Radiation Dose Challenge Lung cancer CT screening and nodule follow-up require repeat scans, creating cumulative radiation exposure over time. Ultra-low-dose CT (ULDCT) at approximately 0.16 mSv reduces this burden - compared to 1.77 mSv for standard-dose CT - but conventional reconstruction methods produce noisy images that compromise nodule detection.
Reconstruction Algorithms Hybrid iterative reconstruction (HIR) is the standard approach for reducing noise in low-dose CT images. Deep learning reconstruction (DLR), which uses neural networks trained on paired high- and low-dose CT images, is a newer approach that may more effectively remove noise while preserving diagnostically important image details.
Study Objective This prospective study directly compared DLR and HIR in chest ULDCT, evaluating image quality, nodule detection rate, and nodule measurement accuracy using a commercially certified AI-assisted nodule evaluation system, with standard-dose CT as the reference standard.
Key Finding DLR significantly outperformed HIR in all dimensions tested: lower image noise, higher radiologist quality scores, higher nodule detection rate (83.4% vs 74.0%), and more accurate volumetric measurement. DLR at ULDCT achieved image quality comparable to standard-dose CT.
Prospective Design Eighty-four participants who had incidental or suspected pulmonary nodules underwent both standard-dose CT (SDCT) and ULDCT sequentially on the same visit. SDCT at 120 kV with automatic mA served as the reference, and ULDCT at 120 kV and fixed 20 mA delivered an effective dose of 0.16 mSv - a 91% radiation dose reduction versus SDCT's 1.77 mSv.
Reconstruction Groups All ULDCT images were reconstructed using four settings: HIR standard (HIR-Std), HIR strong (HIR-Str), DLR standard (DLR-Std), and DLR strong (DLR-Str). The DLR algorithm used was Canon's Advanced Intelligent Clear-IQ Engine (AiCE). For comparison against SDCT, the best-performing HIR setting (HIR-Str) and best DLR setting (DLR-Str) were selected.
AI Nodule Evaluation System All images were analyzed using the Yitu AICare CT Chest system (Deepwise Inc.), a commercially certified AI platform approved by China's National Medical Products Administration. The AI system automatically detected, segmented, and measured all nodules, with a senior radiologist reviewing and confirming detections against the SDCT reference.
Measurement Metrics Nodule measurement accuracy was assessed by absolute percentage error (APE) - the unsigned deviation of ULDCT measurements from SDCT reference values - for both long diameter and volumetric measurements across nodule size subgroups (3-6 mm and 6 mm or larger) and type subgroups (subsolid, solid, calcified).
Objective Noise Reduction Lung tissue noise was 61.4 HU for SDCT, 61.5 HU for HIR-Str, and 55.1 HU for DLR-Str. DLR-Str achieved statistically lower noise than both HIR settings and even below standard-dose CT levels, demonstrating the superior denoising capability of the neural network-based algorithm.
Subjective Quality Scores Radiologist quality scores on a 5-point Likert scale showed SDCT scored 4.9-5.0, HIR-Str scored 3.6-3.7, DLR-Std scored 4.3-4.4, and DLR-Str scored 4.6, which was comparable to SDCT (p=0.054-0.16). Interobserver agreement was excellent (kappa=0.91), confirming consistent quality assessment.
DLR vs. HIR Quality Gap Both subjective and objective measures confirmed a substantial quality advantage for DLR over HIR at the same radiation dose. This gap explains the downstream differences in nodule detection performance - cleaner images allow the AI system to identify more subtle nodules.
BMI Independence DLR maintained its quality advantage in both normal weight and overweight/obese participants, with no significant difference in noise or quality scores between BMI subgroups. This confirms that DLR's noise reduction benefit is not limited to thin patients.
Overall Detection Rate Of 535 nodules detected by SDCT in 79 participants, DLR-Str detected 446 (83.4%) versus 396 (74.0%) for HIR-Str - a 9.4 percentage point advantage (p less than 0.001). DLR's detection rate exceeded 80%, while HIR-Str fell short of this threshold, with an odds ratio of 2.02 in multivariable analysis.
Small and Subsolid Nodule Advantage DLR's advantage was consistent across nodule sizes (3-6 mm: 80.9% vs 70.6%, p less than 0.001; 6 mm or larger: 93.4% vs 87.7%, p=0.03) and types (subsolid: 62.0% vs 43.8%, p less than 0.001; solid: 87.4% vs 80.1%, p less than 0.001). The particularly large gap for subsolid nodules is clinically important because these are at highest risk of malignant transformation.
Volumetric Measurement Accuracy DLR-Str showed significantly lower absolute percentage error in volume measurement than HIR-Str: 17.9% versus 19.5% (p less than 0.001). This improvement was consistent across both BMI groups, both nodule size groups, and for solid and calcified nodule types, though not for subsolid nodules.
Diameter Measurement No significant difference was found in long diameter measurement accuracy between DLR-Str and HIR-Str (p=0.70), suggesting that the volumetric measurement benefit of DLR is specific to 3D volume quantification rather than linear diameter assessment, which is less sensitive to reconstruction quality.
Near-Chest-Radiograph Radiation Doses At 0.16 mSv, ULDCT with DLR delivers radiation close to that of a chest X-ray (approximately 0.03-0.1 mSv). This opens the possibility of replacing annual chest radiography with ULDCT in follow-up programs for known nodule carriers, combining substantially lower dose than standard LDCT with better detection than plain radiography.
Subsolid Nodule Detection The marked detection advantage of DLR for subsolid nodules - ground-glass opacities and part-solid nodules - is especially clinically relevant. These nodules have higher malignant potential per unit size than solid nodules, and missing them at ULDCT would have the greatest consequences for patient outcomes.
Volumetric Growth Monitoring The improved volumetric measurement accuracy of DLR versus HIR matters for longitudinal nodule surveillance. Volume doubling time is a key criterion for determining whether a nodule is growing and needs intervention. More accurate baseline and follow-up volume measurements directly improve the reliability of growth rate calculations.
AI-DLR Synergy The combination of DLR's superior image quality and an AI nodule detection system creates a multiplicative benefit: cleaner images allow the AI to detect more nodules with fewer false positives, while the AI reduces radiologist reading time. This synergy supports the feasibility of high-volume ULDCT screening programs.
Single Vendor Limitation DLR was tested using only Canon's AiCE algorithm. Different manufacturers implement DLR differently, and performance characteristics - including noise reduction characteristics and their effect on AI detection - may vary across scanner platforms and vendor-specific DLR implementations.
BMI Range Limitation Only one participant had a BMI above 30 kg/m2, making it impossible to assess DLR performance in obese patients, who present the greatest noise challenge at ultra-low doses due to increased photon attenuation through larger body habitus.
AI-Only Nodule Assessment The study evaluated nodule detection using AI alone rather than combined AI plus radiologist visual reading. In clinical practice, AI typically serves as a second reader, and combined reading would likely improve detection rates for both HIR and DLR, potentially narrowing the gap.
Clinical Implementation Studies Demonstrating improved image quality and nodule detection in a research setting is only the first step. Prospective studies that track whether ULDCT DLR-based follow-up programs reduce the number of unnecessary interval scans, biopsies, and missed cancers compared to standard LDCT are needed to establish clinical value.