The Radiation Dose Problem PET/CT is a powerful tool for lung cancer screening and staging, but the combined radiation from both the PET radiotracer and the CT component exposes patients to meaningful doses. For screening programs targeting healthy or high-risk individuals, minimizing radiation exposure while preserving diagnostic quality is a key safety and feasibility requirement.
CNN Denoising as the Solution This study tested whether convolutional neural network (CNN)-based denoising of low-dose (decimated) PET/CT images could restore image quality to levels sufficient for reliable lung lesion detection. If confirmed, patients could receive significantly lower radiation doses without sacrificing the diagnostic information clinicians need.
Study Design 49 patients underwent standard-dose PET/CT. Images were mathematically decimated (count-reduced) to simulate low-dose acquisitions at various dose fractions. A CNN denoiser was trained to map low-dose images back to standard-dose quality. Three radiologist observers then evaluated lesion detectability across all image types in a blinded observer study.
Key Result Across 588 reconstructions evaluated, CNN-denoised low-dose images achieved lesion detectability comparable to standard-dose scans at meaningful dose reductions. This observer-based validation is a critical step toward clinical adoption because it demonstrates real-world diagnostic utility, not just technical image quality metrics.
PET/CT's Role in Lung Cancer [18F]-FDG PET/CT combines metabolic information from the PET component with anatomical detail from CT, making it valuable for detecting, staging, and monitoring lung cancer. Compared to CT alone, PET/CT improves sensitivity for mediastinal lymph node involvement and distant metastases.
Radiation Dose Concerns A standard whole-body PET/CT delivers an effective dose of approximately 12-25 mSv - considerably higher than CT alone. For lung cancer screening programs where thousands of scans are performed annually on healthy at-risk populations, this cumulative dose is a legitimate concern that has slowed PET adoption in screening protocols.
Prior Approaches to Dose Reduction Previous strategies for PET dose reduction included using fewer administered activity (MBq), shorter scan durations, or improved reconstruction algorithms. However, each approach degrades image quality and reduces lesion signal-to-noise ratio. CNN-based denoising offers an alternative pathway - acquire at low dose and computationally restore quality afterward.
Observer-Based vs. Technical Metrics Image quality can be measured technically using metrics like SNR (signal-to-noise ratio) or SSIM (structural similarity), but the clinically relevant question is whether radiologists can actually see what they need to see. This study used a human observer framework specifically to answer that clinical question rather than a technical one.
Count Decimation Protocol Standard-dose PET images were decimated by randomly discarding detected photon counts to simulate acquisitions at fractions of the original dose (e.g., 50%, 25%, 12.5%). This mathematical simulation avoids the ethical and practical challenges of actually irradiating subjects at multiple dose levels while producing physically realistic low-dose noise characteristics.
CNN Denoiser Architecture The denoising network was trained as a supervised regression task: pairs of low-dose (decimated) and standard-dose images were used to teach the CNN to predict the clean image from the noisy input. The architecture was optimized for PET image characteristics including Poisson noise statistics that differ from the Gaussian noise assumed by many general-purpose denoisers.
Observer Study Protocol Three radiologists with subspecialty PET/CT experience independently rated lung lesion detectability on blinded image sets that included: original standard-dose images, decimated low-dose images, and CNN-denoised versions of those low-dose images. A 5-point confidence scale was used for each lesion at each dose level.
Statistical Analysis Receiver Operating Characteristic (ROC) analysis was applied to observer ratings, with AUC used to compare detectability across image types. A mixed-effects model accounted for correlations within observers and within patients across multiple lesions and dose levels. Inter-observer agreement was quantified to assess reliability of the observer ratings.
Detectability Preservation Across the 588 reconstructions evaluated (49 patients x multiple dose levels x image types), CNN-denoised images maintained lesion detectability significantly closer to standard dose than unprocessed low-dose images at equivalent dose levels. The gap in observer AUC between denoised and standard-dose images was substantially smaller than the gap between raw low-dose and standard-dose.
Dose-Dependent Performance At modest dose reductions (50%), CNN denoising essentially closed the performance gap completely - denoised images were radiologist-equivalent to full-dose scans. At more aggressive reductions (25% dose), a small but statistically significant detectability gap remained, indicating a practical lower bound on dose reduction without quality compromise.
Lesion Size Effects Smaller lesions (sub-centimeter) showed greater sensitivity to dose reduction than larger lesions, as expected from PET physics. CNN denoising provided the greatest benefit for these small lesions where noise is most problematic, suggesting the tool may be especially valuable for the challenging sub-centimeter nodules most relevant to early lung cancer detection.
Inter-Observer Agreement Radiologist agreement was higher on standard-dose and CNN-denoised images than on raw low-dose images, indicating that denoising reduced the ambiguity that causes observer disagreement - a practical quality indicator beyond raw detectability statistics.
Screening Program Impact If PET/CT dose can be halved without diagnostic loss, the radiation burden of PET-based lung cancer screening becomes far more defensible for population-level programs. The effective dose could approach that of standard diagnostic CT, removing the primary safety barrier to wider PET/CT adoption in screening settings.
Patient-Level Benefits For individual patients undergoing surveillance imaging over multiple years - a common scenario for lung cancer survivors or high-risk individuals - cumulative dose savings are substantial. Reducing each scan by even 30-40% translates into several full chest CT equivalents saved over a 5-year surveillance period.
Healthcare System Scalability Lower administered PET radiotracer doses reduce scanner occupancy time and radiotracer preparation costs. This has system-level implications for PET center capacity, potentially enabling more patients to be scanned with the same cyclotron and scanner infrastructure.
Workflow Considerations CNN denoising can be applied as an automated post-processing step without radiologist intervention, integrating into existing PACS and reconstruction workflows. The additional processing time (seconds per scan on modern GPU hardware) does not meaningfully impact clinical turnaround time.
Simulated vs. Acquired Low Dose Count decimation mathematically simulates low-dose acquisition but does not perfectly replicate physically acquired low-dose scans, which have additional noise sources from scattered radiation and detector physics. Validation with prospectively acquired true low-dose scans is needed to confirm that simulation-trained denoisers generalize to real acquisitions.
Sample Size 49 patients is a modest cohort for training and validating a CNN denoiser. While the observer study provided robust statistical power through multiple lesion and dose-level observations, the CNN itself may not have seen sufficient diversity in lesion types, sizes, and backgrounds. Larger training datasets including different scanners and patient body types are needed.
Single Scanner and Protocol All scans were acquired on the same scanner with the same protocol. PET/CT image quality varies substantially across scanner generations, manufacturers, and reconstruction software. The CNN may not generalize without retraining when deployed on different scanner hardware, which is a practical barrier to multicenter adoption.
Future Directions Future work should include prospective acquisition of truly low-dose scans to validate simulation-based findings, expand to multiple scanner models, test on lesion subtypes beyond nodules (e.g., ground-glass opacities, pleural lesions), and explore whether CNN-enhanced low-dose PET impacts clinical management decisions compared to standard dose in a controlled trial.