Skin Lesion Segmentation Using Deep Learning with Auxiliary Task.

Area of Science:

• Dermatology
• Medical Image Analysis
• Computer Vision

Background:

• Skin lesion segmentation is crucial for analysis and classification.
• Challenges include fuzzy boundaries and similar lesion colors.
• Existing methods use ensembling, multi-scale aggregation, or multi-task learning.

Purpose of the Study:

• To propose a novel Convolutional Neural Network (CNN) architecture for skin lesion segmentation.
• To address the limitations of multi-task learning requiring extra labeled data.
• To improve segmentation accuracy by incorporating auxiliary information.

Main Methods:

• A novel CNN architecture is proposed, integrating edge prediction as an auxiliary task.
• A cross-connection layer module facilitates feature map exchange between tasks.
• A multi-scale feature aggregation module enhances performance.

Main Results:

• The proposed method achieved a Jaccard Index (JA) of 79.46, Accuracy (ACC) of 94.32, and Sensitivity (SEN) of 88.76.
• Outperformed state-of-the-art methods.
• The model is learned in an end-to-end manner.

Conclusions:

• The novel CNN architecture with auxiliary edge prediction significantly improves skin lesion segmentation.
• The cross-connection and multi-scale modules enhance feature utilization and boundary focus.
• This integrated model offers a robust and efficient solution for skin lesion analysis.