LSSF-Net: Lightweight segmentation with self-awareness, spatial attention, and focal modulation

Area of Science:

• Dermatology
• Medical Imaging
• Computer Vision

Background:

• Accurate skin lesion segmentation is vital for early skin cancer detection in computer-aided diagnosis on mobile platforms.
• Challenges include varied lesion shapes, indistinct boundaries, and occlusions like hair and markers.
• Existing models like U-Nets struggle with subtle skin lesion variations and contextual information.

Purpose of the Study:

• To develop a novel, lightweight network for efficient skin lesion segmentation on mobile devices.
• To improve segmentation accuracy by addressing challenges posed by complex lesion characteristics.
• To achieve state-of-the-art performance with minimal computational resources.

Main Methods:

• Proposed a lightweight encoder-decoder network with 0.8 million parameters.
• Incorporated conformer-based focal modulation attention, self-aware local and global spatial attention, and split channel-shuffle.
• Evaluated the model on ISIC 2016, ISIC 2017, ISIC 2018, and PH2 benchmark datasets.

Main Results:

• The proposed network demonstrated state-of-the-art performance in skin lesion segmentation.
• Achieved a high Jaccard index, indicating superior segmentation accuracy.
• The lightweight design is suitable for deployment on mobile platforms.

Conclusions:

• The novel lightweight network effectively segments skin lesions, outperforming existing methods.
• The model's architecture successfully captures fine-grained details and contextual information for improved accuracy.
• This advancement facilitates more reliable computer-aided diagnosis of skin cancer on mobile devices.