Exploring the diagnostic performance of low-dose CT for colorectal cancer assessment: feasibility of artificial intelligence iterative reconstruction

Introduction

Colorectal cancer (CRC) is the third most frequently diagnosed malignancy worldwide and the second leading cause of cancer-related death (1-4). Epidemiological data from 2022 indicated that CRC accounted for 9.6% (1.9 million) of new cancer cases and 9.3% (0.9 million) of cancer deaths globally (2). Current guidelines emphasize the essential role of imaging examinations in the comprehensive management of CRC, encompassing initial diagnosis, staging, and follow-up. Among available techniques, contrast-enhanced long-range abdominal computed tomography (CT) has been widely adopted in clinical practice for assessing the primary CRC tumor, regional lymph node metastasis (RLNM), and distant metastases, given its high availability and fast execution (5-7).

The accurate identification of visceral peritoneal invasion (VPI) is critical in CRC assessment, as it is a key diagnostic criterion for distinguishing pT4a from earlier stages and an independent prognostic factor for poor survival outcomes (8,9). However, the inherent limitations of CT imaging in soft-tissue contrast resolution pose significant challenges to the reliable visualization of CRC and its surrounding tissues (10). These challenges are particularly pronounced in high-risk pT4 staging cases, where desmoplastic and neoplastic pericolonic fat infiltration can appear nearly indistinguishable on CT images (11). Consequently, additional examinations and advanced image reconstruction algorithms are frequently required for a conclusive diagnosis.

With the advancement of artificial intelligence (AI), several deep learning-based reconstruction (DLR) algorithms have been developed. Investigations into the application of DLR in cancer imaging have demonstrated its superior ability to improve image quality and lesion characterization compared to routinely available hybrid iterative reconstruction (HIR) (12,13). Notably, a recent preliminary study (14) reported that DLR outperformed HIR in diagnosing VPI of CRC, reducing the false-positive rate (5.6% vs. 20.4%) without increasing the false-negative rate. Nevertheless, most existing evidence is derived from routine-dose (RD) CT protocols, leaving the efficacy of DLR at low-dose (LD) settings, a critical consideration for radiation dose optimization according to the “as low as reasonably achievable” principle, largely unexplored.

In this study, we hypothesized that DLR applied to LD CRC imaging would achieve image quality and diagnostic reliability comparable to those of RD images reconstructed with HIR, which remains the standard approach for CRC evaluation. To test this hypothesis, a latest-generation DLR algorithm featuring a newly designed backbone, namely artificial intelligence iterative reconstruction (AIIR, United Imaging Healthcare, Shanghai, China), was employed. The aim of this study was to explore the feasibility of applying AIIR to LD CRC diagnosis, compared to that of standard-of-care CT examination. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-aw-2244/rc).

Methods

Study design and patient cohort

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of the Sixth Affiliated Hospital of Sun Yat-sen University (No. 2023ZSLYEC-313) and informed consent was taken from all individual participants.

From March to June 2024, a total of 393 consecutive patients scheduled for contrast-enhanced abdominopelvic CT due to suspected or confirmed CRC at the Sixth Affiliated Hospital of Sun Yat-sen University were considered for inclusion in this study. The exclusion criteria were: (I) lack of pathology reports or follow-up; (II) pathologically confirmed non-CRC; (III) previous treatments of resection surgery, radiotherapy or chemotherapy causing limitations in image analysis; and (IV) presence of severe motion/metal artifacts on CT images. Finally, 203 patients with CRC were enrolled. Figure 1 shows the flowchart of study design.

Figure 1 Flowchart of study design. AIIR, artificial intelligence iterative reconstruction; CRC, colorectal cancer; HIR, hybrid iterative reconstruction.

CT image acquisition and reconstruction

All patients underwent contrast-enhanced abdominal CT on a 64-row CT scanner (uCT 760, United Imaging Healthcare, Shanghai, China). The LD CT was obtained immediately after RD portal venous phase (PVP) CT in the same breath hold. The scanning parameters were as follows: 120 kVp tube voltage; 0.62 s rotation time; 1.0 pitch; 40 mm collimation; automatic tube current modulation with a reference of 200 and 20 mAs for the RD and LD scanning, respectively (15). The contrast medium (400 mg I/mL, Iomeprol, Bracco Imaging Italia S.r.L.) with a weight-dependent volume of 1.2 mL/kg was intravenously administered at an injection rate of 3.0 mL/s, followed by a 30 mL saline chase at the same rate. A region of interest (ROI) for bolus tracking was placed on the descending aorta with a triggering threshold of 200 Hounsfield units (HU).

