Artificial intelligence iterative imaging reconstruction with low-dose CT scans for hepatic metastases Authors: Xiang Liu, Zhi-Lei Zhang, Xiao-Hui Qi, Jiang-Yang Pan, Xue-Li Fan, Jing Liu, Yi-Ting Sun, Qi Wang Published: 2026 ABSTRACT AbstractContrast-enhanced CT is commonly used in the evaluation of hepatic metastatic lesions. This prospective study aimed to assess the capability of artificial intelligence iterative reconstruction (AIIR) in low-dose CT for detection of hepatic metastases. Thirty-two patients with hepatic metastases were enrolled and underwent dual-phase CT scans in the venous phase. Each patient received a standard-dose (SD) scan (120 kVp, 180 mAs) followed by a reduced-dose (RD) scan (120 kVp, 30 mAs). All data were analyzed. Compared with the SD CT scan, the volume CT dose index (CTDIvol), dose length product (DLP), and effective dose (ED) were significantly lower in the RD CT, with an average radiation reduction of approximately 80%. Radiation dose reduction significantly increased the image noise and degraded the overall image quality. However, higher reconstruction levels resulted in an increasingly prominent “waxy” texture on CT images. There were significant (P < 0.05) differences in liver noise, subcutaneous fat noise, liver parenchymal SNR, portal vein SNR, liver parenchymal contrast-to-noise ratio (CNR), and portal vein CNR among the groups, with the ranking of noise, SNR and CNR: RD AIIR Group 1 < RD AIIR Group 3 < RD AIIR Group 5 < SD Karl group < RD Karl group. In the SD Karl group, 218 metastatic nodules were identified in 32 patients, whereas in the RD Karl group, 176 (80.7%) nodules were detected. However, the RD AIIR 3 group identified 204 (93.6%) nodules. As the metastatic nodules size decreased, the detection rate in the RD Karl significantly declined, with the detection rate 14.29% in the RD Karl for nodules < 5 mm. In RD AIIR 3, the detection rate reached 100% for lesions ≥ 10 mm and 57.14% for lesions < 5 mm, which was significantly higher than in the RD Karl group (P < 0.05). The CNR of intrahepatic nodules in the SD Karl, RD Karl, and RD AIIR 3 was 2.68 (2.17, 3.80), 2.18 (1.64, 2.79), and 6.34 (4.34, 8.36), respectively, with significant differences among them. The signal-to-noise ratio of the lesions in the RD AIIR 3 group was the highest. AIIR-assisted low-dose CT reconstruction yields image quality comparable to standard-dose CT, while preserving diagnostic performance for the detection of small hepatic metastases.Keywords: Artificial intelligence, Liver metastasis, Iterative reconstruction, Low-dose CT scanSubject terms: Cancer, Diseases, Gastroenterology, Medical research, Oncology Introduction The liver is one of the most common sites of cancer metastasis, accounting for nearly 25% of all cases1–3. Metastatic hepatic tumors are more prevalent than primary hepatocellular or biliary malignancies4. Early and accurate detection of liver metastases is critical for treatment planning and prognostic assessment. Contrast-enhanced computed tomography (CT) is widely used to monitor treatment response in hepatic metastases5,6. However, the rapid increase in CT examinations over recent decades has raised concerns regarding radiation exposure7. To minimize the risk of radiation-induced cancer, the lowest feasible radiation dose should be used in CT scans. Filtered back projection (FBP) is a standard algorithm for reconstructing CT images from raw data8. However, with a lower radiation dose, the FBP algorithm may result in higher noise, degraded image quality and reduced diagnostic accuracy. With constant exposure strength, significant streak artifacts along the lateral direction are often distinct on FBP images. To reduce noise and enhance image quality especially at low radiation doses, iterative reconstruction (IR) techniques, such as hybrid iterative methods, have been introduced9. These methods are based on the FBP architecture and apply iterative noise reduction algorithms in both the projection domain and image domain to reduce the streak artifacts and image noise. However, due to inherent limitations of FBP, the ability to reduce noise and streak artifacts is limited with reduced dose. To overcome the limitation, the model-based iterative reconstruction (MBIR) has been developed as a fully iterative reconstruction technique incorporating a sophisticated model to describe the data acquisition process and a general prior term to regularize the final reconstruction. Despite the high potential in image quality improvement, application of the MBIR method in clinics is still challenging. First, the algorithm architecture complexity of MBIR is higher than FBP and hybrid iterative reconstruction methods. Second, different regularization strength can lead to significant differences in image quality, and high regularization strength in the MBIR images may have unnatural “plastic texture”. This artifact is particularly pronounced in parenchymal organs such as the liver, where it can obscure subtle structural details and impair visualization of small focal lesions. Additionally, soft tissue interfaces, vascular structures, and regions with subtle density differences may appear abnormally smooth, potentially impacting