Deep learning performance compared to healthcare experts in detecting wrist fractures from radiographs: A systematic review and meta-analysis

OBJECTIVE

To perform a systematic review and meta-analysis of the diagnostic accuracy of deep learning (DL) algorithms in the diagnosis of wrist fractures (WF) on plain wrist radiographs, taking healthcare experts consensus as reference standard.

METHODS

Embase, Medline, PubMed, Scopus and Web of Science were searched in the period from 1 Jan 2012 to 9 March 2023. Eligible studies were patients with wrist radiographs for radial and ulnar fractures as the target condition, studies using DL algorithms based on convolutional neural networks (CNN), and healthcare experts consensus as the minimum reference standard. Studies were assessed with a modified QUADAS-2 tool, and we applied a bivariate random-effects model for meta-analysis of diagnostic test accuracy data.

RESULTS

Our study was registered at PROSPERO with ID: CRD42023431398. We included 6 unique studies for meta-analysis, with a total of 33,026 radiographs. CNN performance compared to reference standards for the included articles found a summary sensitivity of 92% (95% CI: 80%-97%) and a summary specificity of 93% (95% CI: 76%-98%). The generalized bivariate I-squared statistic indicated considerable heterogeneity between the studies (81.90%). Four studies had one or more domains at high risk of bias and two studies had concerns regarding applicability.

CONCLUSION

The diagnostic accuracy of CNNs was comparable to that of healthcare experts in wrist radiographs for investigation of WF. There is a need for studies with a robust reference standard, external data-set validation and investigation of diagnostic performance of healthcare experts aided with CNNs.

CLINICAL RELEVANCE STATEMENT

DL matches healthcare experts in diagnosing WFs, which potentially benefits patient diagnosis.

Nordic radiographers' and students' perspectives on artificial intelligence - A cross-sectional online survey

INTRODUCTION

The integration of artificial intelligence (AI) into the domain of radiography holds substantial potential in various aspects including workflow efficiency, image processing, patient positioning, and quality assurance. The successful implementation of AI within a Radiology department necessitates the participation of key stakeholders, particularly radiographers. The study aimed to provide a comprehensive investigation about Nordic radiographers' perspectives and attitudes towards AI in radiography.

METHODS

An online 29-item survey was distributed via social media platforms to Nordic students and radiographers working in Denmark, Norway, Sweden, Iceland, Greenland, and the Faroe Islands including items on demographics, specialization, educational background, place of work and perspectives and knowledge on AI. The items were a mix of closed-type and scaled questions, with the option for free-text responses when relevant.

RESULTS

The survey received responses from all Nordic countries with 586 respondents, 26.8% males, 72.1% females, and 1.1% non-binary/self-defined or preferred not to say. The mean age was 37.2 with a standard deviation (SD) of ±12.1 years, and the mean number of years since qualification was 14.2 SD ± 10.3 years. A total of 43% (n = 254) of the respondents had not received any AI training in clinical practice. Whereas 13% (n = 76) had received AI during radiography undergrad training. A total of 77.9% (n = 412) expressed interest in pursuing AI education. The majority of respondents were aware of the potential use of AI (n = 485, 82.8%) and 39.1% (n = 204) had no reservations about AI.

CONCLUSION

Overall, this study found that Nordic radiographers have a positive attitude toward AI. Very limited training or education has been provided to the radiographers. Especially since 82.8% reports on plans to implement AI in clinical practice. In general, awareness of AI applications is high, but the educational level is low for Nordic radiographers.

IMPLICATION FOR PRACTICE

This study emphasises the favourable view of AI held by students and Nordic radiographers. However, there is a need for continuous professional development to facilitate the implementation and effective utilization of AI tools within the field of radiography.

Pill or pump? Nitroglycerin 0.5 mg tablet vs 0.8 mg spray: Effect on proximal vessel diameters at Coronary CT Angiography (CCTA)

INTRODUCTION

The administration of sublingual Nitroglycerin (NTG) prior to CT coronary angiography (CCTA) can be perfomed using pump spray or tablets. Choice of method seems to be based on local preference, rather than published guidelines. This retrospective analysis tested whether proximal coronary diameters differed dependent on the sublingual administration of 0.5 mg Nitroglycerin (NTG) tablets or 0.8 mg NTG spray.

