Influence of deep learning image reconstruction algorithm for reducing radiation dose and image noise compared to iterative reconstruction and filtered back projection for head and chest computed tomography examinations: a systematic review

Background The most recent advances in Computed Tomography (CT) image reconstruction technology are Deep learning image reconstruction (DLIR) algorithms. Due to drawbacks in Iterative reconstruction (IR) techniques such as negative image texture and nonlinear spatial resolutions, DLIRs are gradually replacing them. However, the potential use of DLIR in Head and Chest CT has to be examined further. Hence, the purpose of the study is to review the influence of DLIR on Radiation dose (RD), Image noise (IN), and outcomes of the studies compared with IR and FBP in Head and Chest CT examinations. Methods We performed a detailed search in PubMed, Scopus, Web of Science, Cochrane Library, and Embase to find the articles reported using DLIR for Head and Chest CT examinations between 2017 to 2023. Data were retrieved from the short-listed studies using Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) guidelines. Results Out of 196 articles searched, 15 articles were included. A total of 1292 sample size was included. 14 articles were rated as high and 1 article as moderate quality. All studies compared DLIR to IR techniques. 5 studies compared DLIR with IR and FBP. The review showed that DLIR improved IQ, and reduced RD and IN for CT Head and Chest examinations. Conclusions DLIR algorithm have demonstrated a noted enhancement in IQ with reduced IN for CT Head and Chest examinations at lower dose compared with IR and FBP. DLIR showed potential for enhancing patient care by reducing radiation risks and increasing diagnostic accuracy.


Introduction
Computed tomography (CT) plays an important role in modern diagnostic radiology and assists in the identification of various complex disorders.Over the past ten years, CT scan utilization has increased significantly globally as new clinical reasons are continually identified.An estimated 375 million CT examinations are continually performed annually worldwide, with a 3-4% annual growth rate.The demands of physicians and other health care providers, as well as technology developments, have had a considerable impact on the world market for CTs.Compared to other traditional imaging modalities, CT scans offer significantly higher radiation doses (RD) despite having significant diagnostic benefits for specific patients.Adult CT scans dramatically raise cancer risk.[3][4][5] A recent study reported seventeen-fold variations in high-dose CT examinations among different countries.There is a four-fold variation in effective dose [ED] for Chest and abdomen examinations with less variation for CT head in adults and suggested optimization of radiation doses. 6The most recommended practice in the CT sector is to reduce CT radiation exposure as low as reasonably achievable while maintaining the Image Quality (IQ).Reducing the exposure factors of tube voltage (kVp) and tube current (mA) reduces RD but increases image noise. 7,8Up until ten years ago, Filtered back projection (FBP) was the only technique used for image reconstruction in CT.Although this method produces high-quality images it has noise issues at low doses and is prone to artifacts.Although an iterative reconstruction (IR) method was proposed in 1970, computational power restrictions prevented its widespread use in clinical settings.The Hybrid Iterative reconstruction (HIR) method was introduced in 2009 which had low computation time and allowed it to be implemented in clinical practice.The HIR combines iteratively reconstructed images in the raw data domain with FBP images to reduce image noise (IN).The first complete model-based iterative reconstruction (MBIR) received FDA approval in 2011.Compared to the HIR technique, this reconstruction minimizes artifacts and noise.However, it requires a greater computational power demand, which results in lengthy reconstruction times.[18] Deep learning image reconstruction algorithms (DLIR) are the most recent developments in CT image reconstruction technology.DLIRs are increasingly replacing IR techniques due to their disadvantages such as negative image texture and nonlinear spatial resolutions.DLIR is based on deep convolution neural networks (CNN) which learn from the input data sets.It gains the ability to distinguish actual signal and IN through training using pairs of low and high-quality images.In comparison to FBP and IR, the trained CNNs can distinguish between noise and signal much better, allowing for better dose reduction while preserving the image quality.DLIR produces an image texture similar to that of FBP even at low doses and high strengths.0][21][22] More research is required to determine the potential applications of DLIR in clinical settings.Our literature search showed there is no systematic review performed in head and chest CT examinations using Deep learning reconstruction algorithm for reducing RD and improving IQ in CT.Hence, the purpose of the article is to review the influence of DLIR on RD, IN, and outcomes of the studies compared with IR and FBP in CT Head and Chest examinations.

