RPS 2005 - Recent AI advancements in MRI for precision imaging

Author Block: M. T. Nevalainen¹, O. Nykänen², J. Järvinen¹, V. Casula¹, L. Räsänen¹, M. Nissi², M. T. Nieminen¹; ¹Oulu/FI, ²Kuopio/FI
Purpose: To compare the diagnostic performance of the magnetic resonance fingerprinting (MRF)-derived synthetic MR images of the knee against conventional MR images.
Methods or Background: MRF is an emerging technique to rapidly produce quantitative MRI maps, but its clinical feasibility remains low. However, through deep-learning (DL) synthetic conventional contrasts can be now derived. In this single-center prospective study 78 subjects with knee osteoarthritis were scanned on 3T scanners with isotropic T2-weighted fat-saturated (T2w fs) and PD-weighted (PDw) sequences, and with sagittal MRF sequence. U-net’s were trained to synthesize contrasts from the MRF raw data. Thirty nine cases (1014 images) were used for DL training/validation and 39 cases (1008 images) for diagnostic comparison. Two experienced fellowship-trained musculoskeletal radiologists performed the MOAKS grading and Likert scale assessment of image quality. The inter-rater and inter-method reliability were evaluated using Cohen's kappa and percentages of exact matches.
Results or Findings: The inter-rater reliability was 0.720 (CI 0.625-0.814) for synthetic images and 0.755 (CI 0.678-0.832) for conventional images. Varying inter-method reliability was observed between the synthetic and conventional images: for cartilage grades 𝛋-values varied between 0.418-0.862 (mean 0.747; exact matches 85.9%), for bone marrow lesions between 0.196-0.945 (mean 0.637; exact matches 91.0%), for osteophytes between 0.701-0.885 (mean 0.729; exact matches 60.3%), for meniscus pathology between 0.505-0.782 (mean 0.621; exact matches 78.2%). For effusion and Baker’s cyst 𝛋-values were 0.868 and 0.846, and exact matches 71.8% and 92.3%, respectively. The average Likert scores were better for the conventional than synthetic images (4.6 vs 3.2 for T2w fs) and (4.8 vs. 3.9 for PDw).
Conclusion: The MRF-derived DL-based synthetic clinical contrasts provide good interchangeability with state-of-the-art conventional MR sequences; however, further development is needed to enhance the image quality.
Limitations: Small number of patients, only sagittal images.
Funding for this study: The funding of this study was received from The Finnish Medical Foundation, The Finnish Cultural Foundation, The Terttu Foundation.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the regional medical research ethics committee of the Wellbeing services county of North Ostrobothnia

Author Block: J. Lian, V. Vardhanabhuti; Hong Kong/HK
Purpose: Neurodegenerative diseases, including Parkinson's disease (PD) and Alzheimer's disease (AD), pose a significant healthcare burden to the aging population. Timely detection of these diseases, even before symptom onset, is crucial for early intervention. Although prior studies have shown that wearable movement-tracking data holds promise as a PD biomarker, the potential of combining it with brain MRI for improved prediction remains unexplored. This study aimed to investigate the role of brain MRI as a biomarker for predicting neurodegenerative diseases, alongside accelerometry data, using the UK Biobank dataset.
Methods or Background: This study comprises 19,793 participants with brain MRI scans (T1 and T2), accelerometry data, polygenic risk scores (PRS), and questionnaire-derived lifestyle information. In total, 56 participants were classified as positive cases if diagnosed with any of AD, other forms of dementia, or PD at least one year after the initial assessment. We preprocessed the brain MRI data by segmenting it into 144 segments (normalized by intracranial volume). Disease incidence prediction was employed using XGBoost models, with results reported using the Area Under the Receiver Operator Characteristic Curve (AUROC).
