Style Transfer as Data Augmentation: Evaluating Unpaired Image-to-Image Translation Models in Mammography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models show promise for breast cancer detection in mammograms.
• Overfitting and poor generalizability limit clinical application due to domain differences.
• Data augmentation, including image-to-image translation, can enhance model generalizability.

Purpose of the Study:

• To evaluate style transfer algorithms for image-to-image translation in mammography.
• To identify key aspects for assessing style transfer model performance, especially without ground truth.
• To compare CycleGAN and SynDiff models for unpaired image translation across diverse mammography datasets.

Main Methods:

• Utilized cycle-consistent generative adversarial networks (CycleGAN) and diffusion-based SynDiff models.
• Performed unpaired image-to-image translation across three distinct mammography datasets.
• Analyzed the advantages, disadvantages, and unique contributions of various performance metrics.

Main Results:

• Model performance evaluation is complex, particularly without ground truth.
• Certain metrics may be unsuitable due to undesirable model performance aspects.
• Various metrics assess distinct facets of model performance, necessitating a multi-metric approach.

Conclusions:

• Comprehensive assessment of image-to-image translation models requires multiple metrics.
• Effective evaluation is crucial for trustworthy augmentation of mammography datasets.
• Improved model generalizability, equity, and performance can enhance patient outcomes.