RPS 1712 - Imaging in children: realistic benefits of artificial intelligence (AI)

Author Block: A. S. Brendlin, S. Afat, U. Schmid, I. Tsiflikas; Tübingen/DE
Purpose or Learning Objective: To explore the potential impact of an AI-based denoising software solution on image quality and workflows in paediatric thorax ultra-low-dose computed tomography (ULDCT) scans.
Methods or Background: From 01.04.2021-01.10.2021, 60 consecutive paediatric patients with ultra-low-dose CT were included. Images were reconstructed using weighted filtered back projection (wFBP), advanced modelled iterative reconstruction (ADMIRE) strength 2, and wFBP with a novel denoising postprocessing algorithm (PixelShine). Three readers with varying experience levels independently rated subjective image quality on a 5-point Likert scale (1=worst - 5=best). An intraclass correlation coefficient was used to measure inter-rater-agreement. Image noise (standard deviation of Hounsfield Units from regions of interest in homogenous paraspinal muscles) was compared for objective image quality. An additional phantom scan was performed to investigate the dose-reduction capabilities of PixelShine. Statistical analysis ensued using a mixed-effects model. Time to diagnosis was measured for all reconstruction methods.
Results or Findings: Subjective image quality was significantly higher for PixelShine vs ADMIRE 2 vs wFBP (4 IQR [4-5] vs 3 IQR [2-3] vs 2 IQR [2-3]; each pConclusion: Compared to conventional reconstruction methods, AI postprocessing denoising may positively impact subjective and objective image quality in pediatric ULDCT thorax scans and improve radiological workflows.
Limitations: This is a retrospective study with 60 patients. Dose reduction capabilities were only measurable with a phantom.
Ethics committee approval: IRB approved.
Funding for this study: No funding was received for this study.

Author Block: M. Caruso, C. Ricciardi, G. D. Paoli, F. Di Dato, V. Romeo, M. Petretta, R. Iorio, A. Brunetti, S. Maurea; Naples/IT
Purpose or Learning Objective: Biliary atresia (BA), a rare obliterative cholangiopathy, leads to a progressive biliary cirrhosis and Kasai portoenterostomy (KP) represents its first-line treatment. Our objective was to compare the accuracy of quantitative parameters extracted from the laboratory, ultrasound (US) and magnetic resonance (MR) using machine learning (ML) algorithms in predicting long-term outcome of native liver survivor patients with BA after KP.
Methods or Background: Twenty-four patients were evaluated according to clinical and laboratory data at initial evaluation (median follow-up=9.7 years) after KP as with ideal (n=15) or non-ideal (n=9) medical outcome, successively they were re-evaluated after additional 4 years and classified in Group 1 (n=12) as stable and Group 2 (n=12) as non-stable in the disease course. Laboratory and quantitative US and MR imaging parameters were merged to test ML algorithms.
Results or Findings: The only statistically significant parameters between Group 1 and 2 were total and direct bilirubin (TB and DB) as laboratory parameters, while US stiffness as an imaging parameter. The values of TB and DB were still in the normal range, but tend towards the upper limit. Naïve Bayes was the best algorithm in terms of accuracy, sensitivity, specificity values and AUCROC to predict long-term outcome, selecting only laboratory parameters (TB and DB), while Random Forest and k-Nearest Neighbors algorithms reached the same results using either laboratory or imaging parameters.
Conclusion: The ML evaluation, merging laboratory and quantitative imaging findings, shows the pivotal role of TB and DB values in predicting long-term outcome of BA patients after KP, even though their values may be within normal range. Therefore, clinicians should be alert when the values of these laboratory parameters show subtle changes.
Limitations: The small sample size and the retrospective type of the investigation.
Ethics committee approval: The ethics approval was obtained.
Funding for this study: No funding was received for this study.

Author Block: R. Maarek, A-L. Hermann, A. Kamoun, R. Khelifi, A. Marchi, M. Collin, A. Jaillard, T. Nguyen, H. Ducou Le Pointe; Paris/FR
Purpose or Learning Objective: The number of conventional X-ray examinations in paediatric emergency departments is constantly increasing, leading to avoidable errors in interpretation by the radiologist. The use of artificial intelligence (AI) could improve the interpretation workflow by prioritising pathological radiographs and providing assistance in fracture detection.
Methods or Background: A cohort of 300 anonymised radiographs performed for peripheral skeletal fracture detection in patients aged 2 to 21 years was retrospectively collected. The gold standard was established for each examination after an independent review by two radiologists experts in musculoskeletal imaging. In case of disagreement, a consensual review with a third expert radiologist was performed. Out of the 300 examinations, 150 presented at least a fracture. All radiographs were then read by 3 senior radiologists and 5 junior radiologists in training between the 2nd and 4th year of residency without and immediately after with the help of an AI. Poor quality radiographs were excluded from the cohort. Sensitivity and specificity for each group of radiologists were calculated without and with the help of AI.
Results or Findings: The standalone sensitivity and specificity of the AI were respectively 91% and 90%. The mean sensitivity for all groups was 73.3% without AI, it increased by almost 10% to 82.8% with the aid of the AI. For the junior radiologists, it increased from 71.9% to 82.2% (+10.3%) and for the seniors from 75.6% to 83.8% (+8.2%). On average, the specificity increased from 89.6% to 90.3% (+0.7%) and from 86.2% to 87.6% (+1.4%) for juniors. For senior radiologists, the average specificity slightly decreased from 95.1% to 94.9% (-0.2%).
Conclusion: The aid of the AI increased sensitivity by an average of 10% without decreasing specificity.
Limitations: No limitations identified.
Ethics committee approval: IRB approval n°20202256.
Funding for this study: No funding was received for this study.

