Brain Tumor Classification in MRI Images Based on EfficientNetB0: Impact of Learning Rate, Optimizer and Batch Size

Brain tumor classification based on Magnetic Resonance Imaging (MRI) images is essential for early diagnosis, yet traditional manual interpretation remains time-consuming and error-prone. This research delves into the ways in which deep learning models for brain tumor classification are impacted by changes in learning rate, optimizer, and batch size. The backbone model was created by training on a Kaggle brain MRI dataset that contained 3,264 grayscale images of four categories: glioma tumor, meningioma tumor, pituitary tumor, and no tumor. Preprocessing included normalization, grayscale-to-RGB conversion, and data augmentation. Using weights pre-trained by ImageNet, transfer learning was implemented. The model's final layer was adjusted for four-class output with Softmax activation. Experiments were conducted across four learning rates, two optimizers, and five batch sizes. There were 50 training iterations for each configuration, with validation loss serving as the basis for early termination. The following metrics were used to assess performance: recall, accuracy, precision, F1-score, and confusion matrices. Experimental results show that Adam generally delivers more stable and accurate outcomes than Stochastic Gradient Descent (SGD) under similar conditions. Higher learning rates combined with moderate batch sizes led to optimal performance. Pituitary tumors were accurately identified, while no_tumor and meningioma were often confused. These findings demonstrate how crucial hyperparameter adjustment is to enhance the accuracy of deep learning-based Identifying brain tumors.