Results or Findings: The comprehensive model, which incorporated all modalities, achieved an AUROC of 0.760 on the test dataset. Among the top 20 predictive features, 17 were related to MRI data, pertaining to different regions in the basal ganglia (e.g., ventral striatum, amygdala), hippocampus, and white matter hyperintensities. In contrast, an ablation study with a model excluding MRI data only achieved an AUROC of 0.645. Other important features include average activity acceleration between 5-7 p.m., and duration of sleep.
Conclusion: Our findings suggest the potential of brain MRI in conjunction with activity tracking data as predictive biomarkers for early neurodegenerative disease detection.
Limitations: Positive cases were not large enough, causing imbalanced dataset.
Funding for this study: No funding was obtained for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the author’s institution's local ethics board (UW-20814) at the University of Hong Kong. The population cohort in this study was from the UK Biobank 20 [Application Number 78730] which received ethical approval from the North West Multicentre Research Ethics Committee (REC reference: 11/NW/03820). All participants gave written informed consent before enrollment.

Author Block: R. Haase, T. Pinetz, E. Kobler, Z. Bendella, C. Gronemann, D. Paech, A. Effland, K. Deike, A. Radbruch; Bonn/DE
Purpose: The main objective of this study was to test a new DNN approach to synthesize artificial T1w-full-dose images from corresponding non-contrast and low-dose images and compare its performance with two reimplemented state-of-the-art approaches (referred to as setting-A and B).
Methods or Background: Two hundred and thirteen participants received an MRI brain scan with an adapted imaging protocol including a T1w-low-dose after 20% of the standard dose of a gadolinium-based contrast agent. Fifty participants were randomly chosen as test set. The new approach is referred to as setting-C. Synthesized artificial T1w-full-dose images were assessed using a reader-based study. Two readers scored the overall image quality, the interchangeability with the true T1w-full-dose in regard to the clinical conclusion, the contrast enhancement of lesions and their conformity to the respective true reference lesions.
Results or Findings: The overall image quality was rated lower in setting-A than in the two remaining settings, whose image quality did not differ from each other and the true full-dose images. The average counts of false positives per case were 0.33±0.93, 0.07±0.33, and 0.05±0.22 for the settings A-C, respectively. The proportion of scans scored as fully or mostly interchangeable was significantly higher in setting-C (70%) than in settings A (40%) and B (57%). The contrast enhancement was significantly reduced in all settings compared to the original T1w-full-dose. Using a five-point Likert-scale, there was no significant difference between the contrast signal reduction of setting-A (-1.10±0.98) and setting-B (-0.91±0.67), but between both settings and setting-C (-0.50±0.55). The average scores of conformity were 1.75±1.07, 2.19±1.04, and 2.48±0.91 for settings A-C, respectively.
Conclusion: The new approach showed a significantly better qualitative performance than published alternatives. Nevertheless, a relevant proportion of cases with inadequate synthesis of the contrast signal remains using a low-dose of 20% of the standard dose.
Limitations: Monocentric study.
Funding for this study: R.H. is funded by a research grant (BONFOR; O-194.0002.1) of the Medical Faculty of the University of Bonn. A.E. and T.P. are funded by the German Research Foundation under Germany's Excellence Strategy (EXC-2047/1, 390685813 and EXC-2151, 390873048).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the Ethics Committee for Clinical Trials on Humans and Epidemiological Research with Personal Data of the Faculty of Medicine of the Rheinische Friedrich-Wilhelms University Bonn: reference no. 450/20.

Author Block: N. Zhao, Y. Pan, Z. Xue, F. Gao, F. Shi, D. Shen; Shanghai/CN
Purpose: The aim of this study was to develop a model for estimating brain age from MR image on a large-scale normal aging population covering entire lifespan, and to assess its potential for early screening of cognitive impairment patients.
