Improving Efficiency of Brain Tumor Classification Models Using Pruning Techniques.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Computational Neuroscience

Background:

• Convolutional Neural Networks (CNNs) are crucial for medical image analysis, including MRI brain tumor classification.
• High computational complexity can limit the efficiency and deployment of CNNs in clinical settings.
• Pruning techniques offer a method to optimize CNNs by reducing model size and computational load.

Purpose of the Study:

• To investigate the impact of pruning on the computational complexity of a CNN for MRI brain tumor classification.
• To identify optimal pruning percentages that balance reduced complexity with classification performance.
• To enhance the efficiency of CNN models for brain tumor diagnosis.

Main Methods:

• A five-layered CNN model was developed for MRI brain tumor classification.
• Systematic pruning of weights and neurons was applied, ranging from 0% to 99%.
• Classification accuracy was recorded at each pruning level to evaluate performance trade-offs.

Main Results:

• The CNN model's weights could be pruned by up to 70% while retaining acceptable accuracy.
• Neuron pruning up to 10% did not significantly compromise the model's classification accuracy.
• A clear trade-off between model complexity reduction and classification performance was observed.

Conclusions:

• Pruning is an effective technique for reducing the computational complexity of CNNs used in MRI brain tumor classification.
• Judicious pruning of weights and neurons can significantly improve inference time without sacrificing diagnostic accuracy.
• Optimized CNN models through pruning hold promise for more efficient clinical applications in neuro-oncology.