Stacked Cross Validation with Deep Features

Detection of malignant skin lesions is important for early and accurate diagnosis cancer. In this work, a hybrid method lesion detection from dermoscopy images proposed. The combines the feature extraction process convolutional neural networks (CNN) with an ensemble learner called stacked cross-validation (CV). features extracted by three different CNN architectures, namely, ResNet50, Xception, VGG16 are used training four baseline classifiers, which support vector machines, k-nearest neighbors, artificial networks, random forests. outputs these classifiers to train logistic regression model as meta-classifier. performance proposed compared trained individually well AdaBoost classifier, another learner. Feature Xception architecture, outperforms all other benchmark models achieving scores 0.909, 0.896, 0.886, 0.917 accuracy, F1-score, sensitivity, AUC, respectively.