RD images were reconstructed with both HIR (Karl 3D, United Imaging Healthcare, Shanghai, China) and AIIR, referred to as RD-HIR and RD-AIIR, respectively. LD images were reconstructed with AIIR (LD-AIIR). All reconstructions were 1 mm in slice thickness and 1 mm in slice interval. Finally, a total of 609 image sets (203 patients × 3 reconstructions) were generated for the image analysis.

Technically, AIIR adapts the model-based iterative reconstruction (MBIR) framework by replacing its traditional regularization term with a dedicated denoising convolutional neural network (CNN) for noise suppression. While both operate in the projection domain, their data optimization differs. The data optimization process of MBIR can be formulated as:

U*=argminU(||AU−Y||w2+βR(U))[1]

where U* is the final optimized image, U is the CT image, Y is the measured projection data, A is the well calibrated system matrix, and β is a constant. The term ||AU−Y||w2 and βR(U) represent the data fidelity and regularization term, respectively. In MBIR, βR(U) only focuses on a very simple description of clinical image characteristics, resulting in unnatural image structures, such as “plastic-like” image appearances, which is usually more severe when the dose is lower. To address this problem, a dedicated CNN was used to replace βR(U) to extract noise from LD images without changing anatomic structure appearance by learning the relationship between signal and noise. Hence, AIIR combines the strengths of MBIR in characterizing image details with the superior capability of CNN in handling image noise and texture (14-19). Accordingly, AIIR has demonstrated promising performance in various LD CT applications, including pediatric chest CT (16), lung screening (17), hepatic lesion detection (18), aortic CT angiography (19), abdominopelvic CT (20) and CT urography (15).

VPI and RLNM assessment

VPI is defined as either: (I) tumor penetration through the visceral peritoneum; or (II) perforated tumor with tumor cells connecting to the visceral peritoneum through inflammation (9). In the areas of the colon and rectum without peritoneal covering, specifically the posterior aspects of the ascending and descending colon and lower rectum, T4a is not applicable. In these anatomical regions, tumors that invaded adjacent organs or structures, such as bladder, prostate and ureter (T4b), were statistically considered T4a-positive in this study.

Malignant clinical mesorectal lymph nodes (MLNs) were defined by three malignant imaging features in accordance with the European Society of Gastrointestinal and Abdominal Radiology (ESGAR) criteria (21): irregular margins, heterogeneous signal intensity, and round configuration. RLNM was regarded as positive if the lymph nodes met any of the following conditions (21): short-axis (SA) diameter ≥9 mm; SA diameter of 5–8 mm accompanied by at least two malignant features; or SA diameter <5 mm presenting all three malignant features. Detection of RLNM provides essential information for determining the need for adjuvant chemotherapy in patients with locally advanced disease (22).

Two board-certified radiologists (with over 15 years of experience in abdominal diagnosis) were asked to independently diagnose and assess the VPI and RLNM. The likelihood of the VPI and RLNM was evaluated solely according to the manifestation on the images, using a four-point scale: 1, definitely absent; 2, likely absent; 3, likely present and 4, definitely present. With pathological results in surgical patients (n=150) and imaging results [positron emission tomography (PET)/CT, magnetic resonance imaging (MRI), or follow-up CT] in non-surgical patients (n=53) serving as the reference standard, the diagnostic performance for VPI and RLNM was evaluated using a receiver operating characteristic (ROC) analysis by varying the threshold of confidence score (1-4). Each reader evaluated LD-AIIR, RD-HIR and RD-AIIR images in three separate sessions, with an interval period of 3 weeks to minimize memory recall. For each session, readers were presented with anonymized images in a random order, and they were permitted to scroll, zoom and adjust the display window settings. As discrepancy occurred in each session, consensus was obtained from the two radiologists upon discussion.

Image quality evaluation

Qualitative analysis

In the process of diagnostic evaluation, the two aforementioned radiologists also scored the qualitative image quality for each patient, regarding the primary tumor conspicuity, tumor margin sharpness, the diagnostic certainty of hepatic metastasis and extramural venous invasion, using a five-point Likert scale system. The specific grading criteria were as follows: 1= poor, 2= low, 3= average, 4= good and 5= excellent for the first two metrics; 1= unable to estimate, 2= weak, 3= moderate, 4= strong and 5= definite for the last two metrics.

Quantitative analysis

Three circular ROIs of 10–20 mm² were drawn on the solid component of the primary tumor, the normal intestinal wall, and the subcutaneous abdominal fat by the third radiologist. Care was taken to avoid vessel edges, calcification plaques, and peristalsis artifacts that may be present in each ROI. The copy-paste function was used to ensure consistency on ROI placement throughout the three image sets. Image noise was defined as the standard deviation (SD) of the CT value (in HU) within the ROI. The contrast-to-noise ratio (CNR) of primary tumor relative to intestinal wall (CNRtumortointestinalwall) and primary tumor relative to subcutaneous abdominal fat (CNRtumortosubcutaneousfat) were calculated as

CNRtumortointestinalwall=HUtumor−HUintestinalwallSDintestinalwall[2]

CNRtumortosubcutaneousfat=HUtumor−HUsubcutaneousfatSDsubcutaneousfat[3]

where HU denotes the mean CT value within the ROI.

