logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-06282021-105720


Tipo di tesi
Tesi di laurea magistrale
Autore
LONOCE, FABIANA
URN
etd-06282021-105720
Titolo
Discrimination between Nevus and Melanoma Skin Lesion using Siamese Deep Neural Networks and Dermatological Features
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
INGEGNERIA BIOMEDICA
Relatori
relatore Prof. Ciuti, Gastone
Parole chiave
  • Barlow twins
  • dermatological features
  • lesions similarity
  • melanoma detection
  • Siamese neural networks
Data inizio appello
16/07/2021
Consultabilità
Non consultabile
Data di rilascio
16/07/2091
Riassunto
Malignant melanoma (MM) and non-melanoma skin cancer (NMSC) are the two main skin tumor. The NMSC incidence is higher than MM but the mortality rate is more worrying in the second one, since the 5-year survival rate decreases significantly as the melanoma stage increases. Therefore, frequent self-screening and screening by the dermatologist is advisable. Once several features are observed, the dermatologist decides whether to perform biopsy and then histopathological examination. To assist the clinician in skin cancer detection, convolutional neural networks (CNN) have been implemented. Their use is still very limited due to a skepticism concerning the black-box nature of the neural networks. However, several studies have shown CNN are able to outperform the average of dermatologists in lesions classification tasks. The aim of this work was twofold: dermatological features were extracted from melanomas and nevi to provide objective information and overcome semi-quantitative evaluations. To assess the rightness of the implemented model, probability densities of the extracted parameters were generated. These show high consistency with the ABCD criteria. Further investigations will be conducted to evaluate the results generalizability. Secondly, a Siamese deep neural network was implemented to discriminate between melanomas and nevi. This work is part of a larger industrial project, where the network was created to support an ensemble of multiclass classification models showing suboptimal performance in discriminating these two types of lesions, due to their similarity. Limitations related to the large amount of labelled data needed to train it, the exigency of classifier and the possibility to exploit also non-labelled data from all the devices developed within the project led to the implementation of a new network considered as Siamese evolution, called Barlow twins, which showed great potential. For the best of my knowledge, this is the first time these networks have been used to do this task.
File