DynBrainNet: A dynamic lightweight residual network with Grad-CAM for brain tumor classification

Abstract

Brain tumor classification is crucial for prompt treatment and improved patient prognosis. Deep learning techniques have shown promising results in medical image classification; however, challenges remain in balancing the accuracy, computational efficiency, and model interpretability. This study proposes a novel lightweight deep learning architecture, DynBrainNet, designed for accurate and interpretable brain tumor classification using magnetic resonance imaging (MRI). DynBrainNet incorporates dynamic residual convolutional layers, Learned Group Convolutions (LGConv), early exit strategies, and hyperparameter optimization through grid search to achieve efficient and scalable performance. The model was assessed using a publicly accessible dataset consisting of 24,832 MRI images categorized into four categories: glioma, meningioma, pituitary tumor, and no tumor. DynBrainNet outperformed established transfer learning models, including ResNet50, InceptionV3, VGG16, EfficientNetV2B2, and Xception, achieving the highest classification accuracy of 95.3% among the models tested. The Gradient-weighted Class Activation Mapping (GradCAM) method was used to create heatmaps that clarify the model's predictions and can be used for clinical applications. The experimental results verified that the proposed DynBrainNet model provides a computationally efficient, interpretable, and robust framework for the diagnosis of brain tumors.