Kidney cancer is increasingly detected as small, localized masses - often discovered accidentally during CT scans performed for other reasons. Many of these small renal masses (SRMs), defined as tumors less than 4 cm in size, grow slowly and may be managed without immediate surgery through a strategy called active surveillance: periodic imaging to monitor whether the tumor is growing and to trigger treatment only if it does.
Active surveillance is now endorsed by major guidelines for selected patients, including older or medically frail individuals and, increasingly, younger patients with small, low-risk tumors. The standard imaging tool for surveillance is contrast-enhanced CT, which provides reliable measurements of tumor size and shape. However, repeated CT scans over years of surveillance deliver cumulative radiation doses that may themselves carry a small but meaningful cancer risk.
Estimates suggest that a 10-year active surveillance regimen involving 13 CT scans could deliver a cumulative dose of approximately 83 millisieverts (mSv) - exceeding a 55-mSv threshold above which cancer risk from radiation begins to increase. There is therefore a genuine medical need to reduce radiation exposure during surveillance while preserving the imaging quality necessary to accurately measure small changes in tumor size.
This multi-institutional study analyzed CT scans from 70 patients undergoing active surveillance for SRMs at a Norwegian university hospital. Rather than exposing patients to additional radiation to capture real low-dose scans, the researchers used a validated computational technique to simulate reduced-dose images from existing routine-dose scans. This generated images equivalent to 75% dose reduction (LD75) and 90% dose reduction (LD90) without any additional patient radiation exposure.
To address the image noise that increases at lower dose levels, the researchers applied a commercially available deep learning-based denoising (DLD) algorithm called ClariCT.AI. This software uses a U-Net type neural network trained on over 1 million pairs of noisy and clean CT images from 24 different scanner models. It was applied to both the 75% and 90% reduced-dose image sets, creating two additional datasets: LD75-DLD and LD90-DLD.
Nine radiologists from three university hospitals independently evaluated all 350 CT datasets (70 patients across 5 image sets) for tumor diameter, tumor proximity to the collecting system, and tumor shape irregularity - the key parameters used to guide surveillance decisions. This multiobserver design allowed robust assessment of how much measurement variability exists at each dose level, expressed as the limits of agreement with the mean (LOAM).
The key finding was that LD75 CT scans (routine dose reduced by 75%) produced tumor size measurements in agreement with routine-dose scans, with a LOAM of plus or minus 2.4 mm compared to plus or minus 2.2 mm for routine dose. These overlapping values indicate that the slight increase in measurement variability at 75% reduced dose is clinically acceptable - it remains within the 5-mm-per-year growth threshold commonly used to trigger intervention decisions in surveillance protocols.
At 90% dose reduction without denoising, measurement reproducibility deteriorated significantly (LOAM of plus or minus 3.0 mm), indicating that this level of dose reduction alone is too aggressive for reliable surveillance. However, when the deep learning denoising algorithm was applied to the 90% reduced-dose images, measurement reproducibility was restored to plus or minus 2.4 mm - matching the performance of 75% reduction without denoising.
Observer agreement for tumor proximity to the collecting system and tumor shape irregularity was similar across all dose levels (no statistically significant differences). Subjective image quality ratings were highest for LD75-DLD images, which observers actually rated better than routine-dose scans - a remarkable finding that reflects the deep learning algorithm's ability to reduce noise while maintaining image texture. This means that applying deep learning denoising to a 75% reduced-dose scan produces images that radiologists find easier to read than standard routine-dose CT.
The clinical significance of these findings is substantial for patients undergoing long-term active surveillance. Using the average routine dose of 6.4 mSv per scan from this study, a 10-year surveillance program with 13 scans would deliver 83 mSv total. Applying a 75% dose reduction would reduce this to approximately 20.8 mSv, and a 90% reduction with deep learning denoising would bring it down to approximately 8.3 mSv - a dramatic reduction in cumulative radiation exposure.
This matters because the patients most likely to benefit from active surveillance are also those who will be monitored for the longest periods. Younger patients who are candidates for surveillance face decades of potential exposure if radiation doses are not reduced. The ability to cut lifetime radiation burden by as much as 90% without sacrificing diagnostic accuracy represents a meaningful improvement in the safety profile of surveillance-based management.
The deep learning denoising algorithm used in this study is vendor-agnostic - it works with CT scanners from multiple manufacturers and does not require any special hardware. This makes it practical for implementation across diverse clinical settings, including hospitals that have older CT equipment that cannot achieve deep learning-based reconstruction natively within the scanner itself.
This rigorous multiobserver study provides strong evidence that radiation dose during CT-based active surveillance of small kidney tumors can be safely and substantially reduced. A 75% dose reduction using standard iterative reconstruction is already feasible with conventional CT technology. Adding deep learning denoising enables a 90% dose reduction while maintaining measurement quality comparable to routine-dose scans.
For patients and families, this research offers reassurance that choosing active surveillance over immediate surgery does not necessarily mean accepting significant cumulative radiation exposure. As these reduced-dose protocols are incorporated into clinical guidelines and practice, the safety balance of active surveillance versus intervention will shift further in favor of surveillance for appropriate patients.
The authors acknowledge limitations including reliance on simulated rather than real low-dose scans, use of CT scanners from a single manufacturer, and exclusion of cystic renal masses. Future research should validate these findings using real low-dose CT acquisitions, across multiple scanner types, and including additional tumor types that are also monitored via active surveillance.