Methods or Background: We proposed a novel approach to build a brain age prediction model in lifespan datasets using T1-weighted MR images. This approach consists of extracting three-level hierarchical information through neural networks and fusing them with the cross-attention mechanism, to capture inherent brain age fingerprinting in MR images. Specifically, the hierarchical information included: (1) brain volumes and ratios of 106 parcellations derived from a pre-trained segmentation model, (2) 2D image slices selected from specific brain regions, which potentially contain brain lesion information such as white matter hypointensities, lacunes, and perivascular spaces, (3) 3D CNN features with input of MR image.
Results or Findings: This study included 3,711 subjects aged 6-96 years from in-house datasets, with 3,372 cognitively normal (CN), 207 late MCI (LMCI), and 132 AD. Based on the proposed model, CN subjects achieved a mean absolute error of 2.72 years. Furthermore, when applying this model to cognitively impaired subjects, AD group had higher brain age gap (BAG) compared to both LMCI and CN groups (4.43 vs. 2.47 vs. -0.5 years; P < .001). Finally, combing BAG with learned age-related features as inputs of multi-layer perceptron for differentiating between CN, LMCI, and AD yielded predictive accuracy of 91% for CN vs. LMCI, 91% for CN vs. AD, and 96% for LMCI vs. AD.
Conclusion: The BAG from prediction model appears to be highly correlated with cognitive impairment and could be used for screening of cognitive impairment patients.
Limitations: The model utilises 3D high-resolution images while the extension to clinically low-resolution MRI scans should be studied.
Funding for this study: This work was supported in part by National Natural Science Foundation of China (62131015), Science and Technology Commission of Shanghai Municipality (STCSM) (21010502600), Key R&D Program of Guangdong Province, China (2021B0101420006), STI2030-Major Projects (2022ZD021 3100), The China Postdoctoral Science Foundation (Nos. BX2021333, 2021M703340), and National Key Research and Development Program of China (2022YFE02 05700). Data collection and sharing for this project was funded by Shanghai Zhangjiang National Innovation Demonstration Zone Special Funds for Major Projects (ZJ2018-ZD-012), Shanghai Pilot Program for Basic Research (JCYJ- SHFY-2022-014), and Shanghai Pujiang Program (21PJ1421400).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by Autism Brain Imaging Data Exchange (ABIDE), Attention Deficit Hyperactivity Disorder (ADHD-200), Alzheimer’s Disease Neuroimaging Initiative, Open Access Series of Imaging Studies (OASIS), Consortium for reliability and reproducibility (CoRR), Consortium of Chinese Brain Molecular and Functional Mapping, HUASHAN Hospital, and RENJI Hospital, China, approved this study.

Author Block: V. De Villedon De Naide¹, K. Narceau¹, N. Brillet¹, V. Nogues¹, J. H. Zhang², M. Villegas-Martinez¹, M. Stuber², H. Cochet¹, A. Bustin¹; ¹Pessac/FR, ²Lausanne/CH
Purpose: Myocardial T1-rho mapping is a promising biomarker, allowing for contrast-agent-free myocardial injury detection and quantification. However, its operator-dependant processing induces operator variability and radiologist workload rise. The aim of this study was to explore the feasibility of artificial intelligence-driven analysis for efficient and automated myocardial T1-rho mapping.
Methods or Background: A cohort of 573 patients presenting various cardiomyopathies was divided into a training set (n=500) and a distinct test set (n=73). CMR imaging was conducted using a 1.5T Siemens Aera scanner. For each patient, pre-contrast breath-held 3-slice T1-rho mapping was performed, and contrast-enhanced LGE images were acquired in short-axis 12min post-injection of gadolinium. For each patient, a transformer-based model automatically segmented the left ventricular wall on T1-rho images. Then, the right-ventricle insertion points were detected using a U-Net. A 16-segment AHA model was then created for segmental T1-rho values analysis. Segmentation quality was assessed. T1-rho values were quantitatively retrieved for both manual and automated processing across patient, slice and segment levels. The concordance between methods was gauged. Processing times were measured.
