A hybrid CNN architecture for skin lesion classification using deep learning

The majority of cases of skin cancer can be identified using a combination of a doctor's ocular inspection, a dermoscopy, and other diagnostic methods. There are a number of issues that make automated skin lesion detection from dermoscopic pictures challenging, including artifacts (hairs), irregularity, lesion form, and irrelevant feature extraction. As a result of issues like these, the segmentation and classification procedure are more challenging than they should be. Several skin lesion categorization approaches using deep learning based on convolution neural network (CNN) and annotated skin photos show enhanced outcomes because visual observation provides the possibility to utilize artificial intelligence to intercept various skin images. Most current deep learning approaches to skin lesion segmentation derive their predictions from an assembly of various convolutional neural networks (CNN), the aggregation of multi-scale information, or a multi-task learning framework. The key goal is to utilize as much data as possible for accurate forecasting. The skin lesion segmentation task, which is typically paired with the skin lesion classification task, has been shown to benefit from a multi-task learning framework. In this work, we present a new convolutional neural network (CNN) architecture called Densenet and residual network that makes use of contextual data. To test how well our model performed, we looked at examples from the official Human Against Machine dataset (HAM10000), a collection of images from multiple sources. In order to increase the classifier's effectiveness, we up-sampled the data and supplemented it with additional information. The experimental results show that the approaches proposed here enhance the automatic classification of skin lesions with an accuracy of 95%.