An Ensemble Approach for Multiclass Skin Lesion Classification from Dermoscopic Images

Skin cancers, especially melanomas and carcinomas, are among the most fatal diseases affecting the integumentary system (skin). Effective diagnosis of these ailments at an early stage opens the door to proper care, effective treatment plans, and higher life expectancy. However, diagnosis of malignancy in dermatology is a time-consuming process and is subject to a lot of human errors because of striking visual resemblance among distinct categories of lesions. Therefore, computational methods are being employed nowadays to increase the accuracy of diagnostic processes. An automated AI-based detection system can help dermatologists worldwide diagnose and classify skin lesions at the earliest possible time. In this study, we introduce a technique for classifying skin lesions that is built on an ensemble of eight deep learning architectures. The classifiers VGG-16, VGG-17, ResNet50, InceptionV3, InceptionRestnetV2, Xception, DenseNet121, and MobileNetV2 were developed using the training dataset along with fine-tuning of the parameters. These models were later integrated using the stacking approach and combination rule of majority voting. Test accuracy and test loss of 85.26% and 0.4985 were attained using our suggested model. According to our findings, our ensemble model yielded superior accuracy for both overall and across individual classes, compared to existing individual learners.