Multi-Channel-ResNet: An integration framework towards skin lesion analysis

This paper proposes a feasible framework towards skin lesion analysis, named Multi-Channel-ResNet. The basic idea is to assemble multiple residual neural networks (ResNets), in which the training data has been pretreated with different methods. For different situations with practical applications, we put forward two training methods. Our method performs better than a single ResNet or a simple ensemble of ResNets. The validity of the framework is verified on two data sets: dermoscopic images and skin surface photos. For dermoscopic images, we completed the third part of a public competition, called “ISIC 2017: Skin Lesion Analysis Towards Melanoma Detection”. The metric is the mean value of the area under the curve (AUC) for the melanoma and seborrheic keratosis classifications. Our framework achieves a result of 0.917 on the test set, which is 0.046 higher than a single ResNet. For skin surface photos, we collected images of four diseases, including eczema, heatrash, subitum, and varicella. The framework achieves 82.4% accuracy on the test set, which is 3% higher than the baseline. This implies that the proposed framework is applicable in practice and achieves excellent performance.