METHODS

287 ECG-gated CCTA studies with optimal image quality and Agatston scores<400 were included in this retrospective analysis. 143 of the patients were dosed with NTG tablets at a dose of 0.5 mg prior to CCTA. 144 patients received 2 puffs of 0.4 mg NTG spray for a total dose of 8 mg. All were scanned on a second-generation Dual Source CT. Diameters of proximal segments of Left Main (LM), Right (RCA), Left Anterior (LAD) and circumflex (CX) coronary arteries were measured using semi-automatic electronic callipers by two blinded readers. Results were summarised as the mean of maximum and minimum diameters. Sex-specific analysis of diameters was carried out using repeated-measures ANOVA for each vessel. Agreement between readers was examined with Bland-Altman analysis and intra-class-correlation coefficient (ICC).

RESULTS

No significant differences in coronary diameters were found except in the RCA for women and LM for men. In both cases, diameters were smaller in the spray group (11 and 9%, respectively). Reader agreement was excellent, with ICC>0.96 for all vessels, and no significant bias, except in CX (0.03 mm).

CONCLUSIONS

We found no evidence for the systematic superiority of either administration method in proximal coronary vessels.

IMPLICATIONS FOR PRACTICE

Choosing between tablet or spray NTG prior to CCTA can be guided by practical, economical and hygienic considerations alone.

Collimation border with U-Net segmentation on chest radiographs compared to radiologists

INTRODUCTION

Chest Radiography (CXR) is a common radiographic procedure. Radiation exposure to patients should be kept as low as reasonably achievable (ALARA), and monitored continuously as part of quality assurance (QA) programs. One of the most effective dose reduction tools is proper collimation practice. The purpose of this study is to determine whether a U-Net convolutional neural networks (U-CNN) can be trained to automatically segment the lungs and calculate an optimized collimation border on a limited CXR dataset.

METHODS

662 CXRs with manual lung segmentations were obtained from an open-source dataset. These were used to train and validate three different U-CNNs for automatic lung segmentation and optimal collimation. The U-CNN dimensions were 128 × 128, 256 × 256, and 512 × 512 pixels and validated with five-fold cross validation. The U-CNN with the highest area under the curve (AUC) was tested externally, using a dataset of 50 CXRs. Dice scores (DS) were used to compare U-CNN segmentations with manual segmentations by three radiographers and two junior radiologists.

RESULTS

DS for the three U-CNN dimensions with segmentation of the lungs ranged from 0.93 to 0.96, respectively. DS of the collimation border for each U-CNN was 0.95 compared to the ground truth labels. DS for lung segmentation and collimation border between the junior radiologists was 0.97 and 0.97. One radiographer differed significantly from the U-CNN (p = 0.016).

CONCLUSION

We demonstrated that a U-CNN could reliably segment the lungs and suggest a collimation border with great accuracy compared to junior radiologists. This algorithm has the potential to automate collimation auditing of CXRs.

IMPLICATIONS FOR PRACTICE

Creating an automatic segmentation model of the lungs can produce a collimation border, which can be used in CXR QA programs.

The effect of Gaussian noise on pneumonia detection on chest radiographs, using convolutional neural networks

INTRODUCTION

Chest X-rays (CXR) with under-exposure increase image noise and this may affect convolutional neural network (CNN) performance. This study aimed to train and validate CNNs for classifying pneumonia on CXR as normal or pneumonia acquired at different image noise levels.

METHODS

The study used the curated and publicly available "Chest X-Ray Pneumonia" dataset of 5856 AP CXR classified into 1583 normal, 4273 viral and bacterial pneumonia cases. Gaussian noise with zero mean was added to the images, at 5 image noise variance levels, corresponding to decreasing exposure. Each noise-level dataset was split into 80% for training, 10% for validation, and 10% for test data and then classified using custom trained sequential CNN architecture. Six classification tasks were developed for five Gaussian noise levels and the original dataset. Sensitivity, specificity, predictive values and accuracy were used as evaluation performance metrics.

