MambaLiteUNet: Compact Segmentation Model for Skin Lesions
A team of researchers has introduced MambaLiteUNet, an innovative model designed to improve skin lesion segmentation. This model incorporates Mamba state space modeling within a U-Net structure and features three main components: Adaptive Multi-Branch Mamba Feature Fusion, Local-Global Feature Mixing, and Cross-Gated Attention. These elements enhance the interaction between local and global features while preserving spatial details. MambaLiteUNet has demonstrated impressive performance, recording an average Intersection over Union (IoU) of 87.12% and a Dice score of 93.09% across various datasets including ISIC2017 and HAM10000, surpassing leading models in the field. The research is available on arXiv under the ID 2604.20286.
Key facts
- MambaLiteUNet integrates Mamba state space modeling into U-Net architecture.
- Includes three modules: AMF, LGFM, and CGA.
- Achieves average IoU of 87.12% and Dice score of 93.09%.
- Tested on ISIC2017, ISIC2018, HAM10000, and PH2 benchmarks.
- Outperforms state-of-the-art models.
- Paper published on arXiv with ID 2604.20286.
- Focuses on skin lesion segmentation for early cancer diagnosis.
- Model is compact with reduced parameter count.
Entities
Institutions
- arXiv