DermaDL: Advanced Convolutional Neural Networks for Automated Melanoma Detection

Jose Rodrigues (1) , Bruno Brandoli (2) , Sihem Amer-Yahia (3)

1 USP - Universidade de São Paulo = University of São Paulo
2 Dalhousie University [Halifax]
3 LIG - Laboratoire d'Informatique de Grenoble

In this paper we use state-of-the-art deep convolu-tional neural networks for computer-aided melanoma detection. As a result, we present the DermaDL mobile application, where dermatologists can use neural network modules for automated lesion analysis with the aim of identifying and classifying skin lesions with regard to malignancy. The proposed methodology includes a preprocessing step for data organization, normaliza-tion, augmentation and image segmentation; after that, we will employ transfer learning from state-of-the-art models previously elaborated for the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), using extensively trained Inception-type neural network models. Finally, the models will be optimized for mobile processors, allowing for mobility and convenient use. This method can classify several types of skin lesions present in the International Skin Imaging Collaboration (ISIC) archive with at least 90% accuracy, purposefully documenting and triaging clinical cases before further thorough examination.

Keywords: Index Terms-deep learning; dermatology; dermoscopy; early detection; melanoma; mobile; skin cancer

Domains: Computer Science [cs]

Jose Rodrigues, Bruno Brandoli, Sihem Amer-Yahia. DermaDL: Advanced Convolutional Neural Networks for Automated Melanoma Detection. 2020 IEEE 33rd International Symposium on Computer-Based Medical Systems (CBMS), Jul 2020, Rochester, France. pp.504-509, ⟨10.1109/CBMS49503.2020.00101⟩. ⟨hal-02972535⟩