RESULTS

CNN evaluation on the different datasets revealed no performance drop from the original dataset to the five datasets with different noise levels. Sensitivity, specificity and accuracy for the normal datasets were 98.7%, 76.1% and 90.2%. For the five Gaussian noise levels the sensitivity, specificity and accuracy ranged from 96.9% to 98.2%, 94.4%-98.7% and 96.8%-97.6%, respectively. A heat map was used for visual explanation of the CNNs.

CONCLUSION

The CNNs sensitivity maintained, and the specificity increased in distinguishing between normal and pneumonia CXR with the introduction of image noise.

IMPLICATIONS FOR PRACTICE

No performance drops of CNNs in distinguishing cases with and without pneumonia CXR with different Gaussian noise levels was observed. This has potential for decreasing radiation dose to patients or maintaining exposure parameters for patients that require additional radiographs.

Low-cost 3D-printed anthropomorphic cardiac phantom, for computed tomography automatic left ventricle segmentation and volumetry - A pilot study

INTRODUCTION

Accurate cardiac left ventricle (LV) delineation is essential to CT-derived left ventricular ejection fraction (LVEF). To evaluate dose-reduction potential, an anatomically accurate heart phantom, with realistic X-ray attenuation is required. We demonstrated and tested a custom-made phantom using 3D-printing, and examined the influence of image noise on automatically measured LV volumes METHODS: A single coronary CT angiography (CCTA) dataset was segmented and converted to Standard Tessellation Language (STL) mesh, using open-source software. A 3D-printed model, with hollow left heart chambers, was printed and cavities filled with gelatinized contrast media. This was CT-scanned in an anthropomorphic chest phantom, at different exposure conditions. LV and "myocardium" noise and attenuation was measured. LV volume was automatically measured using two different methods. We calculated Spearmans' correlation of LV volume with noise and contrast-noise ratio respectively om 486 scans of the phantom. Source images were compared to one phantom series with similar parameters. This was done using Dice coefficient on LV short-axis segmentations.

RESULTS

Phantom "Myocardium" and LV attenuation was comparable to measurements on source images. Automatic volume measurement succeeded, with mean volume deviation to patient images less than 2 ml. There was a moderate correlation of volume with CNR, and strong correlation of volume with image noise. With papillary muscles included in LV volume, the correlation was positive, but negative when excluded. Variation of volumes was lowest at 90-100 kVp for both methods in the 486 repeat scans. The Dice coefficient was 0.87, indicating high overlap between the single phantom series and source scan. Cost of 3D-printer and materials was 400 and 30 Euro respectively.

CONCLUSION

Both anatomically and radiologically the phantom mimicked the source scans closely. LV volumetry was reliably performed with automatic algorithms.

IMPLICATIONS FOR PRACTICE

Patient-specific cardiac phantoms may be produced at minimal cost and can potentially be used for other anatomies and pathologies. This enables radiographic phantom studies without need for dedicated 3D-labs or expensive commercial phantoms.

Pericoronary adipose tissue attenuation in low-risk asymptomatic individuals, sex-differences and association with markers of cardiovascular disease

Funding Acknowledgements

Type of funding sources: None.

Introduction

Pericoronary adipose tissue (PCAT) attenuation by coronary computed tomography angiography (CCTA) is a marker of coronary inflammation and predicts clinical outcomes in symptomatic patients undergoing CCTA. Sex-differences in PCAT CT attenuation among asymptomatic individuals are not previously described.

Purpose

To evaluate PCAT CT attenuation according to sex and markers of cardiovascular disease (CVD).

Methods

Cross-sectional cohort study including asymptomatic individuals, 50- or 60-year of age, not taking any medicine and without known CVD or type-2 diabetes. At baseline and 5-year follow-up smoking habits, blood pressures and biochemistry (lipids, CRP, fibrinogen, D-dimer, t-PA, PAI-1, vWF) were recorded and Agatston Score measured. At follow-up, CCTA was achieved. Quantitative coronary plaque analysis was performed and PCAT CT attenuation within a radial distance of 3 mm from the outer vessel wall 10–50 mm distal to the origin of the right coronary artery measured. A validated PCAT CT attenuation threshold (high vs low risk) of -70.1 Hounsfield units was applied.