Design
This review was carried out as per the "Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA)" guidelines. 23terature search strategy A comprehensive literature search was performed using databases such as "PubMed, Scopus, Web of Science, Cochrane Library, and Embase" to find the relevant original studies (Table 1).The MeSH terms such as "Deep Learning Image Reconstruction" "Radiation dose" "Image quality", "Head and Chest Computed Tomography" were used (Table 2).The search was limited to the English language including both adult and paediatric populations of Head and Chest CT examinations.

Selection criteria
Articles were screened considering the Participant's Intervention Comparison and Outcome (PICO) methodology.Case studies, case reports, conference abstracts, letters, editorial reviews, meta-analyses, or surveys were not included.
The title and abstract of all the articles were independently and blindly screened by the two researchers.The articles that described a comparison of DLIR algorithms with the IR technique or FBP were included in the final review.The exclusion criteria were phantom studies, physics-based performance of DLIR, other language than English, articles with no comparison of DLIR with FBP, HIR/MBIR, and articles with no Hounsfield Unit (HU), Contrast to Noise Ratio (CNR), Signal to Noise Ratio (SNR).

Data extraction
Data from each article was assessed independently by two researchers and any differences were solved by the third researcher.

Quality assessment
To evaluate the quality of all the included articles, the custom-made Quality Assessment (QA) scale was used. 24The list of all the questions for the quality assessment (underlying data).A score of 1 was given if the answer to the question was "yes" and each study was assigned a score ranging from 0 to 18. Based on the total scores obtained by each study, the studies were classified into three quality levels: Low-quality studies (score of 6 or lower), Moderate-quality studies (score between 7 and 11), High-quality studies (score of 12 or more).

Study selection
The search in PubMed, Scopus, Web of Science, Cochrane Library and Embase resulted in 196 studies.101 duplicates were removed.The title and abstract of 95 studies were assessed and 76 studies were excluded as they did not meet the inclusion criteria.A total of 19 reports were sought for retrieval.A total of the full text of 19 articles were assessed for eligibility.Among them, 4 articles were excluded (3 studies were excluded due to no comparison of DLIR with FBP and HIR/MBIR, and 1 article were excluded due to lack of HU, CNR, and SNR).Finally, 15 articles were included in the systematic review.

Characteristics of selected studies
CT imaging has increased recently with the advancement in CT technology.The studies included in the review covered different countries such as China (n = 8), Japan (n = 1), France (n = 1), Korea (n = 3), Netherland (n =1), Sweden (n = 1).
The RD data and IQ parameters were collected from different CT vendors such as General Electric (GE) Health care (128, 256, 512-slice, and dual-energy CT), Siemens Healthineers (256-slice), Canon Medical system (320 and 640-slice).A total sample size of 1292 was collected from the included studies.4 studies used prospective data collection, and 11 studies used retrospective data collection.The characteristics of the study and the outcomes of each study are summarized in Table 3.

Quality assessment
The results of the quality assessment are summarized in Table 4.All studies compared DLR to hybrid iterative reconstruction techniques.5 studies compared DLIR with IR and FBP algorithms.A total of 14 studies were rated as high and 1 study as moderate quality.

Discussion
This systematic review focussed on investigating the influence of DLIR on RD, IN, and outcomes of the studies compared with IR and FBP in Head and Chest CT examinations.

CT head
Our review noted that for CT Brain examination, DLIR (Medium and High) showed reduced IN (18-52%), improved IQ (GM-WM differentiation) with better detection of cerebral lesions, and reduced RD (25%).In the Pediatric CT brain, a study by Sun et al. noted that higher strength DLIR reduced image noise and noted better detection of cerebral lesions in 0.625 mm compared to 5 mm slice thickness.The thinner sections of DLIR-H were able to identify micro-hemorrhages of less than 3 mm. 29 for the patients with interstitial lung disease (ILD) and noted that LDCT DLIR showed better visualization of honeycombing and assessment of bronchiectasis. 39Kim et al. noted that DLIR-H yielded higher scores in determining the prominence of the lungs main structures of the lungs. 34Jiang et al. noted that ULD-CT with DLIR under or overestimated the long diameter and sub-solid nodules compared with CECT Thorax.DLIR-H overestimated the solid and calcified nodules while underestimating the long diameter and amount of subnodules. 31Wang et al. noted LDCT with DLIR provides higher scores for assessing pulmonary lesions except for subsolid nodules or ground glass opacity nodules (GGN) compared to SD with HIR, whereas GGN greater than 4 mm can be picked up on LDCT DLIR images. 38Tian et al. reported that DLIR-H appeared to be slightly smoothed and DLIR M provides higher structures on visualization of smoother structures. 36Ferri et al. reported DLIR reconstruction series provided the smallest volume of emphysema compared with Adaptive statistical iterative reconstruction-V (ASIR-V) and FBP and also observed the increase in strength of DLIR led to a decrease in the size of emphysema. 30e study has a few limitations.Firstly, we did not include phantom studies.Secondly, we did not perform meta-analysis due to heterogeneity in terms of scanners and protocols used for head and chest examinations.The adoption of DLIR algorithms holds promise for improving IQ, reducing RD, and mitigating IN in Head and Chest CT examinations compared to traditional IR and FBP techniques.Healthcare providers may consider incorporating DLIR into their imaging protocols to enhance patient care by reducing radiation risks while maintaining diagnostic accuracy.Furthermore, future research efforts should focus on optimizing DLIR algorithms, investigating their long-term effects on patient outcomes, and evaluating cost-effectiveness compared to conventional reconstruction methods.Additional studies exploring the application of DLIR in other anatomical regions and patient populations could further expand its utility and impact on healthcare delivery.

