Automated Histopathological Evaluation of Tumor Images using CNNs
Abstract: Histopathology is the procedure used in medicine to optically examine microscope-images of tissue samples (biopsies) in order to study the manifestation of disease. It is used, for example, to diagnose the spread and type of cancer tumors, which have implications on the chosen treatment. Currently this type of analysis is done manually by trained professionals. It is time consuming and the diagnostic agreement between professionals varies. Due to this, there is a potential use for an automation of the process, using for example neural networks. Convolutional Neural Networks (CNNs) has become increasingly popular for image analysis within the medical field. They have proven them self to be among the best techniques. CNNs has for example been successfully used to classify different types of lung cancer tissue in microscope-images. This thesis evaluates three different CNN architectures (InceptionV3, VGG16 and compactVGG) on classification of tiles, from medical whole-slides of sliced tumor biopsies. The biopsies are from lymph node metastases, from patients with malignant melanoma (i.e. skin cancer). The data consists of 19 whole-slides, where four different tissue components, to be classified by the models, has been manually annotated by a pathologist. Furthermore, the thesis examines the effect of the chosen size of the image/tile being classified (magnification or tile size). It is concluded that InceptionV3 for a tile size of 224 give the best results. With a prediction accuracy of 89.6%, 90.0%, 97.5% (the last result did not include ambiguous tiles) on 3 different test data sets. Its performance is very similar to VGG16.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)