Dataset Information

Perceived Realism of High-Resolution Generative Adversarial Network-derived Synthetic Mammograms.

ABSTRACT:

Purpose

To explore whether generative adversarial networks (GANs) can enable synthesis of realistic medical images that are indiscernible from real images, even by domain experts.

Materials and methods

In this retrospective study, progressive growing GANs were used to synthesize mammograms at a resolution of 1280 × 1024 pixels by using images from 90 000 patients (average age, 56 years ± 9) collected between 2009 and 2019. To evaluate the results, a method to assess distributional alignment for ultra-high-dimensional pixel distributions was used, which was based on moment plots. This method was able to reveal potential sources of misalignment. A total of 117 volunteer participants (55 radiologists and 62 nonradiologists) took part in a study to assess the realism of synthetic images from GANs.

Results

A quantitative evaluation of distributional alignment shows 60%-78% mutual-information score between the real and synthetic image distributions, and 80%-91% overlap in their support, which are strong indications against mode collapse. It also reveals shape misalignment as the main difference between the two distributions. Obvious artifacts were found by an untrained observer in 13.6% and 6.4% of the synthetic mediolateral oblique and craniocaudal images, respectively. A reader study demonstrated that real and synthetic images are perceptually inseparable by the majority of participants, even by trained breast radiologists. Only one out of the 117 participants was able to reliably distinguish real from synthetic images, and this study discusses the cues they used to do so.

Conclusion

On the basis of these findings, it appears possible to generate realistic synthetic full-field digital mammograms by using a progressive GAN architecture up to a resolution of 1280 × 1024 pixels.Supplemental material is available for this article.© RSNA, 2020.

SUBMITTER: Korkinof D

PROVIDER: S-EPMC8043361 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:PurposeDeveloping deep learning algorithms for breast cancer screening is limited due to the lack of labeled full-field digital mammograms (FFDMs). Since FFDM is a new technique that rose in recent decades and replaced digitized screen-film mammograms (DFM) as the main technique for breast cancer screening, most mammogram datasets were still stored in the form of DFM. A solution for developing deep learning algorithms based on FFDM while leveraging existing labeled DFM datasets is a generative algorithm that generates FFDM from DFM. Generating high-resolution FFDM from DFM remains a challenge due to the limitations of network capacity and lacking GPU memory.MethodIn this study, we developed a deep-learning-based generative algorithm, HRGAN, to generate synthesized FFDM (SFFDM) from DFM. More importantly, our algorithm can keep the image resolution and details while using high-resolution DFM as input. Our model used FFDM and DFM for training. First, a sliding window was used to crop DFMs and FFDMs into 256 × 256 pixels patches. Second, the patches were divided into three categories (breast, background, and boundary) by breast masks. Patches from the DFM and FFDM datasets were paired as inputs for training our model where these paired patches should be sampled from the same category of the two different image sets. U-Net liked generators and modified discriminators with two-channels output, one channel for distinguishing real and SFFDMs and the other for representing a probability map for breast mask, were used in our algorithm. Last, a study was designed to evaluate the usefulness of HRGAN. A mass segmentation task and a calcification detection task were included in the study.ResultsTwo public mammography datasets, the CBIS-DDSM dataset and the INbreast dataset, were included in our experiment. The CBIS-DDSM dataset includes 753 calcification cases and 891 mass cases with verified pathology information, resulting in a total of 3568 DFMs. The INbreast dataset contains a total of 410 FFDMs with annotations of masses, calcifications, asymmetries, and distortions. There were 1784 DFMs and 205 FFDM randomly selected as Dataset A. The remaining DFMs from the CBIS-DDSM dataset were selected as Dataset B. The remaining FFDMs from the INbreast dataset were selected as Dataset C. All DFMs and FFDMs were normalized to 100μm × 100μm in our experiments. A study with a mass segmentation task and a calcification detection task was performed to evaluate the usefulness of HRGAN.ConclusionsThe proposed HRGAN can generate high-resolution SFFDMs from DFMs. Extensive experiments showed the SFFDMs were able to help improve the performance of deep-learning-based algorithms for breast cancer screening on DFM when the size of the training dataset is small.