A Comparative Study of Neural Network Architectures for Lesion Segmentation and Melanoma Detection

Melanoma is a fast spreading and deadly form of skin cancer accounting for majority of deaths due to this type of disease. If not treated early, it can spread quickly to other organs. Luckily melanoma symptoms become visible to victims creating an opportunity to detect it at an early stage. As people know less about specific symptoms of it and due to the shortage of expert physicians, automating the detection of melanoma has become an important issue for public health. Several computer-aided diagnostic approaches have been proposed so far. Along with traditional image processing based techniques, lately researchers have successfully used deep learning for many different purposes. Deep neural networks are vastly being used in segmentation, object detection, classification etc. This paper shows how deep learning applied with augmentation can save us from complex pre-processing steps and it studies the performance of different neural network architectures for lesion segmentation as an integral part of image processing based techniques and finally evaluates network architectures for melanoma detection on dermoscopic images using transfer learning and presents a comparative view of those studies.