diagnostic confidence in the detection of small metastases9–11. Studies have shown that approximately 10% of hepatic metastases from pancreatic ductal carcinoma may be missed, with up to 32% labeled as indeterminate compared with MRI. Additionally, CT scans have limited sensitivity in detecting lesions smaller than 10 mm in diameter6. Therefore, balancing the radiation dose and diagnostic accuracy remains a major challenge in CT imaging of hepatic metastases. Deep learning (DL) technique, as a branch of artificial intelligence, represents a new reconstruction method based on the convolutional neural networks12 and enhances image quality by comparing low-dose scan data with optimal data obtained during the training phase. It greatly reduces imaging noise and significantly improves both image quality and reconstruction speed13. However, at ultra-low radiation doses, anatomical structures are often obscured by mottled noise and noise-induced streak artifacts, rendering reconstructed images non-diagnostic. Artificial intelligence iterative reconstruction (AIIR) combines the strengths of DL technique and MBIR14,15, integrating the noise/artifact suppression capability of DL with the structural detailed preservation of IR. In low-dose CT reconstruction, AIIR incorporates 3D DL denoisers into the MBIR iteration, allowing these denoisers to implicitly learn more effective regularization functions. In each iteration, the AIIR method first minimizes the data-fidelity cost function and subsequently applies a deep learning denoising engine to reduce noise. Image quality improves progressively with more iterations, as finer details are recovered while noise is continually suppressed. The iterative process terminates once the reconstructed image meets the clinical standards and user-defined presets (Levels 1–5). AIIR Levels 1 to 5 reflect a gradual increase in iterative reconstruction strength and noise reduction efficacy. Level 1 uses relatively mild iteration, resulting in slightly higher image noise but better preservation of spatial details. By contrast, Level 5 applies stronger iteration, achieving more pronounced noise suppression and smoother image texture. This approach effectively reduces noise and artifacts, enhances spatial resolution, and improves the detectability of low-contrast lesions in chest and vascular imaging with low-dose scans16–21. However, it remains unknown whether AIIR can effectively detect small low-contrast hepatic metastases in ultra-low-dose CT scans. A key indicator for evaluating CT image quality and lesion detectability is the contrast-to-noise ratio (CNR), which quantifies the ability to distinguish target tissues (e.g., hepatic metastases) from surrounding background tissues (e.g., normal liver parenchyma) while accounting for image noise. It is defined as (CT value of lesion - CT value of liver parenchyma) / standard deviation of background noise. A higher CNR indicates better contrast between target and background tissues, facilitating the detection of low-contrast lesions such as small hepatic metastases, which is crucial for resolving the trade-off between radiation dose reduction and diagnostic accuracy. In this study, we evaluated the AIIR’s performance in detecting small (< 3 cm) hepatic metastases and its image quality under low radiation doses, hypothesizing that AIIR combined with low-dose CT reconstruction can achieve image quality and diagnostic performance comparable to that of standard-dose CT reconstruction for the detection of small hepatic metastases. Material and Method Subjects This prospective single-center study was approved by the Ethics Committee of the Fourth Hospital of Hebei Medical University, and informed consent was obtained from all patients prior to participate. All methods were carried out in accordance with relevant guidelines and regulations. Tumor patients who underwent contrast-enhanced hepatic CT scans between December 2023 and June 2024 were prospectively enrolled. Patients were enrolled if they presented with hepatic metastases ≤ 3 cm in diameter and fewer than 20 lesions per patient, confirmation of hepatic metastasis by imaging (confirmed by at least two imaging modalities, with characteristic rim enhancement in the arterial phase and relative hypoenhancement in the portal venous phase) or by pathology examination, and lesions enlargement on follow-up imaging. Patients were excluded if they had an age < 18 years old, presence of severe cardiac or renal dysfunction, motion artifacts or metal artifacts on CT images, and interventional treatment before enrollment. CT scanning protocol CT scans were conducted using the United Imaging uCT968 (United Imaging Healthcare, Shanghai, China), with standardized breath‑hold training before imaging and bolus-tracking software to detect the optimal scan delay for arterial phase in each patient. All patients received an iodine contrast injection (concentration 350 mg/ml at a dose of 1.5 ml/kg body weight) via an arm vein at a rate of 3.0 ml per second using a power injector, followed by 30 mL saline flush injection at the same injection rate. When the CT value of the aorta reached 150 HU, scanning was triggered to acquire arterial-phase images (Approximately 