Pre-Trained Generative Adversarial Network for Limited-Labeled Brain MRI Segmentation

Abstract

Accurate brain MRI segmentation is challenging due to subtle tissue boundaries, inter-subject variability, and the limited availability of annotated data. To address these challenges, we propose a multi-class brain MRI segmentation framework that integrates transfer learning with generative adversarial networks (GANs) to improve robustness and accuracy in low-data settings. The proposed model employs a GAN architecture with a ResNet-50 generator pre-trained on ImageNet, enabling effective feature transfer while stabilizing adversarial training. Brain MRI volumes are processed in a slice-wise 2D manner for computational efficiency, and grayscale slices are mapped to three-channel representations to ensure compatibility with pre-trained backbones. Adversarial learning further enforces local anatomical plausibility in the predicted segmentation maps. Experimental results using six-fold cross-validation demonstrate that the proposed approach consistently outperforms state-of-the-art segmentation models, achieving an accuracy of 98.41%, a Dice Average (excluding background) of 93.43%, and a mean IoU of 90.14%. These results highlight the effectiveness of combining transfer learning and adversarial regularization for reliable brain MRI segmentation under limited data conditions.