Results

Included were 123 participants (60 women). Independent of co-variation, PCAT CT attenuation (median, [IQR]) was lower in women (-71.0, [-77.2- -67.0]) vs men (-64.5, [-69.9- -57.4]), p &lt; 0.001. No associations between PCAT CT attenuation (high vs low) and risk factors of CVD, CAC or coronary plaque volumes were demonstrated (Table). Variations in blood pressures, biochemical markers and CAC over five years were not associated with PCAT CT attenuation.

Conclusion

In low-risk asymptomatic individuals, PCAT CT attenuation was lower in women compared to men, irrespective of markers of CVD.

Table. Patient characteristics stratified by PCAT CT attenuation PCAT CT attenuation ≤ -70.1 HU (n = 49) PCAT CT attenuation &gt; -70.1 HU (n = 74) p-value Risk factors Age65-years55-years 2623 3143 0.32 SexMenWomen 1534 4826 &lt;0.001 TobaccoNeverCurrent/previous 1732 3737 0.10 Systolic BP, mmHgDiastolic BP, mmHgTotal cholesterol, mmol/lLDL-cholesterol, mmol/lHDL-cholesterol, mmol/lTriglycerides, mmol/l 137 (17)76 (10)5.61 (0.92)3.50 (0.93)1.41 (0.30)1.65 (0.99 - 2.22) 136 (16)77 (10)5.42 (0.82)3.30 (0.82)1.45 (0.35)1.35 (1.03 - 2.11) 0.660.590.230.220.570.52 Biochemistry CRP, mg/lFibrinogen, μmol/lD-dimer, mg/lvWFt-PAPAI-1 1.16 (0.99 - 2.22)9.5 (8.5 - 10.7)0.40 (0.30 - 0.49)128 (102 - 154)7.1 (5.8 - 8.8)20.5 (16.2 - 31.8) 0.61 (0.30 - 1.14)9.0 (7.8 - 10.0)0.32 (0.24 - 0.47)116 (92 - 146)6.3 (5.1 - 8.8)20.3 (14.7 - 26.3) &lt;0.010.10&lt;0.050.110.200.34 Coronary plaque data Agatston ScoreTotal plaque volume, mm&sup3;NCP volume, mm&sup3;CP volume, mm&sup3;LD-NCP volume, mm&sup3; 1 (0 - 36)15.7 (0 - 143.3)0 (0 - 128.1)0 (0 - 14.6)0.5 (0 - 18.7) 8 (0 - 115)15.6 (0 - 268.2)13.5 (0 - 220.5)1.7 (0 - 31.7)1.8 (0 - 21.8) 0.300.450.490.360.74 Values are n (%), mean (SD) or median (IQR).Abbreviations: HU =Hounsfield unit; LDL =low-density lipoprotein; HDL =high-density lipoprotein; BP =blood pressure; CRP = c-reactive protein: vWF =von-Willebrand Factor; t-PA =tissue plasminogen activator; PAI-1 =plasminogen activator inhibitor -1; NCP =non-calcified plaque; CP =calcified plaque; LD-NCP =low-density non-calcified plaque.

Anode heel effect: Does it impact image quality in digital radiography? A systematic literature review

INTRODUCTION

The anode heel effect can be used to optimize image quality and/or patient dose in digital radiography (DR). In film-screen radiography, the effect can equalize optical density in regions of varying attenuation. Clinical experience suggests that the implementation of DR has led to less awareness of anode orientation. Post-processing is assumed to compensate, but may also alter image impression and potentially obscure image details. Published evidence was examined for the influence of the anode heel effect on image quality in DR.

METHOD

A systematic literature search was carried out using PubMed, Embase, and Web of Science databases. Title and abstracts were screened blinded by three authors, according to in-/exclusion criteria, followed by full-text analysis for final inclusion. Studies where technical and/or visual image quality were reported, was included. All studies were analyzed and assigned quality scores, according to relevant questions. The authors devised a scoring system based on reported information pertaining to reproducibility, interpretation, and generalizability of the methods and conclusions.

RESULTS

Five studies were included of heterogeneous design, each with methodological shortcomings. Only a few anatomical areas were covered. Very few patients were examined, and in no studies were images evaluated by radiologists or reporting radiographers. Relevant information such as post-processing, image quality criteria and analysis was insufficient in most studies, making reproduction difficult. Results were contradictory, especially concerning technical vs visual image quality.