25 to 30 s after contrast agent injection to minimize the additional radiation burden, the arterial phase was acquired with a tube current 30% lower than that used for the venous phase). Thirty seconds later, a standard-dose (SD) scan of the venous phase was performed with a standard dose (120 kVp fixed, tube current modulation level 3 with a reference tube current of 162 mAs). After the standard-dose scanning, the scanning with a reduced-dose (RD) was started immediately (120 kVp fixed, tube current modulation level 1 with a reference tube current of 40 mAs). Other scanning parameters, including slice thickness 0.5 mm, rotation time 0.5 s per rotation, pitch 0.9938 and collimation width 80 mm, were the same between the standard-dose scan and lower-dose scan. The scan range extended from the superior margin of the diaphragm to the inferior margin of the pubic symphysis. CT image reconstruction The primary data from contrast-enhanced CT scan were uploaded to the United Imaging Explorer platform (Uinnovation‑CT, R001, United Imaging Healthcare Co., Ltd., Shanghai, China) for reconstruction. The reconstruction was processed with 1 mm slice thickness and 1 mm interval. The primary data from standard-dose CT scans was analyzed and reconstructed using the commercial iterative reconstruction (Karl algorithm, United Imaging Healthcare, Shanghai, China) with standard intensity level 5 and core algorithm B‑SOFT‑B. The primary data from low-dose CT scan was reconstructed using the Karl algorithm (standard intensity level 5) and AIIR at levels 1, 3, and 5.The radiation dose in each patient was determined as ED = k × DLP, where ED represented the effective dose, k was equaled to 0.015mSv/(mGy.cm), and DLP represented the dose length product. CT image evaluation Subjective Image Analysis Subjective image assessment was performed independently by two radiologists (with 12 and 5 years of experience, respectively) who were blinded to the reconstruction methods, dose levels, and full scan context and evaluated individual image slices from all sequences. Each image was randomly presented and evaluated by both reviewers for noise, “waxy” appearance, and overall quality, following the predefined criteria in Table 1. Table 1Image Quality Scoring Criteria.ScoreNoise LevelWaxy AppearanceOverall Image Quality1Severe noise with coarse granularityPronounced waxy appearance, completely unacceptable with severe loss of detailAnatomical structures are indistinct, rendering the image non-diagnostic2High noise with coarse granularityUnacceptable waxy appearance, blurred image detailsAnatomical details are blurred with significant noise, impairing diagnostic confidence3Moderate noise, slightly coarse granularityAcceptable waxy appearance, moderate detail clarityMost structures are clear and diagnostically acceptable, albeit with reduced sharpness4Mild noise, fine granularityMild waxy appearance, relatively clear detailsAnatomical details are fairly clear with only slightly reduced sharpness and mild noise5Minimal to no noise, fine and smooth textureNegligible waxy appearance, sharp and clear detailsAnatomical structures are sharply defined with minimal noise, excellent for diagnosis Image Quality Scoring Criteria. Objective Image Analysis For the five image groups, CT values were measured at three locations in the liver parenchyma (one in the left lobe and two in the right lobe, avoiding vessels and bile ducts), as well as in the paravertebral muscle, portal vein, and intrahepatic metastases (at the level of T12-L1). The noise level, represented by the standard deviation (SD), was measured in the liver and the subcutaneous fat of the anterior abdominal wall. The position and size of the regions of interest (ROIs) were kept consistent across different reconstruction series, carefully avoiding vascular structures. The signal-to-noise ratio (SNR) for the liver parenchyma and portal vein, along with their contrast-to-noise ratio (CNR) relative to the back muscle, were calculated as follows: SNRliver/portal vein = CT valueliver/portal vein/ SDfat, and CNR liver/portal vein = (CT valueliver/portal vein – CT valueparaspinal muscle) / SDfat. Comparative Performance of Lesion Assessment To evaluate nodule detection and measurement accuracy, one optimal RD AIIR reconstruction was selected a priori among RD AIIR 1, RD AIIR 3, and AIIR 5 reconstructions based on the best balance of low noise and high subjective quality. The performance of this selected RD AIIR dataset was then benchmarked against the SD Karl and RD Karl groups. The evaluation metrics included the lesion detection rate and the consistency of lesion size measurements. Additionally, the lesion contrast-to-noise ratio (CNR) was compared across all groups using the formula: CNRlesion = (CT valuesurrounding liver parenchyma – CTlesion) / SDliver parenchyma. Statistical analysis All the statistical analyses were performed using the SPSS version 26.0 (IBM, Chicago, IL, USA). The Shapiro–Wilk test was used to assess the normality and variance homogeneity in each data set. Continuous measurement data in the normal distribution were presented as mean ± standard deviation and tested with the One-way ANOVA followed by LSD post-hoc tests, whereas