[论文解读] High-Resolution Mammogram Synthesis using Progressive Generative Adversarial Networks

Dimitrios Korkinof, Tobias Rijken+4 more|arXiv (Cornell University)|Jul 9, 2018

Generative Adversarial Networks and Image Synthesis参考文献 27被引用 41

一句话总结

该论文展示了通过 progressively trained GANs 生成的高分辨率合成全场数字化乳腺摄影（FFDMs），达到最高 1280x1024 像素，并解决医学图像合成中的训练稳定性问题。

ABSTRACT

The ability to generate synthetic medical images is useful for data augmentation, domain transfer, and out-of-distribution detection. However, generating realistic, high-resolution medical images is challenging, particularly for Full Field Digital Mammograms (FFDM), due to the textural heterogeneity, fine structural details and specific tissue properties. In this paper, we explore the use of progressively trained generative adversarial networks (GANs) to synthesize mammograms, overcoming the underlying instabilities when training such adversarial models. This work is the first to show that generation of realistic synthetic medical images is feasible at up to 1280x1024 pixels, the highest resolution achieved for medical image synthesis, enabling visualizations within standard mammographic hanging protocols. We hope this work can serve as a useful guide and facilitate further research on GANs in the medical imaging domain.

研究动机与目标

• Show feasibility of generating realistic high-resolution mammograms using progressively trained GANs.
• Improve training stability for high-resolution medical image synthesis.
• Provide guidance on data preprocessing, training strategies, and evaluation for FFDM generation.

提出的方法

• Adopt progressive growing of GANs to progressively increase image resolution during training.
• Use Wasserstein GAN with gradient penalty to stabilize training.
• Condition the generator on mammogram view (CC vs MLO) to improve realism.
• Preprocess data to fit 1280x1024 with hanging protocol preservation.
• Increase training data size from 150k to 450k images and adjust learning rate and discriminator updates to enhance stability.
• Evaluate outputs using metrics like Sliced Wasserstein distance and qualitative artifact analysis.

实验结果

研究问题

• RQ1 Can progressively trained GANs produce realistic, high-resolution FFDMs suitable for clinical visualization?
• RQ2 What training strategies stabilize GAN optimization at very high resolutions for medical images?
• RQ3 How does conditioning on mammogram view affect the realism of generated CC and MLO views?

主要发现

• Generated FFDMs reach 1280x1024 pixels, the highest resolution reported for medical image synthesis in the paper.
• The synthetic images exhibit substantial inter-image variability and realism, with identifiable artifacts and some view-specific challenges.
• MLO views are harder to model due to higher anatomical variation.
• Artifacts such as transitions to larger sizes and implant reproduction occur but can be mitigated with training adjustments.
• Increasing training data and conditioning on view improve generation stability and realism.
• Training on multi-GPU infrastructure (8× V100) can achieve the target resolution within extended training times.