High-performance brain tumor classification via transfer learning with fully fine-tuned and modified CNN architectures

Abstract

Accurate and timely brain tumor diagnosis remains challenging due to the limitations of manual MRI interpretation. This study introduces a transfer learning framework using six fully fine–tuned and modified CNN architectures DenseNet169, InceptionV3, Xception, ResNet50, InceptionResNetV2, and EfficientnetB0 each equipped with a redesigned classification head. The proposed head combines global average pooling with four dense layers (1024, 512, 256, 128 neurons) incorporating ReLU activation, Batch Normalization, and Dropout to enhance generalization. The models were trained on the Kaggle Brain Tumor MRI dataset containing 7,023 images across four tumor classes, using data augmentation and five–fold stratified cross–validation for robust evaluation. Results show that the modified CNNs achieve state–of–the–art performance, with InceptionResNetV2 obtaining the highest accuracy (98.7%) and F1–score (98.7%) with minimal overfitting. Xception and DenseNet169 also demonstrated strong performance with lower computational cost. Overall, the proposed fully fine–tuned architectures provide an efficient and reproducible solution for automated brain tumor classification, supporting radiologists and improving diagnostic workflow efficiency.