Results or Findings: Automated processing of the T1-rho slices revealed a reduced processing time (~3 s vs. 1 min 51 s±22 s) in comparison to manual processing. Automated segmentation quality yielded favourable results (global DICE of 81.9±9.0%). Analysis of T1-rho values indicated no difference between manual and automated processing (54.9±4.6 ms vs. 55.4±5.1 ms, P=0.098). Strong correlations (ICC>0.8) were found with minimal biases at patient and slice levels, while agreement was lower at the segment level. All AHA segments did not differ significantly between manual and automated T1-rho measurements.
Conclusion: Automated processing of myocardial T1-rho maps shows strong agreement with manual processing and comparable segmentation quality with enhanced time efficiency, highlighting its promising clinical application.
Limitations: Further clinical studies in various cardiomyopathies are warranted.
Funding for this study: This project was supported by funding from the French National Research Agency under grant agreements Equipex MUSIC ANR-11-EQPX-0030, ANR-22-CPJ2-0009-01, ANR-21-CE17-0034-01, Programme d’Investissements d’Avenir ANR-10-IAHU04-LIRYC. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No101076351).
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not applicable

Author Block: J. Steiglechner, B. Bender, G. Lohmann, K. Scheffler, U. Ernemann, T. Lindig; Tübingen/DE
Purpose: This study sought to formulate a generalised classifier for differentiating progressive mild cognitive impairment (pMCI) from stable MCI (sMCI) by integrating brain MRI segmentation and longitudinal atrophy analysis, essential for early Alzheimer's disease (AD) intervention.
Methods or Background: Utilising a segmentation model [AIRAmed] on 3D T1w high-field MRI, which labels 30 anatomical regions, we generated z-statistics through age- and sex-adjusted comparisons with a reference cohort concerning total intracranial volume (TIV) measures. Atrophy rates were quantified using a time-scaled approach. Z-statistics and atrophy rates were inputted into a classifier trained via an 80:20 train-test partition of the Alzheimer’s disease national initiative (ADNI) using logistic regression to distinguish pMCI from sMCI.
Results or Findings: Model generalisation was evaluated on four cohorts (retained test set of ADNI (53 pMCI/93 sMCI), Australian Imaging Biomarkers and Lifestyle Flagship Study of Ageing [AIBL] (15/42), Minimal Interval Resonance Imaging in Alzheimer’s Disease [MIRIAD] (4/8), and Open Access Series of Imaging Studies [OASIS] (27/80)) achieving ROC AUCs of 0.86, 0.78, 0.90, and 0.85 respectively. Including atrophy rates improved overall performance from AUC 0.826 to 0.847, enhancing specificity from 70.4% to 75.8%.
Conclusion: This research underscores the efficacy of leveraging both static (volumetric measures) and dynamic (atrophy rate) features to predict MCI to AD progression. The novel biomarker introduced, combining MRI segmentation and longitudinal atrophy evaluation, not only presents high sensitivity and specificity, as validated by AUC ROC across various datasets but also emerges as a robust tool applicable in future research and diagnostic assistance.
Limitations: Limitations are not applicable for this study.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study used anonymised data.

Author Block: E. O'Shaughnessy; Paris/FR
Purpose: Orbital tumours pose a diagnostic challenge due to their diverse locations and histopathological variations. In recent years, advancements in imaging have enhanced diagnosis, but classification remains challenging. The application of artificial intelligence in radiology and ophthalmology has shown promising results. The primary aim of this study was to develop and assess the performance of machine learning in accurately identifying malignant orbital tumors based on multiparametric 3 Tesla MRI data.
Methods or Background: This single-center prospective study enrolled patients with orbital masses who underwent pre-surgery 3 Tesla MRI scans between December 2015 and May 2021. We employed a Random Forest model with nested stratified cross-validation, considering various combinations of explanatory features. SHAP (SHapley Additive exPlanations) values have been used to evaluate feature contributions, and multiple metrics assessed model performance.
