Effect of Hyperparameter Tuning on Transfer Learning Models for Brain Tumor Detection and Classification

A mass of abnormal cell growth in the brain is known as a brain tumor, these are of two types benign and malignant. Currently, medical professionals use Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) scans to analyze and identify the presence of tumors. Finding brain tumors is an arduous but crucial endeavor. The emergence of Artificial Intelligence techniques has led to the detection of brain tumors using various Machine Learning algorithms and techniques, particularly Deep Learning. These algorithms are advantageous as there are fewer errors and greater precision which in turn which makes decision-making and selection of the patient's best course of action more accurate. In this paper, we have employed Magnetic Resonance Images (MRI) images that were applied to a variety of transfer learning models, including VGG19, InceptionV3, and DenseNet121. To determine the optimal values of the learning rate, batch size, optimizer, and the number of epochs we performed hyperparameter tuning for each model. The Wilcoxon-Mann-Whitney test was performed to assess whether the hyperparameters have made a significant impact on each metric used. After performing this test and determining the accuracy we discover that DenseNet121 gives us the highest accuracy of 96%, followed by InceptionV3 which displays an accuracy ranging from 89 to 91%, and VGG19 which provides an accuracy of 85% but with a rather high penalty on error.