Auto-detection of motion artifacts on CT pulmonary angiograms with a physician-trained AI algorithm

1 1 Title: 2 Auto-detection of motion artifacts on CT pulmonary angiograms with a physician-trained AI 3 algorithm 4 Giridhar Dasegowda1,2 5 Bernardo C Bizzo1,2 6 Parisa Kaviani 1,2 7 Lina Karout1,2 8 Shadi Ebrahimian1 9 Subba R Digumarthy1 10 Nir Neumark2 11 James Hillis1,2 12 Mannudeep K Kalra1,2 13 Keith J Dreyer1,2 14 15 1- Massachusetts General Hospital and Harvard Medical School 16 2- Mass General Brigham Data Science Office 17 Corresponding author: Dr. Bernardo C Bizzo 18 Email: [email protected] . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice. 2 19 Abstract 20 Purpose: Motion-impaired CT images can result in limited or suboptimal diagnostic 21 interpretation (with missed or miscalled lesions) and patient recall. We trained and tested an 22 artificial intelligence (AI) model for identifying substantial motion artifacts on CT pulmonary 23 angiography (CTPA) that have a negative impact on diagnostic interpretation. 24 Methods: With IRB approval and HIPAA compliance, we queried our multicenter radiology 25 report database (mPower, Nuance) for CTPA reports between July 2015 - March 2022 for the 26 following terms: "motion artifacts," "respiratory motion," "technically inadequate," and 27 "suboptimal" or "limited exam." All CTPA reports belonged to two quaternary (Site A, n= 335; B, 28 n= 259) and a community (C, n= 199) healthcare sites. A thoracic radiologist reviewed CT 29 images of all positive hits for motion artifacts (present or absent) and their severity (no 30 diagnostic effect or major diagnostic impairment). Coronal multiplanar images belonging to 793 31 CTPA exams were de-identified and exported offline into an AI model building prototype 32 (Cognex Vision Pro, Cognex Corporation) to train an AI model to perform two-class classification 33 ("motion" or "no motion") with data from the three sites (70% training dataset, n= 554; 30% 34 validation dataset, n= 239). Separately, data from Site A and Site C were used for training and 35 validating; testing was performed on the Site B CTPA exams. A 5-fold repeated cross-validation 36 was performed to evaluate the model performance with accuracy and receiver operating 37 characteristics analysis (ROC). . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 3 38 Results: Among the CTPA images from 793 patients (mean age 63 ± 17 years; 391 males, 402 39 females), 372 had no motion artifacts, and 421 had substantial motion artifacts. The statistics 40 for the average performance of the AI model after 5-fold repeated cross-validation for the two- 41 class classification included 94% sensitivity, 91% specificity, 93% accuracy, and 0.93 area under 42 the ROC curve (AUC: 95% CI 0.89-0.97). 43 Conclusion: The AI model used in this study can successfully identify CTPA exams with 44 diagnostic interpretation limiting motion artifacts in multicenter training and test datasets. 45 Clinical relevance: The AI model used in the study can help alert the technologists about 46 the presence of substantial motion artifacts on CTPA where a repeat image acquisition can help 47 salvage diagnostic information. 48 49 50 51 52 53 54 55 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 4 56 Introduction 57 The chest represents one of the most frequently scanned body parts in CT, but also ranks high 58 among the most challenging parts to obtain an optimal diagnostic quality. Recent technical 59 advancements in multidetector-row CT scanners have led to tremendous improvements in 60 diagnostic quality with lower noise, higher contrast, and fewer motion artifacts. Although with 61 faster scanning times and better reconstruction techniques, most chest CT exams are generally 62 optimal for diagnostic interpretation, several artifacts can still have a negative impact on 63 diagnostic interpretations of chest CT [1]. The common artifacts include beam hardening, 64 photon starvation, partial volume, metal, motion, and cone-beam artifacts [2, 3]. Such artifacts 65 can hinder the optimal evaluation of lung abnormalities as well as mediastinal and vascular 66 findings. 67 Prior studies have reported that motion artifacts are frequent, especially on legacy scanners, 68 and can limit the diagnostic information from both routine chest CT as well as CT pulmonary 69 angiography (CTPA). Motion artifacts can result in misinterpretation as the artifacts can mimic 70 an embolus, or the artifact can cause an apparent abrupt vessel cut-off [4, 5]. In the lung 71 parenchyma, artifacts can reduce the ability to detect and characterize both focal (such as lung 72 nodules) and diffuse parenchymal processes. Such artifacts are especially common in CTPA of 73 critically ill patients and patients with shortness of breath and/or persistent cough [6]. The use 74 of wide-area detector scanners and scan capabilities such as high non-overlapping pitch and 75 faster rotation time reduce scanning duration and have lower artifacts [3, 7]. Such fast 76 acquisition modes are not compatible with dual-energy CT and in patients with large body . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 5 77 habitus. Furthermore, such scanners and advanced techniques still represent a minority of 78 clinically deployed scanners even in developed countries such as the United States [8]. 79 An automated method of detecting substantial motion artifacts at the time of scanning can 80 help re-scan patients during the same imaging session with better coaching or faster scanning 81 techniques. Recent advances in machine and deep learning (DL) have led to the creation of 82 several AI algorithms in medical imaging including in the areas of image reconstruction, triaging, 83 quality control and pathology detection [9-14]. Therefore, we trained and tested an artificial 84 intelligence (AI) model to identify substantial motion artifacts on CTPA that have a negative 85 impact on diagnostic interpretation. 86 87 Materials and Methods 88 Study Design 89 Our retrospective study was conducted after receiving approval from Institutional Review 90 Board (IRB). The study was Human Insurance Portability and Accountability Act (HIPAA) 91 compliant. The study methodology has been described in accordance to CLAIM guidelines [15]. 92 The model was trained and tested by physicians without any prior knowledge of, or training in 93 machine learning or coding. 94 95 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 6 96 Data Definitions 97 The study included adult patients (≥19 years) who underwent CTPA at one of the three 98 hospitals (quaternary hospitals: Massachusetts General Hospital, Brigham and Women's 99 Hospital; community hospital: Cooley Dickinson Hospital) within an integrated health system. A 100 commercial radiology report search engine, Nuance mPower, was used to identify radiology 101 reports with a mention of motion artifacts in CTPA examinations performed between Jan 2015 102 to November 2021. The following keywords were used to identify the eligible CTPA: "motion 103 artifacts," "respiratory motion," "technically inadequate," and "suboptimal" or "limited exam." 104 The search was optimized to identify positive CTPA with the keyword hits (without mention of 105 “absence” or “no” within a few words of the keywords). A negative search with the exact 106 keywords was used to identify the control CTPA exams without motion artifacts. 107 All CTPA exams were performed on one of the 28 CT scanners in the three participating sites 108 with single or dual-energy CT protocols using the standard of care scan protocols. For each 109 CTPA, a thin slice coronal multiplanar image at the level of descending thoracic aorta was de- 110 identified and exported offline from the PACS workstation. 111 112 Ground truth 113 In addition to the radiology reports, a thoracic radiologist (MKK with 16 years of subspecialty 114 experience) reviewed all CTPA and opined on the presence or absence of substantial motion 115 artifacts. Each CTPA thus had the opinions of two radiologists, the reporting radiologist, and the . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 7 116 study coinvestigator radiologist. Substantial motion artifacts were defined as CTPA exams with 117 the presence of motion artifacts involving both lungs (at least 50% of each lung) and limiting 118 the ability to assess pulmonary embolism or parenchyma. Minor motion artifacts involving a 119 single lobe or smaller portions of lungs without effect on the diagnostic evaluation of 120 pulmonary embolism or parenchyma were labeled as negative for substantial motion artifacts. 121 Although extremely common, minor artifacts are not as important since they should not trigger 122 repeat imaging. CTPA with evidence of “white lungs” (diffuse parenchymal opacities), a 123 substantial bilateral lung volume loss, or pneumonectomies were excluded from the study (n= 124 50 CTPA exams). These cases were excluded since it is difficult to assess motion artifacts’ 125 impact on the pulmonary evaluation in such cases. CTPA exams without the complete inclusion 126 of lung apex and bases were excluded from the study. 