Efficient Video Polyp Segmentation by Deformable Alignment and Local Attention.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate Video Polyp Segmentation (VPS) is crucial for early colorectal cancer detection and polyp treatment.
• Existing VPS methods struggle with spatial-temporal modeling, motion variations, and computational overhead.
• Challenges include noise, reduced accuracy, and the complexity of optical flow models.

Purpose of the Study:

• To propose a novel VPS framework, Deformable Alignment and Local Attention (DALA), addressing limitations of current methods.
• To improve the modeling of spatial-temporal relationships in colonoscopy videos for enhanced polyp segmentation.
• To achieve accurate and efficient polyp segmentation without significant computational cost.

Main Methods:

• A shared encoder jointly encodes feature representations of paired video frames.
• A Multi-Scale Frame Alignment (MSFA) module using deformable convolution estimates inter-frame motion.
• Local Attention (LA) selectively aggregates aligned features for precise spatial-temporal representations.

Main Results:

• DALA demonstrates superior performance on the SUN-SEG and PolypGen datasets compared to state-of-the-art models.
• The framework effectively handles scale variations and motion complexities in colonoscopy videos.
• Achieved enhanced segmentation accuracy and efficiency in video polyp segmentation.

Conclusions:

• DALA offers a significant advancement in Video Polyp Segmentation technology.
• The proposed method provides a more accurate and computationally efficient solution for polyp detection.
• This framework has the potential to improve early diagnosis and treatment of colorectal cancer.