An aseptic approach towards skin lesion localization and grading using deep learning and harris hawks optimization

Abstract

AbstractSkin cancer is the most common form of cancer. It is predicted that the total number of cases of cancer will double in the next fifty years. It is an expensive procedure to discover skin cancer types in the early stages. Additionally, the survival rate reduces as cancer progresses. The current study proposes an aseptic approach toward skin lesion detection, classification, and segmentation using deep learning and Harris Hawks Optimization Algorithm (HHO). The current study utilizes the manual and automatic segmentation approaches. The manual segmentation is used when the dataset has no masks to use while the automatic segmentation approach is used, using U-Net models, to build an adaptive segmentation model. Additionally, the meta-heuristic HHO optimizer is utilized to achieve the optimization of the hyperparameters of 5 pre-trained CNN models, namely VGG16, VGG19, DenseNet169, DenseNet201, and MobileNet. Two datasets are used, namely "Melanoma Skin Cancer Dataset of 10000 Images" and "Skin Cancer ISIC" dataset from two publicly available sources for variety purpose. For the segmentation, the best-reported scores are 0.15908, 91.95%, 0.08864, 0.04313, 0.02072, 0.20767 in terms of loss, accuracy, Mean Absolute Error, Mean Squared Error, Mean Squared Logarithmic Error, and Root Mean Squared Error, respectively. For the "Melanoma Skin Cancer Dataset of 10000 Images" dataset, from the applied experiments, the best reported scores are 97.08%, 98.50%, 95.38%, 98.65%, 96.92% in terms of overall accuracy, precision, sensitivity, specificity, and F1-score, respectively by the DenseNet169 pre-trained model. For the "Skin Cancer ISIC" dataset, the best reported scores are 96.06%, 83.05%, 81.05%, 97.93%, 82.03% in terms of overall accuracy, precision, sensitivity, specificity, and F1-score, respectively by the MobileNet pre-trained model. After computing the results, the suggested approach is compared with 9 related studies. The results of comparison proves the efficiency of the proposed framework.