Regression analysis

To evaluate the independent effect of imaging protocols on diagnostic accuracy for VPI and RLNM, two separate generalized estimating equation (GEE) models with a logit link function and an exchangeable correlation structure were constructed. In each model, the observation unit was one image set per patient, and the patient identifier was specified as the subject-level clustering variable to account for within-patient correlations arising from the three image sets per patient. The dependent variable was diagnostic correctness (correct vs. incorrect) for the given diagnostic task. The primary independent variable was imaging protocol (LD-AIIR, RD-AIIR, RD-HIR, with RD-HIR as reference). Covariates included age, body mass index, and gender. Image quality metrics (e.g., SD, CNR, edge sharpness) were not included in the primary models to avoid both collinearity and overadjustment. Odds ratios (ORs) and 95% confidence intervals (CIs) were calculated for each independent variable, representing the odds of correct diagnosis. Statistical significance was set at P<0.05.

Statistical analysis

All statistical analysis was performed using commercial software (SPSS version 27.0, IBM Corp., Armonk, NY, USA). Continuous variables were presented as mean ± SD or median and interquartile range (IQR) according to their normality, which was examined by the Kolmogorov-Smirnov test. Using the reference standard, the diagnostic performance for VPI and RLNM was compared in terms of the area under the curve (AUC), where the sensitivity, specificity and accuracy were reported using values at the maximum Youden index. Differences in AUC were compared by the DeLong test, and differences in sensitivity, specificity and accuracy were examined using the McNemar test. The qualitative scores and quantitative metrics (noise and CNR) were compared using the paired t-test or Wilcoxon signed-rank test, as appropriate. Inter-observer agreement for qualitative evaluation was tested by kappa analysis, where 0–0.20, 0.21–0.40, 0.41–0.60, 0.61–0.80, and 0.81–1 were considered as slight, fair, moderate, substantial, and perfect agreement, respectively. A P value of <0.05 was considered to indicate statistical significance.

Results

Patient characteristics

The final cohort consisted of 203 CRC patients (male/female: 125/78, colon/rectal cancer: 90/113) with a mean age of 59.2±13.0 years, and a mean BMI of 23.0±3.3 kg/m². According to the reference standard, 43 (21.2%) and 113 (55.7%) cases were classified as having VPI and RLNM among the included patients, respectively. The mean CTDIvol, DLP and effective dose [ED; ED conversion factor (23), 0.015 mSv·mGy⁻¹·cm⁻¹] for the RD PVP scanning were 14.7±2.9 mGy, 995.2±159.0 mGy·cm and 14.9±2.4 mSv, respectively. The corresponding values for LD PVP scanning were 1.5±1.1 mGy, 96.6±16.3 mGy·cm and 1.5±0.2 mSv, respectively, showing a dose reduction of up to 90.3%. Detailed information about patient characteristics is presented in Table 1.

Diagnostic performance for VPI

As shown in Table 2, AIIR significantly improved the diagnostic performance for VPI compared to HIR at RD setting (P<0.05), yielding an AUC of 0.89 (95% CI: 0.84–0.93) versus 0.80 (95% CI: 0.74–0.86). Although the AUC of LD-AIIR was inferior to that of RD-AIIR as expected (P<0.05), it remained comparable to that of RD-HIR (P>0.05). There were no significant differences in sensitivity among the three reconstructions (P>0.05). Nevertheless, significant improvement was observed in the specificity by the use of AIIR compared to RD-HIR, even at LD (P<0.05). Consequently, RD-AIIR achieved the highest accuracy (91.13%; 95% CI: 86.35–94.66%), followed by LD-AIIR (82.96%; 95% CI: 77.37–87.65%) and RD-HIR (78.33%; 95% CI: 72.02–83.79%). The inter-observer agreement was substantial for RD-AIIR (κ=0.781), RD-HIR (κ=0.711) and LD-AIIR (κ=0.798). All P values were derived from DeLong tests for AUC comparisons and McNemar tests for sensitivity, specificity and accuracy comparisons.

