Related Experiment Video
Updated: Mar 18, 2026
09:10
Correlative Light and Electron Microscopy CLEM as a Tool to Visualize Microinjected Molecules and their Eukaryotic Sub-cellular Targets
Published on: May 4, 2012
19.8K
Generating synthetic CEM from low-energy images using deep learning: A future without contrast media? A
Konstantinos Zormpas-Petridis1,2,3, Reza Kalantar4,5, Ludovica Iaccarino6
1Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Rome, Italy. konstantinos.zormpaspetridis@policlinicogemelli.it.
European Radiology Experimental
|March 16, 2026
Summary
Deep learning can create synthetic iodine-enhanced mammograms from low-energy images, potentially allowing radiologists to perform tasks like lesion detection without contrast injection.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Deep Learning for Mammography
Background:
- Contrast-enhanced mammography (CEM) provides valuable diagnostic information but requires iodine contrast injection.
- Low-energy CEM images lack the contrast enhancement needed for optimal lesion detection.
- Artifacts, such as the 'halo' artifact, can complicate the interpretation of clinical iodine-enhanced mammograms.
Purpose of the Study:
- To develop and validate a deep learning model for generating synthetic iodine-enhanced mammograms from low-energy CEM images.
- To assess the visual similarity and diagnostic utility of synthetic images compared to clinical iodine-enhanced images.
- To evaluate the model's ability to improve image quality by reducing artifacts.
Main Methods:
- A two-dimensional cycle-generative adversarial network was trained on 390 paired low-energy and iodine-enhanced CEM images from 100 patients.
- The model was validated on 40 test patients, evaluating image quality using contrast-to-noise ratio (CNR), mean absolute error (MAE), and similarity index metric (SSIM).
- Radiologists assessed background parenchymal enhancement (BPE) and lesion detection on both clinical and synthetic images, noting artifact presence.
Main Results:
- High correlation (MAE r=0.99, SSIM r=0.80) was observed between clinical and synthetic iodine-enhanced images regarding changes from low-energy images.
- The synthetic images demonstrated improved CNR compared to clinical images and effectively corrected the 'halo' artifact present in over 50% of clinical images.
- Radiologist performance on synthetic images showed high accuracy for BPE (85.8%) and comparable lesion detection (79.4% vs. 89.4% accuracy, 72.1% vs. 87.4% sensitivity, 90.0% vs. 92.3% specificity) to clinical images.
Conclusions:
- Deep learning effectively generates visually realistic synthetic iodine-enhanced mammograms from low-energy CEM images.
- These synthetic images enable radiologists to perform key clinical tasks, including lesion detection and BPE evaluation, without contrast administration.
- The developed model offers a promising approach to leverage the benefits of CEM while eliminating the need for contrast media and improving image quality.