127 128 Model 129 The AI model was trained on a deep learning model-building platform, Vision Pro Deep Learning 130 (VPDL, COGNEX Corporation, Natick, MA). The software enables users to train DL models based 131 on a labeled image dataset using a vision-optimized deep neural network. The users require no 132 formal programming or coding knowledge or experience. The VPDL platform has two different 133 options for training classification models: High detail Mode (HDM) and Focused Mode (FM). The 134 HDM enables model training for challenging or complex applications (such as pixel-level 135 information in image domain) and provides higher accuracy. A heat map is also generated in 136 HDM that indicates the image region that was most influential in the classification decision. FM . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 8 137 enables fast training of models for simple applications (such as the distinction between 138 different image types or body parts). We used the HDM classification model for identifying 139 CTPA examinations with substantial motion artifacts. 140 141 Training 142 A physician co-investigator (GD with one-year post-doctoral research fellowship experience in 143 thoracic imaging) trained the AI model on the VPDL platform without prior programming or 144 data science knowledge. The images were de-identified, exported from the PACS workstation 145 (Visage), and then uploaded onto the software platform which was installed on a virtual 146 machine within the hospital intranet to maintain data security and privacy. Within the platform, 147 the study coinvestigator labeled each uploaded image as "motion" (with substantial motion 148 artifacts) or "no motion" (without substantial motion artifacts) based on the assessment from 149 the thoracic radiologist. In the first training, all CTPA examinations from three hospitals were 150 included. The software randomized the training and validation data set with a 70%-30% 151 distribution, respectively. We performed five-fold repeated cross-validation to evaluate the 152 robustness of the model. The output was recorded and separately analyzed. 153 To establish the inter-institutional generalizability of the model, we trained a model using 154 images from two hospitals (A and C, after excluding the data from site B), and then tested the 155 algorithm on CTPA data from the third hospital (site B). 156 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 9 157 Statistical Analysis 158 Information on the distribution of true positive, true negative, false positive, and false negative 159 CTPA was recorded in Microsoft Excel worksheets (Microsoft Inc). Data were analyzed with 160 SPSS statistical software, version 26 (IBM Inc.). The performance of the AI model was evaluated 161 using sensitivity, specificity, accuracy, and area under the curve (AUC) for the receiver operating 162 characteristic (ROC) analysis. For the five-fold repeated cross-validation, the average of the five 163 models was considered. We also estimated the F-score for AI model performance, as a measure 164 of the harmonic average of precision and recall/sensitivity. 165 166 Results 167 Our study included 793 CTPA examinations from 793 adult patients (mean age= 63 ± 17; 391 168 men, 402 women). The distribution of CTPA across each site is as follows: Site A, n= 335; Site B, 169 n= 259; Site C, n= 199. A total of 455/793 (57%) CTPA examinations were performed during the 170 emergency visit, while 277/793 (35%) and 111/793 (14%) examinations were among inpatients 171 and outpatients, respectively. Most CTPA with substantial motion artifacts were either from the 172 emergency department (n = 213/455) or inpatients (n = 146/227) with a minority of patients 173 coming with outpatient referrals (n = 33/111). There was no discrepancy between the reporting 174 radiologists and the research radiologist for the presence of substantial motion artifacts. 175 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 10 176 Model validation at three sites 177 Among CTPA exam datasets from all three sites in model training, 471 CTPA (n= 421/793, 178 53.1%) had substantial motion artifacts and 372 CTPA (n= 372/793, 46.9%) were without 179 substantial artifacts. For the average performance after 5-fold repeated cross-validation for the 180 two-class classification (as either with or without substantial motion artifacts), on the 30% 181 validation data set, the AI model had a sensitivity of 94%, specificity of 91%, 93% accurate with 182 AUC of 0.93 (95% confidence interval [95% CI] 0.89 - 0.97). The best performing model had F- 183 scores of 96% and 95% for identifying CTPA with and without substantial motion artifacts, 184 respectively. The two-class classification of the AI model is summarized in Fig 1. 185 186 Fig 1. Coronal MPR images of CTPA examinations in two patients with (A) and without 187 substantial motion artifacts. The AI algorithm correctly classified the images as with motion (A: 188 with 100% confidence score) and “no motion” (B: with 99.4% confidence score). 