Results or Findings: We analysed 113 patients (50.4% female, 49.6% male) with an average age of 51.5 [19-88]. Among the eight machine learning models evaluated, the one that incorporated all 46 explanatory features (morphology, DWI, DCE, and IVIM) achieved an AUC of 0.9 [0.73-0.99], while the "10-features signature" model attained an AUC of 0.88 [0.71-0.99]. The ten most influential features for the Random Forest model included three quantitative IVIM features, four quantitative DCE features, one quantitative DWI feature, one qualitative DWI feature, and age.
Conclusion: This study suggests that machine learning, when combining multiparametric MRI data, including DCE, DWI, and IVIM, can yield highly effective models for classifying orbital tumors. The "10-features signature" model may be preferred due to its strong performance, simplicity, and adherence to the parsimony principle.
Limitations: The study's limitations encompass its single-center design and the relatively small number of subjects.
Funding for this study: This study was publicly funded.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by an Institutional Research Ethics Board and adhering to Declaration of Helsinki.

Author Block: J. Herrmann, S. Gassenmaier, S. Ursprung, H. Almansour, S. Werner, S. Afat; Tübingen/DE
Purpose: The evaluation of bone disease in multiple myeloma (MM) is an important topic in oncologic imaging. The objective was to determine the impact of a Deep Learning (DL)-reconstruction for whole body (WB) diffusion-weighted-imaging (DWI) for staging MM patients at 3 Tesla compared to standard DWI on reducing acquisition time and improving image quality.
Methods or Background: Thirty patients (mean age, 61±11 years; range, 35–82; 16 men, 14 women) were consecutively included in this retrospective, monocentric study between February and August 2023. Inclusion criteria were standard DWI (DWI_S) in clinically indicated MRI at 3 Tesla, and DL-reconstructed WB-DWI (DWI_DL). All patients were examined using the institution's standard MRI protocol for staging MM including DWI with two different b-values (0 s/mm² and 800 s/mm²) and calculation of apparent diffusion coefficient (ADC) maps. Image quality was qualitatively assessed by two radiologists using a visual 5-point Likert scale (5=best).
Results or Findings: The overall image quality was evaluated to be significantly superior in DWI_DL as compared to DWI_S for b=0 s/mm², b=800 s/mm², and ADC maps by all readers (p<0.05). The extent of noise was evaluated to be significantly less in DWI_DL as compared to DWI_S for b=0 s/mm², b=800 s/mm², and ADC maps by all readers (p<0.001). No significant differences were found regarding artifacts, lesion detectability, sharpness of organs, and diagnostic confidence (p>0.05). Acquisition time for DWI_S was 7:45 min and simulated acquisition time for DWI_DL was 5:03 min.
Conclusion: DWI_DL enhances image quality for WB-MRI in staging of multiple myeloma patients at 3 Tesla. Simulation results suggest a potential reduction in acquisition time of 35 %, highlighting the promise of DL in advancing clinical efficiency.
Limitations: Priliminary results of this study with a small sample size and retrospective design.
Funding for this study: No funding was obtained for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the University of Tuebingen

Author Block: K. Jannusch¹, H. A. Peters¹, N-M. Bruckmann¹, J. Morawitz¹, F. Dietzel¹, L. Umutlu², G. Antoch¹, J. Kirchner¹, J. Caspers¹; ¹Düsseldorf/DE, ²Essen/DE
Purpose: The objective of this study was to evaluate if a machine-learning prediction model based on clinical- and imaging features derived from baseline breast [18F]FDG-PET/MRI staging can predict pathologic complete response (pCR) in patients with newly diagnosed breast cancer prior to neoadjuvant systemic therapy (NAST).
Methods or Background: This study retrospectively enrolled 143 women with newly diagnosed breast cancer. All women underwent a breast [18F]FDG-PET/MRI, histopathological workup of their breast cancer lesions, and evaluation of clinical data. Fifty-six features derived from PET, MRI, sociodemographic/anthropometric, histopathologic, and clinical data were generated and used for input into an extreme-Gradient-Boosting model (XGBoost) to predict pCR. The model was evaluated in a five-fold nested-cross-validation and reduced the risk of overoptimistic estimations. Diagnostic model-performance was assessed by determining ROC-AUC, sensitivity, specificity, PPV, and NPV. Feature importances of XGboost were evaluated to assess which features contributed most to distinguish between pCR and non-pCR.
