A novel hybrid Res2Net-UNet model for accurate brain tumor segmentation in MRI
Chelli N. Devi, Thomas Samraj Lawrence, Prabhakar Rontala Subramaniam · International Journal of Cognitive Computing in Engineering · 2025
• Accurate diagnosis of brain tumor from MR images improves patient longevity • Novel hybrid Res2Net-UNet model with sequential attention module proposed • Dice value of 0.9630, 0.9513 and 0.9240 for whole, core and enhancing tumor obtained • Segmentation results superior to traditional methods and state-of-the-art literature • First-ever hybrid Res2Net-UNet model for brain tumor segmentation The automated segmentation and analysis of brain tumor from magnetic resonance images (MRI) is clinically significant and has immense social potential given the high incidence and mortality rate of brain tumors. Many atlas-based and deep learning techniques in the recent literature achieve good segmentation of the whole tumor. However, the segmentation of other tumor regions (like core and enhancing tumor) is less accurate. To address this issue, the paper proposes a novel hybrid Res2Net-UNet model for brain tumor segmentation. This consists of a modified Res2Net network in the encoder side of the U-Net. Likewise, a sequential channel and spatial attention mechanism is included in the decoder side. The segmentation results are evaluated both qualitatively and quantitatively. The following metrics are used for quantitative evaluation: Dice ratio, Jaccard index, sensitivity, specificity and precision. The proposed model obtained Dice values of 0.9630, 0.9513 and 0.9240 for the whole, core and enhancing tumor regions, respectively. To the best of our knowledge, this is the first combined Res2Net-UNet model for brain tumor segmentation. The segmentation results by the proposed method are higher than other traditional networks like U-Net, ResNet, DenseNet and EfficientNet. The results are superior to recent works in the state-of-the-literature, especially for the core and enhancing tumor regions. Thus, the model has potential for accurate diagnosis of brain tumors from clinical MR images.