RPS 1505b - Technical advances and imaging biomarkers

Author Block: A. DELL'ORSO; Empoli/IT
Purpose or Learning Objective: Artefacts in MRI represent a significant problem leading to loss of diagnostic information and substantial costs. A post-processing algorithm, for orthogonal phase encoding reduction of artefact (OPERA) is proposed and tested in the clinical setting.
Methods or Background: The OPERA procedure is based on the key concept that noise-induced signal intensity alterations are randomly distributed, whereas the position of ghosts and aliasing is predictable along with columns or rows of pixels. OPERA combines the intensity values of two images acquired with the same parameters, but with orthogonal phase encoding directions, to correct artefacts. The efficacy of the OPERA procedure on MRI artefacts reduction was tested at a medium-sized general public hospital by using an Espree Siemens 1.5T MR scanner for a total period of 14 months. The procedure was tested on a total of 1003 MR images [55 randomly selected patients (56.4% females; mean age 54.6±16.7 years)]. OPERA corrected images were compared with the corresponding reference-image (Ri) by computing signal-to-noise (SNR) and contrast-noise-ratio (CNR). Images (OPERA vs Ri) were shown in blind at two radiologists with a long-standing MRI expertise by using a better-worse Likert-type scale response, to evaluate artefacts, SNR and CNR.
Results or Findings: OPERA application did not significantly affect SNR (+4.3%; IQR:2.61-5.27%) and CNR (+4.30%; IQR: 2.86-6.04%). The two radiologists observed: artefact reduction (responses 4 and 5 of the Likert-scale) between 82.4% and 83.4% (inter-rater agreement, weighed K=0.766); perceived SNR improvement (82.8% to 88.5% K=0.714) and contrast improvement (86.9% to 88.9% K=0.722).
Conclusion: The testing of OPERA in the daily MRI practice indicates the efficacy of the algorithm in reducing MRI artefacts and improving perceived image quality.
Limitations: Not tested on the heart and superior abdomen. Single-centre study.
Ethics committee approval: OSS_15_145, December 14, 2015.
Funding for this study: No funding has been received for this study.

Author Block: A. Dennis, A. Roca-Fernandez, J. Mcgonigle, A. Jandor, G. Ralli, V. Carapella, R. Banerjee; Oxford/UK
Purpose or Learning Objective: In a prospective longitudinal observational study in individuals who had recovered from COVID-19, we set out to assess the degree of organ impairment in the heart, lungs and visceral organs using quantitative MRI and explore potential links with ongoing symptoms.
Methods or Background: Quantitative MRI data were collected with CoverScanMD across two sites in the UK (Siemens 1.5T and 3T). The 30 min scan assesses: inflammation of the heart, kidneys, liver and pancreas with T1-relaxation; lung function with a dynamic structural T2-weighted scan measuring the difference between max inspiration and expiration; fat in the liver and pancreas using PDFF. Impairment for each organ was considered when the metric was outside of pre-defined reference ranges. Associations between organs and symptoms were explored with logistic regression adjusted by time from first symptoms to MRI.
Results or Findings: In N=451 (44yrs, 74% female, 89% white, median 179 days following infection), inflammation was observed in the heart (14%), liver (12%), pancreas (6%) and kidney (4%); fat in liver (22%) and pancreas (30%); estimated lung capacity was reduced in 11%. 21% had evidence of abnormality in 2 or more organs and the number of abnormal metrics negatively correlated with the length of time between initial symptoms and scan (r = -0.23, P<.001) suggesting some recovery with time. 66% reported ongoing severe breathlessness or fatigue which was significantly associated with increased white cell count (WCC) (P=.03) and marginally with myocarditis (P=0.06).
Conclusion: Coronavirus is associated with multi-organ dysfunction 6 months after infection. Organ inflammation was associated with symptoms. Multi-organ MRI may provide a diagnostic tool to stratify patients with long COVID and aid clinical management.
Limitations: Study population was limited by ethnicity.
Ethics committee approval: The study protocol was approved by a UK ethics committee (20/SC/0185).
Funding for this study: Funding was received from Perspectum Ltd.