Author Block: B. Elders1, H. Sadafi2, J. Costa2, J. de Backer2, H. A. W. M. Tiddens1, P. Wielopolski1, B. Pullens1, P. Ciet1; 1Rotterdam/NL, 2Kontich/BE
Purpose or Learning Objective: The aim of this study was to use Magnetic Resonance Imaging (MRI) based Computational Fluid Dynamic (CFD) modelling of the upper airways after digital surgery to predict the effect of various surgical interventions on airflow patterns and resistance in children post Laryngotracheal Stenosis (LTS) repair.
Methods or Background: After open airway surgery for LTS several complex anatomical changes of the airway can remain, leading to altered upper airway airflow patterns and increased airway resistance for which re-operation might be needed. In this study, CFD analyses were performed on free-breathing (FB) and inspiratory (Insp) MRI scans of a healthy volunteer, and of two patients post LTS repair. Digital surgery was executed to predict the effect of 1) widening of the vocal cords, 2) removal of the tracheal deformation (TD), 3) both widening of the vocal cords and TD removal.
Results or Findings: Patient 1 had severe vocal cord stenosis and mild TD. Patient 2 had severe vocal cord stenosis and severe TD. Compared to the healthy volunteer during FB and Insp, patient 1 showed an increased total airway resistance of 269% and 180%, and patient 2 showed an increase of 735% and 1548%. In patient 1 the best airway resistance was achieved when just the vocal cords were widened (FB:-20.2 Pa.s/L (45.6%), Insp:-17.9 Pa.s/L (68.8%)). In patient 2 the best results were obtained when both the vocal cords were widened and the TD was removed (FB:-71.1 Pa.s/L (71.0%), Insp:-133.9 Pa.s/L (87.3%)).
Conclusion: Our findings suggest that CFD modelling can be used to study the effect of virtual surgical upper airway interventions in patients with complex airway pathology.
Limitations: Proof of concept study in a low number of subjects.
Ethics committee approval: IRB approved.
Funding for this study: Funding was received from Vrienden van het Sophia (SSWO).

Author Block: M. Esser1, I. Tsiflikas1, J. Jago2, L. Rouet3, A. Stebner4, J. F. Schäfer1; 1Tübingen/DE, 2Bothell, WA/US, 3Suresnes/FR, 4Münsterlingen/CH
Purpose or Learning Objective: To evaluate the interrater agreement of kidney volume segmentation by three-dimensional ultrasound (3D-US) in children with hydronephrosis and comparison to hydronephrosis grading.
Methods or Background: 48 kidney volumes were acquired in 45 patients with hydronephrosis by freehand 3D-US (6-1 MHz volumetric sector array, electronic rotation) in prone position (35 male; median age, 4 years; range, one month to 16 years). Semi-automated kidney segmentation (prototype software) was performed afterwards by two readers providing volumes for total kidney volume, dilated collective system and renal parenchyma (total kidney volume minus dilated collective system), as well as hydronephrosis index (renal parenchyma divided by total kidney volume). Interrater agreement was evaluated with a two-way intraclass correlation coefficient (ICC). The maximum anteroposterior diameter of renal pelvis was measured on transverse 2D B-mode images, and hydronephroses were classified grade 1-4.
Results or Findings: Most hydronephroses were grade 2 (n=29), followed by grade 3 (n=15) and grade 1 (n=4). The most frequent patient history included pelvic ureteric junction stenosis (n=21). Interrater agreement for total kidney volume, collective system, hydronephrosis index and renal parenchyma was good to excellent with ICC of 0.94, 0.87, 0.83 and 0.91 respectively (pConclusion: Novel 3D-US volumetric analysis has a high degree of interrater agreement providing parenchyma volume in hydronephrotic kidneys. Volumes of the collective system and hydronephrosis index correlate with the extent of hydronephrosis.
Limitations: No interrater agreement was assessed.
Ethics committee approval: This prospective study was approved by the local ethics committee.
Funding for this study: The study was supported by a research grant in the framework of a collaboration contract with Philips Ultrasound, Inc.