Our subgroup analysis revealed that 158 cases were correctly diagnosed across all three reconstruction groups. Among these, 144 cases were rated with limited confidence (score 2 or 3) on RD-HIR images. With AIIR reconstruction, 91 (63.19%) of these were graded in the same direction but with higher confidence (score 1 or 4) at RD. No significant difference in the scoring proportion was found between RD-HIR and LD-AIIR. Notably, among 160 patients without VPI, 37 cases (23.13%) were misdiagnosed as positive on RD-HIR images. AIIR improved the characterization of negative cases, where the misdiagnosis rate decreased to 18.75% (30/160) and 7.5% (12/160) on LD-AIIR and RD-AIIR, respectively. A representative case is shown in Figure 2.

Figure 2 Venous phase CT and histopathological findings of an 83-year-old male with a tumor of pathological stage T3 in the upper rectum (all CT images: WW/WL =300/70 HU). On RD-HIR image, the boundary between the tumor and the bladder wall was ill-defined, suggesting VPI and organ invasion (stage T4b). On both RD-AIIR and LD-AIIR images, radiologists classified VPI as absent (stage T3) due to the distinguishable thin, clear fat space (arrows) between the tumor boundary and bladder wall. Histopathological findings (hematoxylin-eosin staining, ×40 magnification) showed tumor (arrows) infiltrated to the muscular layer but not to the visceral peritoneum (stage T3). The red circle demarcates the tumor area of the surgical gross specimen. CT, computed tomography; HU, Hounsfield unit; LD-AIIR, low-dose with artificial intelligence iterative reconstruction; RD-AIIR, routine-dose with artificial intelligence iterative reconstruction; RD-HIR, routine-dose with hybrid iterative reconstruction; VPI, visceral peritoneal invasion; WL, window level; WW, window width.

Diagnostic performance for RLNM

For RLNM diagnosis, the highest values for all diagnostic metrics were obtained on RD-AIIR. The diagnostic performance of LD-AIIR was inferior to that of RD-AIIR (all P<0.05). Compared to RD-HIR, the AUC and accuracy were significantly lower (both P<0.05) on LD-AIIR, although sensitivity and specificity did not differ significantly (both P>0.05) (Table 3, Figure 3). The inter-observer agreement was substantial for RD-AIIR (κ=0.762), RD-HIR (κ=0.795) and LD-AIIR (κ=0.773). All P values were derived from DeLong tests for AUC comparisons and McNemar tests for sensitivity, specificity and accuracy comparisons.

Figure 3 Venous phase CT and histopathological findings of a 46-year-old female with upper rectal cancer at lymph node stage N1a (WW/WL =400/70 HU). Both radiologists missed the RLNM (arrow) on LD-AIIR image, while correctly diagnosed it on both RD-AIIR and RD-HIR images due to more heterogeneous enhancement and ill-defined margins of the lymph node. Histopathological findings (hematoxylin-eosin staining, ×40 magnification) demonstrated the tumor infiltration (red arrows) within the lymph node. The red circle demarcates the tumor area of the surgical gross specimen. CT, computed tomography; HU, Hounsfield unit; LD-AIIR, low-dose with artificial intelligence iterative reconstruction; RD-AIIR, routine-dose with artificial intelligence iterative reconstruction; RD-HIR, routine-dose with hybrid iterative reconstruction; RLNM, regional lymph node metastasis; WL, window level; WW, window width.

Image quality

Qualitative analysis

The results of qualitative analysis are shown in Table 4 and Figure 4. RD-AIIR provided superior tumor conspicuity as compared to the other two reconstructions (both P<0.05). The mean score of tumor conspicuity was slightly lower on LD-AIIR than that on RD-HIR, without statistical difference (P>0.05). Similar results were observed for the evaluation of tumor margin sharpness and the diagnostic certainty of hepatic metastasis. Regarding the diagnostic certainty of extramural venous invasion, there were no significant differences among the three reconstructions (P>0.05). The scoring results of extramural venous invasion were all lower than those of hepatic metastasis on three reconstructions, indicating the difficulties in identifying such characteristics on CT images. The inter-observer agreement for all qualitative scoring was substantial to perfect for RD-AIIR (κ=0.729–1), RD-HIR (κ=0.657–0.888) and LD-AIIR (κ=0.654–0.944). All P values were derived from the Wilcoxon signed-rank test.

Figure 4 Comparison of qualitative image quality among three reconstructions in (A) a 59-year-old male with a tumor of stage T3 in the upper rectum (WW/WL =400/70 HU), and (B) a 65-year-old male with hepatic metastases secondary to colon cancer (WW/WL =220/80 HU). (A) The scores of the tumor conspicuity and tumor margin sharpness for RD-AIIR, RD-HIR and LD-AIIR were 5, 4 and 3, respectively. (B) The scores of the diagnostic certainty of hepatic metastases (arrows) for RD-AIIR, RD-HIR and LD-AIIR were 5, 4 and 4, respectively. HU, Hounsfield unit; LD-AIIR, low-dose with artificial intelligence iterative reconstruction; RD-AIIR, routine-dose with artificial intelligence iterative reconstruction; RD-HIR, routine-dose with hybrid iterative reconstruction; WL, window level; WW, window width.