Author Block: D. C. Bastos¹, M. Rosa¹, H. M. H. Lee¹, E. P. Reis¹, G. Szarf¹, A. Gupta², V. K. Venugopal², V. Mahajan²; ¹São Paulo/BR, ²New Delhi/IN
Purpose or Learning Objective: To evaluate the bias in the diagnostic performance of a deep learning-based chest X-ray classification algorithm on previously unseen external data.
Methods or Background: 632 chest X-rays were randomly collected from an academic centre hospital and anonymised selectively, leaving out fields needed for the bias estimation (manufacturer name, age, and gender). They were from six different vendors AGFA (388), Carestream (45), DIPS (21), GE (31), Philips (127), and Siemens (20). The male and female distribution was 376 and 256. The X-rays were read for consolidation ground truth establishment on CARING analytics platform (CARPL). These X-rays were run on open-sourced chest X-ray classification model. Inferencing results were analysed using Aequitas, an open-source python-based package to detect the presence of bias, fairness of algorithms. Algorithms’ performance was evaluated on the three metadata classes gender, age group, and brand of equipment. False omission rate (FOR) and false-negative rate (FNR) metrics were used for calculating the inter-class scores of bias.
Results or Findings: AGFA, 60 to 80 age group, and male were the dominant entities and hence considered as baseline for evaluation of bias towards other classes. Significant false omission rate (FOR) and false negative rate (FNR) disparities were observed for all vendor classes except Siemens as compared to AGFA. No gender disparity was seen. All groups show FNR parity whereas all classes showed disparity with respect to false omission rate for age.
Conclusion: We demonstrate that AI algorithms may develop biases, based on the composition of training data. We recommend bias evaluation check to be an integral part of every AI project. Despite this, AI algorithms may still develop certain biases, some of those difficult to evaluate.
Limitations: Limited pathological classes were evaluated.
Ethics committee approval: IRB approved.
Funding for this study: None.

Author Block: F. K. YEUNG, W. Y. Ip, V. Vardhanabhuti; Hong Kong/HK
Purpose or Learning Objective: Virtual non-contrast (VNC) as a concept has been used to create non-contrast images from contrast studies. In the context of spectral CT, accurate VNC helps with the accurate material decomposition (MD) properties. The aim of the study is to evaluate the performance of VNC images generated from pre-clinical photon-counting CT in a phantom.
Methods or Background: Commercially available iodine contrast agent (Iopamiro 370) was diluted into various concentrations of iodine solutions (185, 93, 46, 23, 10, 8, 6, 4, 2, 1, 0.5 mg/ml), placed in a phantom and scanned using a pre-clinical photon-counting spectral CT scanner (MARS Bioimaging Ltd., Christchurch, New Zealand) with pre-calibrated MD protocol of multi-energy range. Iodine was decomposed by the scanner’s reconstruction algorithm and the measured iodine concentration was compared with the known diluted concentrations as reference. The images were further processed by a custom-made MATLAB (Mathworks, Natick, MA) program to explore the relationship between CT number (in HU) and iodine concentration with subsequent correlation analysis. After iodine removal by VNC post-processing, the corresponding CT number of iodine vials of VNC and original contrast images were compared with the ground truth water vial respectively.
Results or Findings: For iodine images, measured iodine concentration was comparable with calculated ones (mean absolute difference of 9.1%). The HU values and iodine concentration was linearly and highly correlated (adjusted R squared of 0.9974 with p-value of <0.001). At 32-49.9 keV energy range, the mean absolute HU difference between water and VNC image of typical 10 mg/ml iodine solution was 27±13 HU while comparing to contrast image of 605±12 HU.
Conclusion: Virtual non-contrast technique is feasible in spectral CT scanner with good accuracy and correlation with known concentrations in phantom study.
Limitations: No limitations were identified.
Ethics committee approval: Not applicable.
Funding for this study: Not applicable.

