Distinguish Melanoma from Benign Melanocytic and Non-Melanocytic Lesions

Mónica Andreia Teixeira do Amaral

Informações chave

Autores:

Mónica Andreia Teixeira do Amaral (Mónica Andreia Teixeira do Amaral)

Orientadores:

Ana Catarina Fidalgo Barata (Ana Catarina Fidalgo Barata); Jorge Dos Santos Salvador Marques (Jorge Dos Santos Salvador Marques)

Publicado em

28 de novembro de 2019

Resumo

Melanoma is a very aggressive type of cancer. Consequently, the detection of melanoma at early stages is crucial to prevent it from metastasizing, increasing survival rate. However, distinguishing between melanoma and other types of lesions, especially non-melanocytic ones, remains a difficult task. Dermoscopy is a procedure used by physicians for skin cancer diagnosis, that has shown to be more accurate than naked-eye analysis, since it can magnify the size of the inspected lesions by up to 100x. Several medical procedures are used to differentiate lesions, mostly for melanoma diagnosis, such as ABCD rule or Menzies scoring method. Several Computer-Aided Diagnosis systems have been developed in the interest of helping physicians better detect melanomas. Nonetheless, most of CAD systems aren't able to distinguish between melanocytic and non-melanocytic lesions, which is still an open problem and poorly addressed in literature. This thesis proposes a deep neural network to automatically discriminate melanoma from melanocytic (benign nevi) and non-melanocytic (seborrheic keratosis) lesions. To achieve this goal several regularization techniques were investigated in order to improve the performance of the developed model. The best evaluated configuration achieved a Balanced Accuracy (BACC) of 59,9% for the validation dataset and 53,1% for the test dataset. Furthermore, the best two models at correctly detecting melanomas, reached a same Sensitivity (SE) of 56,7% in validation dataset and SE=32,5% and SE=49,6% for test dataset. The dataset used in this thesis was collected from the 2017 ISIC Challenge database.