Developed Newton-Raphson based deep features selection framework for skin lesion recognition

Melanoma is the fatal form of skin cancer; however, its diagnosis at the primary stages significantly reduces the mortality rate. These days, the increasing numbers of skin cancer patients have boosted the requirement for a care decision support system - capable of detecting the lesions with high accuracy. In this work, a method is proposed for skin cancer localization and recognition by implementing a novel combination of a deep learning model and iteration-controlled Newton-Raphson (IcNR) based feature selection method. The proposed framework follows three primary steps - lesion localization through faster region based convolutional neural network (RCNN), deep feature extraction, and feature selection by IcNR approach. In the localization step, a new contrast stretching approach based on bee colony method (ABC) is being followed. The enhanced images along with their ground truths are later plugged into Fast-RCNN to get segmented images. A pre-trained model, DenseNet201, is utilized to extract deep features via transfer learning, which are later subjected to selection step using proposed IcNR approach. The selected most discriminant features are finally utilized for classification using multilayered feed forward neural networks. Tests are performed on ISBI2016 and ISBI2017 datasets to achieving an accuracy of 94.5% and 93.4%, respectively. Simulation results reveal that the proposed technique outperforms existing methods with greater accuracy, and time.