DermaDL: Advanced Convolutional Neural Networks for Computer-Aided Skin-Lesion Classification

• Published in: SN computer science Vol. 2; no. 4; p. 253
• Main Authors: Lima, Daniel M., Rodrigues-Jr, Jose F., Brandoli, Bruno, Goeuriot, Lorraine, Amer-Yahia, Sihem
• Format: Journal Article
• Language: English
• Published: Singapore Springer Singapore 01.07.2021; Springer Nature B.V; Springer
• Subjects: AI and Deep Learning Trends in Healthcare; Artificial neural networks; Carcinogens; Classification; Computer Imaging; Computer Science; Computer Systems Organization and Communication Networks; Data Structures and Information Theory; Datasets; Dermatology; Design; Information Retrieval; Information Systems and Communication Service; Lesions; Neural networks; Original Research; Pattern Recognition and Graphics; Skin; Skin cancer; Software Engineering/Programming and Operating Systems; Vision; Deep learning; Skin lesion classification; Melanoma detection; Dermatology; Skin cancer
• ISSN: 2662-995X; 2661-8907
• DOI: 10.1007/s42979-021-00641-5

Early identification of the type of skin lesion, some of them carcinogenic, is of paramount importance. Currently, the use of Convolutional Neural Networks (CNNs) is the mainline of investigation for the automated analysis of such lesions. Most of the existing works, however, were designed by transfer learning general-purpose CNN architectures, adapting existing methods to the domain of dermatology. Despite effective, this approach poses inflexibility and high processing costs. In this work, we introduce a novel architecture that benefits from cutting-edge CNN techniques Aggregated Transformations combined to the mechanism of Squeeze-and-Excite organized in a residual block; our architecture is designed and trained from scratch to solve both the binary melanoma detection problem, as well as the multi-class skin-lesion classification problem. Our results demonstrate that such an architecture is competitive to major state-of-the-art architectures adapted to the domain of skin-lesion diagnosis. Our architecture is prone to evolve and to provide low processing cost for real-world in situ applications using a much smaller number of weights if compared to previous works.