A Fine-Funed CNN for Multiclass Classification Of Brain tumors On Figshare CE-MRI and its Raspberry Pi Deployment

Abstract

This paper introduces a fine-tuned Convolutional Neural Network (CNN) for multiclass classification of brain tumors on contrast-enhanced T1-weighted MRI scans. The proposed model integrates batch normalization, dropout, and lightweight convolutional blocks to extract discriminative features while maintaining computational efficiency suitable for embedded deployment. Experiments were conducted on the Figshare dataset comprising 3,064 MRI slices from 233 patients with gliomas, meningiomas, or pituitary tumors. Images were preprocessed through resizing, normalization. The model was trained using the Adam optimizer with a learning rate of 1e-4, a batch size of 32, and 100 epochs. Evaluation metrics included accuracy, precision, recall, and F1-score. The fine-tuned CNN achieved an overall accuracy of 94.08%, with class-specific performance indicating strong results for pituitary tumors (precision 95.65%, recall 95.96%) and meningiomas (precision 90.20%, recall 88.81%), while glioma classification showed high sensitivity (recall 96.85%) but lower precision (75.00%). To validate real-world applicability, the model was converted to TensorFlow Lite and deployed on a Raspberry Pi 4, achieving an inference time of approximately 60 ms per image. These findings demonstrate that fine-tuned CNNs can offer a competitive and resource-efficient solution for computer-aided diagnosis of brain tumors, balancing accuracy and practicality in clinical environments with limited computational resources.