Summary:
This thesis deals with a system for processing, analyzing and classifying medical images. It focuses on the diagnosis of skin cancer using machine learning techniques, specifically the algorithm used is Bag-of-Features (BoF).
Such a system can be used by the attending physician adjunctively during the diagnosis by assisting him in the examination in order to identify suspicious areas of the skin. Thus, in this way utilizing prior knowledge from already diagnosed medical images, our system is trained with the ultimate goal of increasing the effectiveness of skin cancer diagnosis and reducing diagnosis time for the benefit of the patient.
The images used to develop the specific system were drawn from the scientific database of Harvard University (https://dataverse.harvard.edu)
The theoretical framework of the present thesis is analyzed within the first three chapters where the diagnostic methods for skin lesions and the two types of dermoscopes, the non-polarized dermoscopy (NPD) and the polarized dermoscopy (PD) are analyzed. Extensive reference is made to the anatomy of the skin, the structure of the cell, the cell cycle, the difference between a healthy cell and a cancerous one, as well as seven types of skin cancer. Finally, protective measures to prevent skin cancer are also presented. After the analysis of dermatoscopies and cancerous skin lesions, Machine Learning, Artificial Neural Networks and the Bag-of-features algorithm are presented.
The following chapters present the experimental part of the thesis, in the programming environment of MATLAB and the results from the application of the algorithm using the image data from the above mentioned scientific database.
Based on the results the main conclusions are:
1. Due to the fact that the number of images was not equal in all categories, but varied widely instead, it was necessary to limit the number of images used by each group. This resulted in the training of the CNN (Convolutional Neural Networks) algorithm failing, since there were not enough images needed.
2. In contrast, the bag-of-features algorithm achieved satisfactory results, with a small number of images.
3. In order to have more available images for the Machine Learning algorithms, we propose that images need to be pre-processed without losing diagnostic information.
Keywords:
machine learning, artificial neural networks, skin cancer, dermatoscopy, cell
