Hyperparameter Optimization for Deep Learning-based Automatic Melanoma Diagnosis System

Abstract

Deep learning is widely used in the development of automatic diagnosis systems for melanoma. However, there are some parameters called hyperparameters which should be set arbitrarily. Optimum setting of hyperparameters is challenging. The dermoscopic images on the database are trained on GoogLeNet. The hyperparameters verified in this study were random seed, solver type, base learning rate, epoch, and batch size. By using a genetic algorithm, these hyperparameters were optimized to obtain higher validation accuracy than other methods such as brute force or Bayesian optimization. The highest validation accuracy was 89.75%. The best hyperparameter settings were: 2 for random seed, RMSProp for solver type, 0.0001 for base learning rate, 30 for epoch, 32 for batch size, and 368 seconds for training time. Using the genetic algorithm, we successful-ly set the hyperparameters for efficient deep learning. Using the system developed in this study, we plan to search for a broader range of hyperparameters and identify multiple groups including lesions other than melanoma.