[Paper Review] MelanoGANs: High Resolution Skin Lesion Synthesis with GANs

Christoph Baur, Shadi Albarqouni+1 more|arXiv (Cornell University)|Apr 12, 2018

Generative Adversarial Networks and Image Synthesis18 references44 citations

TL;DR

The paper compares DCGAN, LAPGAN, and novel DDGAN variants to synthesize 256x256 skin lesion images from a small ISIC 2017 dataset, and demonstrates use of synthetic high-resolution samples to address class imbalance in skin lesion classification.

ABSTRACT

Generative Adversarial Networks (GANs) have been successfully used to synthesize realistically looking images of faces, scenery and even medical images. Unfortunately, they usually require large training datasets, which are often scarce in the medical field, and to the best of our knowledge GANs have been only applied for medical image synthesis at fairly low resolution. However, many state-of-the-art machine learning models operate on high resolution data as such data carries indispensable, valuable information. In this work, we try to generate realistically looking high resolution images of skin lesions with GANs, using only a small training dataset of 2000 samples. The nature of the data allows us to do a direct comparison between the image statistics of the generated samples and the real dataset. We both quantitatively and qualitatively compare state-of-the-art GAN architectures such as DCGAN and LAPGAN against a modification of the latter for the task of image generation at a resolution of 256x256px. Our investigation shows that we can approximate the real data distribution with all of the models, but we notice major differences when visually rating sample realism, diversity and artifacts. In a set of use-case experiments on skin lesion classification, we further show that we can successfully tackle the problem of heavy class imbalance with the help of synthesized high resolution melanoma samples.

Motivation & Objective

• Motivate high-resolution skin lesion synthesis with GANs using small datasets in medical imaging.
• Compare DCGAN, LAPGAN, and a proposed DDGAN architecture for 256x256 synthetic skin lesion generation.
• Assess how synthetic high-resolution lesions affect classifier performance on imbalanced data.
• Evaluate whether learned upsampling and different discrimination strategies influence realism and diversity of samples.

Proposed method

• Implement DCGAN, LAPGAN, and DDGAN variants to generate 256x256 skin lesion images from noise.
• DDGAN uses a single noise source and image discrimination with multi-level discriminators to enable end-to-end training.
• DDGAN variants include upsampling (learned via non-parametric upsampling) and deconvolution-based upsampling.
• LAPGAN relies on multiple noise sources and residual discrimination at pyramid levels.
• Evaluation uses histogram-based distribution similarity (JS Divergence, Wasserstein Distance) and qualitative visual assessment.

Experimental results

Research questions

• RQ1 Can high-resolution (256x256) skin lesion images be realistically synthesized from a small dataset across GAN architectures?
• RQ2 How do DCGAN, LAPGAN, and DDGAN compare in terms of distribution similarity, diversity, sharpness, and artifacts in generated lesions?
• RQ3 Do synthesized high-resolution melanoma samples help mitigate severe class imbalance in skin lesion classification?
• RQ4 Is learned upsampling via deconvolution advantageous over standard upsampling in DDGAN for this task?

Key findings

• All models approximate the real data distribution; however, visual realism, diversity, and artifact patterns differ across models.
• DCGAN best matches intensity distribution but shows checkerboard artifacts and less diversity.
• LAPGAN yields high sample diversity but introduces high-frequency artifacts.
• DDGAN with image discrimination (upsampling) closely matches the real distribution with fewer high-frequency artifacts than LAPGAN and is easier to train than LAPGAN.
• In use-case classification, synthetic melanoma samples from LAPGAN and DDGAN variants improve performance under class imbalance, sometimes surpassing models trained on full real data.