Segmentation of Attributes of the Skin Lesion Using Deep Ensemble Models

In this paper, the utilization of various deep learning models is conducted in the form of ensemble learning for the detection of skin cancer. A number of different methods for segmenting and classifying skin lesions, both conventional and deep, have been discussed in the literature. The segmentation is conducted on dermoscopic images that are considered essential for the diagnosis of artefacts present in the input skin images. In order to increase the performance of the segmentation, we use ensemble-based learning from various deep learning algorithms, including various CNNs: AlexNet, ResNet, and GoogLeNet. Multiple datasets were used for this research, namely, ISBI 2016 Dataset, ISIC 2017 Dataset, ISBI 2018 Dataset, and HAM10000 Dataset, with more than 15,000 images. For the training phase, 0.001 learning rate is chosen with the size of batches as 28. The simulation is conducted using 16 GB RAM on a Matlab simulation tool and to test the efficacy of the model against various performance metrics. The final results of each classifier are passed to a majority voting ensemble that combines the predicted outcomes of the deep neural networks.