LAMA: Lesion-Aware Mixup Augmentation for Skin Lesion Segmentation

Area of Science:

• Dermatology and Medical Image Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Deep learning models show promise in diagnosing skin conditions from images.
• Current segmentation methods struggle with images containing multiple skin lesions, leading to information loss.
• Existing datasets lack sufficient examples of multi-lesion images, hindering model training.

Purpose of the Study:

• To develop a novel data augmentation technique to improve skin lesion segmentation accuracy, particularly for images with multiple lesions.
• To address the limitations of current machine learning methods in handling complex multi-lesion dermoscopic images.

Main Methods:

• Proposed the Lesion-Aware Mixup Augmentation (LAMA) technique to generate synthetic multi-lesion images by combining existing lesion images.
• Trained a deep neural network using the LAMA method on the ISIC 2017 Challenge dataset.
• Created and utilized a new multi-lesion (MuLe) segmentation dataset from ISIC 2020 images for testing.

Main Results:

• LAMA significantly improved segmentation performance on the MuLe dataset, increasing Jaccard score by 8.3% (0.687 to 0.744) and Dice score by 5% (0.7923 to 0.8321).
• The method also enhanced performance on single-lesion images from the ISIC 2017 dataset, with Jaccard score increasing by 0.08% and Dice score by 0.6%.
• Experimental results demonstrated LAMA's effectiveness in improving segmentation accuracy for both single and multiple skin lesion images.

Conclusions:

• The proposed LAMA data augmentation technique effectively enhances deep learning model performance for skin lesion segmentation.
• LAMA shows promise in overcoming challenges associated with multi-lesion images, improving diagnostic information retrieval.
• Further research into LAMA is warranted to explore its full potential in dermatological image analysis.