Author Block: S. Ellmann, K. Hellwig, M. Eckstein, A. Hartmann, R. Fietkau, M. Hecht, M. Uder, T. Bäuerle; Erlangen/DE
Purpose or Learning Objective: Dynamic contrast enhancement (DCE) evaluation in MRI is subject to various difficulties, as the signal intensities reflect contrast media concentration non-linearly. Moreover, individual outliers in the time-intensity curves may compromise the semiquantitative values of the enhancement dynamics. In addition, the plethora of analysis methods are hardly comparable with each other. Thus, the aim of this work was the establishment of a web-based tool for the quantitative and semiquantitative evaluation of DCE-MRI.
Methods or Background: An interactive web application was programmed using R to load DCE-MRI studies and semiautomatically fit the raw data to a Brix model using the Levenberg-Marquardt method. Goodness-of-fit was assessed by R2. From the fitted curves, the Brix-equation was extracted, along with (semi-)quantitative parameters: A, kel, kep, time to peak, peak-enhancement, area-under-the-curve. The maximum and minimum of a fitted curve’s first derivative were defined as wash-in and wash-out, respectively. In an exemplary case series, 22 neck tumour patients were retrospectively analysed with regard to their response to immunotherapy. For this purpose, pre-therapeutic DCE-MRIs covering the tumour were assessed using the developed online tool. The above parameters were compared between treatment responders and non-responders.
Results or Findings: Fitting the Brix model to the raw DCE-MRI data yielded a mean R2 of 0.939 (95%CI 0.922-0.956). Regarding the case series, treatment responders’ tumours featured a significantly higher time to peak (p=0.048) and wash-out (p=0.031) compared to non-responders.
Conclusion: Raw DCE-MRI data can be fit to pharmacokinetic models semiautomatically with the presented online tool. In a case series, fitting was accurate, and the resulting parameters could be used to identify responders to immunotherapy in neck cancer patients.
Limitations: Proof-of-concept-study with a fully functional web application, but only supported by a small case series.
Ethics committee approval: Approved by the local ethics committee.
Funding for this study: Internal funding.

Author Block: I. Molwitz¹, G. Campbell¹, J. Yamamura¹, T. Knopp¹, R. Fischer¹, J. Wang², A. Busch¹, M. Grosser¹, P. Szwargulski¹; ¹Hamburg/DE, ²Dallas, TX/US
Purpose or Learning Objective: To develop a material decomposition algorithm for detector-based dual-layer spectral CT (dlsCT) fat quantification, which so far has only been implemented for source-based dual-energy CT techniques, in phantoms and validate it in first patients.
Methods or Background: Phantoms were created with 0, 5, 10, 25, 40, 100% fat and 0, 4.9, 7.0 mg/ml iodine, respectively. Scans were performed with the IQon Spectral CT (Philips, The Netherlands), and 3T MR chemical-shift relaxometry (MRR). Based on maps of the photoelectric effect and Compton scattering, three-material decomposition, including fat and iodine, was done in the image space. After written consent, n=10 patients (mean age 55 years ± 18; six men) in need of a CT staging were prospectively included, received contrast-enhanced abdominal dlsCT scans at 120kV, and MRI scans for MRR. As reference tissue for the liver and the skeletal muscle, retrospectively available non-contrast-enhanced spectral CT data sets were employed. Agreement between dlsCT and MR was evaluated for the phantoms, three hepatic and two muscular regions of interest per patient by intraclass correlation coefficients (ICC) and Bland-Altman analyses.
Results or Findings: The ICC was excellent in the phantoms (0.978[95%CI 0.937-0.993]) and the skeletal muscle (0.956[95%CI 0.890-0.982]). The ICC was moderate for log-transformed liver fat values (0.75[95%CI 0.48-0.881]). The Bland-Altman analysis yielded a mean difference of -0.714%[95%CI -4.512-3.085] for the liver and 0.501%[95%CI -4.319-5.321] for the skeletal muscle. Interobserver and intraobserver agreements were excellent (>0.9).
Conclusion: DlsCT fat quantification delivers reliable results for the liver and the skeletal muscle, which provide information about steatosis hepatis and muscle quality as parameters of prognostic relevance, in retrospectively available spectral data sets of routine exams.
Limitations: Experimental development and proof of concept study with low patient numbers.
Ethics committee approval: Available.
Funding for this study: Funded by Hamburg Research Center for Medical Technology (04fmthh2020).