data which were not in the normal distribution were presented as median and interquartile range and tested with the Friedman test. Paired Student’s t test was used to determine differences within the same group receiving different treatments. The Chi-square test was used to determine the difference in the detection rate. The P value less than < 0.05 was considered statistically significant. Agreement was assessed using Kappa concordance correlation coefficient. K < 0.2: no agreement, 0.2–0.4: poor agreement, 0.4–0.6: moderate agreement, 0.6–0.8: substantial agreement, and 0.8-1.0: strong agreement. Results Patients and radiation dose Thirty-two patients were enrolled with a mean age of 61.9 years and a mean body mass index of 22.8 ± 3.4 kg/m² (range: 17.1–31.8 kg/m²) (Fig. 1; Table 2). Compared with the standard-dose CT scan, the volume CT dose index (CTDIvol), DLP, and ED were significantly lower in the low-dose CT scan with an average radiation reduction of approximately 80% (Table 3). The standard-dose CT scan revealed 218 intrahepatic metastatic lesions, with an average diameter 10.48 mm. Flow chart of patient enrollment. Table 2Demography of the patients.VariablesValuesAge (y)36–78(mean, 61.9)SexM20F12Cancer sourcesPulmonary cancer8Hepatic cancer5Uterine derived cancer2Colorectal cancer5Gastric cancer2Esophageal cancer3Pancreatic cancer5Breast cancer2Body mass index (kg/m²)22.8 ± 3.4Hepatic metastatic lesion no.208Lesion size (mm)4.2–29.4(mean, 10.48) Demography of the patients. Table 3Scanning doses in the standard and reduced dose groups.GroupsCTDIvol (mGy)DLP (mGy/cm)ED (mSv)Standard dose13.34 ± 2.11726.90 ± 241.0510.9 ± 3.62Reduced dose2.76 ± 0.39123.12 ± 52.251.85 ± 0.78t32.9616.1216.12P<0.01 *<0.01 *<0.01 *Note: DLP, dose length product; ED, effective dose; CTDI, CT dose index. Scanning doses in the standard and reduced dose groups. Note: DLP, dose length product; ED, effective dose; CTDI, CT dose index. CT subjective evaluation Reducing the radiation dose significantly increased the image noise and degraded overall image quality in the RD Karl group. The lowest noise was observed in the RD AIIR‑1 group (Table 4). However, higher reconstruction levels introduced an increasingly pronounced “waxy” texture on CT images most notably in the RD AIIR‑1 group, and no difference was found among other groups. For the overall quality, there were no significant (P > 0.05) differences between the SD Karl and RD AIIR‑3 or between RD Karl and RD AIIR‑1. SD Karl and RD AIIR groups achieved the highest image quality scores. These results together suggested that AIIR at level 3 provided the best balance, effectively reducing the noise from low-dose scans while avoiding excessive “waxy” texture, which made RD AIIR‑3 the optimal protocol for high-quality, low-dose CT imaging. In the subjective analysis of image quality, substantial-to-strong agreement was noted between the two readers (κ value ranged from 0.71 to 0.92). Table 4Subjective comparison of scores for abdominal scanning with different reconstruction algorithms.Reconstruction algorithmNoiseImage waxy textureOverall imaging qualitySD KarlRD KarlP3.0(3.0,4.0)2.0(2.0,2.0)<0.015.0(5.0,5.0)5.0(5.0, 5.0)1.05.0(5.0,5.0)3.0(2.0,3.0)<0.01SD KarlRD AIIR 1P3.0(3.0,4.0)5.0(5.0,5.0)<0.015.0(5.0,5.0)2.0(2.0,3.0)<0.015.0(5.0,5.0)3.0(2.0,3.0)<0.01SD KarlRD AIIR 3P3.0(3.0,4.0)4.0(4.0,4.0)0.0175.0(5.0,5.0)4.0(3.0,4.0)<0.015.0(5.0,5.0)5.0(4.0,5.0)1.0SD KarlRD AIIR 5P3.0(3.0,4.0)3.0(3.0,4.0)1.05.0(5.0,5.0)4.0(4.0,5.0)0.2335.0(5.0,5.0)4.0(4.0,4.0)0.015RD KarlRD AIIR 1P2.0(2.0,2.0)5.0(5.0,5.0)<0.015.0(5.0,5.0)2.0(2.0,3.0)<0.013.0(2.0,3.0)3.0(2.0,3.0)1.0RD KarlRD AIIR 3P2.0(2.0,2.0)4.0(4.0,4.0)<0.015.0(5.0,5.0)4.0(3.0,4.0)<0.013.0(2.0,3.0)5.0(4.0,5.0)<0.01RD KarlRD AIIR 5P2.0(2.0,2.0)3.0(3.0,4.0)<0.015.0(5.0,5.0)4.0(4.0,5.0)0.1563.0(2.0,3.0)4.0(4.0,4.0)<0.01RD AIIR 1RD AIIR 3P5.0(5.0,5.0)4.0(4.0,4.0)0.0412.0(2.0,3.0)4.0(3.0,4.0)0.0463.0(2.0,3.0)5.0(4.0,5.0)<0.01RD AIIR 1RD AIIR 5P5.0(5.0,5.0)3.0(3.0,4.0)<0.012.0(2.0,3.0)4.0(4.0,5.0)<0.013.0(2.0,3.0)4.0(4.0,4.0)<0.01RD AIIR 3RD AIIR 5P4.0(4.0,4.0)3.0(3.0,4.0)0.0084.0(3.0,4.0)4.0(4.0,5.0)0.0055.0(4.0,5.0)4.0(4.0,4.0)0.046 Subjective comparison of scores for abdominal scanning with different reconstruction algorithms. SD Karl RD Karl P 3.0(3.0,4.0) 2.0(2.0,2.0) <0.01 5.0(5.0,5.0) 5.0(5.0, 5.0) 1.0 5.0(5.0,5.0) 3.0(2.0,3.0) <0.01 SD Karl RD AIIR 1 P 3.0(3.0,4.0) 5.0(5.0,5.0) <0.01 5.0(5.0,5.0) 2.0(2.0,3.0) <0.01 5.0(5.0,5.0) 3.0(2.0,3.0) <0.01 SD Karl RD AIIR 3 P 3.0(3.0,4.0) 4.0(4.0,4.0) 0.017 5.0(5.0,5.0) 4.0(3.0,4.0) <0.01 5.0(5.0,5.0) 5.0(4.0,5.0) 1.0 SD Karl RD AIIR 5 P 3.0(3.0,4.0) 3.0(3.0,4.0) 1.0 5.0(5.0,5.0) 4.0(4.0,5.0) 0.233 5.0(5.0,5.0) 4.0(4.0,4.0) 0.015 RD Karl RD AIIR 1 P 2.0(2.0,2.0) 5.0(5.0,5.0) <0.01 5.0(5.0,5.0) 2.0(2.0,3.0) <0.01 3.0(2.0,3.0) 3.0(2.0,3.0) 1.0 RD Karl RD AIIR 3 P 2.0(2.0,2.0) 4.0(4.0,4.0) <0.01 5.0(5.0,5.0) 4.0(3.0,4.0) <0.01 3.0(2.0,3.0) 5.0(4.0,5.0) <0.01 RD Karl RD AIIR 5 P 2.0(2.0,2.0) 3.0(3.0,4.0) <0.01 5.0(5.0,5.0) 4.0(4.0,5.0) 0.156 3.0(2.0,3.0) 4.0(4.0,4.0) <0.01 RD AIIR 1 RD AIIR 3 P 5.0(5.0,5.0) 4.0(4.0,4.0) 0.041 2.0(2.0,3.0) 4.0(3.0,4.0) 0.046 3.0(2.0,3.0) 5.0(4.0,5.0) <0.01 RD AIIR 1 RD AIIR 5 P 5.0(5.0,5.0) 3.0(3.0,4.0) <0.01 