RPS 2511 - Artificial intelligence (AI) meets the brain

Author Block: S. Pistocchi¹, T. Hilbert¹, D. Rodriguez¹, B. Clifford², T. Feiweier³, Z. Hosseini⁴, V. Dunet¹, S. Cauley², T. Kober¹; ¹Lausanne/CH, ²Boston, MA/US, ³Erlangen/DE, ⁴Atlanta, GA/US
Purpose or Learning Objective: The duration of MRI acquisitions is a major limitation of this imaging modality, but especially for time-sensitive applications such as stroke or when imaging non-compliant or very young patients. Here we aim to evaluate the image quality of a new fast AI-enhanced protocol utilising a prototype multi-shot, multi-contrast EPI sequence and compare it to the standard imaging protocol at our institution.
Methods or Background: Between the 1st and 31st of June 2021, the AI-enhanced multi-shot multi-contrast EPI prototype sequence was added to our standard protocol in 30 brain MRI examinations with mixed clinical indications. The prototype sequence provided five contrasts (2D sagittal T1, axial FLAIR, T2GE, DWI) in a total of two minutes of scan time. Images were prospectively reviewed and independently compared to the standard 7:30 min: sec protocol by two experienced neuroradiologists. Six items (overall image quality, grey-white matter interface, basal ganglia delineation, sulci, motion, and susceptibility artefacts) were assessed on each generated contrast using a 4-point Likert scale. Inter-observer concordance was assessed using the Gwet AC1 coefficient.
Results or Findings: The AI-enhanced multi-shot multi-contrast EPI protocol allowed a 73% reduction of acquisition time and showed good to excellent overall image quality (mean score ≥3). Inter-observer concordance was good to excellent (Gwet AC1: 0.52 to 1.0). Motion and susceptibility artefacts were mostly rated as absent or minor with no adverse effect on diagnostic use, but with more heterogeneous inter-observer concordance (Gwet AC1: 0.27 to 0.83).
Conclusion: The AI-enhanced multi-shot multi-contrast EPI protocol demonstrated good image quality with a 73% reduction in acquisition time. Further studies evaluating diagnostic performance in time-sensitive clinical applications should be planned.
Limitations: This study is monocentric and has a small sample size.
Ethics committee approval: Not applicable.
Funding for this study: No funding has been used for this study.

Author Block: K. Murayama¹, Y. Ohno¹, H. Ikeda¹, H. KImata², N. Akino², K. Fujii², Y. Kataoka¹, A. Katagata¹, H. Toyama¹; ¹Toyoake/JP, ²Otawara/JP
Purpose or Learning Objective: To directly compare the capability for image quality improvements on brain contrast-enhanced CT angiography (CE-CTA) for ultra-high-resolution CT (UHR-CT) in intracranial aneurysms patients among deep learning reconstruction (DLR) and hybrid-type iterative reconstruction (IR) and model-based IR.
Methods or Background: 21 intracranial aneurysm patients underwent brain CE-CTA and reconstructed by DLR, hybrid-type IR and model-based IR using a UHR-CT system with super-high resolution mode (SHR: 0.25mm×160 rows/1792 channels). CT values at MCA were assessed by ROI measurements. Image J software was used to generate the profile curves. To assess the capability for improvement of spatial resolution with UHR-CT and DLR, full width at half maximum (FWHM), the width of the edge rise distance (ERD) and the edge rise slope (ERS) were measured at each vessel. For qualitative assessment, overall image quality, artefact, aneurysm, and vascular depiction levels were assessed by 5-point scales by two board-certified radiologists. CT values, ERS and all qualitative indexes were compared by Tukey’s HSD test. Inter-observer agreements of each method were evaluated by kappa statistics with χ2 test.
Results or Findings: CT values and ERS of model-based IR and DLR were significantly higher than those of hybrid-type IR at MCA (p<0.05). Inter-observer agreement of each index by all methods was determined as moderate, substantial or excellent (0.51≤κ≤0.92, p<0.001). In addition, overall image quality and artefact of DLR were significantly improved as compared with others (p<0.05). Aneurysm and vascular depiction levels had no significant difference among all methods (p>0.05).
Conclusion: DLR has a potential for image quality improvements than hybrid-type and model-based IR on brain CE-CTA for UHR-CT.
Limitations: Not applicable.
Ethics committee approval: This retrospective study was approved by the Institutional Review Board of Fujita Health University.
Funding for this study: This study was financially supported by Canon Medical Systems Corporation.

