Deep Neural Network for Melanoma Classification in Dermoscopic Images

Abstract

Melanoma classification in dermoscopic images is a very challenging task on account of the low contrast of skin lesions, the various forms of melanomas, the high degree of visual similarity between melanoma and non-melanoma lesions and artifacts of dermoscopic images such as dark lighting. In this paper, we investigate pathological course of outlier lesions developing to be melanoma and try to meet the above challenges by proposing a novel neural network based on Efficient-B5. Compared with existing approaches, our deeper, wider and higher resolution network can capture far more complex and more fine-grained feature representations for melanoma classification. In order to evaluate model performance, we conduct a variety of experiments. The experimental results on a large publicly available dataset ISIC 2020 Challenge Dataset, which is generated by the International Skin Imaging Collaboration and images of it are from several primary medical sources, have demonstrated the significant performance gains of our proposed network compared with prior popular melanoma classifiers, ranking the first in melanoma classification.