with the aim of designing new methods for automatic decision-making on medical wound images.