Generative adversarial networks based skin lesion segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Accurate skin cancer diagnosis relies on precise segmentation of skin lesions from dermoscopic images.
• Computer-aided diagnosis (CAD) tools can assist clinicians by automating this segmentation process.

Purpose of the Study:

• To introduce a novel adversarial learning-based framework, Efficient-GAN (EGAN), for automated skin lesion segmentation.
• To develop a lightweight version, Mobile-GAN (MGAN), for efficient deployment in resource-constrained environments.

Main Methods:

• Proposed EGAN framework utilizing an unsupervised generative network for lesion mask generation.
• Generator module features a top-down squeeze excitation-based compound scaled path and an asymmetric lateral connection-based bottom-up path.
• Implemented a discriminator module and a morphology-based smoothing loss to ensure accurate and smooth lesion boundaries.

Main Results:

• EGAN achieved state-of-the-art performance on the International Skin Imaging Collaboration Lesion Dataset.
• Achieved a Dice coefficient of 90.1%, Jaccard similarity of 83.6%, and accuracy of 94.5%.
• Mobile-GAN (MGAN) demonstrated comparable performance with significantly fewer training parameters and faster inference.

Conclusions:

• The proposed EGAN framework significantly enhances skin lesion segmentation accuracy.
• MGAN offers a computationally efficient alternative for real-time applications and low-resource settings.
• These advancements hold promise for improving computer-aided diagnosis in dermatology.