ULFAC-Net: Ultra-Lightweight Fully Asymmetric Convolutional Network for Skin Lesion Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Artificial intelligence in dermatology

Background:

• Accurate skin lesion segmentation is crucial for diagnosis but challenging due to irregular shapes and noise.
• Current deep learning methods are computationally expensive, limiting their use in portable dermoscopic devices.

Purpose of the Study:

• To propose an ultralightweight fully asymmetric convolutional network (ULFAC-Net) for efficient skin lesion segmentation.
• To address the computational cost limitations of existing deep learning models for deployment on low-power devices.

Main Methods:

• Introduced Parallel Asymmetric Convolution (PAC) and Attentive PAC (Att-PAC) modules for feature extraction and enhancement.
• Developed a lightweight textual information submodule and an asymmetric encoder-decoder architecture.
• Validated ULFAC-Net on ISIC2018, ISBI2017, ISIC2016, and PH2 datasets.

Main Results:

• ULFAC-Net achieved competitive segmentation performance across multiple datasets.
• The proposed network has only 0.842 million parameters and 3.71 GFLOPs.
• Demonstrated effectiveness and robustness compared to state-of-the-art methods.

Conclusions:

• ULFAC-Net offers an efficient and effective solution for skin lesion segmentation.
• The ultralightweight design makes it suitable for deployment on dermoscopic devices with limited computational power.