Polar contrast attention and skip cross-channel aggregation for efficient learning in U-Net

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Existing lesion segmentation models require significant computational resources, limiting their use on edge devices.
• Advancements in attention, skip connections, and deep supervision increase computational complexity.
• There is a need for efficient models deployable on smartphones and point-of-care devices.

Purpose of the Study:

• To develop a lightweight and computationally efficient lesion semantic segmentation model.
• To improve the performance of lesion segmentation on resource-constrained devices.
• To achieve high accuracy with a low parameter count.

Main Methods:

• Introduced polar transformations to simplify image data for faster processing.
• Developed a learning-efficient polar-based contrast attention (PCA) mechanism using Hadamard products.
• Implemented a novel skip cross-channel aggregation (SC^2A) with Gaussian depthwise convolution.

Main Results:

• The model achieved state-of-the-art performance on ISIC 2018 and Kvasir datasets with approximately 25K parameters.
• Demonstrated strong generalization across diverse datasets (PH^2, CVC-Polyp).
• Outperformed other lightweight models in mobile settings regarding IoU, Dice score, and running time.

Conclusions:

• The proposed model offers a highly efficient and accurate solution for lesion segmentation.
• Its low computational requirements make it suitable for deployment on edge and mobile devices.
• This work advances the feasibility of advanced medical image analysis in resource-limited clinical settings.