Random Deep Neural Network for Melanoma Recognition

The paper presents a new method for building the CNN architecture with improved generalization capabilities. The network is formed from two deep structures differing by the activation functions. The dense layer signals are also separated into two branches and the verdicts of both are fused randomly into final decision using the classical cross-entropy loss formulation and softmax approach. In the learning process, the parameters of the locally connected CNN layers are adapted based on random combinations of the signals of these two parallel CNN subnetworks. The randomness introduced in all stages of the network structure fulfills the role of automatic implicit regularization at a very small number of learning data. The positive role of such regularization has been confirmed in numerical experiments of image recognition representing melanoma or non-melanoma cases. The results obtained by using the proposed method showed a significant improvement in numerical quality measures, including accuracy, sensitivity, F1 measure, and the area under the ROC curve.