Pooling and Convolution Layer Strategy on CNN for Melanoma Detection

Abstract

Melanoma is the most lethal skin cancer nowadays, but an early diagnosis can considerably boost the survival rate and cure rate. Convolutional neural network (CNN) can be implied on medical image classification and can help the early melanoma detection. Our study mainly focuses on the influence of the CNN model on the melanoma detection. AlexNet and VGG are used as the network and the size of the pooling layer and the number of convolution layers are adjusted. According to the changes of parameters, the metrics of different models are compared. After training the models, the results illustrated that the pooling size of 3 contributed to better performance than that of 2, and the VGG model with 5 convolutional layers, max pooling, and pooling size of 3 is discovered to reach the lowest loss value of about 0.38. In conclusion, the specific recognition performance varies depending on the CNN structure, and what causes this variation also includes the number of convolutional layers and the pooling kernel size. Our main contribution also includes the identification of the best structure for detecting melanoma by comparing the model performance.