189 190 Model testing 191 External testing with Sites A and C training datasets and Site B as the test site, the model 192 performance statistics were 85% sensitivity, 90% specificity, 86% accuracy, and an AUC of 0.87 193 (95% CI 0.82 - 0.92). One hundred forty-seven CTPA exams were correctly classified into those 194 with substantial motion artifacts (true positive), and nine CTPA were mislabeled as positive for 195 motion artifacts (false positive). Seventy-seven CTPA exams were correctly classified as without . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 11 196 motion artifacts (true negative), and twenty-six CTPA exams were misclassified as without 197 motion (false negative). The confusion matrix of model performance is represented in Fig 2 and 198 the model performance for validation and test data is summarized in Table 1. 199 200 Fig 2: Confusion matrix generated by the AI model for the testing (A true positive rate 201 [sensitivity] – AI correctly labeled images with motion; B false-negative rate – AI incorrectly 202 labeled “motion” images as without motion; C false positive rate – AI incorrectly labeled “no 203 motion” images as with substantial motion artifacts; D true negative rate [specificity]– AI 204 correctly identified images without motion artifacts). 205 206 Table 1. Summary of our AI model’s performance for motion artifacts detection in CTPA exams 207 from the three participating sites (A, B, C). (key: + present; - absent; CI confidence interval). 208 Training Test Sensitivity Specificity Accuracy AUC 95% CI (AUC) Motion + 295 126Model validation (training + validation data from all sites) Motion - 259 113 94% 91% 93% 0.93 0.89-0.97 Motion + 196 173Model testing (training from A, C and testing on B) Motion - 231 86 85% 90% 86% 0.87 0.82-0.92 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 12 209 Discussion 210 We report high accuracy, sensitivity, and specificity of physician-trained and tested AI model for 211 identifying substantial motion artifacts on CTPA examinations. Prior studies have reported on 212 the ability of AI algorithms to identify anatomic regions with motion artifacts [16]. Our study 213 uses a single coronal multiplanar reformatted image per CT exam and therefore could be more 214 time-efficient, while ignoring region-specific, sparse motion artifacts which do not require 215 repeat acquisition [17-19] . To our best knowledge, there are no peer-reviewed reports on the 216 use of AI models for identifying motion artifacts in CTPA or chest CT examinations. 217 Beri et al. reported that their AI algorithm trained to identify the motion affected regions by 218 segmenting the entire image series had an AUC of 0.81 [18]. We achieved a similar AUC of 0.88 219 with a single coronal MPR image per CTPA examination. Other studies [17, 19] on motion 220 artifact detection in CT images have focused in coronary CT angiography (CCTA) rather than 221 CTPA examinations. Ma et al. reported 91% sensitivity and 71% specificity with 87% accuracy 222 for detecting motion artifacts in CCTA examinations [17]. Likewise, Elss et al. trained an AI 223 algorithm for identifying motion artifacts on CCTA and reported an accuracy of 94% [19]. Xu et 224 al. reported a fully automatic AI for grading image quality (motion artifacts) of CCTA using semi- 225 automatic labeling and tracking of the coronary arteries [20] Based on the identification and 226 estimation of motion artifact, other investigators have reported on motion artifact correction 227 and compensation solutions for head and cardiac CT examinations [21, 22]. 228 Although not yet cleared by the US Food and Drug Administration (FDA), our proof-of-concept 229 study and the AI model used in the study may have potential clinical implications. Firstly, given . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 13 230 the high frequency of motion artifacts in chest CT and CTPA examinations, our study highlights 231 the role of AI models in identifying motion artifacts. If integrated with CT scanners, such AI 232 models can efficiently detect and alert the CT technologist to the presence of motion artifacts 233 that are likely to have a substantial effect on the diagnostic interpretation. Secondly, such 234 artifacts can be present on chest CT exams acquired on both the older, less advanced scanners 235 and newer, faster scanners. On the latter, modification of scanning parameters can enable 236 faster scanning of the entire chest in under 1 second [23]. On the older scanner, the AI- 237 generated surveillance for motion-impaired CTPA or chest CT examinations can prompt 238 technologist to give