Enhancing Brain Tumor Classification and Generalization Using DDPM-Generated MRI, Mutual Information and Ensemble...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep generative models enhance deep learning in medical imaging by creating synthetic data.
• Limited training data can hinder the generalization of AI models for brain tumor classification.
• Denoising Diffusion Probabilistic Models (DDPM) offer a promising approach for generating realistic synthetic medical images.

Purpose of the Study:

• To evaluate if DDPM-generated synthetic MRI, with and without mutual information (MI) regularization, improves brain tumor classification.
• To assess the generalization capabilities of DDPM-generated data across heterogeneous datasets.
• To compare the performance of DDPM with MI against baseline models and traditional augmentation techniques.

Main Methods:

• Trained DDPM models to generate synthetic low-grade glioma (LGG) and high-grade glioma (HGG) MRI images, including a variant with MI regularization.
• Assessed synthetic image quality using metrics like Frechet-Inception-Distance (FID) and Inception-Score (IS).
• Trained a 2D ResNet-152 classifier on real data, augmented data, DDPM-generated data, and DDPM+MI-generated data, evaluating performance with accuracy and F1-score using cross-dataset validation.

Main Results:

• DDPM models generated high-quality synthetic MRI images, with the MI variant showing improved realism and diversity (lower FID, higher IS).
• The DDPM+MI approach achieved superior cross-dataset classification accuracy (0.89 BraTS-to-TASMC, 0.85 TASMC-to-BraTS).
• DDPM+MI consistently outperformed baseline models, traditional augmentation, and standard DDPM without MI.

Conclusions:

• DDPM with MI regularization, combined with ensemble learning, significantly enhances brain tumor classification generalization across diverse datasets.
• This method provides a robust solution for improving AI model performance in cross-institutional clinical settings.
• Synthetic data generation using DDPM+MI is a valuable strategy for overcoming data limitations in medical imaging AI.