2.0(2.0,3.0) 4.0(4.0,5.0) <0.01 3.0(2.0,3.0) 4.0(4.0,4.0) <0.01 RD AIIR 3 RD AIIR 5 P 4.0(4.0,4.0) 3.0(3.0,4.0) 0.008 4.0(3.0,4.0) 4.0(4.0,5.0) 0.005 5.0(4.0,5.0) 4.0(4.0,4.0) 0.046 Objective evaluation No significant (P > 0.05) differences existed in the CT value in liver and portal vein among the groups. However, there were significant (P < 0.05) differences in liver noise, subcutaneous fat noise, liver parenchymal SNR, portal vein SNR, liver parenchymal CNR, and portal vein CNR among the groups (Tables 5 and 6), with the ranking of noise, SNR and CNR: RD AIIR Group 1 < RD AIIR Group 3 < RD AIIR Group 5 < SD Karl group < RD Karl group. There were no significant (P > 0.05) differences in liver parenchymal SNR, portal vein SNR, liver parenchymal CNR and portal vein CNR between the SD Karl group and the RD AIIR 5 group, but significant (P < 0.05) differences in the comparisons of the remaining items among all groups,. Based on the objective and subjective results, the RD AIIR 3 generated the best image quality for low-dose scans. Inter-rater reliability between two reviewers was substantial for objective qualitative image assessment (κ values = 0.65–0.73). Table 5Comparison of liver parenchyma and portal vein CTV and standard deviation of the CTV of the liver and subcutaneous fat.VariablesSD KarlRD AIIR 1RD AIIR 3RD AIIR 5RD Karl P Hepatic CTV104.32 ± 13.41103.20 ± 13.32103.22 ± 13.20103.13 ± 13.61105.56 ± 11.110.919Portal vein CTV135.49 ± 32.09148.67 ± 33.01148.60 ± 31.86148.42 ± 33.31138.48 ± 29.500.231Hepatic CTVS-D14.62 ± 6.691.55 ± 1.186.81 ± 1.4811.67 ± 2.8517.43 ± 2.31<0.01Subcutaneous fat CTVS-D14.62 ± 6.695.81 ± 5.288.61 ± 5.1711.84 ± 5.3416.06 ± 2.94<0.01Note: CTV, computed tomography value; AIIR, artificial intelligence iterative reconstruction; SD, standard dose; RD, reduced dose; CTVS-D, standard deviation of the CTV. Comparison of liver parenchyma and portal vein CTV and standard deviation of the CTV of the liver and subcutaneous fat. Note: CTV, computed tomography value; AIIR, artificial intelligence iterative reconstruction; SD, standard dose; RD, reduced dose; CTVS-D, standard deviation of the CTV. Table 6Comparison of hepatic and portal vein SNR and CNRVariablesHepatic SNRPortal vein SNRHepatic CNRPortal vein CNRSD KarlRD AIIR 1P8.74(7.59,9.73)25.03(17.28,34.72)<0.0112.57(10.51,15.58)31.90(18.49,56.58)<0.014.21(3.62,5.35)10.23(8.24,16.38)<0.016.20(4.71,7.87)17.8(10.15,34.46)<0.01SD KarlRD AIIR 3P8.74(7.59,9.73)14.86(10.59,17.78)<0.0112.57(10.51,15.58)19.93(13.78,27.11)0.0254.21(3.62,5.35)6.43(5.22,8.32)0.0096.20(4.71,7.87)12.80(7.38,16.31)<0.01SD KarlRD AIIR 5P8.74(7.59,9.73)9.27(8.10,13.40)1.012.57(10.51,15.58)12.73(9.60,19.13)1.04.21(3.62,5.35)4.16(3.07,5.73)1.06.20(4.71,7.87)7.36(4.93,11.64)1.0SD KarlRD KarlP8.74(7.59,9.73)6.67(5.89,7.60)0.00512.57(10.51,15.58)8.88(6.93,10.25)0.0044.21(3.62,5.35)2.86(2.2.0,3.54)0.0046.20(4.71,7.87)4.99(2.99,6.12)0.025RD AIIR 1RD AIIR 3P25.03(17.28,34.72)14.86(10.59,17.78)0.02131.90(18.49,56.58)19.93(13.78,27.11)0.04110.23(8.24,16.38)6.43(5.22,8.32)0.01517.8(10.15,34.46)12.80(7.38,16.31)0.015RD AIIR 1RD AIIR 5P25.03(17.28,34.72)9.27(8.10,13.40)<0.0131.90(18.49,56.58)12.73(9.60,19.13)<0.0110.23(8.24,16.38)4.16(3.07,5.73)<0.0117.8(10.15,34.46)7.36(4.93,11.64)<0.01RD AIIR 1RD KarlP25.03(17.28,34.72)6.67(5.89,7.60)<0.0131.90(18.49,56.58)8.88(6.93,10.25)<0.0110.23(8.24,16.38)2.86(2.2.0,3.54)<0.0117.8(10.15,34.46)4.99(2.99,6.12)<0.01RD AIIR 3RD AIIR 5P14.86(10.59,17.78)9.27(8.10,13.40)0.01219.93(13.78,27.11)12.73(9.60,19.13)0.0196.43(5.22,8.32)4.16(3.07,5.73)0.01512.80(7.38,16.31)7.36(4.93,11.64)0.015RD AIIR 3RD KarlP14.86(10.59,17.78)6.67(5.89,7.60)<0.0119.93(13.78,27.11)8.88(6.93,10.25)<0.016.43(5.22,8.32)2.86(2.2.0,3.54)<0.0112.80(7.38,16.31)4.99(2.99,6.12)<0.01RD AIIR 5RD KarlP9.27(8.10,13.40)6.67(5.89,7.60)<0.0112.73(9.60,19.13)8.88(6.93,10.25)0.0054.16(3.07,5.73)2.86(2.2.0,3.54)0.0027.36(4.93,11.64)4.99(2.99,6.12)<0.01Note: CNR, contrast-to-noise ratio; SNR, signal-to-noise ratio; SD, standard dose; RD, reduced dose. Comparison of hepatic and portal vein SNR and CNR SD Karl RD AIIR 1 P 8.74(7.59,9.73) 25.03(17.28,34.72) <0.01 12.57(10.51,15.58) 31.90(18.49,56.58) <0.01 4.21(3.62,5.35) 10.23(8.24,16.38) <0.01 6.20(4.71,7.87) 17.8(10.15,34.46) <0.01 SD Karl RD AIIR 3 P 8.74(7.59,9.73) 14.86(10.59,17.78) <0.01 12.57(10.51,15.58) 19.93(13.78,27.11) 0.025 4.21(3.62,5.35) 6.43(5.22,8.32) 0.009 6.20(4.71,7.87) 12.80(7.38,16.31) <0.01 SD Karl RD AIIR 5 P 8.74(7.59,9.73) 9.27(8.10,13.40) 1.0 12.57(10.51,15.58) 12.73(9.60,19.13) 1.0 4.21(3.62,5.35) 4.16(3.07,5.73) 1.0 6.20(4.71,7.87) 7.36(4.93,11.64) 1.0 SD Karl RD Karl P 8.74(7.59,9.73) 6.67(5.89,7.60) 0.005 12.57(10.51,15.58) 8.88(6.93,10.25) 0.004 4.21(3.62,5.35) 2.86(2.2.0,3.54) 0.004 6.20(4.71,7.87) 4.99(2.99,6.12) 0.025 RD AIIR 1 RD AIIR 3 P 25.03(17.28,34.72) 14.86(10.59,17.78) 0.021 31.90(18.49,56.58) 19.93(13.78,27.11) 0.041 10.23(8.24,16.38) 6.43(5.22,8.32) 0.015 17.8(10.15,34.46) 12.80(7.38,16.31) 0.015 RD AIIR 1 RD AIIR 5 P 25.03(17.28,34.72) 9.27(8.10,13.40) <0.01 31.90(18.49,56.58) 12.73(9.60,19.13) <0.01 10.23(8.24,16.38) 4.16(3.07,5.73) <0.01 17.8(10.15,34.46) 7.36(4.93,11.64) <0.01 RD AIIR 1 RD Karl P 25.03(17.28,34.72) 6.67(5.89,7.60) <0.01 