A team at the Technion in Haifa has rebuilt a dynamic breast MRI image in about a second, where the standard exam produces one every 1-2 minutes, and the work is now public, the university said on June 22, 2026. The method, called ELITE, was published in Nature Communications on May 19, 2026, and the lead author, Dr. Eddy Solomon, says the speed alone is not the point. The new frames let radiologists watch a contrast agent flow through tissue in near real time, which is the part that decides a diagnosis.
Breast MRI is the most sensitive imaging tool for women at high genetic risk, with accuracy above 90 percent against roughly 50-60 percent for ultrasound and mammography combined, according to the Technion announcement. What it gives up is time: a single dynamic scan can stretch past an hour. The Technion team and collaborators at Weill Cornell Medical College and NYU want the scan to keep up with the biology, not the other way around.
A One-Second Frame Inside an Hour-Long Scan
ELITE stands for Enhanced Locally low-rank Imaging for Tissue contrast Enhancement, and the team describes it as a radial MRI reconstruction framework for dynamic contrast-enhanced imaging. The idea is to rebuild a usable picture one image per second, instead of once every minute or two, by combining a mathematical model of how tissue behaves over time with a small deep-learning network that cleans up what the model cannot finish on its own. The work is in the open-access ELITE MRI paper in Nature Communications, dated May 19, 2026.
Solomon is the corresponding author and runs the work from the Technion’s Faculty of Biomedical Engineering in Haifa. The other authors are Sungheon Gene Kim and Jonghyun Bae at Weill Cornell Medical College, and Linda Moy, Laura Heacock, and Li Feng at New York University’s Center for Advanced Imaging Innovation and Research. Funding came from the NIH and the Radiological Society of North America, listed in the paper’s acknowledgements.
The Technion presented the work publicly on June 22, 2026, in an announcement that tied the publication to a year of follow-up work on a 300-scan breast MRI dataset Solomon and his NYU collaborators released in Radiology: Artificial Intelligence. The dataset, the team says, was built to give AI methods a fair test in this corner of imaging, and ELITE is the framework that plugs into the workflow the data was designed for.
Why Speed Is MRI’s Hardest Problem
An MRI does not take a photograph. It collects raw frequency data, point by point, into a grid called k-space, and only afterward does a computer turn that grid into the image a radiologist reads. Filling the whole grid is what makes a scan slow, and a full exam can stretch past an hour. The shortcut is to collect less data and let a computer fill in the rest, a move called undersampling.
The danger is that a computer filling in gaps can do more than blur the picture. It can invent. In the 2020 fastMRI challenge, where 19 teams competed on the same NYU Langone brain data, the results paper described what the organizers called common failure modes. Several leading systems produced reconstructions that scored well on standard automated quality checks but contained structures that were not in the original anatomy, a behaviour the team labelled hallucination.
We also identify common failure modes across the submissions, highlighting areas of need for future research in the MRI reconstruction community.
That warning comes from the 2020 fastMRI challenge paper, written by a team led by Matthew Muckley of NYU Langone Health, in IEEE Transactions on Medical Imaging. The risk sits at the heart of dynamic breast MRI. The diagnostic information is in the contrast curve, the way a tissue lights up after dye is injected and fades out, and a hallucinated vessel or a smoothed-over tumor is not a small error. It is the wrong answer in the part of the scan a radiologist is supposed to trust the most.
Math First, AI to Clean Up
ELITE’s design is a reaction to that history. Instead of handing the whole reconstruction job to a neural network and asking it to invent missing pixels, the team built the system around a mathematical model of how tissue behaves over time. A technique called locally low-rank subspace modeling captures the slow, structured parts of the contrast curve directly from the undersampled data. A smaller ResNet deep-learning network then comes in to remove the noise and streak artifacts the model cannot handle on its own.
The ability to track the movement of the contrast agent almost continuously will allow physicians to identify small tumors more accurately, better distinguish between benign and malignant tumors, and more precisely characterize biological tumor properties such as blood flow and vascular permeability.
That is the lead author, Dr. Eddy Solomon of the Technion’s Faculty of Biomedical Engineering, speaking in the university’s June 22, 2026 announcement. The sequencing he describes, physics first, AI second, is a deliberate choice for a clinical setting. The mathematics carries the diagnostic weight, and the AI tidies up the rest. the 2020 fastMRI challenge results are the reason the team is being careful, not because the method cannot be more aggressive, but because aggressive reconstruction in breast MRI can quietly change the answer.
| Feature | Standard dynamic breast MRI | ELITE (Technion method) |
|---|---|---|
| Temporal resolution | One image every 1-2 minutes | One image per second |
| Core reconstruction | Compressed sensing or GRASP variant | Locally low-rank subspace modeling |
| Role of the neural network | Optional denoising | ResNet denoising and artifact cleanup only |
| Primary risk in undersampling | Blur, residual aliasing | Residual artifacts; physics limits hallucination |
| Clinical use today | Standard in high-risk screening | Research validation; not yet routine |
54 Patients, Visible Tumors
The team tested the framework on 54 patients using the publicly available fastMRI breast dataset, and the results, published May 19 in Nature Communications, describe substantial improvements in contrast-to-noise ratio and noise reduction, with flexible temporal resolution that reaches the one-second mark. The paper also reports that the method extends to neck and brain imaging, where dynamic contrast-enhanced MRI is also used.
