Small renal masses (SRMs), frequently detected through imaging, are often managed through active surveillance. This approach involves serial monitoring using contrast-enhanced CT, which is widely used due to its diagnostic precision, availability and speed. However, the repeated imaging required over time raises concerns about cumulative radiation exposure. Balancing the need for accurate assessment with the imperative to reduce radiation dose has become a key objective in optimising surveillance protocols. A recent multiobserver study explored whether lower-dose CT scans, both with and without deep learning–based denoising (DLD), could provide reliable tumour assessments comparable to routine-dose (RD) scans. The findings offer insights into how advanced image processing may support safer and more sustainable monitoring strategies.
Agreement in Tumour Size Assessment
The study assessed CT images from 70 patients under active surveillance for SRMs, using five imaging datasets: RD, 75% reduced dose (LD75), 90% reduced dose (LD90), and both LD levels with DLD enhancement (LD75-DLD and LD90-DLD). Radiologists evaluated each scan for tumour size, tumour nearness to the collecting system (TN) and tumour shape irregularity (TSI). Agreement between observers was analysed using limits of agreement with the mean (LOAM) and Gwet’s AC2 coefficient.
Results showed that LD75 scans were comparable to RD in measuring maximum tumour diameter, with LOAM values of ±2.4 mm and ±2.2 mm respectively. When DLD was applied, the measurements remained consistent even at the 90% reduced dose level, with LOAM at ±2.4 mm for LD90-DLD. Only the LD90 dataset without denoising displayed wider variability (±3.0 mm), suggesting that the absence of noise reduction impairs measurement reproducibility. These findings demonstrate that a 75% reduction in radiation does not adversely affect measurement consistency, and that denoising can compensate for the degradation observed at more aggressive dose reductions.
Image Quality and Observer Confidence
Beyond size measurements, the study evaluated subjective image quality and diagnostic confidence. Observers rated LD75-DLD highest overall, exceeding RD in terms of clarity, contrast and noise levels. LD90, in contrast, received the lowest scores, reflecting a noticeable decline in image quality due to increased noise. However, applying DLD to LD90 images substantially improved both perceived and measured quality. Quantitative analysis confirmed these ratings, with noise levels in renal tissue, cortex, muscle and fat significantly reduced in DLD-enhanced datasets. For instance, noise in the renal mass dropped from 51.1 HU in LD90 to 20.3 HU in LD90-DLD, approaching the RD level of 14.5 HU.
Must Read: AI-Driven Renal Function Estimation Using Non-Contrast CT
Contrast-to-noise ratio, a critical factor in distinguishing tumour boundaries, also improved with DLD. LD75-DLD achieved a ratio of 12.5, outperforming RD at 10.3. These enhancements translated into higher observer confidence, particularly in delineating tumour contours. Agreement for TN and TSI was consistently high across all datasets, including those with DLD, indicating that even ordinal features of renal mass morphology remain reliably assessable at reduced doses when denoising is applied.
Clinical Relevance and Study Considerations
The implications of these findings are substantial for clinical practice. Patients undergoing CT-based active surveillance may undergo more than a dozen scans over a decade, with RD protocols exposing them to an estimated cumulative dose of over 80 mSv. This exceeds the widely acknowledged safety threshold of 55 mSv, above which cancer risk is considered significantly increased. By implementing LD75 or LD90-DLD protocols, cumulative exposure could be reduced to 20.8 mSv or even 8.3 mSv, without compromising diagnostic reliability.
Nevertheless, the study recognises several limitations. The use of simulated LD datasets, while validated and vendor-agnostic, may not fully reflect the nuances of real-world scanning conditions. Additionally, only two CT systems from a single vendor were used, although the DLD model was trained across multiple platforms and reconstruction settings. The study did not address surveillance of cystic renal masses, which require a different diagnostic approach. Also, maximum axial diameter was used to measure tumour size, which may not fully capture complex tumour morphology. Despite these factors, the consistency across observers and close alignment with established clinical thresholds support the findings’ applicability.
The study demonstrated that a 75% dose reduction in CT imaging can be implemented safely for SRM surveillance when using conventional iterative reconstruction. Furthermore, incorporating deep learning–based denoising enables even more substantial reductions—up to 90%—while preserving measurement accuracy, image quality and diagnostic confidence. These results suggest that denoising technologies can support safer imaging regimens for patients requiring long-term monitoring. Clinical protocols should consider integrating such technologies to reduce radiation exposure while maintaining the high standards needed for effective surveillance of renal tumours.
Source: Radiology: Imaging Cancer
Image Credit: iStock