Conclusion
In conclusion, DLIR is a versatile and valuable technology that consistently improves IQ, enhances lesion detection, reduces radiation exposure, and mitigates image artifacts across a wide range of medical imaging applications compared with IR and FBP.A careful selection of strengths of DLIR, slice thickness and radiation dose levels are required for evaluation of tiny lesions, which can overcome with next generation DLIR algorithms.Overall, DLIR holds promise for improving patient care and diagnostic accuracy in various clinical settings.
quality in CT brain and Chest examinations.The review provides valuable insights into the advancements of DLIR in CT imaging, its benefits in reducing radiation dose and image noise while improving Image quality, and highlights the potential for further research and application of DLIR in clinical settings.
Are the rationale for, and objectives of, the Systematic Review clearly stated?Yes Are sufficient details of the methods and analysis provided to allow replication by others?Yes

Are the conclusions drawn adequately supported by the results presented in the review? Yes
If this is a Living Systematic Review, is the 'living' method appropriate and is the search schedule clearly defined and justified?('Living Systematic Review' or a variation of this term should be included in the title.)

Not applicable
Competing Interests: No competing interests were disclosed.

Reviewer Expertise: Radiation Imaging and Dosimetry
I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.

Thayalan Kuppusamy
Dr Kamakshi Memorial Hospital, Chennai, Tamil Nadu, India Questions/Comments 1.What is the estimated CT examinations number in India, including pediatric?
2. Why abdomen CT is not taken for the study since it involves larger dose variation .
3.Why phantom study is excluded, it is basic method and gold standard of dose estimation?

Major comments
The review article provides comprehensive review of the Deep Learning Image Reconstruction (DLIR) and its application on radiation dose and image noise.This review particularly concentrating on the application and outcomes in head and chest Computed Tomography examinations compared to traditional Iterative (IR) and Filtered Back Projection (FBP) techniques.
In the methodology section the authors follow the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA), which is an evidence-based minimum set of items for reporting in systematic reviews and meta-analyses.It has been clearly explained.
In the results the author clearly explains the key outcome in related to the aim and objectives.In particular DLIR's influence on radiation dose reduction, image noise reduction, and enhancement of image quality in CT head and chest examinations were clearly explained.The discussion and conclusion effectively explain the results, Minor comments Some more references may be added to support the review article.The article may include the data of various other anatomical part like thorax.

Suggestions to editor:
On a whole , it is an excellent review article that represents a valuable contribution for the dose reduction in the CT scan using proper image reconstruction technology.Also this article provides the impact of Deep Learning Image Reconstruction technology on improvement of image quality.

Minor suggestions:
Please proofread for grammatical errors.Additionally, consider including points related to the smaller sample size of pediatric CT cases to provide a more comprehensive overview.

Suggestions to editor:
Overall, it is an excellent review article that represents a valuable contribution to the field of CT image reconstruction technology, providing insights into the impact of Deep Learning Image Reconstruction technology on dose reduction and improvement of image quality.
Are the rationale for, and objectives of, the Systematic Review clearly stated?Yes

Are sufficient details of the methods and analysis provided to allow replication by others? Yes
Is the statistical analysis and its interpretation appropriate?Yes Are the conclusions drawn adequately supported by the results presented in the review?Yes If this is a Living Systematic Review, is the 'living' method appropriate and is the search schedule clearly defined and justified?('Living Systematic Review' or a variation of this term should be included in the title.)