31.90(18.49,56.58) 8.88(6.93,10.25) <0.01 10.23(8.24,16.38) 2.86(2.2.0,3.54) <0.01 17.8(10.15,34.46) 4.99(2.99,6.12) <0.01 RD AIIR 3 RD AIIR 5 P 14.86(10.59,17.78) 9.27(8.10,13.40) 0.012 19.93(13.78,27.11) 12.73(9.60,19.13) 0.019 6.43(5.22,8.32) 4.16(3.07,5.73) 0.015 12.80(7.38,16.31) 7.36(4.93,11.64) 0.015 RD AIIR 3 RD Karl P 14.86(10.59,17.78) 6.67(5.89,7.60) <0.01 19.93(13.78,27.11) 8.88(6.93,10.25) <0.01 6.43(5.22,8.32) 2.86(2.2.0,3.54) <0.01 12.80(7.38,16.31) 4.99(2.99,6.12) <0.01 RD AIIR 5 RD Karl P 9.27(8.10,13.40) 6.67(5.89,7.60) <0.01 12.73(9.60,19.13) 8.88(6.93,10.25) 0.005 4.16(3.07,5.73) 2.86(2.2.0,3.54) 0.002 7.36(4.93,11.64) 4.99(2.99,6.12) <0.01 Note: CNR, contrast-to-noise ratio; SNR, signal-to-noise ratio; SD, standard dose; RD, reduced dose. Metastatic nodules detection Among the 32 enrolled patients, a total of 218 intrahepatic metastatic lesions were identified, with a mean diameter of 10.48 mm (range 4.2–29.4 mm). Lesion confirmation was achieved through a composite reference standard consisting of: (1) histopathological confirmation from biopsy/resection specimens (15 lesions, 6.9%), (2) confirmation by at least one alternative imaging modality (MRI and/or PET-CT) within 4 weeks of the study CT (120 lesions, 55.0%), and (3) evidence of lesion progression or response to treatment on follow-up imaging at 3–6 months (83 lesions, 38.1%). All lesions were confirmed through consensus review by two experienced radiologists who had access to all available clinical, pathological, and imaging data. Metastatic nodules demonstrated peripheral enhancement during the arterial phase and appeared hypoenhancing relative to the liver parenchyma in the portal venous phase. 176 (80.7%) nodules were detected in the RD Karl group, with the detection rate significantly (P < 0.05) lower than the SD Karl. However, The RD AIIR 3 group detected 204 nodules with a detection rate of 93.6%, which was significantly higher than that of the RD Karl group (P < 0.05). The overall detection rate remained slightly lower than that of the SD Karl group, likely attributable to the small size or subcapsular location of some nodules. (Table 7). Table 7The detection rate in the SD and RD Karl and RD AIIR 3.Lesion sizeSD KarlRD KarlRD AIIR 3P for RD Karl vs. RD AIIR 3Lesion no.Lesion no.Detection rateLesion no.Detection rate<5 mm7114.29%457.14%<0.015–10 mm1088074.07%9789.81%<0.0110–20 mm999191.92%99100%<0.0520–30 mm44100%4100%1.0Note: AIIR, artificial intelligence iterative reconstruction; SD, standard dose; RD, reduced dose. The detection rate in the SD and RD Karl and RD AIIR 3. Note: AIIR, artificial intelligence iterative reconstruction; SD, standard dose; RD, reduced dose. To evaluate size-dependent detection, lesions were stratified as ≥ 20 mm, 10–20 mm, 5–10 mm, and < 5 mm. Similar detection rates were detected for metastatic nodules of 20–30 mm in either the RD Karl or the RD AIIR 3 (Fig. 2). However, as the metastatic nodules size decreased, the detection rate in the RD Karl group was significantly decreased, with the detection rate for nodules with diameter < 5 mm accounting for only 14.29% in the RD Karl grouop, significantly lower than the SD Karl group. However, AIIR significantly increased the detection rate for lower-dose scan. In RD AIIR 3, the detection reached 100% for lesions ≥ 10 mm. Even for lesions < 5 mm, RD AIIR 3 achieved a 57.14% detection rate, which was significantly higher than the RD Karl group (P < 0.05). These findings indicate that AIIR level 3 reconstruction significantly enhances the detection capability of reduced-dose scans for small metastatic nodules. Specifically, for metastatic nodules larger than 10 mm, the detection performance of the RD AIIR 3 group was comparable to that of the SD Karl group. For nodules measuring 5–10 mm, the detection rate of RD AIIR 3 (89.81%) was lower than that achieved by SD Karl, while for nodules smaller than 5 mm, the detection rate approximated only half of that observed in the SD Karl group. For lesion detection, the diagnostic confidence between the two reviewers was strong agreement for lesions larger than 10 mm (κ value = 0.93) and substantial agreement (κ value = 0.67) for lesions smaller than 10 mm. Standard-dose (SD), reduced dose (RD) Karl, and RD AIIR reconstruction in patients with hepatic metastasis from breast cancer and hepatic cancer with multiple intrahepatic metastases. A-C. Female patients with hepatic metastasis from breast cancer. A. The SD Karl reconstruction image showed low density nodules in the right anterior lobe of the liver with unclear edges. The hepatic metastatic lesions were not shown in the RD Karl reconstruction image (B) or the RD AIIR reconstruction image (C). D-F. Female patients with hepatic cancer and multiple intrahepatic metastases. D. The SD Karl reconstruction image showed low-density nodules in the right posterior lobe of the liver with unclear edges. E. The RD Karl reconstruction image demonstrated unclear nodules. F. The RD AIIR reconstruction image revealed that the lesions were low-density nodules with clear edges. The lesion size and contrast-to-noise ratio (CNR) were evaluated among the three groups of SD Karl, RD Karl, and RD AIIR 3 (Fig. 3). The average