CONCLUSION

Limited published evidence was found quantifying the influence of the anode heel effect on image quality using DR technology. More methodologically, robust studies are needed. The published evidence neither proves nor disproves the impact of the heel effect on image quality in DR.

IMPLICATIONS FOR PRACTICE

Based on a systematic review, no firm recommendations for anode orientation relating to image quality in DR can be provided.

Image quality of bedside chest radiographs in intensive care beds with integrated detector tray: A phantom study

INTRODUCTION

Vendors offer intensive care beds with integrated detector trays for bedside radiography, promoting better ergonomics and patient comfort. However, no documentation of the effects on diagnostic image quality has been located. This study examines measured and subjective image quality of supine bedside chest radiographs with and without use of such a detector tray.

METHODS

A contrast-detail phantom (CDRAD 2.0) was exposed using standard supine chest exposure parameters. Plexiglass plates of 16 and 21 cm were placed in front to simulate patient attenuation for standard and adipose patients. Exposures were repeated with the detector placed in tray and directly in bed. Images were analysed using dedicated software giving a figure-of-merit IQFinv. Results were compared using ANOVA. Then an anthropomorphic chest phantom (Lungman) was exposed using the same parameters, and the same placements of the detector. Exposures were done with and without extra conformal tissue to simulate varying patient sizes, and with and without added typical intensive care equipment. Images were analysed by two radiologists using a three-point scale, on five image quality criteria. Radiologist also stated whether the images were sufficient for diagnosis. Results were compared using Visual Grading Characteristics, using dedicated software, resulting in Areas Under the Curve (AUC-VGC) for each combination and criteria. Inter- and intra-rater reliability were assessed with kappa statistics. Composite Visual Grading Analysis (VGAS) scores were calculated for each image. Both IQFinv and were normalized and compared.

RESULTS

For all criteria both IQFinv and AUC-VGC was significantly better when exposing the detector directly in bed, than with the detector placed in the tray across all exposures. When stratified into thin and adipose patients, IQFinv decreased significantly for thin patients, while VGAS-scores did not. For adipose patients, both figures were significantly lower with the detector in the tray.

CONCLUSION

Use of detector tray for bedside chest imaging decreases image quality.

IMPLICATIONS FOR PRACTICE

Radiographers should critically evaluate image quality and experimentally determine optimal exposure factors, when taking equipment with integrated trays into use.

P6175Prediction of coronary revascularization by coronary computed tomography angiography derived fractional flow reserve - different algorithms for interpretation

Introduction

Coronary CT angiography (CTA) derived fractional flow reserve (FFRct) is increasingly used for guiding referral to invasive procedures in patients with stable chest pain. However, optimal interpretation of FFRct-analysis in terms of location and threshold of applied FFRct-values is unclear.

Purpose

To evaluate the clinical performance of various vessel-specific physiological FFRct derived measures of ischemia for prediction of standard of care guided coronary revascularization in patients with stable chest pain and coronary artery disease as determined by coronary CTA.

Methods

Retrospective study in patients with stable chest pain referred for coronary angiography based on coronary CTA. Standard acquired coronary CTA data sets were transmitted for core-laboratory analysis at HeartFlow. Any FFRct value in the major coronary arteries ≥1.8 mm in diameter, including side branches, were registered. Lesions were categorized as positive for ischemia using 6 different algorithms: Lowest in vessel FFRct-value (1) ≤0.75 or (2) ≤0.80; 2 cm distal-to-lesion FFRct-value (3) ≤0.75 or (4) ≤0.80; ΔFFRct (5) ≥0.06 or a combination of 2 and 5. The personnel responsible for downstream patient management had no information regarding FFRct test results.

Results

A total of 172 patients were included. Revascularization was performed in 62 (35%) patients. The diagnostic performance of different FFRct algorithms for predicting standard of care guided coronary revascularization is shown in the Table.