Not applicable
Competing Interests: No competing interests were disclosed.
Reviewer Expertise: Radiation Biology, Radiation dose related research in CT, X-rays, Fluoroscopy etc.. Radiography.
I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.
The benefits of publishing with F1000Research: Your article is published within days, with no editorial bias • You can publish traditional articles, null/negative results, case reports, data notes and more • The peer review process is transparent and collaborative • Your article is indexed in PubMed after passing peer review • Dedicated customer support at every stage • For pre-submission enquiries, contact research@f1000.com

Figure 1 .
Figure 1.Flow chart for study selection.

Reviewer Report 15
May 2024 https://doi.org/10.5256/f1000research.161532.r268965© 2024 Kuppusamy T. This is an open access peer review report distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

©
2024 Shetty S. This is an open access peer review report distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.Shashi Kumar Shetty NITTE university, Karnataka, IndiaMajor comments:The article provides an insightful and comprehensive review of the influence of Deep Learning Image Reconstruction (DLIR) on radiation dose, image noise, and outcomes in head and chest CT examinations compared to traditional Iterative (IR) and Filtered Back Projection (FBP) techniques.The methodology section is meticulously detailed, following PRISMA guidelines.The results section offers a clear summary of the inclusion and exclusion criteria, key outcomes related to DLIR's influence on radiation dose reduction, image noise reduction, and enhancement of image quality in CT head and chest examinations.The inclusion of quality assessment scores adds validation to the review's findings.The discussion and conclusion effectively explain the results, offering potential applications of DLIR in clinical settings.

Table 1 .
Study retrieval method from database.

Table 2 .
Participants intervention comparison and outcome methodology for determining study selection criteria.

Table 3 .
Characteristics of selected studies for Head and Chest CT examinations.

Table 4 .
Quality scores of the selected studies.

Table 5 .
Radiation dose parameters and percentage reduction in radiation dose, image noise of CT Head and Chest.

Are the rationale for, and objectives of, the Systematic Review clearly stated? Yes Are sufficient details of the methods and analysis provided to allow replication by others? Partly Is the statistical analysis and its interpretation appropriate? Yes Are the conclusions drawn adequately supported by the results presented in the review? Yes If this is a Living Systematic Review, is the 'living' method appropriate and is the search schedule clearly defined and justified? ('Living Systematic Review' or a variation of this term should be included in the title.) Yes Competing Interests:
4.Image noise (IN) depends on type of noise filters used-will all 15 articles used had same noise filter? 5. Is slice thickness is considered as a parameter in deciding radiation dose?6.The RD data and IQ parameter collected from various CT vendors shows variation, means not homogeneous?Comparing the RD with them will result a meaningful data or not?Why Phillips vendor is not included since large number of installation is here in India 7.CT radiation dose and cancer risk is debatable, it can not be taken as hypothesis.Probability of cancer induction is less if the dose is < 100 mSv.No CT scan offer such magnitude of radiation now a days.8.There was no mention about diagnostic reference level (DRL), especially in India.Is the stated DLIR brings the dose below the DRL or not ?9. CDDI, focal spot, slice thickness, filtration (bow type filter), tube rotation time, pitch, current modulation, AEC, low dose protocol, kV, mAs, scan area may differ in each CT and influence doseall the parameter are accounted?No competing interests were disclosed.

have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.
This is an open access peer review report distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
© 2024 S D. Dr. Tamijeselvan S College of MRIT, Mother Teresa Postgraduate and Research institute of Health sciences, Indira Nagar, Puducherry, India

the rationale for, and objectives of, the Systematic Review clearly stated? Yes Are sufficient details of the methods and analysis provided to allow replication by others? Yes Is the statistical analysis and its interpretation appropriate? Yes Are the conclusions drawn adequately supported by the results presented in the review? Yes If this is a Living Systematic Review, is the 'living' method appropriate and is the search schedule clearly defined and justified? ('Living Systematic Review' or a variation of this term should be included in the title.)
1. Lee JE, Choi SY, Hwang JA, Lim S, et al.: The potential for reduced radiation dose from deep learning-based CT image reconstruction: A comparison with filtered back projection and hybrid iterative reconstruction using a phantom.Medicine (Baltimore).2021; 100 (19): e25814 PubMed Abstract | Publisher Full Text Are

have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.
Reviewer Report 30 April 2024 https://doi.org/10.5256/f1000research.161532.r268958