Brain tumor classification using CNN: Difference between optimizers

Abstract

The classification of brain tumor using artificial intelligence has long been a heated topic and never failed to arouse public attention, being both efficient and useful. Each year, around nearly 12000 people are diagnosed with brain tumor, and the AI approach allows a number of specific cases to be identified quickly with few errors, contributing in saving millions of lives. This paper uses CNN (ResNet50 architecture) for classification and evaluates the performances of three kinds of optimizers adaptive moment estimation (Adam), stochastic gradient descent (SGD), and genetic algorithm (GA) when being applied to the model. The resulting accuracy scores are, respectively, 93%, 90%, and 95%, which demonstrates that genetic algorithm performs the best, suggesting a fine choice of utilization in practical scenarios. The results as well show the most precise diagnosis on pituitary tumor and the least on meningioma tumor, providing a direction for future improvement on dataset and training parameters.