Brain Tumor Classification Using a Combination of Variational Autoencoders and Generative Adversarial Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Brain tumors have low survival rates, necessitating accurate and timely diagnosis.
• Magnetic Resonance Imaging (MRI) is crucial for brain tumor diagnosis.
• Deep learning models require large datasets, which are scarce in medical imaging.

Purpose of the Study:

• To propose a framework using unsupervised deep generative neural networks to overcome data limitations in brain tumor MRI datasets.
• To enhance the performance of deep learning models for brain tumor classification by augmenting training data with generated images.

Main Methods:

• A framework combining Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) was developed.
• A swapped encoder-decoder network generated informative noise vectors from MR images.
• Cascaded GANs sampled from these noise vectors to generate realistic brain tumor MRI images, avoiding mode collapse.

Main Results:

• The proposed method successfully generated realistic brain tumor MRI images, augmenting limited datasets.
• Classification accuracy improved from 72.63% to 96.25% when using generated images for training.
• For glioma, recall, specificity, precision, and F1-score reached 0.769, 0.837, 0.833, and 0.80, respectively.

Conclusions:

• The generative framework effectively addresses the challenge of small medical datasets.
• The generated images significantly improve deep learning model performance for brain tumor classification.
• This approach offers a valuable clinical tool for medical experts and can be applied to other medical imaging domains.