Improving Skin Disease Classification using Residual Attention Network

The most substantial organ of the human body is the skin. It plays an essential role in the sustenance of life and health. It helps in providing an airtight, watertight and flexible barrier between the internal body organs and the adverse elements from outside environment. Skin conditions contribute 1.79% of the global burden of disease worldwide. Development in techniques to visually inspect a skin disease is essential to fasten diagnosis and minimise life-threatening situations. Automated classification of skin disorders via image processing and various machine learning algorithms have been proposed in the literature. Previous research has demonstrated that Convolutional Neural Networks (CNNs) have great ability to recognise specific regions in images without providing the annotated bounding boxes of those specific regions. Hence, we plan to compare a custom CNN model along with the Residual Attention Network model and a custom CNN model based on ResNet without any attention layers for skin classification problems. The attention layer would improve the localisation ability of a CNN model and consider only the relevant regions from the images. Moreover, the residual network works better for small sample learning problems. So, a combination of residual and attention units is suitable to tackle the concerned problems.