Results or Findings: Nested-cross-validation yielded a mean ROC-AUC of 80.4±6.0% for prediction of pCR. Mean sensitivity, -specificity, -PPV and -NPV of 54.5±21.3%, 83.6±4.2%, 63.6±8.5% and 77.6±8.1% could be achieved. Histopathological data were the most important features for classification of the XGBoost-model followed by PET-, MRI-, and sociodemographic/anthropometric features.
Conclusion: The evaluated multi-source XGBoost model shows promising results for reliably defining pCR in breast cancer patients prior to NAST. However, the yielded performance is yet insufficient for the algorithm to be implemented in the clinical decision-making process.
Limitations: Inhomogeneity, especially with regard to the histopathological characteristics or NAST therapy regimes might be one limitation. Nonetheless, the risk of inhomogeneity is consistent with the clinical reality of breast cancer patients and underlines the need for a large number of patients to be included in a machine-learning based approach. Thus, as many combinations as possible can be learned by the model.
Funding for this study: The study has been funded by Deutsche Forschungsgemeinschaft (DFG), the German Research Foundation (DFG (BU3075/2‑1 and KI2434/1-2).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: All procedures performed were in accordance with the ethical standards of the institutional research committee and with the principles of the 1964 Declaration of Helsinki and its later amendments. University Duisburg-Essen (study number 17-7396-B0) and University Düsseldorf (study number 6040R).

Author Block: M. A. Mutke, T. Rusche, A. Lonak, K. Blackham, M-N. Psychogios, J. M. Lieb; Basle/CH
Purpose: AI-enhanced MRI sequences promise improved image resolution at reduced acquisition times but are often proprietary products, with limited evaluation. The goal of this study was to investigate whether an AI-enhanced FLAIR sequence of the brain is non-inferior to a standard sequence, as assessed qualitatively by experienced human radiologists at both general and individual levels.
Methods or Background: Patients underwent both standard FLAIR-sequence and an AI-enhanced FLAIR product sequence with reduced acquisition time of 40% (deep-resolve-boost, Siemens Healthineers), post contrast injection on 1.5T and 3T MRIs. Two experienced neuroradiologists conducted a side-by-side comparison. Using a five-point Likert scale (ranging from non-diagnostic to excellent), images were assessed for signal-to-noise ratio, anatomic clarity, overall quality, imaging artefacts, and diagnostic confidence. Potential lesions (pseudolesions) unique to one sequence and not readily identifiable as artifacts were noted and compared (McNemar's test). A mixed-model analysis determined the sequence type's effect on ratings, adjusting for rater and patient variances. Non-inferiority was tested (lower bound of the confidence interval not exceeding -0.5 points).
Results or Findings: Thirty patients were assessed for cerebral metastases (n=11), glioma (n=9), meningioma (n=5) and other pathologies (n=5). In the mixed model, AI sequences significantly outperformed by 0.4 points (Likert scale) in anatomic clarity but underperformed by -0.5 points for artifacts, failing to meet non-inferiority. In all other categories, the AI sequence matched the standard. Both readers identified pseudolesions on a total of 5/30 patients.
Conclusion: The faster, AI-enhanced FLAIR sequence is non-inferior to the standard except for an increase in imaging artifacts. These artefacts can result in pseudolesions, potentially causing unnecessary follow-up imaging. We emphasize the need for critical and rigorous evaluation of new sequences and suggest a training phase for radiologists.
Limitations: Small sample size, single-centre with qualitative, subjective evaluation limit this study.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: The imaging was part of a quality control program, informed consent was waived by the local ethics committee.