Semi-Supervised Learning with GANs for Melanoma Detection

Melanoma is an aggressive skin cancer that can be curable if caught early. Typically, the diagnosis involves a basic skin examination. In case of need, biopsy and histopathological examination are also performed. In addition to reducing a dermatologist's workload, computer-aided diagnosis provides an independent score regardless of a clinician's experience. The effectiveness of deep-learning-based methods has increased melodramatically over the past decade, and they are outperforming traditional image-processing methods in segmentation tasks. But, this type of learning approach has the main drawback, namely, it requires a large number of labeled samples per class for training. However, a new technique may ameliorate this constraint. To automate the learning process of feature representation of melanoma samples in a semi-supervised manner, this research work proposes Deep Convolution Adversarial Networks (DCGAN). In extensive experiments, the proposed feature learning method proves effective. Although some work on this had been done before, the proposed method works better than others in terms of training time and performance. With only 200 labeled images, it achieved precision, accuracy, and recall scores of 0.7689, 0.7525, and 0.7384. Additionally, the proposed method can generate real-world dermoscopy images.