According to the Technion announcement, the clinical evaluation found improved tumor visibility compared with existing methods, exceptionally high image quality, and high diagnostic sensitivity. Shorter scan times, the team argues, will let a given MRI machine serve more women per day, a constraint that matters most where scanners are scarce. The headline numbers to keep in mind:
- One second per image, against 1-2 minutes in a standard dynamic scan
- 54 patients in the validation study
- 90% accuracy for dynamic breast MRI in high-risk screening
- 50-60% accuracy for mammography and ultrasound combined
- 1-2 minutes per image in a standard dynamic breast MRI
Why Breast Cancer Is the Right Test Case
Dynamic contrast-enhanced breast MRI works by injecting a contrast agent and watching how tissue takes it up. Malignant tissue tends to draw the agent in quickly and release it quickly. Benign tissue fills more slowly and steadily. The diagnostic curve sits in that difference, and reading it well needs fine detail and fine timing at the same moment. A scan that returns one image every 1-2 minutes can show the curve; a scan that returns one per second can show its shape.
The global stakes are large. In 2022, an estimated 2.3 million women were diagnosed with breast cancer and 670,000 died from it, according to the World Health Organization, and the gap between rich and poor countries is wide. Mortality in countries with a low Human Development Index is nearly twice as high as in countries with a very high one, even though the disease itself is no rarer.
The Technion’s framing of the new method is pointed: the limit on MRI access has rarely been the science, and it has often been scanner time. Cutting a dynamic exam by an order of magnitude, even with the validation work still ahead, is the kind of capacity gain that can shift who gets screened. the global 2022 breast cancer fact sheet from the WHO is the source for both numbers.
An Israeli Pipeline, Not a Single Lab
The advance lands inside a country that has quietly built a serious footprint in medical-imaging AI. The Israel Innovation Authority counts nine Israeli companies with FDA approvals in the field, totaling around 60 clearances, close to 6 percent of the global total. Aidoc, an Israeli clinical AI company, said the FDA had cleared its foundation-model triage tool, CARE, which combines 11 newly cleared indications with 3 existing ones to flag acute findings in a single abdomen CT workflow; the same release puts Aidoc in more than 1,600 medical centers worldwide. Other Israeli companies ship AI imaging products in the same export lanes.
The Technion itself now ranks first in Europe, first in Israel, and 21st worldwide for AI research, according to the international CSRankings index, based on 2005-2025 data, published by the university on February 18, 2026. In the subfield of machine learning, the Technion ranks first in Israel, first in Europe, and tenth worldwide. Solomon’s lab sits in the Faculty of Biomedical Engineering in Haifa, and his collaborators on the ELITE paper are at Weill Cornell and NYU.
The Israeli cluster is the part of the story that the headline number does not capture. the Israeli medical imaging AI landscape as a whole is one of the densest in the world, and the same export pipeline that ships Aidoc is now also shipping peer-reviewed methods in Nature Communications.
That is also why the Technion’s own announcement of ELITE reads as a recruiting moment, not just a press release. A frontier researcher choosing Haifa over the usual coastal hubs is the kind of detail that turns a method into an institutional signal.
The Conservative Choice Built into ELITE
ELITE’s design is conservative in a way the broader AI-in-imaging wave is not. The neural network does not decide what the image is. It cleans up what the physics cannot finish, and the physics is grounded in a model of tissue behavior that does not learn from training data on its own.
That choice has a cost. A purely physics-driven reconstruction is slower to ship than a one-size-fits-all deep-learning shortcut, and it does not promise the dramatic acceleration that some AI marketing has implied for MRI. The 2020 fastMRI challenge made the field cautious, and ELITE’s design is the cautious answer.
For breast cancer, the trade is the right one. A method that misses a small tumor by hallucinating it away is worse than a slower method that does not. ELITE is the framework the team is now testing across hospitals, and the validation studies the team has flagged are the next step. Technion’s June 22 announcement of the work lays out the schedule.
Frequently Asked Questions
What does ELITE actually do in a breast MRI?
It rebuilds the dynamic contrast-enhanced image about once a second, using a mathematical model of how tissue takes up and releases the contrast agent over time, then runs a small ResNet network to remove noise and streak artifacts from the undersampled raw data. The diagnostic information is read off the contrast curve in the rebuilt frames.
How much faster is it than a standard dynamic breast MRI?
The Technion team says a traditional dynamic breast MRI exam produces one image every 1-2 minutes at best. ELITE produces one image per second. The team demonstrated the gap on 54 patients in the paper published in Nature Communications on May 19, 2026.
Is ELITE approved for routine hospital use?
No. The work is a peer-reviewed research result, published in Nature Communications on May 19, 2026 and presented by the Technion on June 22, 2026. Wider clinical use will depend on the validation studies the team has flagged as the next step.
Why does time resolution matter in breast MRI?
The diagnostic information in a dynamic breast MRI is the contrast curve, the way a tissue lights up after dye is injected and fades out. A malignant lesion tends to take up the dye quickly and release it quickly, and reading that difference well is what the scan is for. A faster frame rate captures the shape of the curve more faithfully.
Can the same framework be used in other parts of the body?
Yes. The Nature Communications paper reports that the method also improves neck and brain imaging, where dynamic contrast-enhanced MRI is also used, and the team describes ELITE as a viable alternative for other DCE-MRI applications.
Disclaimer: This article is for informational purposes only. Medical imaging research and its clinical adoption are subject to regulatory approval, professional medical judgment, and individual patient circumstances. The figures and results described reflect the state of the research as of publication. Readers should consult a qualified medical professional for guidance on screening, diagnosis, and treatment options.
