An Integrated Deep Ensemble Approach for Skin Lesion Segmentation and Classification

Abstract

Skin cancer has emerged as one of the fastest-growing diseases across the globe, and hence, timely and reliable identification is crucial for increasing survival rates. In the past few years, deep neural network automated analysis has been found to be beneficial in aiding doctors in diagnosing skin cancer by providing accurate segmentation and classification results. The current study presents a new way to examine skin lesions. It uses several neural network models to increase accuracy using the ISIC 2018, and HAM10000 dermoscopic image datasets. The ROIs in dermoscopy images are identified by the semantic segmentation models such as U-Net, SegNet, and DeepLabV3 in the current research work. Following identification of the ROI, neural network models based on the use of the convolution layer, such as VGG16, ResNet50, and InceptionV3, are applied for lesion classification. The use of many models as an ensemble instead of only one model is useful in making the system work consistently using the strengths of the models employed. The performance was evaluated with standard metrics for both segmentation and classification, including Dice coefficient, Intersection over Union (IoU), accuracy, precision, recall, F1-score, and AUC. From experiments, it was seen that using multiple models together works better than using a single model. It was seen that U-Net and DeepLabV3 performed better in segmenting the image than SegNet, as they were able to understand the image details and the lesion area at various scales. The proposed ensemble framework achieved a Dice coefficient of 0.93, an IoU of 0.87, and a segmentation accuracy of 0.97, while the classification ensemble achieved an AUC of 0.99. The overall system worked well for the automatic skin lesion analysis task. It can be used to create a computer tool that helps doctors identify skin cancer early.