Improving Skin Lesion Diagnosis Employing Attention-Based LSTM Networks

Diagnosing skin lesions is, therefore, an exercise full of challenges even for the most seasoned dermatologists, more so when several factors are present. Examples of features about this condition include the minimal variation between healthy and tumor skin and the high level of similarity among different types of skin lesions, sometimes coupled with blurred edges of the lesion. The merits of this involve the help that it accrues to a physician trying to achieve accuracy at very early stages in malignant skin lesion detection, with an automated computer-aided detection system trained on a dataset consisting of images. Indeed, because of recent developments in deep learning, it is at this moment that new techniques, such as dilated convolution, have recently been introduced. Concerning conventional ConvNets, these claim better accuracy with identical, similar levels of computational complexity. We are introducing a method of classification of skin lesions based on an attention-based long short-term memory network in this paper. The proposed methodology was inspired from the source of attention mechanisms in people's visual systems. These mechanisms concentrate on the essential parts of an image and filter away auditory information, which is not eventually indispensable to an image from what the human 'visual' system has to attend to. When compared to the existing approaches proposed technique achieves with an accuracy of 96.93%.