Reducing radiation exposure in computed tomography (CT) remains a core objective in clinical imaging, particularly where patients require repeated scans. In oncological imaging, however, this must be carefully balanced against the need to detect subtle abnormalities, such as small liver lesions. Deep learning reconstruction (DLR) algorithms have emerged as a potential tool to reduce noise and improve image clarity, thereby supporting dose reduction. A recent prospective study compared the diagnostic performance of high-strength DLR with iterative reconstruction (IR) across standard and reduced-dose abdominal CT, focusing specifically on the detection of liver lesions.
Image Quality Improvements with DLR
DLR demonstrated considerable improvements in quantitative image quality parameters across all dose levels. Compared with IR, DLR reduced image noise by approximately 40% and enhanced contrast-to-noise ratio (CNR) by 67%. These metrics remained consistent across standard-dose, medium-dose and low-dose CT scans, with no significant changes in mean CT numbers between reconstruction methods at any dose level. Slight reductions in liver CT numbers were noted at lower doses, attributed to minor delays in scan timing rather than algorithmic limitations. Despite this, image noise and CNR were consistently better with DLR, underscoring its capability to enhance visual image quality, particularly in lower dose conditions.
Must Read: AI-Driven Imaging in Liver Disease
Diagnostic Performance Across Dose Levels
Although DLR effectively improved image quality, it did not offer superior diagnostic performance compared to IR when it came to liver lesion detection. At standard-dose levels, both reconstruction methods yielded similar detection rates of 83% for all lesions ≤ 20 mm. However, lesion detectability declined significantly at reduced-dose levels. Medium-dose DLR detected only 65.2% of lesions, while low-dose DLR identified just 58.9%. This pattern was particularly pronounced for small lesions. For lesions ≤ 10 mm, detection fell from 79.4% at standard-dose IR to 54.9% at medium-dose DLR and 45.1% at low-dose DLR. Similar reductions were observed in the subset of malignant lesions ≤ 10 mm, with detection rates dropping from 79.7% to 43.2% across the same range. Notably, detection performance remained stable for lesions > 10 mm, suggesting that DLR’s limitations are most apparent in the identification of smaller, lower-contrast targets.
Clinical Implications and Protocol Considerations
The study’s findings highlight important considerations for clinical CT protocol optimisation. While DLR can enhance image presentation and reduce subjective noise, these benefits do not necessarily equate to improved clinical performance when evaluating small liver lesions. In oncological imaging, where early identification of metastases is critical, the reduced sensitivity at lower doses may impact patient care decisions. Therefore, efforts to lower radiation dose using DLR must be approached cautiously in this context.
Nevertheless, the technology does show promise for certain clinical scenarios. When lesion size is less critical, such as in the monitoring of stable disease or in non-oncological evaluations, DLR may permit modest dose reductions without compromising clinical outcomes. Additionally, in cases involving multiple-phase imaging or frequent follow-up, particularly in younger or more radiosensitive populations, DLR offers an opportunity to minimise exposure without entirely sacrificing diagnostic confidence for larger lesions.
It is also important to note that this study evaluated one specific high-strength DLR algorithm from a single vendor. As such, the results may not be directly generalisable to other algorithms or clinical settings. Furthermore, although the comparison was made within consistent dose levels to mitigate confounding factors, a slight reduction in liver contrast enhancement due to delayed scanning for the lower-dose images could have influenced the results. However, the consistent performance decline across both reduced-dose scans suggests radiation dose, rather than timing, was the dominant influence.
While high-strength deep learning reconstruction improves image noise and contrast-to-noise ratio at standard and reduced radiation doses, it does not enhance the detection of small liver lesions compared to iterative reconstruction. Furthermore, diagnostic performance deteriorates significantly for lesions ≤ 10 mm as dose levels decrease, even with DLR. These findings reinforce the need for careful protocol tailoring when considering dose reduction strategies in CT imaging. Where smaller lesion detection is essential, particularly in oncological follow-up, caution is warranted in implementing low-dose protocols, even when using advanced reconstruction techniques. Conversely, where larger lesions are the primary concern, DLR may offer a viable path toward safer, lower-dose imaging without major compromises in diagnostic accuracy.
Source: European Radiology
Image Credit: iStock
