Anti-Aliasing Attention U-net Model for Skin Lesion Segmentation.

Area of Science:

• Biomedical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Accurate biomedical image segmentation is crucial but challenged by limited data and low image quality.
• Existing deep learning models are computationally expensive, requiring large parameters and extensive processing time.
• Developing efficient and lightweight segmentation algorithms is essential for practical applications.

Purpose of the Study:

• To introduce a novel, lightweight segmentation model named the mobile anti-aliasing attention u-net (MAAU) model.
• To address the challenges of limited data and low image quality in biomedical image segmentation.
• To improve the efficiency and reduce the computational cost of segmentation algorithms.

Main Methods:

• The MAAU model utilizes an encoder-decoder architecture with an anti-aliasing layer and attention mechanisms.
• Data augmentation techniques including flip, rotation, shear, translate, and color distortions were employed to enhance model robustness.
• The model was evaluated on the ISIC 2018 and PH2 datasets for skin image segmentation.

Main Results:

• The MAAU model demonstrated superior performance compared to state-of-the-art segmentation methods.
• The proposed model has significantly fewer parameters (4.2 million) compared to traditional deep learning models.
• Data augmentation effectively improved segmentation efficiency on the tested datasets.

Conclusions:

• The MAAU model offers a lightweight and effective solution for biomedical image segmentation.
• The approach successfully overcomes limitations related to data scarcity and computational complexity.
• MAAU shows promise for real-world applications requiring efficient and reliable image segmentation.