Abstract
<jats:p>Accurate brain tumor segmentation from magnetic resonance imaging (MRI) remains a challenging task because supervised deep learning models require large quantities of annotated data, which are expensive and time-consuming to obtain. This study investigates whether synthetic MRI images generated using a Deep Convolutional Generative Adversarial Network (DCGAN) can improve U-Net-based brain tumor segmentation using synthetic data augmentation. Experiments were performed on the LGG-MRI dataset comprising 3,929 image-mask pairs. A baseline U-Net was first trained using the original training dataset. Synthetic MRI images were subsequently generated using a DCGAN, and threshold-derived pseudo-masks were assigned to the generated images to construct an augmented training dataset. The same U-Net architecture was then retrained using the augmented dataset and evaluated on an identical held-out test set. Compared with the baseline model, DCGAN-based augmentation increased the Dice coefficient from 0.2067 to 0.3037 and the Intersection over Union (IoU) from 0.1243 to 0.1918, while reducing the final test loss from 0.0474 to 0.0275. These results indicate that synthetic MRI augmentation was associated with improved segmentation performance under the experimental conditions of this study. However, the reliance on threshold-derived pseudo-labels and evaluation on a single dataset limit the generalizability of the results. The proposed workflow provides a reproducible implementation for evaluating DCGAN-based synthetic data augmentation in supervised brain tumor segmentation and establishes a baseline for future studies employing more reliable annotation strategies and broader experimental validation.</jats:p>