Improving Performance of Convolutional Neural Network Models for Brain Tumor Classification from MRI Images through Hyperparameter Tuning

Abstract

Brain tumors are a common and severe disease, requiring timely diagnosis and treatment to reduce mortality rates and improve patient survival chances. This research presents a method for classifying brain tumors using deep neural networks. The Br35H dataset, consisting of 3,000 images (1,500 diseased and 1,500 non-diseased), was used for model training, divided into 70% training, 20% validation, and 10% testing sets. Additionally, 253 Brain MRI Images and 4,600 Brian Tumor images were used to test the model's accuracy. The experiments were divided into three groups: Group 1 employed popular architectures from existing literature without modification, using default hyperparameter settings. Specifically, the configurations included the Adam optimizer with a learning rate of 0.001 and a batch size of 32. The top five architectures with the highest accuracy were selected for further experiments. The architectures DenseNet201, Xception, InceptionResNetV2, MobileNetV2, and NasNetMobile accuracy of 98.00%, 98.00%, 98.00%, 97.67%, and 97.33% and loss of 0.08, 0.06, 0.08, 0.11, and 0.11 respectively. Group 2: Tuning three hyperparameters separately: Batch size, Optimizer, and Learning rate. The best results were Xception (Batch size 16) accuracy of 98.34% and loss of 0.08, InceptionResNetV2 (Learning rate 0.01) accuracy of 98.67% and loss of 0.13, and Xception (Optimizer Adamax) accuracy of 98.67% and loss of 0.06. Group 3: Tuning pairs of hyperparameters: The best architecture was InceptionResNetV2 (Batch size 32, Optimizer Adam, Learning rate 0.01) accuracy of 100.00% and loss of 0.001.