Dermatoscopy Using Multi-Layer Perceptron, Convolution Neural Network, and Capsule Network to Differentiate Malignant Melanoma From Benign Nevus

2. Literature Review

Computer-assisted melanoma detection from dermoscopy images is more efficient than manual approach though it a very challenging task. A key challenge here is the enormous intra-class similarities of melanomas and nevus. Melanoma and nevus share the visual similarities in terms of surface texture, size, color, shape, and position in the dermatoscopy images. Conversely, the comparatively poor contrasts and vague boundaries among skin lesions and usual skin sections may further challenge the computer-assisted identification job. Sometimes, the existence of image degradation factors, such as the typical veins and hair or sometimes artificially imposed images due to noise or blur degrades the quality of skin lesions, thereby making the identification task harsher (Yu et al., 2017).

To date, researchers have instituted clever ways to resolve this challenging task. Alcon et al. (2009) have proposed a decision support system for melanoma classification. This work has combined the result of the image classification with some context information such as age, gender, skin type and suffered body portion. The system effectively categorized melanoma images with a precision of 86%. More recently, Cavalcanti & Scharcanski (2011) present an even more innovative system, which comprises image pre-processing, region segmentation, feature extraction, and lesion classification stages. Moreover, a two-stage classification model has been suggested that reconfirm the lesions labeled as benign to reduce the false negative cases. These authors also test their work with two publicly available datasets and achieve an accuracy of 96.71%.