RPS 2011 - Hot Topic: AI-assisted neuroimaging and image analysis

Author Block: N. Snoeijink, J. D. J. Slotman, S. Kamerbeek, I. M. Nijholt, M. F. Boomsma, E. Wiegman; Zwolle/NL
Purpose: To assess inter-observer variability of expert-based radiation target definition with and without AI assistance in gamma knife radiosurgery for brain metastases on MRI scans.
Methods or Background: A deep learning-based model was developed for automatic detection and delineation of brain metastases using contrast-enhanced T1-weighted (CE-T1w) and black blood (BB) MR sequences. Data from 224 consecutively treated patients were randomly split into training (n=162) and testing (n=64) sets. The AI model achieved an overall F1-score of 0.93, sensitivity of 0.88, PPV of 0.98, and a DICE score of 0.82. Subsequently, it was tested on an independent dataset consisting of 15 patients, with in total 93 brain metastases, classified as S (< 0.1 cc), M (≥0.1 cc & <1 cc), or L (≥1 cc). Four observers evaluated each case: two radiotherapists independently, AI model alone, and combination of AI with a radiotherapist able to refine results manually. Inter-observer variability was quantified using the DICE-score using clinical delineation by one of the radiotherapists as reference.
Results or Findings: Inter-observer variability in DICE-scores was observed among radiotherapists (0.88) and between radiotherapist and the AI model (0.73), particularly for smaller metastases (<0.1 cc). Adjustments made by the radiotherapist to the AI-generated delineations increased the DICE-score to 0.84. The largest improvement was seen for small lesions (S: 0.54 --> 0.76, M: 0.81 --> 0.86, L: 0.87 --> 0.88).
Conclusion: The AI model is trained on delineations from multiple radiotherapists, who do not always agree with each other. This inter-expert variability may limit AI’s ability to achieve a maximal DICE, potentially challenging effective implementation of AI in clinical practice.
Limitations: Small sample size
Funding for this study: Innovation and Science fund Isala, The Netherlands
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: T. Han, J. Zhou; Lanzhou/CN
Purpose: This study aimed to evaluate the value of intratumoral and peritumoral radiomics models based on multiparametric MRI in predicting meningiomas recurrence/progression (R/P).
Methods or Background: A total of 623 patients from Hospital A were divided into a training set (n=294) and a test set (n=127); 202 patients from Hospital B comprised the external validation set. A Clinical-Radiological (C-R) model was developed using significant clinical/MRI features. Radiomics features were extracted from intratumoral (Intra) and peritumoral regions at 5 mm (Peri-5) and 10 mm (Peri-10). Models were constructed using Lasso and evaluated via ROC, calibration curves, and decision curve analysis (DCA). Interpretability was assessed with SHAP (SHapley Additive exPlanations) plots.
Results or Findings: The training set AUCs for C-R, Intra, Peri-5, and Peri-10 models were 0.820, 0.880, 0.860, and 0.840, respectively. Test set and external validation AUCs ranged from 0.820–0.910 and 0.620–0.750. The results of the DeLong test indicated that, in external validation set, the predictive performance of the Intra model was significantly superior to that of the C-R, Peri-5 mm, and Peri-10 mm models (P=0.046, P=0.024, P=0.035).
Conclusion: The Intra, Peri-5 mm, and Peri-10 mm radiomics models and C-R model, demonstrated good predictive performance in predicting meningioma R/P, providing a theoretical reference for formulating personalized treatment plans for meningioma patients.
Limitations: Firstly, it is a retrospective study. Secondly, this study only analyzed the tumor and peritumoral regions, while the predictive value of the peritumoral edema region remains to be further explored. Lastly, manual segmentation methods were used to delineate the VOI of the tumor in this study. In future studies, we aim to explore the automatic segmentation method and larger prospective datasets to improve enhance and validate the model's performance and robustness.
Funding for this study: This study was supported by grants of Natural Science Foundation of China (82371914).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This retrospective study received approval from the Ethics Review Committees of the Hospital A (2023A-169) and Hospital B (2024262K), with the need for informed consent waived.

Author Block: N. Zdanovskis, K. Šneidere, K. Kalva, Z. A. Litauniece, A. Usacka, Z. Freibergs, A. Platkajis, A. Stepens; Riga/LV
Purpose: To assess whether machine learning algorithms can predict Montreal Cognitive Assessment (MoCA) scores from MRI morphometry in patients with cognitive impairment.
Methods or Background: Eighty subjects were included, with 70 used for training and 10 for testing. From structural MRI, 101 morphometric features were extracted, including cortical thickness (68 regions), subcortical volumes (20 bilateral structures), corpus callosum subdivisions (5), and global volumetric measures (8). Six supervised regression models were evaluated: Linear Regression, Random Forest, Gradient Boosting, AdaBoost, Support Vector Machine (SVM), and k-Nearest Neighbors (kNN). Performance was assessed using mean squared error (MSE), mean absolute error (MAE), and R².
Results or Findings: Random Forest achieved the best accuracy (MSE = 32.7, MAE = 4.55, R² = 0.655), followed by Gradient Boosting (MSE = 47.8, MAE = 5.91, R² = 0.496) and AdaBoost (MSE = 58.4, MAE = 6.80, R² = 0.384). Linear Regression and kNN showed weak predictive value (R² = 0.001 and 0.10), while SVM performed poorly (R² = –0.145). Neural Networks failed to converge (R² = –3.14). At the subject level, Random Forest predictions were closest to actual scores; for example, a patient with MoCA = 24 was predicted as 25, while another with MoCA = 7 was predicted as 13, and one with MoCA = 30 as 24.
Conclusion: Machine learning models demonstrated the ability to approximate MoCA scores from MRI-derived morphometric features with clinically meaningful accuracy. These findings suggest that in future ML-based approaches could be applied in clinically relevant scenarios, such as supporting early detection of cognitive impairment and stratifying patients for further testing.
Limitations: Single-center retrospective study with a modest sample size. External validation is required to confirm generalizability. MoCA, while widely used, may not capture the full spectrum of cognitive domains.
Funding for this study: Modifiable bio and life-style markers in predicting cognitive decline (MOBILE-COG) No: RSU-PAG-2024/1-0014 is financed by the investment of the European Union Recovery and Resilience Facility and the state budget within the project "RSU internal and RSU with LASE external consolidation" No. 5.2.1.1.i.0/2/24/I/CFLA/055.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approval was obtained from the institutional ethics committee, and all participants provided informed consent.

