Majority Voting Ensemble of Deep CNNs for Robust MRI-Based Brain Tumor Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate brain tumor classification is vital for patient treatment and prognosis.
• Deep convolutional neural networks (CNNs) show potential in medical image analysis.
• Limited studies compare CNN architectures or use ensemble methods for brain tumor classification.

Purpose of the Study:

• Evaluate multiple CNN models for brain tumor classification.
• Optimize classification performance using a majority voting ensemble.
• Assess performance on T1-weighted MRI brain images.

Main Methods:

• Fine-tuned seven pretrained CNN architectures to classify four brain tumor types.
• Trained models using SGDM and ADAM optimizers on public and external datasets.
• Constructed a majority voting ensemble from 14 trained models.

Main Results:

• Individual models achieved high accuracy, with GoogLeNet and Inception-v3 reaching 0.987.
• The ensemble model surpassed individual performance, achieving 0.998 accuracy and 0.997 Kappa coefficient.
• Ensemble approach demonstrated superior sensitivity, precision, and robustness across all tumor classes.

Conclusions:

• A majority voting ensemble of diverse CNNs significantly boosts MRI-based brain tumor classification accuracy.
• Ensemble learning and model diversity are crucial for developing reliable AI diagnostic tools.
• This approach offers a promising advancement for AI-driven neuro-oncology diagnostics.