Author Block: S. Kim, C. Suh, H. Oh, E. P. Hong, S. Park, J. K. SUNG, W. H. SHIM, S. J. Kim; Seoul/KR
Purpose or Learning Objective: To develop and validate a deep learning-based automatic brain volumetry (DLABV) for the differentiation of parkinsonian syndromes using 3D T1-weighted brain MR images.
Methods or Background: A DLABV was trained using a dataset of 3D T1-weighted brain MR images. 2D U-Net model was used for model architecture. The training dataset which contains 300 cognitively normal subjects (CN, 129 men) was labelled with FreeSurfer 6.0 brainstem substructure module. The test dataset consists of 207 CN, 52 progressive supranuclear palsy (PSP) patients, 65 multiple system atrophy (MSA) patients, and 189 Parkinson disease (PD) patients. The volume of the midbrain, pons, medulla, SCP, the midbrain-pons area ratio (MP) and the midbrain-pons volume ratio (MP_vol) were measured for differentiation of parkinsonian syndromes. Normalised volume using intracranial volume (ICV) was also used. To distinguish between each group, the receiver operating characteristic curve and area under the curve (AUC) was calculated and classification accuracy was measured by support vector machine (SVM).
Results or Findings: Compared with simple volumetry, volumetry using ICV normalisation showed more accurate performance in the differentiation of parkinsonian syndromes. The AUC in PSP vs PD using normalised midbrain volume was 0.89. In addition, the AUC in MSA vs PD using normalised pons volume was 0.97. MP_vol in MSA patients were significantly larger than in PSP patients and AUC was 0.98. Using normalised volume and MP showed highest classification accuracy.
Conclusion: The DLABV using ICV normalisation allowed an accurate differentiation of parkinsonian syndromes using 3D T1-weighted brain MR images.
Limitations: It is unclear whether the early parkinsonian syndromes can be differentiated using brain volumetry since our study did not target early parkinsonian syndrome patients.
Ethics committee approval: Our institutional review board approved this study.
Funding for this study: This study has received funding by the National Research Foundation of Korea.

Author Block: Q. D. Strotzer, J. Schlaier, A. Beer; Regensburg/DE
Purpose or Learning Objective: Structural connectivity based on diffusion-weighted magnetic resonance imaging (DWI) is gaining importance in research and clinical use in fields like deep brain stimulation. Individual DWI is often unavailable. Therefore, normative connectomes based on averaged whole-brain tractography are a practical alternative. Comparisons of these concepts are sparse. Here, we compared patient-specific and normative approaches by their ability to predict the effects of deep brain stimulation using a symptom-specific, machine learning-based approach.
Methods or Background: Twenty-one patients who received bilateral subthalamic deep brain stimulation for Parkinson’s disease were included. For every electrode contact (168 in total), we computed tractography patterns based on individual DWI and two normative connectomes (32 healthy individuals, 90 Parkinson’s patients). Connectivity strength to 36 brain structures was calculated for every electrode contact, resulting in a dataset of 168 observations (electrode contacts) with 36 attributes (connectivity strength) for each connectome. Stimulation-associated symptom mitigation and side effects were assessed for every contact. We tested the prediction of stimulation outcomes based on connectivity strength using several supervised learning algorithms.
Results or Findings: Support vector machines yielded overall the best results. Averaged across all clinical classes (symptoms, side effects), the individual connectome achieved the highest area under the receiver operating characteristic curve (AUC-ROC; .81) compared to the normative healthy (.76) and disease-matched connectomes (.74). By clinical class, there were significant differences for paresthesia and autonomous side effects in favour of the individual connectome. Results differed considerably between clinical classes, from a mean AUC-ROC of 0.68 for paraesthesia to 0.91 for hyperkinesia.
Conclusion: Clinical effects may be mediated by different networks, as revealed by tractography methods based on DWI. Individual connectomes may be superior in predicting stimulation effectiveness.
Limitations: This study is done with single-centre data and has a limited sample size.
Ethics committee approval: Approval by the local ethics committee.
Funding for this study: No funding was received for this study.

Author Block: K. G. van Leeuwen¹, R. Becks¹, S. Schalekamp¹, B. Van Ginneken¹, M. J. Rutten², M. De Rooij¹, F. J. A. Meijer¹; ¹Nijmegen/NL, ²'S-Hertogenbosch/NL
Purpose or Learning Objective: The commercially available AI tool (StrokeViewer v2, Nicolab) supports the diagnostic process of stroke by detecting large vessel occlusions (LVO) on CTA. We prospectively evaluated this tool in our department to monitor safety and impact.
Methods or Background: We implemented the software with the goal to improve the diagnosis of LVO and elevate the diagnostic confidence of the radiologist (resident). We used quantitative measures (data from clinical systems, vendor log files) and qualitative measures (user survey) to analyse diagnostic performance, number of users, login attempts, radiologists’ diagnostic confidence, and user experience.
Results or Findings: In total, 226 CTAs with a clinical indication of stroke between January-June 2021 were prospectively evaluated. Thirteen cases of posterior circulation and distal vessel occlusions were excluded as they were outside the intended use of the AI tool. The AI tool missed 12 of the 36 occlusions in the middle cerebral or intracranial internal carotid artery (M1=1, M2=10, ICA=1) resulting in an accuracy of 86.4%. Irrespective of location, the sensitivity was 77.8% and specificity 90.4%. The number of monthly unique users varied between 8 and 24 radiologists/residents. Log in attempts dropped after the initial month (which included training) to a monthly average of 44 attempts. The diagnostic confidence did not increase during the use of the tool. The likelihood that users would recommend StrokeViewer to colleagues was rated 4.5/10.
Conclusion: Over six months, the use of StrokeViewer dropped and users did not sense improvement of diagnostic confidence. Measures have been taken to stimulate adoption for the latter six months of the trial period.
Limitations: Because of the prospective character, no comparison could be made between radiologists supported by AI vs radiologists without AI.
Ethics committee approval: Not applicable.
Funding for this study: Not applicable.