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

Abstract

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.

Popular Abstract

This master thesis investigates the possibility of using machine learning to classify non-pigmented skin cancer lesions on patients in clinics. The most common non-pigmented types of skin cancer are Basal Cell Carcinoma, Squamous Cell Carcinoma, the last one possibly starting from a lesion called Actinic Keratosis. In Sweden, the number of cases is estimated to grow in the next 20 years.

To contain and possibly reverse these trends, more primary and secondary prevention intiatives are required. The Sk ̊ane University Hospital in Lund in collaboration with Lund University has started a project to create an algorithm capable of classifying between the three aforementioned types of lesions.

To find a solution using machine learning, a present problem is the lack of sufficient data which in this project was solved by studying the different sources of data available online that could be used. The data sets chosen were based on the availability of the lesions studied in the data set and the proximity of the ethnicity of the skin to the fair skin of Swedes. Furthermore, during 3 weeks more data was collected at Sk ̊ane University Hospital in Lund, and this was used for further network testing.

Multiple networks were trained with the different data sets and compared in terms of performance. With the best network of each data set, we tested the versatility of it to be used in a different group of images not used during training. From this study, we obtain poor versatility results that could be explained by the fact that the images from one data set to another varied in terms of which device was used to capture the image and the different skin types. Additionally, we studied the parts of the image in which the network was focussing when making a prediction, to give more feedback to the dermatologist. It was found that the model does not look directly at the lesion but rather looks at the surrounding skin.

To conclude, the results of the networks obtained are promising and give hope for a future solution with good prediction capabilities that could be deployed in the clinics to help dermatologists reach a diagnosis.

• author: Aguilera Manzanera, Maria del Pilar
• supervisor: Ida Arvidsson; Anna Gummeson
• organization: Mathematics (Faculty of Engineering)
• course: FMAM05 20221
• year: 2022
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Melanoma, Skin cancer, Dermatoscopy, Image classification, Machine learning, Artificial intelligence, Convolutional neural networks, Dermatology, Squamous cell carcinoma, Basal cell carcinoma, Actinic keratosis, Computer-aided Diagnostics, Digital dermatology
• publication/series: Master’s Theses in Mathematical Sciences
• report number: LUFTMA-3467-2022
• ISSN: 1404-6342
• other publication id: 2022:E17
• language: English
• id: 9087860
• date added to LUP: 2022-06-13 12:32:44
• date last changed: 2022-06-13 12:32:44