A Deep Learning Model for the Segmentation and Classification of Melanoma Skin Lesions
Presenter: Tony Mogoa, SCES
Abstract
Melanoma is the most fatal of skin cancers as it accounts for most skin cancer deaths. However, If melanoma is detected early, it is highly curable. Its diagnosis involves a non-invasive examination of skin lesions and their subjection to biopsy. The non-invasive examination can be challenging even for dermatologists and hence require a lot of experience. Machine learning models, trained on dermoscopic and clinical images, have been applied in the classification of skin lesions to address the shortage of dermatologists and increase the chances of early diagnosis. Though dermoscopy increases accuracy in diagnosis, it is more expensive and needs more technical expertise while clinical images, which are taken using user-grade cameras, are cheaper and more accessible and hence have a greater potential of making Melanoma diagnosis more accessible and support self-examination to increase the chances of early diagnosis. Existing models for classifying melanoma lesions from clinical images have Questionable generalizability and have been trained on very small datasets.
This research used experimental methodology to develop a deep-learning model for lesion segmentation and classification for increased generalizability through a comparative study of deep-learning architectures involving training and testing across five non-dermoscopic datasets.