Quantitative analysis

The results of quantitative analysis are shown in Figure 5. Significant differences were detected between each pair of the three reconstructions. Compared to RD-HIR, the image noise was markedly reduced on RD-AIIR and LD-AIIR images (both P<0.05). Meanwhile, the CNR values of tumor relative to intestinal wall and subcutaneous fat on RD-AIIR and LD-AIIR were both superior to those on RD-HIR images (both P<0.05). All P values were derived from the paired t-test.

Figure 5 Comparison of the noise and CNR among three reconstructions. *, P<0.05. All P values were derived from the paired t-test. CNR, contrast-to-noise ratio; LD-AIIR, low-dose with artificial intelligence iterative reconstruction; RD-AIIR, routine-dose with artificial intelligence iterative reconstruction; RD-HIR, routine-dose with hybrid iterative reconstruction.

Regression results

The multivariable GEE analysis showed that imaging protocol was significantly associated with diagnostic accuracy for both VPI and RLNM when compared to the reference protocol (RD-HIR). For VPI diagnosis, RD-AIIR significantly improved diagnostic accuracy compared to RD-HIR (OR =3.352, 95% CI: 2.071–5.425, P<0.001), whereas LD-AIIR did not show a statistically significant difference (OR =1.160, 95% CI: 0.971–1.387, P=0.102). For RLNM diagnosis, RD-AIIR was also associated with significantly improved diagnostic accuracy compared to RD-HIR (OR =1.190, 95% CI: 1.057–1.351, P=0.047), while LD-AIIR was associated with lower diagnostic accuracy than RD-HIR (OR =0.844, 95% CI: 0.740–0.964, P=0.012). None of the patient characteristics (age, BMI, gender) reached statistical significance in either model (all P>0.05).

Discussion

In this study, we investigated the feasibility of AIIR for CRC diagnosis, specifically for assessing VPI and RLNM at the LD setting. Results revealed that AIIR provides superior qualitative and quantitative image quality compared to HIR. For assessing VPI, LD-AIIR with an aggressive dose reduction of 90.3% reduced false-positive findings and demonstrated significantly higher specificity than RD-HIR. However, the diagnostic performance for RLNM was found to be inferior on LD-AIIR images.

LD CT inevitably increases noise level and compromises image quality, posing significant challenges for clinical diagnosis. Consequently, there is growing interest in applying DLR techniques to improve image quality in LD abdominal CT images (24,25). Jensen et al. (24) reported that DLR improved the quality of PVP abdominal images even at a 65% radiation dose reduction while preserving lesion detection performance. Similarly, Li et al. (25) revealed that, compared to the RD PVP images (15.51 mGy), 40 keV virtual monoenergetic images (7.95 mGy) reconstructed with DLR exhibited superior image quality while maintaining diagnostic sensitivity and specificity of small lesions. However, these studies primarily focused on CRC metastases rather than the primary tumor itself, leaving the impact of LD CT on CRC characterization largely unexplored. To the best of our knowledge, this study is the first to investigate the potential of AIIR for LD CRC imaging. Our results demonstrate that a substantially reduced radiation dose (1.5±1.1 vs. 14.7±2.9 mGy, achieving a 90.3% reduction) combined with AIIR remains feasible for CRC diagnostic imaging, compared to standard-of-care CT examinations.

In CRC staging, the accurate identification of tumor-related features (e.g., VPI) is heavily dependent on image quality, given the inherent limitations in soft-tissue contrast resolution in CT images reconstructed with HIR. The present findings demonstrate that LD-AIIR significantly improves tumor conspicuity, providing a CNR comparable to that of RD-HIR, which was analogous to previous investigations (17,18). This improvement in CNR translated into diagnostic performance for VPI that was comparable to RD-HIR, and LD-AIIR also yielded significantly higher specificity. This may be attributed to AIIR’s superior contrast resolution, which improves recognition of CRC contrast enhancement. This facilitates clear visualization of tumor boundaries and subtle imaging characteristics, such as irregular VPI—a hallmark of T4a staging. By improving the delineation of these critical features, AIIR may reduce false positives and avoid misinterpretation of VPI, even under aggressive dose reduction. The regression analysis corroborated these observations for VPI diagnosis, showing that RD-AIIR significantly outperforms RD-HIR (OR =3.35, P<0.001), whereas LD-AIIR yields comparable diagnostic correctness to RD-HIR (OR =1.16, P=0.102).

