# Generating synthetic CEM from low-energy images using deep learning: A future without contrast media? A proof-of-concept study

## TL;DR

This study shows that deep learning can create synthetic mammograms that look like contrast-enhanced images without using contrast media, potentially reducing the need for injections.

## Contribution

A deep learning model that generates synthetic iodine-enhanced mammograms from low-energy images, avoiding contrast media use.

## Key findings

- Synthetic images showed high similarity to clinical images with MAE r=0.99 and SSIM r=0.80.
- Lesion detection accuracy on synthetic images was 79.4% with 72.1% sensitivity.
- The model reduced artifacts like the 'halo' effect present in clinical images.

## Abstract

We used deep learning to generate synthetic, resembling in appearance, iodine-enhanced, mammograms from low-energy contrast-enhanced mammography (CEM) images.

We retrospectively selected 140 CEM examinations. We trained a two-dimensional cycle-generative adversarial network on 390 images in 100 patients (195 breasts; 102 positive and 93 negative for lesion detection) using paired low-energy and iodine-enhanced images as input and output, respectively. We validated our model in 40 test patients (63 breasts; 37 positive and 26 negative for lesion detection) by calculating the contrast-to-noise ratio (CNR) for low-energy, synthetic, and clinical iodine-enhanced images and the mean absolute error (MAE) and similarity index metric (SSIM) between clinical and synthetic iodine-enhanced images regarding their changes from low-energy. Three radiologists scored (a-to-d) the test set images for background parenchymal enhancement (BPE) and lesion detection (yes/no) on clinical and synthetic images. The presence of artifacts was reported on all images.

We observed a high correlation between clinical and synthetic iodine-enhanced images regarding their changes from low-energy: MAE, r = 0.99; SSIM, r = 0.80. CNR was -0.015/-0.16 ± 0.23/0.05 (mean ± standard deviation) for clinical/synthetic, respectively. A “halo” artifact present in above 50% of the clinical iodine-enhanced images was corrected in the synthetic ones. On synthetic images, BPE (scores a–b versus c–d) was 85.8% accurate. Lesion detection accuracy was 89.4% and 79.4%, sensitivity 87.4 and 72.1%, and specificity 92.3% and 90.0% for clinical and synthetic images, respectively.

Deep learning holds the potential to generate synthetic iodine-enhanced mammograms from low-energy images.

Radiologists could perform some clinical tasks, such as lesion detection and BPE estimation on synthetic iodine-enhanced images, without contrast injection.

Our deep learning model generated synthetic iodine-enhanced images that visually resembled the clinical iodine-enhanced images.Radiologists could use the synthetic images to perform clinical tasks, such as lesion detection and BPE evaluation.Our model can improve image quality by removing common artifacts, including the breast-in-breast (halo).Our method is a way to combine the benefits of CEM while sparing the need for contrast media.

Our deep learning model generated synthetic iodine-enhanced images that visually resembled the clinical iodine-enhanced images.

Radiologists could use the synthetic images to perform clinical tasks, such as lesion detection and BPE evaluation.

Our model can improve image quality by removing common artifacts, including the breast-in-breast (halo).

Our method is a way to combine the benefits of CEM while sparing the need for contrast media.