Optimal Res-UNET architecture with deep supervision for tumor segmentation

Background: Brain tumor segmentation is critical in medical imaging due to its significance in accurate diagnosis and treatment planning. Deep learning (DL) methods, particularly the U-Net architecture, have demonstrated considerable promise. However, optimizing U-Net variants to enhance performance and computational efficiency remains challenging. Objective: To develop an optimized Residual U-Net (Res-UNET) architecture enhanced by deep supervision techniques to improve segmentation accuracy of brain tumors on MRI datasets, specifically addressing challenges of conventional segmentation methods. Methods: The study implemented a detailed evaluation of multiple U-Net variations, including basic U-Net, Res-UNet with Autoencoder regularization, and attention-enhanced U-Net architectures. Training was conducted using the BraTS 2018 public MRI dataset. Deep supervision was integrated to improve gradient propagation and segmentation accuracy. The model employed a Dice loss combined with focal loss to handle data imbalance effectively. The proposed network was evaluated using extensive ablation studies, examining the effects of encoder complexity, convolutional filter count, and strategic post-processing. Results: The proposed Res-UNET with deep supervision outperformed other variants, achieving an average Dice score of 0.9498 through five-fold cross-validation. Post-processing strategies improved the robustness of segmentation, particularly enhancing the accuracy of small tumor regions. Comparatively, conventional U-Net architectures yielded lower Dice scores and required significantly longer training times. The study indicates the benefit of integrating deep supervision and residual connections for enhanced model performance. Conclusion: Optimized Res-UNET with deep supervision significantly enhances segmentation accuracy for brain tumors in MRI images, surpassing traditional U-Net models. This model addresses critical issues such as dataset imbalance, lack of annotated data, and computational inefficiencies. Future studies should consider the broader application of optimized U-Net variants across other medical imaging segmentation tasks.