When evaluating RLNM, LD-AIIR showed a marked decline in AUC and diagnostic accuracy compared with both RD-HIR and RD-AIIR, and regression analysis showed that LD-AIIR was associated with significantly lower diagnostic correctness compared to RD-HIR (OR =0.84, P=0.012). This reduction is primarily attributed to the inherent challenges of CT in characterizing small or subtle lymph node abnormalities, which persist even with advanced reconstruction algorithms. Previous studies have similarly reported suboptimal diagnostic performance for RLNM detection on CT, with sensitivities ranging from 64% to 70% and specificities from 71% to 78% (11). These findings are consistent with current clinical guidelines (26), which recommend combining CT with other imaging modalities, such as MRI or PET/CT, to improve the accuracy of lymph node staging. Consequently, although LD-AIIR enables substantial radiation dose reduction, its application in RLNM detection should be considered supplementary. Comprehensive clinical decision-making should integrate a multimodal imaging approach to ensure precise staging and appropriate treatment planning.

In the present study, the choice of reference standard followed routine clinical practice in CRC management. According to the routine protocol, pathological confirmation was typically obtained for patients with early-stage tumors or tumors that had regressed sufficiently after treatment to meet resectability criteria, which was the case for 150 out of 203 enrolled patients. The remaining 53 patients tended to present with locally advanced disease that precluded immediate surgery and required preoperative neoadjuvant therapy. For this subgroup, surgical pathological staging was unavailable at baseline, as any eventual surgery would reflect post-treatment response rather than pretreatment tumor stage. In accordance with clinical practice guidelines such as those of the National Comprehensive Cancer Network (NCCN) (5,6), imaging-based evaluation [high-resolution magnetic resonance (MR), PET/CT, or follow-up] was therefore adopted as the reference standard for this subgroup. Excluding these non-surgical patients would have skewed the study population toward early-stage disease, and the observed diagnostic performance of the index test would no longer have adequately represented the full clinical spectrum of CRC patients requiring evaluation with LD CT. This clinically driven approach reflects a pragmatic balance between diagnostic ideal and clinical feasibility, preserving the generalizability of the findings to real-world practice.

The rationale for our AIIR framework is supported by parallel advances in AI-driven medical imaging. Conceptually, AIIR’s role in recovering diagnostic information from low‑dose CT aligns with a core principle established in image enhancement research: as systematically reviewed by Lai et al. (27), integrated processing pipelines are essential for improving visibility under challenging acquisition conditions—whether in underwater imaging or in noisy CT data. Methodologically, our blended MBIR and CNN approach draws support from the proven efficacy of hybrid fusion strategies in enhancing image clarity, as demonstrated by Xu et al. (28). Furthermore, the rigorous parameter tuning and robustness validation in our AIIR workflow are grounded in established findings on the necessity of systematic model optimization for clinical task performance, as shown in studies on breast and lung cancer classification (29,30). Ultimately, our observed improvements in tumor conspicuity and margin sharpness using AIIR resonate with the proven capability of advanced CNN architectures to enhance fine-detail lesion characterization, as illustrated by Yang et al. (31).

This study has some limitations that should be acknowledged. First, the AIIR algorithm represents only one implementation of the DLR algorithms, for which the current results are vendor-specific. However, the demonstrated investigative design can be easily translational to other DLR algorithms and CT scanners. Second, though the images were mixed and presented in a randomized order, bias due to the difference in image appearance between two algorithms was inevitable during subjective rating and should be carefully considered when interpreting the results. Third, in our clinical practice, the AIIR algorithm was applied in an indiscriminate manner to all consecutive patients. For the purpose of this foundational analysis, however, two specific cases with non-diagnostic image quality from severe metal or motion artifacts that precluded reliable tumor assessment were excluded. Consequently, the performance of AIIR in such challenging scenarios remains unevaluated and warrants further investigation, potentially through techniques such as simulating low-quality conditions or conducting prospective studies in targeted patient populations. Fourth, the involvement of experienced radiologists may limit generalizability of our findings to less experienced readers, although good to excellent inter-reader agreement was observed. Future studies incorporating readers with varying levels of experience are warranted to assess the robustness of the proposed approach across different clinical settings. Lastly, although this study demonstrated the superior diagnostic performance of AIIR, it was not designed to assess its influence on clinical decision-making. Future prospective studies are thus needed to investigate whether the enhanced visualization of features (e.g., fat plane disruption, nodal heterogeneity) translates into actionable changes in patient management.

