A comparative study of deep learning architectures on melanoma detection. (June 2019)

• Record Type: Journal Article
• Authors: Hosseinzadeh Kassani, Sara; Hosseinzadeh Kassani, Peyman
• Is Part Of: Tissue & cell. Volume 58(2019)
• Page Start: 76
• Page End: 83
• Publication Date: 2019-06
• Subjects: Cancer classification -- Computational diagnosis -- Convolutional neural networks -- Deep learning -- Melanoma detection; Cytology -- Periodicals; 571.5
• DOI: 10.1016/j.tice.2019.04.009
• Languages: English
• ISSNs: 0040-8166

Abstract

Highlights: The performance of five widely-used deep learning architectures are compared on skin cancer classification. Different data augmentation techniques is applied to avoid the negative effect of class imbalances. The dataset used for this research is provided by International Skin Imaging Collaboration (ISIC2018). ResNet50 with data augmentation steps achieves best performance. To the best of our knowledge, this is the first comparative study on deep learning algorithms for melanoma detection. Abstract: Melanoma is the most aggressive type of skin cancer, which significantly reduces the life expectancy. Early detection of melanoma can reduce the morbidity and mortality associated with skin cancer. Dermoscopic images acquired by dermoscopic instruments are used in computational analysis for skin cancer detection. However, some image quality limitations such as noises, shadows, artefacts exist that could compromise the robustness of the skin image analysis. Hence, developing an automatic intelligent system for skin cancer diagnosis with accurate detection rate is crucial. In this paper, we evaluate the performance of several state-of-the-art convolutional neural networks in dermoscopic images of skin lesions. Our experiment is conducted on a graphics processing unit (GPU) to speed up the training and deployment process. To enhance the quality of images, we employ different pre-processing steps. We also apply data augmentation methodology such as horizontal and vertical flipping techniques to address the class skewness problem. Both pre-processing and data augmentation could help to improve the final accuracy.