Revascularization Predictions by FFRct N=172 Diagnostic performance FFRCT false negative FFRCT false positive Values given as (%) No. of revasc vessels No. of abnormal vessels FFRCT Algorithm Sens Spec PPV NPV Acc 1 2 3 1 2 3 Distal FFRCT ≤0.75 77 68 58 84 72 12 2 0 29 5 1 Distal FFRCT ≤0.80 92 43 48 90 61 5 0 0 40 20 3 Lesion-specific FFRCT ≤0.75 68 86 74 83 80 17 3 0 12 3 0 Lesion-specific FFRCT ≤0.80 82 78 68 89 80 10 2 0 21 3 1 ΔFFRCT ≥0.06 98 36 47 98 59 1 0 0 51 19 0 Combinationa 92 54 53 92 67 5 0 0 39 12 0 aDistal FFRCT ≤0.80 and ΔFFRCT ≥0.06. Sens = sensitivity; Spec = specificity; PPV = positive predictive value; NPV = negative predictive value; Acc = accuracy; FFRCT = fractional flow reserve derived from coronary CTA; ΔFFRCT = difference between FFRCT-value immediately proximal and distal to lesion; Revasc = revascularized.

Conclusion

The diagnostic performance of FFRct in terms of predicting standard of care guided coronary revascularization is dependent on the applied algorithm for interpretation of the FFRct-analysis.

P6186Symptomatic effect of coronary revascularization at 1-year follow-up in stable chest pain - prediction by coronary computed tomography angiography derived fractional flow reserve

Introduction

Coronary CT angiography (CTA) derived fractional flow reserve (FFRct) is increasingly being used for guiding referral to invasive procedures in patients with stable chest pain. However, the ability of FFRct to predict the symptomatic effect of revascularization remains unclear.

Purpose

To evaluate the ability of different vessel-specific physiological FFRct derived measures of ischemia for predicting the occurrence of chest pain one year after coronary revascularization in stable patients.

Methods

Retrospective study in patients with stable chest pain referred for coronary angiography based on coronary CTA. Standard acquired coronary CTA data sets were transmitted for core-laboratory analysis at HeartFlow. Patients were categorized as positive for ischemia using 3 different algorithms: Lowest in vessel FFRct-value ≤0.80; ΔFFRct ≥0.06 or a combination of the two. Personnel responsible for downstream patient management had no information on FFRct test results. Classification of revascularization was performed based on the applied FFRct algorithm: complete if all FFRct positive lesions were revascularized; incomplete if ≥1 FFRct positive lesion was not revascularized. Symptomatic status at 1-year follow-up was obtained by a visit in the outpatient clinic or by telephone.

Results

A total of 172 patients were included. Revascularization was performed in 62 (35%) patients. At 1-year follow-up 48 (28%) patients had chest pain; 15 (24%) revascularized vs 33 (30%) non-vascularized patients, p=0.415. No difference in utilization of anti-anginal medicine for patients with and without chest pain was registered at 1-year follow-up. The association between the chosen FFRct algorithm, revascularization and occurrence of chest pain at 1-year follow-up are shown in the Table.

FFRct, Revascularization and Chest pain FFRCT, Algorithm Revascularizationb Patients with chest pain 1-year risk of chest pain p-valuec N (%) OR (95%-CI) Distal FFRCT ≤0.80 Incomplete 32 (34) Ref. Distal FFRCT ≤0.80 Complete 4 (15) 0.34 (0.11, 1.06) Distal FFRCT >0.80 No 11 (24) 0.61 (0.27, 1.35) 0.097 ΔFFRCT ≥0.06 Incomplete 34 (35) Ref. ΔFFRCT ≥0.06 Complete 7 (21) 0.49 (0.19, 1.24) ΔFFRCT <0.06 No 7 (18) 0.41 (0.16, 1.03) 0.074 Combinationa abnormal Incomplete 30 (40) Ref. Combination abnormal Complete 6 (18) 0.32 (0.12, 0.87) Combination normal No 11 (19) 0.35 (0.16, 0.78) 0.009 aDistal FFRCT ≤0.80 and ΔFFRCT ≥0.06. bIncomplete (≥1 FFRCT positive lesion not revascularized); complete (All FFRCT positive lesions revascularized); No (No FFRCT positive lesions and revascularization not performed). cBetween group comparison performed using logistic regression.

Conclusion

Revascularization based on classification by FFRct is associated with symptomatic relief at 1-year follow-up in patients with stable chest pain.