Conclusions

In conclusion, AIIR provides superior image quality improvement compared to HIR. The use of AIIR allows up to 90.3% dose reduction for reliable VPI assessment of CRC while maintaining comparable image quality to RD HIR in abdominal CT.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-aw-2244/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-aw-2244/dss

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-aw-2244/coif). T.W. and S.Z. are employed scientific researchers of United Imaging Healthcare, Shanghai, China. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of the Sixth Affiliated Hospital of Sun Yat-sen University (No. 2023ZSLYEC-313) and informed consent was taken from all individual participants.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Biller LH, Schrag D. Diagnosis and Treatment of Metastatic Colorectal Cancer: A Review. JAMA 2021;325:669-85. [Crossref] [PubMed]
2. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
3. Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin 2024;74:12-49. [Crossref] [PubMed]
4. Xie Y, Shi L, He X, Luo Y. Gastrointestinal cancers in China, the USA, and Europe. Gastroenterol Rep (Oxf) 2021;9:91-104. [Crossref] [PubMed]
5. Benson AB, Venook AP, Al-Hawary MM, Arain MA, Chen YJ, Ciombor KK, et al. Colon Cancer, Version 2.2021, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2021;19:329-59. [Crossref] [PubMed]
6. Benson AB, Venook AP, Al-Hawary MM, Azad N, Chen YJ, Ciombor KK, et al. Rectal Cancer, Version 2.2022, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2022;20:1139-67. [Crossref] [PubMed]
7. Cervantes A, Adam R, Roselló S, Arnold D, Normanno N, Taïeb J, Seligmann J, De Baere T, Osterlund P, Yoshino T, Martinelli EESMO Guidelines Committee. Electronic address: clinicalguidelines@esmo. Metastatic colorectal cancer: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up. Ann Oncol 2023;34:10-32.
8. Dighe S, Blake H, Koh MD, Swift I, Arnaout A, Temple L, Barbachano Y, Brown G. Accuracy of multidetector computed tomography in identifying poor prognostic factors in colonic cancer. Br J Surg 2010;97:1407-15. [Crossref] [PubMed]
9. Weiser MR. AJCC 8th Edition: Colorectal Cancer. Ann Surg Oncol 2018;25:1454-5.
10. Chen CY, Hsu JS, Jaw TS, Kuo YT, Wu DC, Lee CH, Shih MC, Tsai TH, Kuo CH, Chen YT, Yang LH, Liu GC. Lowering radiation dose during dedicated colorectal cancer MDCT: comparison of low tube voltage and sinogram-affirmed iterative reconstruction at 80 kVp versus blended dual-energy images in a population of patients with low body mass index. Abdom Imaging 2015;40:2867-76. [Crossref] [PubMed]
11. Nerad E, Lahaye MJ, Maas M, Nelemans P, Bakers FC, Beets GL, Beets-Tan RG. Diagnostic Accuracy of CT for Local Staging of Colon Cancer: A Systematic Review and Meta-Analysis. AJR Am J Roentgenol 2016;207:984-95. [Crossref] [PubMed]
12. Koetzier LR, Mastrodicasa D, Szczykutowicz TP, van der Werf NR, Wang AS, Sandfort V, van der Molen AJ, Fleischmann D, Willemink MJ. Deep Learning Image Reconstruction for CT: Technical Principles and Clinical Prospects. Radiology 2023;306:e221257. [Crossref] [PubMed]
13. Jensen CT, Liu X, Tamm EP, Chandler AG, Sun J, Morani AC, Javadi S, Wagner-Bartak NA. Image Quality Assessment of Abdominal CT by Use of New Deep Learning Image Reconstruction: Initial Experience. AJR Am J Roentgenol 2020;215:50-7. [Crossref] [PubMed]
14. Li J, Zhu J, Zou Y, Zhang G, Zhu P, Wang N, Xie P. Diagnostic CT of colorectal cancer with artificial intelligence iterative reconstruction: A clinical evaluation. Eur J Radiol 2024;171:111301. [Crossref] [PubMed]
15. Liu H, Jiang X, Wang T, Zhang G, Zhao M, Xiang W, Peng H. Low-dose corticomedullary phase CT urography with artificial intelligence iterative reconstruction for bladder cancer evaluation. Br J Radiol 2026;99:541-9. [Crossref] [PubMed]
16. Zhang F, Peng L, Zhang G, Xie R, Sun M, Su T, Ge Y. Artificial Intelligence Iterative Reconstruction for Dose Reduction in Pediatric Chest CT: A Clinical Assessment via Below 3 Years Patients With Congenital Heart Disease. J Thorac Imaging 2025;40:e0827. [Crossref] [PubMed]
