Project

Ultra-Low Dose CT for Novel Screening Applications

Radiologic screening of the future: the case for computationally enhanced ultra-low dose CT — bringing any existing CT scanner down to the radiation level of a plain X-ray to unlock a new generation of population screening.

Dr. Arnaldo Mayer, PhD · Computational Imaging Lab, Sheba Medical Center · Gray Faculty of Medicine, Tel Aviv University

Axial chest CT slice shown three ways: ultra-low dose (noisy), normal dose, and the AI-denoised ultra-low dose result matching normal-dose quality

The ideal modality for imaging based screening does not exist...

“Screening is the process of using tests on a large scale to identify the presence of disease in apparently healthy people”^[1]. At population scale, the implemented screening technique must be accessible, low-cost, and safe. Yet in half a century the US has fielded only a handful of large-scale, reimbursed imaging-based screening programs — mammography (1976), osteoporosis DEXA (1998), and low-dose CT for lung cancer (2015).

Every existing modality trades strengths against weaknesses. For example, CT offers rich 3D information at low cost and high availability, but its ionizing radiation has kept it off the screening table for most applications.

Modality	Cost	2D / 3D	Repeatability	Ionizing radiation	Availability
X-ray	Low	2D	Medium	Very low	High
Ultrasound	Low	2D (mainly)	Low	None	High
CT	Low	3D	High	High	High
MRI	High	3D	High	None	Low
PET/CT-MR	High	3D	High	High	Low

Our approach: AI-enhanced ultra-low dose CT

The vision is to bring any existing, low-cost CT scanner down to the radiation level of a 2D X-ray — turning radiation exposure into a non-issue — and thereby enable entirely new screening applications.

The motivation is concrete: CT screening for lung cancer was proven to reduce mortality by ~20%^[2], and annual low-dose lung screening CT is reimbursed for at-risk populations in the US, amounting to some 17 million scans per year^[3]. But “low dose” is still ~1 to 1.5 mSv — the equivalent of about 10-15 chest X-rays. Following the ALARA principle (As Low As Reasonably Achievable), the goal is to close that gap.

Our AI-enhanced ULDCT delivers ~0.12 mSv — roughly a single chest radiography — while keeping the scan diagnostic. Much more information, for a similar dose.

Building a unique dataset

Supervised denoising needs paired ultra-low-dose and normal-dose scans of the same patient — a dataset that simply did not exist. There are two ways to obtain one:

Retrospective — add synthetic noise to existing normal-dose scans. One scan per patient, aligned by design, no consent required — the noise is simulated.
Prospective — acquire real pairs of ULD and normal-dose scans. Two scans per patient, requiring informed consent and non-rigid registration — the noise is real.

Real noise matters: we have shown^[4] that models trained on genuinely acquired ULD noise achieve significantly higher Peak-SNR, higher structural similarity (SSIM), and higher S3 sharpness (all statistically significant) than those trained on synthetic noise.

Peak-SNR of real vs synthetic noise training across five cross-validation folds and four denoising architectures — Models trained on genuinely acquired ULD noise (blue) consistently outperform those trained on synthetic noise (red) across five cross-validation folds and four denoising architectures (LD-CNN, DnCNN, RED-CNN, CNN-VGG) — higher Peak-SNR and SSIM (p<0.05) and higher S3 sharpness (p<0.01). Adapted from [4].

Structural similarity (SSIM) of real vs synthetic noise training across five cross-validation folds and four denoising architectures — Models trained on genuinely acquired ULD noise (blue) consistently outperform those trained on synthetic noise (red) across five cross-validation folds and four denoising architectures (LD-CNN, DnCNN, RED-CNN, CNN-VGG) — higher Peak-SNR and SSIM (p<0.05) and higher S3 sharpness (p<0.01). Adapted from [4].

The zero-sum game: the ULD scan uses only ε ≈ 5–10% of a normal dose. By reducing the normal dose by ε, which is too small to significantly affect image quality, the paired protocol keeps the patient's total exposure within a single normal-dose budget — a dose reduction 90-95% — which is what made scanning each patient twice ethically approvable.

Because a patient never holds exactly the same full inspiration twice, each ULD scan is brought into alignment with its normal-dose ground truth through affine + deformable registration (Elastix).

