Lightweight colon polyp segmentation algorithm based on improved DeepLabV3.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Current polyp segmentation models are often complex and lack sufficient accuracy.
• Improving the efficiency and precision of polyp segmentation is crucial for early disease detection.

Purpose of the Study:

• To propose a lightweight and accurate polyp segmentation network, Li-DeepLabV3+.
• To enhance feature representation and fusion for improved segmentation performance.
• To validate the model's generalization capabilities across different datasets.

Main Methods:

• Utilized an optimized MobileNetV2 as the backbone for reduced model complexity.
• Implemented an improved simple pyramid pooling module to replace Atrous Spatial Pyramid Pooling, enhancing training efficiency and reducing parameters.
• Employed an improved Unified Attention Fusion Module for fusing low-level and high-level features with channel and spatial attention to enrich boundary information.

Main Results:

• The Li-DeepLabV3+ model demonstrated superior segmentation accuracy compared to existing methods.
• The proposed model achieved significant improvements in segmentation speed, indicating enhanced efficiency.
• Validation on CVC-ClinicDB and Kvasir SEG datasets confirmed the model's generalization ability.

Conclusions:

• Li-DeepLabV3+ offers a lightweight and effective solution for polyp segmentation.
• The model's architecture successfully balances segmentation accuracy and computational efficiency.
• The enhanced feature fusion mechanism contributes to improved boundary information capture in polyp segmentation.