better breath-hold instructions [24] and modify scan parameters or when 239 possible scan at-risk patients on faster scanners. Thirdly, most diagnostic CT scanners in our 240 institution (MGH) have two scan protocols – one for patients who can hold their breath and the 241 other for those who are unable to hold their breath or have substantial motion artifacts on 242 their initial CT acquisition. Despite instructions to our CT technologists to always review the CT 243 images for motion artifacts before taking the patient off the CT table, most technologists 244 cannot or do not comply with the recommendation especially during pandemic times. As a 245 result, interpreting radiologists either report CT with a disclaimer on motion-limited diagnostic 246 value or request patient recall and rescanning. By automating the detection of motion artifacts, 247 AI models such as the one reported in our study could potentially help address compliance and 248 reacquisition, when appropriate. Fourthly, several modern scanners automatically generate 249 multiplanar reformatted images as soon as the data acquisition is complete, so the use of 250 coronal MPR image for our model is not a rate-limiting step. Although our model would still 251 require the identification of the single image at the descending aorta level. Finally, in hospitals . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 14 252 with high CT volumes, the task of quality assessment for image quality is time-consuming and 253 labor-intensive. In such sites, the AI model used in this study can be used to analyze image 254 quality in a retrospective manner. Derived statistics can then be used to develop faster scan 255 protocols and track their impact on diagnostic evaluability. Our study has limitations. We did 256 not perform a power analysis to determine the number of training and testing cases needed to 257 prove our hypothesis. Although the high level of performance associated with the AI model 258 suggests that our sample size was adequate, it is conceivable that more training data could 259 generate better results. Likewise, for external testing, we had only a single site with completely 260 independent clinical operations, and we would require additional external sites to support the 261 claim of generalizability of the AI model. The AI model’s performance can however vary with 262 the change in scan protocols (low dose chest CT versus CTPA protocols) and scanners (for those 263 scanners without input training data). Exclusion of CTPA with “white lungs” (diffuse 264 parenchymal opacities), a substantial bilateral lung volume loss, or pneumonectomies also 265 limits the application of our AI model in such patients. 266 Another limitation of our study pertains to the use of a single coronal MPR image per CT for 267 assessing substantial motion artifacts as opposed to the entire image series for prior studies 268 [18]. It is therefore possible that the performance of our AI model can differ on the entire 269 image series (transverse or coronal) as compared to its current performance on a single image. 270 Despite a high model performance, it is possible that motion artifacts in other anatomic 271 location can affect evaluation of key findings. Given the full longitudinal coverage of anatomy in 272 the coronal plane, we believe that such “missed motion artifacts in key location” is less likely. 273 The AI building platform at the time of manuscript preparation cannot easily group entire image . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 15 274 series and is limited to 2D image input, which impacted our decision to train on the coronal 275 MPR image instead of an axial series. Finally, our model training and testing were limited to 276 CTPA and might not apply to other chest CT protocols or body regions. 277 In conclusion, the physician-trained and tested AI model can help identify substantial motion 278 artifacts on CT pulmonary angiography. Automatic recognition of such artifacts can help CT 279 technologists apply faster scan protocols and reacquire images to mitigate the impact of 280 substantial motion impairment on diagnostic evaluability. 281 282 283 284 285 286 287 288 289 290 291 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 16 292 References 293 1. Veikutis V, Budrys T, Basevicius A, Lukosevicius S, Gleizniene R, Unikas R, et al. Artifacts in 294 computer tomography imaging: how it can really affect diagnostic image quality and confuse clinical 295 diagnosis? Journal of Vibroengineering. 2015;17(2):995-1003. 296 2. Boas FE, Fleischmann D. CT artifacts: causes and reduction techniques. Imaging Med. 297 2012;4(2):229-40. 298 3. Barrett JF, Keat N. Artifacts in CT: recognition and avoidance. Radiographics. 