DeepPCA Based Objective Function for Melanoma Detection

In this paper, we propose an objective function for the convolutional neural network to acquire the variation separability as opposed to the categorical cross entropy which maximizes according to the target labels. This approach is an unsupervised learning method which tends to separate the classes according to their variation in the subspace. Finally, the features extracted are classified using support vector machines (SVM). The deep representative features from the CNN are directly from the data, and thus additionally increase the variance between the images making it more discriminative. The idea is to build a CNN (Convolutional Neural Network) and perform the principal component analysis on top of this while training it in an end-to-end fashion. The backpropagation is done to update the parameters according to the eigen representation of the training data. Experimental results on the widely used MEDNODE database which consists of clinical (non-dermoscopic) images shows that our approach is efficient for the classification of melanoma skin cancer detection task.