Design, implementation and evaluation of a deep learning prototype to classify non-pigmented malignant skin cancer from dermatoscopic images

Sammanfattning: The current trends for most fair-skinned populations are that the incidence of melanoma and non-pigmented skin lesions is growing, and this growing trend will continue for the upcoming years. The emergence of deep learning networks and their promising results in solving real-world healthcare problems and improving diagnostic accuracy opens new possibilities. This thesis consists of the creation of a preliminary deep learning network to classify non-pigmented skin lesions: Basal cell carcinoma, actinic keratosis, and squamous cell carcinoma. This network could be used to provide feedback to the dermatologist regarding the diagnosis of a lesion at Sk ̊anes University Hospital in Lund. We started studying publicly available data sets that could be used to reach our goal. Once we had the data sets that would be used, we proceeded to train the different networks. The networks were trained using transfer learning technology, in which we used existing pre-trained model architectures to train our model. The project was developed in Python using the Keras library that runs under Tensorflow. The results for each of the experiments were compared in terms of performance, and those that obtained the best results were selected. Additionally, we studied the versatility of the models to be used in other data sets that differed from the one used for training, and compared them in terms of accuracy and bias towards certain classes. Finally, the Grad-CAM algorithm was implemented to visualise the hot spot areas on which the model based its predictions for each of the lesions. The final conclusions of the project show promising results that open the possibility of a future real-world implementation of using a deep learning network in a clinic.