2004;24(6):1679- 299 91. 300 4. Johnson PT. Artifacts mimicking pulmonary embolism. Pearls and Pitfalls in Cardiovascular 301 Imaging: Pseudolesions, Artifacts, and Other Difficult Diagnoses. 2015:134. 302 5. Wittram C, Maher MM, Yoo AJ, Kalra MK, Shepard JA, McLoud TC. CT angiography of pulmonary 303 embolism: diagnostic criteria and causes of misdiagnosis. Radiographics. 2004;24(5):1219-38. 304 6. Jones SE, Wittram C. The indeterminate CT pulmonary angiogram: imaging characteristics and 305 patient clinical outcome. Radiology. 2005;237(1):329-37. 306 7. Aviram G, Levy G, Fishman JE, Blank A, Graif M. Pitfalls in the diagnosis of acute pulmonary 307 embolism on spiral computer tomography. Current problems in diagnostic radiology. 2004;33(2):74-84. 308 8. Global Imaging Market Outlook Report 2019 [Available from: https://imvinfo.com/product/imv- 309 2019-global-imaging-market-outlook-report/. 310 9. Huhtanen H, Nyman M, Mohsen T, Virkki A, Karlsson A, Hirvonen J. Automated detection of 311 pulmonary embolism from CT-angiograms using deep learning. BMC Medical Imaging. 2022;22(1):1-10. 312 10. Soffer S, Klang E, Shimon O, Barash Y, Cahan N, Greenspana H, et al. Deep learning for 313 pulmonary embolism detection on computed tomography pulmonary angiogram: a systematic review 314 and meta-analysis. Scientific reports. 2021;11(1):1-8. 315 11. Ebrahimian S, Digumarthy SR, Homayounieh F, Bizzo BC, Dreyer KJ, Kalra MK. Predictive values 316 of AI-based triage model in suboptimal CT pulmonary angiography. Clinical Imaging. 2022;86:25-30. 317 12. Singh R, Wu W, Wang G, Kalra MK. Artificial intelligence in image reconstruction: The change is 318 here. Physica Medica. 2020;79:113-25. 319 13. Lenfant M, Chevallier O, Comby P-O, Secco G, Haioun K, Ricolfi F, et al. Deep learning versus 320 iterative reconstruction for CT pulmonary angiography in the emergency setting: improved image 321 quality and reduced radiation dose. Diagnostics. 2020;10(8):558. 322 14. Dasegowda G, Bizzo B, Gupta R, Kaviani P, Ebrahimian S, Ricciardelli D, et al. Successful creation 323 of clinical AI without data scientists or software developers: radiologist-trained AI model for identifying 324 suboptimal chest-radiographs. Research Square; 2022. 325 15. Mongan J, Moy L, Kahn Jr CE. Checklist for artificial intelligence in medical imaging (CLAIM): a 326 guide for authors and reviewers. Radiology Artificial Intelligence. 2020;2(2). 327 16. Lossau T, Nickisch H, Wissel T, Bippus R, Schmitt H, Morlock M, et al. Motion artifact recognition 328 and quantification in coronary CT angiography using convolutional neural networks. Medical image 329 analysis. 2019;52:68-79. 330 17. Ma H, Gros E, Baginski SG, Laste ZR, Kulkarni NM, Okerlund D, et al. Automated quantification 331 and evaluation of motion artifact on coronary CT angiography images. Medical Physics. 332 2018;45(12):5494-508. 333 18. BERI P. Detection of Motion Artifacts in Thoracic CT Scans. Available from 334 https://pure.tue.nl/ws/portalfiles/portal/168210888/Puneet_B..pdf . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint 17 335 19. Elss T, Nickisch H, Wissel T, Schmitt H, Vembar M, Morlock M, et al., editors. Deep-learning- 336 based CT motion artifact recognition in coronary arteries. Medical Imaging 2018: Image Processing; 337 2018: International Society for Optics and Photonics. 338 20. Xu Y, Sushmit A, Lyu Q, Li Y, Cao X, Maltz JS, et al. Cardiac CT motion artifact grading via semi- 339 automatic labeling and vessel tracking using synthetic image-augmented training data. Journal of X-Ray 340 Science and Technology. 2022(Preprint):1-13. 341 21. Maier J, Lebedev S, Erath J, Eulig E, Sawall S, Fournié E, et al. Deep learning-based coronary 342 artery motion estimation and compensation for short-scan cardiac CT. Medical Physics. 343 2021;48(7):3559-71. 344 22. Su B, Wen Y, Liu Y, Liao S, Fu J, Quan G, et al. A deep learning method for eliminating head 345 motion artifacts in computed tomography. Medical Physics. 2022;49(1):411-9. 346 23. Doda Khera R, Nitiwarangkul C, Singh R, Homayounieh F, Digumarthy SR, Kalra MK. 347 Multiplatform, Non-Breath-Hold Fast Scanning Protocols: Should We Stop Giving Breath-Hold 348 Instructions for Routine Chest CT? Canadian Association of Radiologists Journal. 2021;72(3):505-11. 349 24. Doda Khera R, Singh R, Homayounieh F, Stone E, Redel T, Savage CA, et al. Deploying clinical 350 process improvement strategies to reduce motion artifacts and expiratory phase scanning in chest CT. 351 Scientific Reports. 2019;9(1):1-7. 352 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 23, 2022. ; https://doi.org/10.1101/2022.06.23.22276818doi: medRxiv preprint