Optimized Pattern Aware Brain Tumor Segmentation Using Enhanced U-net Learning

Image processing

Brain tumors remain a major clinical concern due to their significant incidence and associated mortality rates, underscoring the critical need for accurate and automated segmentation to support diagnosis and treatment. While deep learning has brought notable improvements in segmentation performance, several limitations continue to challenge existing methods. In response, we propose a new architecture termed Dual Encoder Mirror Difference Residual U-Net (DEMD-ResUNet). This model utilizes two parallel encoders to process both the original and its horizontally flipped version of the input image. Furthermore, residual units replace conventional convolutional blocks within the encoder, simplifying the training process and reducing risks such as vanishing gradients or the loss of fine-grained details. To enhance feature discrimination, the architecture incorporates a Multimodal Difference Feature Augmentation (MDFA) module, which emphasizes abnormal regions across both original and mirrored modalities. In addition, a Mirror Difference Feature Fusion (MDFF) block is positioned between the encoder and decoder paths to effectively combine symmetric features from the dual encoders and improve segmentation accuracy. Experimental results, including ablation studies, validate the contribution of each proposed module. The DEMD-ResUNet achieves high Dice similarity scores on the BraTS 2018 and BraTS 2019 benchmarks, reporting 0.862, 0.925, and 0.905 for Enhanced Tumor (ET), Whole Tumor (WT), and Tumor Core (TC) respectively on BraTS 2018, and 0.869, 0.922, and 0.916 for the same metrics on BraTS 2019.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.