Related Experiment Video
Updated: Oct 22, 2025
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
2.8K
Data Augmentation Using Adversarial Image-to-Image Translation for the Segmentation of Mobile-Acquired Dermatological
Catarina Andrade1, Luís F Teixeira2,3, Maria João M Vasconcelos1
1Fraunhofer Portugal AICOS, Rua Alfredo Allen, 4200-135 Porto, Portugal.
Journal of Imaging
|August 30, 2021
Summary
This study enhances skin lesion segmentation for macroscopic images by using deep learning to translate dermoscopic images. This approach improves segmentation accuracy, especially for mobile-acquired images, setting a new state-of-the-art.
Area of Science:
- Dermatology
- Medical Imaging
- Artificial Intelligence
Background:
- Dermoscopic images offer detailed skin lesion analysis but are not universally accessible.
- Macroscopic images are a more accessible alternative, but limited datasets hinder deep learning development.
- Mobile-acquired macroscopic images present unique challenges due to data scarcity.
Purpose of the Study:
- To develop a technique for improving macroscopic skin lesion image segmentation using abundant dermoscopic images.
- To bridge the domain gap between dermoscopic and macroscopic skin lesion imaging.
- To enable robust deep learning models for skin lesion analysis using accessible imaging.
Main Methods:
- Utilized a Cycle-Consistent Adversarial Network (CycleGAN) for image-to-image translation between dermoscopic and macroscopic domains.
- Employed qualitative visual inspection for assessing feature preservation and translation accuracy.
- Applied quantitative metrics, including Fréchet Inception Distance (FID), for objective evaluation.
Main Results:
- Achieved significant improvements in macroscopic skin lesion segmentation performance.
- Demonstrated state-of-the-art Jaccard Index scores of 85.13% on the SMARTSKINS dataset and 74.30% on the Dermofit Image Library.
- Successfully translated features between dermoscopic and macroscopic image domains, validated by visual and quantitative analysis.
Conclusions:
- The proposed method effectively leverages large dermoscopic datasets to enhance macroscopic skin lesion segmentation.
- This technique offers a viable solution for developing deep learning models for skin lesion analysis with limited macroscopic data.
- The results pave the way for more accessible and accurate AI-driven dermatological diagnostics.
