Skin Cancer Classification using Delaunay Triangulation and Graph Convolutional Network

Oftentimes, many people or even medical workers misdiagnose skin cancer, which may lead to malpractice and thus, resulting in delayed recovery or life-threatening complications. In this research, a Graph Convolutional Network (GCN) method is proposed as a classification model and Delaunay triangulation as its feature extraction method to classify various types of skin cancers. Delaunay triangulation serves the purpose of boundary extraction, and this implementation allows the model to focus only on the cancerous lesion and ignore the skin around it. This way, the types of skin cancer can be predicted more accurately. Furthermore, GCN offers many advantages in medical image analysis over traditional CNN models. GCN can model interactions between different regions and structures in an image and perform messaging between nodes, whereas CNN is not explicitly designed to do such thing. Other than that, GCN can also leverage transfer learning and few-shot learning techniques to address the challenges of limited annotated medical image datasets. However, the result shows that the proposed model tends to overfit and is unable to generate correct predictions for new skin cancer images. There are several reasons that could lead the model to overfit, such as imbalance data, incorrect feature extraction, insufficient features for data prediction, or the data containing noise.