Deep Learning Based Encoder-Decoder Model for Skin Cancer Classification

In the medical field, identifying skin cancerous areas and characteristics in dermoscopy images enhances automatic skin cancer classification and diagnosis. However, due to an absence of training data, imbalanced classes, variances within classes, and a failure to concentrate on semantically significant lesion parts, accurately classifying MCS cancer is still challenging. To address these problems, the proposed novel UNet with MobileNet model has been utilized to divide skin cancer into seven different classes. The MobileNet approach is used as a classifier instead of a UNet decoder to avoid overfitting. Additionally, the expert learning framework based on specific domain functions is implemented by the classification model. The performance of test findings with a UNet-MobileNet across seven different Human Machine classes using $\mathbf{1 0, 0 0 0}$ training images (HAM10000) dataset. The overall performance of the HAM10000 dataset’s seven various classes is determined by UNet-MobileNet experiments. The accuracy, Dice, and recall evaluation matrices have been used to assess the proposed scheme. The proposed approach outperforms existing techniques on the skin lesion dataset.