Segmentation and Classification: Application of Fully Convolutional Networks and Densely Connected Networks in Melanoma Detection

In recent years, the rapid development of deep learning has promoted the improvement of medical imaging. This paper uses computer vision technology to alleviate the classification problem of malignant melanoma images. The incidence and mortality of malignant melanoma have gradually increased in recent decades and data show that detecting melanoma early can significantly improve patients' survival rate. We conducted a systematic review of existing literature, finding there are several algorithms that can recognize melanoma. Most of these algorithms classify the original image directly or use the K-NearestNeighbor (KNN) method, but these methods' recognition efficiency and accuracy are not high and need a large number of pictures to train. In this paper, our model first uses Fully Convolutional Networks (FCN) to find and cut out the possible diseased part, then resorts to Convolutional Neural Networks (CNN) classifier for classification and recognition, which improves the performance and accuracy of the model. The effectiveness of this model has been verified on the dataset ISIC. Our work achieves an accuracy of 95.56%, precision of 88.57%, recall of 82.67%, false positive ratio of 2.63%, and F1 score of 85%, which is markedly better than the existing methods.