Improved SqueezeNet Architecture for Brain Age Prediction using MRI

Predicting brain age from MRI scans is crucial for understanding neurological development and aging-related disorders. Robust preprocessing methods and flexible model architectures are required because variations in image quality, resolution, and acquisition methodologies can impact deep learning performance. The improved Squeeze Net architecture shown in this work can accurately predict brain age from MRI scans. The approach combines a bilateral filtering procedure for preprocessing to enhance picture quality, and then an Improved Squeeze Net (Imp-Squeeze Net) model for precise evaluation of brain age. The model includes convolutional layers, fire modules, and global average pooling, ending with a dense layer that employs SoftMax activation. Notably, a novel Focal Huber Loss function is introduced to handle the complex data distributions and enhance predictive performance. Comparing the performance of the proposed model to CNN (0.155), GRU (0.151), Bi-LSTM (0.130), RNN (0.157), Bi-GRU ($\mathbf{0} .160$), and LSTM ($\mathbf{0} .158$), a significant reduction in prediction error is observed. When compared to conventional techniques, experimental results show higher accuracy and robustness, highlighting the method's promise in brain research and clinical diagnostics.