Related Experiment Video
Updated: Jul 14, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
568
Attention-based dual-path feature fusion network for automatic skin lesion segmentation.
Zhenxiang He1,2, Xiaoxia Li1,3, Yuling Chen1,3
1School of Information Engineering, Southwest University of Science and Technology, Mianyang, China.
Biodata Mining
|October 8, 2023
Summary
This study introduces the Attention-based Dual-path Feature Fusion Network (ADFFNet) for improved automatic skin lesion segmentation in melanoma diagnosis. The novel network enhances boundary details and context, achieving superior accuracy in segmenting challenging skin lesions.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Automatic skin lesion segmentation is crucial for melanoma Computer Aided Diagnosis (CAD).
- Challenges include blurred boundaries, uneven color, and low contrast, hindering accurate segmentation.
- Existing methods struggle with precise lesion delineation.
Purpose of the Study:
- To propose an Attention-based Dual-path Feature Fusion Network (ADFFNet) for enhanced automatic skin lesion segmentation.
- To address the limitations of current segmentation techniques for complex skin lesions.
- To improve pixel-level segmentation accuracy for better melanoma detection.
Main Methods:
- Developed an Attention-based Dual-path Feature Fusion Network (ADFFNet).
- Incorporated a Boundary Refinement (BR) module in the spatial path to preserve lesion boundary details.
- Utilized a Multi-scale Feature Selection (MFS) module in the context path with attention to capture relevant multi-scale information.
- Implemented a Dual-path Feature Fusion (DFF) module to integrate semantic and detail features guided by global attention.
Main Results:
- The ADFFNet achieved high performance on the ISIC 2018 and PH2 datasets.
- Achieved F1-scores of 0.890 and 0.933, and SE indices of 0.925 and 0.954, respectively.
- Demonstrated superior segmentation performance compared to state-of-the-art methods.
Conclusions:
- The proposed ADFFNet effectively improves automatic skin lesion segmentation accuracy.
- The network's architecture successfully handles challenges like blurred boundaries and low contrast.
- ADFFNet shows significant potential for advancing melanoma Computer Aided Diagnosis (CAD).
