Purpose: To evaluate the impact of artificial intelligence (AI)-based image optimization algorithm on improving aorta computed tomography angiography (CTA) image quality (IQ) at 80 kVp tube voltage and 40 mL contrast medium (CM) compared with a conventional iterative reconstruction (IR) algorithm at 120 kVp.
Methods and Materials: Sixty patients referred for aorta CTA examination were assigned to one of two groups at random on NeuViz 128 CT. Group A underwent an 80 kVp protocol with 40 mL CM (320 mg I/mL) and divided into two subgroups according to reconstruction algorithm (IR for group A1 and further AI-based image optimization for group A2). Group B was scanned with the standard 120 kVp, 80 mL CM and IR algorithm. The quantitative assessment of IQ included aorta attenuation, noise, signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). A five-point scale (5-excellent,1- low) was used by two radiologists independently for qualitative image analysis.
Results: The image noise was significantly reduced while SNR and CNR were significantly increased in the order of group A1, B and A2 (all P<0.001). Compared with group B, the subjective IQ score of group A1 was significantly lower (P = 0.03) while that of group A2 has no significant difference (P = 0.926). The effective dose and CM volume of group A were reduced by 79.18% and 50%, respectively than that of group B.
Conclusion: The AI-based image optimization for aorta CTA with low kV and reduced CM produced IQ comparable to conventional aortic CTA protocol.
Purpose: Common dosimetric quantities in CT such as the CTDIvol or the DLP do not appropriately represent the actual patient dose. More sophisticated methods are not real-time capable. Therefore, we propose the deep dose estimation (DDE), a deep learning-based approach to estimate patient dose distributions in real time.
Methods and Materials: The gold standard to calculate patient-specific dose distributions is to perform a Monte Carlo (MC) simulation that models the physics of CT dose deposition. Being computationally expensive, MC cannot be applied in real time. To overcome this drawback without losing accuracy we developed DDE: a deep convolutional network for CT dose estimation. DDE uses a U-net architecture that takes a two-channel input consisting of a CT volume and a first-order dose estimate volume, which can be calculated analytically at the cost of a forward projection. Using this input, DDE is trained to reproduce MC dose distributions. Here, we generalized DDE to dual-source CT and demonstrate its potential to derive accurate dose distributions.
Results: Applied to test data, DDE yields dose estimates that differ by less than 3% on average from the ground truth MC simulation. Our experiments demonstrate that DDE applies similarly to single- and dual-source scans, different anatomical regions, different shaped filters, different tube voltages and tube currents.
Conclusion: This study demonstrates the potential of deep convolutional neural networks to derive accurate CT dose estimates. Once trained, a 256×256×48 voxel volume can be processed with DDE in 250 ms while achieving the same accuracy as MC simulations.
Purpose: The aim of the study was to identify radiation dose reference levels for CT exams in our institution.
Methods and Materials: A Dose team was established (two radiologists, a TSRM, a radiology resident, the risk manager, the medical director and the medical physicist) whose first objective was to evaluate the quality of the images on 1000 oncological follow- up and to assess the amount of dose delivered through dedicated dose monitoring software (DoseWatch, GE Healthcare). The Dose length product (DLP) has been evaluated for the different anatomical regions for head, neck, thorax, abdomen-pelvis and thorax-abdomen-pelvis. The quality of the images in the different anatomical districts was subsequently assessed.
Results: The quantitative analysis allowed to identify the average DLP for the different anatomical districts. The mean dose for the different anatomical districts was for the brain (570.35 mGy-cm ± 121.32), for the neck (211.58 mGy-cm ± 93.12), for the thorax (160.91 mGy-cm ± 41.81), for the abdomen-pelvis (425.93 mGy-cm ± 122.90) and for the thorax-abdomen-pelvis (685.89 mGy-cm ± 240.04). In all anatomical districts a good image quality was obtained.
Conclusion: The dose monitoring system and the establishment of a Dose team made it possible to identify radiation dose reference levels for CT for individual anatomical districts.
Purpose: To contribute to the creation of CT radiation dose benchmarks for specific clinical indications for adults (>=15 years). We report the distribution of radiation dose metrics based on CT scans submitted in 2016-2017 to a large international dose registry comprised of 160 institutions from 7 countries. The registry was supported through the US National Institutes of Health (NIH) and Patient Centered Outcomes Research Institute (PCORI).
Methods and Materials: We describe median (target) and 75th percentile (benchmark) doses for CT dose index volume (CTDIvol) and dose length product (DLP) by three clinical indications from the European Study on Clinical DRLs (EUCLID).
