SG-MIAN: Self-guided multiple information aggregation network for image-level weakly supervised skin lesion...

Area of Science:

• Dermatology
• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Skin diseases pose significant health risks, necessitating advanced diagnostic tools.
• Computer-aided diagnosis using deep learning aids medical professionals, with lesion segmentation being crucial.
• Traditional segmentation methods require time-consuming pixel-level labeling, increasing costs and complexity.

Purpose of the Study:

• To develop a novel weakly supervised method for skin lesion segmentation using only image-level labels.
• To reduce the reliance on extensive pixel-level annotations in medical image segmentation.
• To improve the accuracy of lesion localization and boundary detection in computer-aided diagnosis systems.

Main Methods:

• Proposed a Self-Guided Multiple Information Aggregation Network (SG-MIAN) for skin lesion segmentation.
• Utilized a backbone network (MIAN) with Multiple Spatial Perceptrons (MSP) guided by classification information for lesion localization.
• Introduced an Auxiliary Activation Structure (AAS) and auxiliary loss functions for self-guided boundary correction.

Main Results:

• The SG-MIAN method demonstrated accurate localization and activation of skin lesion areas.
• The proposed approach achieved precise boundary activation through self-guided correction mechanisms.
• Experiments on HAM10000 and PH² datasets showed superior performance compared to existing weakly supervised segmentation methods.

Conclusions:

• The developed method effectively segments skin lesions using only image-level labels, significantly reducing annotation efforts.
• SG-MIAN offers a promising approach for enhancing computer-aided diagnosis systems for skin diseases.
• This technique addresses the challenges of lesion localization and boundary perception in weakly supervised segmentation.