Performance Analysis of U-Net and Unetr For 2D BraTS Within Consumer-Grade Hardware Limits

This study investigates the performance of U-Net and UNETR, a hybrid model combining a 2D vision transformer (ViT) with CNNs and skip connections, for brain tumor segmentation in a 2D dataset. U-Net, known for its effectiveness with limited data and computational resources, is compared against UNETR, whose ViT is proven to excel at capturing long-range dependencies and achieving high accuracy with large datasets. Although UNETR achieved a validation Dice coefficient of $\mathbf{0 . 7 0}$ over 50 epochs, U-Net attained a higher score of 0.84 in 65 epochs. However, trainable parameter count of the UNETR was one-thirteenth of that of the U-Net’s. This comparison highlights the potential trade-offs between accuracy and training efficiency when choosing between U-Net and UNETR for resource-constrained medical image segmentation tasks. The results of this investigation serve as a foundation for identifying a model that can be effectively applied with constrained computational resources. The study’s conclusions show that while U-Net and UNETR are extremely resource-hungry models that need a lot of processing power to train, they have the potential to be used in consumer-level or low-power computing devices. The primary goal of the study was to find out how U-Net and UNETR performs when these models are trained in consumer grade hardware and which one is more prominent and has more potential for future implementation in low power hardware.