17. Yang L, Liu H, Han J, Xu S, Zhang G, Wang Q, Du Y, Yang F, Zhao X, Shi G. Ultra-low-dose CT lung screening with artificial intelligence iterative reconstruction: evaluation via automatic nodule-detection software. Clin Radiol 2023;78:525-31. [Crossref] [PubMed]
18. You Y, Zhong S, Zhang G, Wen Y, Guo D, Li W, Li Z. Exploring the Low-Dose Limit for Focal Hepatic Lesion Detection with a Deep Learning-Based CT Reconstruction Algorithm: A Simulation Study on Patient Images. J Imaging Inform Med 2024;37:2089-98. [Crossref] [PubMed]
19. Guo R, Peng S, Zhou W, Zhang G, Rong P, Liang Q, Peng R, Deng J, Hu P. One-stop combined coronary-craniocervical computed tomography angiography with low-dose body coverage using artificial intelligence iterative reconstruction: a clinically feasible solution to multi-territorial atherosclerosis diagnosis. Quant Imaging Med Surg 2025;15:1516-27. [Crossref] [PubMed]
20. Cai X, Han J, Zhou W, Yang F, Liu J, Wang Q, Li R. The utility of low-dose pre-operative CT of ovarian tumor with artificial intelligence iterative reconstruction for diagnosing peritoneal invasion, lymph node and hepatic metastasis. Abdom Radiol (NY) 2025;50:5463-73. [Crossref] [PubMed]
21. Beets-Tan RGH, Lambregts DMJ, Maas M, Bipat S, Barbaro B, Curvo-Semedo L, Fenlon HM, Gollub MJ, Gourtsoyianni S, Halligan S, Hoeffel C, Kim SH, Laghi A, Maier A, Rafaelsen SR, Stoker J, Taylor SA, Torkzad MR, Blomqvist L. Magnetic resonance imaging for clinical management of rectal cancer: Updated recommendations from the 2016 European Society of Gastrointestinal and Abdominal Radiology (ESGAR) consensus meeting. Eur Radiol 2018;28:1465-75. [Crossref] [PubMed]
22. van der Pas MH, Meijer S, Hoekstra OS, Riphagen II, de Vet HC, Knol DL, van Grieken NC, Meijerink WJ. Sentinel-lymph-node procedure in colon and rectal cancer: a systematic review and meta-analysis. Lancet Oncol 2011;12:540-50. [Crossref] [PubMed]
23. The 2007 Recommendations of the International Commission on Radiological Protection. ICRP publication 103. Ann ICRP 2007;37:1-332. [Crossref] [PubMed]
24. Jensen CT, Gupta S, Saleh MM, Liu X, Wong VK, Salem U, Qiao W, Samei E, Wagner-Bartak NA. Reduced-Dose Deep Learning Reconstruction for Abdominal CT of Liver Metastases. Radiology 2022;303:90-8. [Crossref] [PubMed]
25. Li S, Yuan L, Lu T, Yang X, Ren W, Wang L, Zhao J, Deng J, Liu X, Xue C, Sun Q, Zhang W, Zhou J. Deep learning imaging reconstruction of reduced-dose 40 keV virtual monoenergetic imaging for early detection of colorectal cancer liver metastases. Eur J Radiol 2023;168:111128. [Crossref] [PubMed]
26. Caruso D, Polici M, Bellini D, Laghi A. ESR Essentials: Imaging in colorectal cancer-practice recommendations by ESGAR. Eur Radiol 2024;34:5903-10. [Crossref] [PubMed]
27. Lai YL, Ang TF, Bhatti UA, Ku CS, Han Q, Por LY. Color correction methods for underwater image enhancement: A systematic literature review. PLoS One 2025;20:e0317306. [Crossref] [PubMed]
28. Xu J, Kiah MLM, Noor RM, Por LY, Wu Y. MHF-UIE a multi-task hybrid fusion method for real-world underwater image enhancement. Sci Rep 2025;15:18131. [Crossref] [PubMed]
29. Li H, Govindarajan V, Ang TF, Shaikh ZA, Ksibi A, Chen YL, Ku CS, Leong MC, Shabaruddin FH, Wan Ishak WZ, Por LY. MSPO: A machine learning hyperparameter optimization method for enhanced breast cancer image classification. Digit Health 2025;11:20552076251361603. [Crossref] [PubMed]
30. Nabeel SM, Bazai SU, Alasbali N, Liu Y, Ghafoor MI, Khan R, Ku CS, Yang J, Shahab S, Por LY. Optimizing lung cancer classification through hyperparameter tuning. Digit Health 2024;10:20552076241249661. [Crossref] [PubMed]
31. Yang J, Siddique MA, Ullah H, Gilanie G, Por LY, Alshathri S, El-Shafai W, Aldossary H, Gadekallu TR. BrainCNN: Automated brain tumor grading from magnetic resonance images using a convolutional neural network-based customized model. SLAS Technol 2025;34:100334. [Crossref] [PubMed]