Attention learning models using local Zernike moments-based normalized images and convolutional neural networks for skin lesion classification

•A DCNNs-based technique is proposed that uses hand-crafted features for classification.•The proposed method derives local Zernike moments-based normalized images.•The method follows a two-phase retraining process using pre-trained DCNNs.•The feature sets are extracted using layers of both phases of retrained DCNNs.•SVM is used for classification after dimension reduction using WPCA. Classification of dermoscopic skin lesion images is challenging due to their complexities. These complexities include intra-class variation, inter-class similarity, lesion region occupying only a small part compared to normal skin tissues, and data-imbalance. Moreover, the unavailability of large training data for good performance and generalization poses another problem. Many deep convolutional neural network (CNN)-based approaches have been proposed in the literature, but their performance is not as satisfactory as for the general classification problems. Several methods based on residual, attention, and transfer learning have been applied using single CNNs and parallel CNNs. We present two parallel CNN models to deal with the above challenges. The two models use CNN architectures, DenseNet201 and Xception, as their backbone and are pre-trained on ImageNet. We develop an effective attention mechanism through the transfer-learning process, which uses local Zernike moments (LZMs) normalized images, referred to as LZMNs. The LZMN images have been derived at various frequencies, which provide increasing details of local information, leading to better discriminative feature maps. The retrained networks are evaluated on the ISIC 2016, ISIC 2017, and ISIC 2018 Challenge datasets using regular classifiers of the baseline networks and the SVM. To further enhance the performance of the proposed models, ensemble learning is used for the regular classifier, while for the SVM classifier, fusion of features emerging from the last convolution layers of the parallel CNNs is used. Finally, the proposed models have been compared with the state-of-the-art techniques and top-ranking approaches, showing their supremacy.