Results: Sample sizes: sinusitis N=28,946, cervical spine/trauma N=83,919, pulmonary embolism N=83,882. CTDIvol target (50%) and benchmark (75%) values (in mGy), respectively: sinusitis (14 and 20), cervical spine/trauma (18 and 26), pulmonary embolism (10 and 15). Corresponding DLP values (in mGy-cm) were: sinusitis (235 and 339), cervical spine/trauma (878 and 1390), pulmonary embolism (388 and 596). Sinusitis scans were 68% lower in both CTDIvol and DLP compared to routine head scans. Pulmonary embolism scans were 23% higher in CTDIvol and 8% higher in DLP compared to routine chest scans.
Conclusion: Dose metrics from large multi-center studies can help create representative DRLs that can be used for dose optimization and institutional evaluation of CT radiation doses specifically for clinical indications to know if their doses routinely exceed these benchmarks. Having clinical indication DRLs at institutions can lead to indication-specific dose-optimized protocols.
Purpose: This work aims to automate the vessel density and noise measurements in coronary CT angiography as part of quality assessment. So far this is done manually by measuring the mean (density) intensity and standard deviation (noise) within a uniform and consistent region of about 1 square cm containing contrast agent (i.e. aorta).
Methods and Materials: We used 567 coronary CT scans from the randomized DISCHARGE trial across 26 sites. We initially applied the Canny edge detection and then used Hough transform to find circles with pre-specified radii which should correspond to the aorta. Additional heuristics regarding the expected position along or aorta were also applied to eliminated obvious errors. Thereafter density and noise were measured within the circles.
Results: For evaluation, automated measurements that differ from the manual measurements up to a certain threshold, were marked as correct. As such, noise difference tolerances of 5, 10 and 15 HUs led to a noise accuracy of 47%, 67% and 78%, respectively. Similarly, density difference tolerances of 25, 50 and 75 HUs led to a density accuracy of 71%, 90% and 96%. Finally, the combined (noise,density) difference tolerances of (5,25) , (10,50) and (15,75) led to accuracies of 37%, 61% and 75%.
Conclusion: An automatic method for quality assurance of noise and vessel density in coronary CT angiography was feasible and has the potential to reduce time and manual effort.
Purpose: To determine the impact of the ESR iGuide clinical decision support system in reducing the rate of inappropriate emergency paediatric diagnostic examinations.
Methods and Materials: 352 paediatric patients referred to Emergency Department of BLINDED Hospital between 2011 and 2016 were retrospectively evaluated. All patients had presented with abdominal pain. Clinical data were entered into the ESR iGuide system to compare appropriateness scores of suggested versus actually performed examinations, and the agreement between them was assessed using Cohen’s k coefficient.
Results: Out of all ultrasound examinations performed, only 70% (255/352) were considered appropriate by ESR iGuide. Computed tomography (CT) was suggested as appropriate in 70% (249/352) of patients, but only 2 of them went ahead with it. Two other patients underwent CT, although it would have been designated as inappropriate according to ESR iGuide. Abdominal X-ray was suggested as appropriate in 253 patients and performed in 18 of them. Four other patients underwent abdominal X-ray when suggested as inappropriate by ESR iGuide. Finally, 9 patients underwent chest X-ray, which was suggested as appropriate by ESR iGuide in 5 patients. Magnetic resonance imaging and scintigraphy were not assessed because unavailable in our emergency department. The agreement between imaging examinations suggested by ESR iGuide and those actually performed was poor (Cohen k less than 0.2).
Conclusion: Strict adherence to the ESR iGuide system would lead to a significant increase in the number of emergency paediatric CT examinations performed instead of ultrasound studies, potentially resulting in higher patient radiation exposure and overall costs.
Purpose: The quality of CT images depends on various factors, such as radiation dose, convolution kernel and resolution. CNN-based methods are sensitive to small variations of image quality, i.e., small perturbations in noise level, texture, contrast, brightness, etc.
Methods and Materials: To address this problem, we proposed an image standardisation method which is capable of transforming images from different domains to one target domain, without its domain information and without changing its content. The model first decomposes an input image into content and style latent codes, which are independent of one another. The codes are able to reconstruct the original image or standardise the image to the target domain by replacing the style code. For the target domain, we reduced its style code variance and took the mean of style codes for standardisation. A domain discriminator was also applied to determine whether two images belong to the same domain.
Results: To verify effectiveness of the proposed method, our model was trained in a semi-supervised manner. For training, each supervised step updates the model with paired images from collaborating hospital followed by one unsupervised step which learns from unpaired data provided by Infervision. With the same experimental setting, the false positive (fp/np) was reduced to 2.17 from 2.62 after standardisation on 310 test cases, while recall only decreases slightly from 82.4% to 80.9%.