The same subject captured at two different lung inflation levels, illustrating why paired scans never align exactly and must be registered — The balloon allegory: two acquisitions of the same subject's lungs are never at identical end inspirium inflation — deformable registration brings the ULD scan into alignment with its normal-dose ground truth.

The denoising engine: the trick is in the loss!

The denoising network is trained with a perceptual loss — comparing denoised and ground-truth patches inside a learned embedding space rather than pixel-by-pixel. The lab's approach evolved from a convolutional perceptual loss (ISBI 2019^[5]) to the Perceptual Transformer Loss (PeTeL) built on a 3-D MaxViT backbone combining block and grid self-attention, which achieves superior quantitative and qualitative denoising (ISBI 2024^[6]).

Denoising framework: an ultra-low-dose patch passes through the denoising network; the denoised patch and the co-registered ground-truth patch are each embedded by a transformer, and the perceptual loss is the L2 distance between their embeddings — Denoising framework under Perceptual Transformer Loss (PeTeL).

One scan, four indications

A single AI-denoised ultra-low dose acquisition currently supports four different clinical indications.

Lung cancer

Screening & nodule follow-up ·
Cleared by Israeli MoH ^[7]^[8]

Calcium scoring

Coronary artery calcium scoring ·
Cleared by Israeli MoH ^[9]

Fatty liver

Hepatic steatosis assessment^[10]

Pneumonia

Infection in immunocompromised patients^[11]

0.12 mSv

AI-enhanced ULDCT dose — about 1.2 chest radiographs

>95%

Radiation dose reduction vs. normal-dose CT

Applications from a single ULD scan

Regulatory clearances

Bibliography

Bonita, R., Beaglehole, R. and Kjellström, T., 2006. Basic epidemiology. World Health Organization. ↩
The National Lung Screening Trial Research Team, 2011. Reduced lung-cancer mortality with low-dose computed tomographic screening. New England Journal of Medicine, 365(5), pp.395–409. doi:10.1056/NEJMoa1102873. ↩
US Preventive Services Task Force, 2021. Screening for lung cancer: US Preventive Services Task Force recommendation statement. JAMA, 325(10), pp.962–970. doi:10.1001/jama.2021.1117. ↩
Green, M., Marom, E.M., Konen, E., Kiryati, N. and Mayer, A., 2018. Learning real noise for ultra-low dose lung CT denoising. In International Workshop on Patch-based Techniques in Medical Imaging (pp. 3–11). Cham: Springer International Publishing. ↩
Green, M., Marom, E.M., Konen, E., Kiryati, N. and Mayer, A., 2019. Feature aggregation in perceptual loss for ultra low-dose (ULD) CT denoising. In 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI) (pp. 1635–1638). IEEE. ↩
Green, M., Marom, E.M. and Mayer, A., 2024. Perceptual transformer loss for the neural denoising of ultra-low dose CT. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI) (pp. 1–5). IEEE. doi:10.1109/ISBI56570.2024.10635771. ↩
Gorenstein, L., et al., 2023. A novel artificial intelligence based denoising method for ultra-low dose CT used for lung cancer screening. Academic Radiology. ScienceDirect. ↩
Kerpel, A., Marom, E.M., Green, M., Eifer, M., Konen, E., Mayer, A. and Betancourt Cuellar, S.L., 2021. Ultra-low dose chest CT with denoising for lung nodule detection. Israel Medical Association Journal, 23(9), pp.550–555. link. ↩
Klug, M., Shemesh, J., Green, M., Mayer, A., Kerpel, A., Konen, E. and Marom, E.M., 2022. A deep-learning method for the denoising of ultra-low dose chest CT in coronary artery calcium score evaluation. Clinical Radiology, 77(7), pp.e509–e517. doi:10.1016/j.crad.2022.03.005. ↩
Kirshenboim, Z., Marom, E.M., Mayer, A., Green, M., Klug, M. and Tau, N., 2022. Liver attenuation assessment in reduced radiation chest computed tomography. Journal of Computer Assisted Tomography, 46(5), pp.682–687. doi:10.1097/RCT.0000000000001340. ↩
Klug, M., Sobeh, T., Green, M., Mayer, A., Kirshenboim, Z., Konen, E. and Marom, E.M., 2025. Denoised ultra-low-dose chest CT to assess pneumonia in individuals who are immunocompromised. Radiology: Cardiothoracic Imaging, 7(2), e240189. doi:10.1148/ryct.240189. ↩