GCSER-UNet Achieves 95% Dice Score in Brain Tumor Segmentation
Researchers have developed the Global Context-aware Squeeze and Excite Residual UNet (GCSER-UNet), a deep neural network for brain tumor segmentation from multimodal MRI scans. The model integrates spatial and channel-wise attention to capture intricate spatial dependencies and contextual information. Evaluated on benchmark datasets, GCSER-UNet achieved a 94% dice score on the TCGA LGG dataset, surpassing the previous state-of-the-art of 91.8%. On the BraTS 2020 dataset, an ensemble approach yielded dice scores of 95%, 92%, and 90% for different tumor subregions. The study, published on arXiv, addresses the need for automated methods to reduce costs and errors in manual tumor identification.
Key facts
- GCSER-UNet is a novel deep neural network for brain tumor segmentation.
- It uses global context-aware squeeze and excite residual blocks.
- Achieved 94% dice score on TCGA LGG dataset, beating prior 91.8%.
- On BraTS 2020, ensemble method scored 95%, 92%, and 90% dice.
- Model fuses spatial and channel-wise attention mechanisms.
- Designed for multimodal MRI slices.
- Aims to automate brain tumor diagnosis, reducing manual effort.
- Published on arXiv with ID 2605.30510.
Entities
Institutions
- arXiv