Conclusion: In our attempt to address the issue by a cycleGAN-inspired unsupervised model, results have shown a significant improvement of image quality and detection performance on the transformed images.
Purpose: We aim to assess the frequency of protocol repetition in whole-body MR imaging within the multi-center German National Cohort (NAKO) in consideration of local, staff-dependent and technical influences. We further intent to determine its effect on scan time, and whether or not automated image quality assessment is able to predict protocol repetition.
Methods and Materials: All subjects enrolled in the MR substudy of the NAKO until December 31, 2016, were included in the analysis (n=11,347). Whole-body imaging was performed at five different sites, employing a uniform set of twelve protocols. All acquisitions were carried out by trained radiologic technologists (RT), whose decisions for protocol repetition were made without supervision or technological advice. Image quality parameters were derived automatically from the acquired images.
Results: RT acquired at least one repeat protocol in 12% (n=1,365) of subjects. The frequency of repetition differed across protocols (p<0.0001) as well as across sites (range: 5.28%-24.34%, p<0.0001), and varied over time (p<0.0001). The mean total scan time of 62.6min increased by 4.8min (95%CI: 4.5-5.2min) in subjects with protocol repetition(s). Several automatically derived image quality parameters were retrospectively predictive for protocol repetition, particularly image sharpness and signal-to-noise ratio, although their predictive value was not uniform for all protocols.
Conclusion: MR protocol repetitions are remarkably prevalent even in the highly standardized and static setting of a large cohort study. Automated image quality assessment shows predictive value for the RT's decision to perform protocol repetitions and has potential to improve time- and cost-efficiency of MR imaging studies.
Purpose: To assess the accuracy of a convolutional neural network (CNN) for the automatic detection of incorrect underexposure in chest X-ray.
Methods and Materials: A dataset of n=260 chest X-rays was evaluated, n=128 X-rays with correct exposure and n=132 with incorrect x-ray exposure (underexposure). For the neural network training were used 203 X-rays (100 correctly exposed X-rays + 103 underexposed X-rays) using Inception_v3 architecture. Subsequently, testing was performed with the remaining images (28 correctly exposed X-rays + 29 underexposed X-rays). Then sensibility, specificity, receiver operating characteristic curve (ROC) and area under ROC (AUC) were computed.
Results: The neural network correctly classified 27 of 29 underexposed X-rays and 23 of 28 of correctly exposed X-rays. Sensibility and specificity were 0.93 and 0.82, respectively; AUC value was 97.17% (95% CI: 93.11%-100%).
Conclusion: Preliminary results demonstrated that neural network has the potential to distinguish between underexposed X-rays and those correctly exposed with an accuracy of 88%. The goal is to supply auxiliary and reproducible information to assist radiologists and technologists when chest X-ray quality is less obvious.
Purpose: The Prostate Imaging Reporting and Data System (PI-RADS v2) was developed to provide imaging and reporting standards of multi-parametric prostate MRI (mpMRI). Urologist rely heavily on radiological reports when planning prostate biopsies. We investigated whether 1) mapping urosurgically relevant key content to RADLEX terms might be feasible to assess radiological report quality with regard to clinical usability 2) and compared it to a fully-automated guideline-based quality assessment.
Methods and Materials: A single center retrospective cohort study of 1028 consecutive patients (01/2017-09/2018; mean age 67.2,range 22-89yrs) with suspected prostate pathology and mpMRI were retrieved from local RIS archives. All reports were generated using a structured reporting tool (www.smart-radiology.com). Independent blinded urologist reviewed 299 mpMRI reports and defined biopsy relevant key information content (KIC). An automatic, cross-lingual mapping of German reports, clinical KIC, and PI-RADSv2 guidline to RADLEX terms was performed using a proprietary information extraction software (www.empolis.com). Then these RADLEX converted texts were compared using free web-tool (radreport-query.com) generated cosine similarty index and Wilcoxon rank-sum test statistics.
Results: RADLEX mapping of urologist-defined key content identified 267 while PI-RADSv2 guideline revealed 248 terms. The identified terms showed a highly significant correlation (p<0.001). RADLEX similarity scores of the reports were significantly higher (p<0.001) when compared to urosurgical KIC (mean=0.24, 0.11-0.33) than with PI-RADS guideline content (mean=0.18,0.06-0.28).
Conclusion: Biopsy-relevant key RADLEX content could serve as an important quality measure of mpMRI reports of the prostate to improve communication with urologists and to support their planning of invasive interventions.