Body map based wound image classification using deep learning

Abstract

Identifying different types of ulcer and surgical wounds based on their distinct features is a complex task in medical imaging. This involves the classification of ulcer and surgical wound into various labels such as diabetic ulcer, pressure ulcer, venous ulcer and surgical wounds. In order to make this process more efficient and cost-effective, there has been different study in this field. A body map based VGG 16 network is used to implement transfer learning onto two trainable dense layers for classification of wound images into five labels. The five labels include the aforementioned four types of wound and another label "Not a wound" which does not contain any wound image. The study is started with AZHMT dataset containing 4790 images. These images are classified using pre-trained inceptionV3 and VGG 16 network separately. The performance of VGG 16 was found to be better than inceptionV3 by almost 4% which was the reason for selecting VGG 16 for further study in this dataset. Also, inceptionV3 is longer and wider than VGG 16 which will learn unnecessary features from images using higher computing resources. The main aim of this thesis is to show that performance can be increased without learning unnecessary features, using fewer computing resources. and by using body map function.