lesion diameters of the lesion were not significantly different among the three groups 10.69 ± 4.06 mm, 10.75 ± 4.44 mm, and 10.34 ± 4.32 mm, respectively, and the intraclass correlation coefficients (ICC) was 0.945, indicating good consistency among the three groups. The CNR of intrahepatic nodules in the three groups (SD Karl, RD Karl, and RD AIIR 3) were 2.68 (2.17, 3.80), 2.18 (1.64, 2.79), and 6.34 (4.34, 8.36), respectively. There were significant (P = 0.000 < 0.05) differences among the three groups and between any two groups. The CNR of the lesions in the RD AIIR 3 group was the highest. Comparison of the lesion diameter and contrast-to-noise ratio (CNR) in the standard-dose (SD) Karl, reduced-dose (RD) Karl and RD AIIR groups. (A) Comparison of the mean lesion diameter in three groups showed no significant differences in the lesion size among the three groups. (B) Comparison of the lesion CNR in three groups demonstrated significant differences (P < 0.05) between any two groups, with significantly greatest lesion CNR obtained in RD AIIR 3 group. Discussion Although CT is preferred in evaluating liver metastases, concerns have arisen over the potential cancer risk associated with repeated scans due to cumulative radiation exposure. Reducing the scan dose decreases the radiation risk but often results in increased image noise and decreased sensitivity for detecting low-contrast lesions. Previous research demonstrated that deep learning reconstruction can maintain the diagnostic quality of CT images with a 67% dose reduction22. However, dose reduction beyond 70% typically compromises image quality to a level not efficient for diagnosis. In our study, even with an 80% radiation dose reduction, AIIR effectively suppresses image noise to levels comparable to standard-dose iterative reconstruction and reduces the waxy appearance in liver CT images. Furthermore, AIIR achieved similar SNR and CNR in the liver parenchyma and portal vein compared to standard-dose images. Karl is a well-established hybrid iterative reconstruction algorithm that has been clinically validated and routinely used at our institution for over five years. Accordingly, the standard radiation dose combined with Karl reconstruction serves as our current clinical reference protocol for the evaluation of liver metastases. Among the low-dose AIIR reconstructions, images reconstructed with AIIR 3 yielded the highest quality in both subjective and objective assessments. Using the metastatic nodules detected in the SD Karl group as the reference standard, we found that AIIR significantly improved the image quality and lesion detectability for liver metastases in ultra-low-dose CT scans. Notably, the sensitivity of AIIR for detecting hepatic lesions > 10 mm was comparable to that of standard-dose CT, a finding corroborated by a prospective cohort study15. Authors have also proved that AIIR could improve the CT image quality at 50% reduced dose while preserving diagnostic performance and confidence for low-contrast colorectal liver metastasis in all lesions, including lesions ≤ 10 mm, and AIIR may thus serve as a promising tool for follow-up monitoring of colorectal cancer while inflicting less radiation damage15. These results suggested that AIIR can efficiently detect larger liver lesions in reduced-dose CT scans, potentially reducing the radiation exposure without compromising the diagnostic performance. In our study, AIIR significantly enhanced the detection rate of small hepatic metastases, particularly those smaller than 10 mm in diameter. For lesions under 5 mm, the detection rate with RD Karl was notably lower at approximately 14.29%. However, AIIR markedly increased the detection rate to 57.14%. For lesions ranging from 5 to 10 mm, RD Karl achieved a detection rate of 74.07%, which improved to 89.81% with AIIR. Our findings were consistent with a previous report with the detection rates of 72.3% for lesions < 5 mm and 90.3% for those 5–10 mm in liver metastases using AIIR13. Their detection rate for lesions < 5 mm was higher than ours, possibly because we focused exclusively on metastatic lesions while some benign nodules were included in that study13 with benign nodules having higher contrast compared to metastases. Early detection of liver metastases plays a major role in management options and long-term prognosis. Patients with liver metastases may be eligible for metastases resection, thus being able to achieve long-term disease remission and survival. Therefore, adequate radiological staging and risk stratification is crucial for the optimal selection of patients for surgery in order to maximize the benefit-risk ratio of liver metastasectomy and to individualize the treatment strategy. Our present findings demonstrate the potential of AIIR in improving the sensitivity of low-dose CT scanning in detecting small hepatic metastases, which is crucial for accurate staging and treatment planning in oncology. AIIR can provide clearer visualization of lesions and enable more accurate measurement and diagnosis of the lesion size. Metastatic nodules appeared as low-density lesions during the venous phase. We observed that low-dose scans increased the image noise and caused lesion