SL-HarDNet: Skin lesion segmentation with HarDNet.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate segmentation of skin lesions from dermoscopy images is crucial for early skin cancer detection.
• The complexity and fuzzy boundaries of skin lesions present significant challenges for automated segmentation algorithms.

Purpose of the Study:

• To develop a novel and robust skin lesion segmentation network (SL-HarDNet) for improved early diagnosis of skin cancer.
• To enhance feature representation and information fusion for more accurate segmentation of complex skin lesions.

Main Methods:

• Proposed a novel skin lesion segmentation network (SL-HarDNet) utilizing HarDNet as the backbone for robust feature learning.
• Introduced three key modules: cascaded fusion module (CFM) for semantic and location aggregation, spatial channel attention module (SCAM) for key spatial information capture, and feature aggregation module (FAM) for cross-level feature fusion.
• Evaluated the network on ISIC-2016&PH2 and ISIC-2018 datasets, comparing against state-of-the-art methods.

Main Results:

• The proposed SL-HarDNet consistently outperformed existing skin lesion segmentation methods across multiple challenge datasets.
• Demonstrated superior performance in capturing both semantic and location information, leading to improved segmentation accuracy.
• Achieved state-of-the-art results, highlighting the effectiveness of the novel network architecture and integrated modules.

Conclusions:

• The SL-HarDNet presents a significant advancement in automatic skin lesion segmentation, offering improved accuracy for early skin cancer diagnosis.
• The integration of HarDNet backbone with CFM, SCAM, and FAM modules effectively addresses the challenges posed by complex lesion boundaries.
• The superior performance validates the potential of SL-HarDNet for clinical application in dermatological image analysis.