Optimizing Multi-Class Brain Tumor Diagnosis Using Transfer Learning with EfficientNetB4

Certain types of tumors in individuals with brain cancer proliferate rapidly, with their average size doubling within 25 days. Accurately identifying the type of tumor enables physicians to develop effective treatment plans and determine the appropriate dosage. Magnetic Resonance Imaging (MRI) is a critical diagnostic technique for evaluating and diagnosing brain tumors because it provides high-contrast images of brain tissues. This article introduces an innovative approach for multi-classification brain tumors by utilizing deep convolutional neural networks (DCNNs), specifically employing EfficientNet-B4 as the base model, enhanced with fine-tuned, customized layers. Our approach incorporates a Global Average Pooling (GAP) layer to mitigate overfitting, batch normalization, and dropout layers to reduce losses and improve generalization. A series of experiments are performed on an open-access Kaggle dataset to identify the optimal model, utilizing seven optimization algorithms, including Adadelta, RMSprop, Adam, and Nadam. Among all models tested, EfficientNet-B4 with AdamW was the best-performing, achieving a test accuracy of 99.24%, a precision, recall, and F1-score of 99.22% and a specificity of 99.75%. In contrast, EfficientNet-B4 with AdamX had the lowest performance, with a test accuracy of 98.55%, precision of 98.53%, recall of 98.46%, F1-score of 98.49%, and specificity of 98.52%. These innovations can potentially enhance clinical decision-making and improve patient treatment in neuro-oncology.