margins to appear blurred, which could lead to overestimation of the lesion diameter. In contrast, AIIR significantly reduced the SNR, resulting in sharper lesion boundaries and smaller measured sizes. The diameters of the same lesions were comparable across standard-dose iterative reconstruction, low-dose iterative reconstruction, and low-dose AIIR. Since the liver is a common site for metastatic disease1–3, the number, size, and location of liver metastases directly influence clinical decision-making regarding treatment options such as surgery, thermal ablation, stereotactic radiotherapy, or chemotherapy. We emphasize that while ultra-low-dose scans (80% dose reduction) with AIIR are not yet perfectly equivalent to full-dose scans for lesions < 5 mm, they offer a clinically acceptable alternative for follow-up and screening with significantly reduced radiation risks. Future strategies, including multi-centric data training and advanced algorithm optimization, may further close this gap. Our study has several strengths, including its focus on evaluating image quality and lesion detectability under ultra-low-dose conditions, as well as the implementation of a dual-venous phase scanning protocol. To minimize the additional radiation burden, the arterial phase was acquired with a reduced tube current (30% lower than that used for the venous phase), and the total radiation dose from both phases in the low-dose protocol remained substantially lower than that of a standard single-phase dose. Consequently, the additional low-dose scan did not increase the overall radiation exposure. Although most hypovascular liver metastases are best visualized during the portal venous phase, we incorporated both arterial and venous phase imaging for two important reasons. First, a subset of liver metastases—particularly those from neuroendocrine tumors, melanoma, sarcoma, and some breast and renal cell carcinomas—may be hypervascular and are better depicted in the arterial phase. Second, the standard protocol at our institution for liver metastases includes biphasic acquisition to ensure comprehensive evaluation, especially in patients with unknown primaries or those at risk for hypervascular metastases. Furthermore, our analysis included only liver metastases, excluding other types of small intrahepatic lesions; this may have yielded more robust results due to the inclusion of lesions with consistent texture and pathological characteristics. Some limitations existed in our study, including the single-center study design, a small cohort of patients, lack of subgroup analysis based on the primary origin or histology of the metastases, Chinese patients enrolled only, and inclusion of only metastatic hepatic lesions. All these issues may produce some publication bias to affect the outcome of our study. There remains a gap in the performance of AIIR reconstruction compared with standard-dose images for detecting metastatic nodules smaller than 10 mm. To address this limitation, future iterations of the AIIR algorithm could integrate deep learning-based noise suppression techniques to better preserve the textural fidelity and edge sharpness, thereby enhancing the detectability of small and subtle lesions. Moreover, the development of reconstruction algorithms tailored to specific pathological types may further improve diagnostic accuracy. Large-scale prospective trials involving diverse patient populations and multiple imaging platforms are warranted to compare LD-AIIR with standard-dose protocols, with the goal of establishing its generalizability and identifying specific clinical scenarios in which low-dose protocols can be reliably implemented. In conclusion, AIIR significantly improves the image quality of liver metastases in ultra-low-dose CT scans. Despite an 80% reduction in the radiation dose, AIIR significantly increases the detection rate of metastatic nodules smaller than 10 mm or has an equally-better detection rate for nodules larger than 10 mm compared to standard-dose iterative reconstruction, substantially reducing the radiation exposure and benefiting patients in the diagnosis and treatment of liver metastases. This low-dose protocol is particularly suitable for young patients, who have a higher lifetime risk of radiation exposure. We can attempt to combine it with MRI as a first-line screening tool to further reduce radiation exposure while maintaining diagnostic accuracy. Author contributions Design of the study: Qi WangData collection: Xiang Liu, Zhi-Lei Zhang, Xiao-Hui Qi, Jiang-Yang Pan, Xue-Li Fan, Jing Liu, Yi-Ting SunData analysis: Xiang Liu, Qi WangSupervision: Qi WangWriting of the original version: Zhi-Lei ZhangRevision: Bu-Lang GaoReview and approval of the paper: All authors. Funding National Key R&D Program of China (grant number: 2023YFC2412205); Medical Science Research Project of Hebei (grant number: 20210403). Data availability The datasets generated and/or analyzed during the current study are not publicly available due to the restriction by the hospital policy but are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available due to the restriction by the hospital policy but are available from the corresponding author on reasonable request.