Employing Squeeze-and-Excitation Architecture in a Fine-Tuned Convolutional Neural Network for Magnetic Resonance Imaging Tumor Classification.

In 2021, there were 182,520 cases of brain and central nervous system (CNS) cancers in the U.S. and 25,400 new cases of brain cancer in 2024. Early detection via magnetic resonance imaging (MRI) significantly improves patient outcomes. This study fine-tunes a residual neural network 50 version 2 (ResNet50V2), a convolutional neural network (CNN), with squeeze-and-excitation (SE) attention mechanisms to enhance MRI-based tumor classification compared to a base ResNet50V2 model. By incorporating SE blocks, the model improves feature prioritization, effectively distinguishing glioma (n = 901), meningioma (n = 913), pituitary tumor (n = 844), and no tumor (n = 438). Trained on a publicly available Kaggle dataset (N = 3,096), the proposed model achieved a 98.4% classification accuracy and an area under the receiver operating characteristic curve (AUC) of 0.999, outperforming the base model's 92.6% accuracy and 0.987 AUC. Statistically significant improvements were observed in meningioma (p = 0.013) and pituitary tumor (p = 0.015) classification accuracy, highlighting the SE model's superior ability to differentiate tumor types. SE attention mechanisms enhance diagnostic precision by addressing feature extraction limitations and long-range dependencies in medical imaging. However, challenges such as dataset size constraints, overfitting risks, and potential representation bias remain. Future research will focus on expanding dataset diversity, exploring vision transformers (ViTs) for improved feature extraction, and employing generative adversarial networks (GANs) for data augmentation.