Aleksandra Dzieniszewska
supervisor: Ryszard Piramidowicz
Melanoma is one of the deadliest kinds of skin cancer, but it might be completely cured if detected early. Therefore, monitoring and early diagnosis of skin lesions have a crucial meaning in preventing cancer diseases. The major problem and the major challenge, however, is patients’ resistance to being diagnosed towards cancer and also limited access to specialists.
The automation of the diagnosis process might to some extent help in solving these problems by providing easy access to mass and sufficiently reliable testing methods, which patients might have access to in a family doctor’s office, or at home.
This work is focused on developing a new solution for a comprehensive automated skin analysis system allowing the classification of changes based on multimodal image data and deep learning models.
In particular, the application of neural networks to enhance the recognition of human skin lesions based on dermatoscopy images with additional polarization channel was investigated. The main research problem identified is to find a method of using available (open) RGB image data sets for the initial training of neural network models to compensate for a small amount of polarization data. To solve it, two main approaches were tested – early and late fusion of information. The architectures were tested on RGB data to find the most promising candidates for transfer learning with polarization images. A multi-path neural network was created to combine results from different images pro