Improving CNN-based brain tumor classification via differential learning rate and data augmentation strategies

Diagnosing brain tumors accurately at an early stage plays an important role in treatment planning and patient survival. Since manual interpretation of medical images is time-consuming and depends heavily on expert experience, automated methods have become increasingly valuable. This study evaluates five transfer learning-based convolutional neural network models, namely DenseNet121, EfficientNetB0, InceptionV3, MobileNetV2, and ResNet50, for multi-class brain tumor classification. To improve classification performance, several data augmentation techniques, including brightness and contrast adjustment, blurring, flipping, and rotation, were applied together with a differential learning rate strategy. This strategy allowed the lower layers of the pretrained networks to preserve general visual representations, while the upper layers adapted more effectively to tumor-specific patterns. The results show that this training approach improved all evaluation metrics compared with fixed learning rate training. EfficientNetB0 achieved the highest performance, with 98.78% accuracy and an F1-score of 0.9878.