Author Block: E. Akl¹, E. Beller¹, D. Cantré¹, F. G. Meinel¹, M-A. Weber¹, S. Langner², W. Hermann¹, M. Lütgens¹, A-C. Klemenz¹; ¹Rostock/DE, ²Greifswald/DE
Purpose: The Alberta Stroke Program Early CT Score (ASPECTS) and advances in CT reconstruction, such as Adaptive Statistical Iterative Reconstruction-V (ASIR-V), and Deep Learning Image Reconstruction (DLIR) play an important role in the neurodiagnostic workflow. This study examines the effect of these reconstruction techniques on automated ASPECTS.
Methods or Background: In this retrospective study, 173 patients (mean age 77 ± 12 years, 39% female) with suspected middle cerebral artery infarction underwent non-contrast CT, reconstructed with FBP, ASIR-V (at 30% and 60%), and DLIR (low, medium, and high). Automated ASPECTS were analyzed, with FBP as the reference standard.
Results or Findings: Compared to FBP, ASIR and DLIR reconstructions resulted in a mild overall underestimation of automated ASPECTS, with the least pronounced underestimation for ASIR-V 30% and DLIR-L, and a more marked underestimation for ASIR-V 60% and DLIR-M/H. Most re-classifications involved shifts of ASPECTS from moderate (5–7) to high (8–10), with DLIR-H showing the greatest effect. DLIR more frequently up-classified patients from ≤5 to ≥6, whereas down-classifications were rare. Regionally, the insular ribbon was most underestimated and M3 most overestimated, with DLIR-H exhibiting the largest total regional discrepancies. Agreement with expert consensus was highest for DLIR-M, followed by ASIR-V, while FBP and DLIR-H showed lower concordance.
Conclusion: Both ASIR-V and DLIR showed generally minor underestimation of ASPECTS compared to FBP. Occasional overestimation, however, resulted more often in reclassifications of the ASPECTS score, which affected patient eligibility for endovascular therapy (ASPECTS ≥ 6 vs. ASPECTS ≤ 5). DLIR-M was most accurate compared with expert consensus. Careful selection, optimization, and standardization of reconstruction parameters are essential for consistent and reliable stroke imaging.
Limitations: No limitations were identified.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Not applicable - retrospective study.

Author Block: Z. Cui, J. Zhang; Shanghai/CN
Purpose: To develop and validate a reliable machine learning (ML) model combining carotid plaque and white matter hyperintensity (WMH) radiomics features with radiological characteristics to assess acute ischemic stroke (AIS) severity.
Methods or Background: Retrospective data were collected from patients with symptomatic carotid artery stenosis (CAS) and AIS between January 2017 and December 2023, and a prospective cohort from October 2023 to October 2024. Based on admission NIHSS scores, patients were categorized into NIHSS score ≤ 1 and > 1 groups. All patients underwent high-resolution vessel wall MRI and brain MRI. Conventional imaging features of carotid plaques and WMH were used to build a conventional imaging model. Radiomics features were extracted from carotid plaque (T1WI and contrast-enhanced T1WI) and WMH (FLAIR), and multiple ML algorithms were applied to build radiomics models. A hybrid model was subsequently developed by combining radiomics and conventional imaging features. Model performance was evaluated in the test set and prospective validation cohort.
Results or Findings: Three cohorts were included: retrospective training (140 patients), testing (59 patients), and prospective validation (71 patients). The conventional imaging model achieved areas under the curve (AUC) of 0.94, 0.88, and 0.75 in the training, test, and validation cohorts, respectively. The radiomics model achieved corresponding AUCs of 0.87, 0.79, and 0.73. The hybrid model, which incorporated 9 conventional imaging features and 23 radiomics features (16 plaque-derived and 7 WMH-derived), yielded superior performance, with AUCs of 0.97, 0.96, and 0.87. Decision curve analysis demonstrated that the hybrid model provided greater net clinical benefit.
Conclusion: The ML-based hybrid model enables accurate, non-invasive AIS severity assessment and demonstrates strong clinical potential.
Limitations: First, the retrospective design may have introduced selection bias. Second, the lack of multicenter validation and the modest sample size remain limitations.
Funding for this study: This work was supported by the National Key R&D Program of China (Grant no. 2022YFF0708700), Shanghai Science and Technology Program (Grant no. 22S31905300, 22YF1405000), the National Natural Science Foundation of China Youth Program (Grant no. 82402393).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study protocol was approved by the Institutional Review Board of our hospital, and written informed consent was obtained from all participants.

Author Block: Y. Huang¹, J. Chu¹, L. Zhao², R. Xu², W. Zhou³, H. Wei³; ¹Guangzhou/CN, ²Beijing/CN, ³Shanghai/CN
Purpose: Accurate preoperative prognosis remains challenging in glioblastoma (GBM). This study developed a multimodal nomogram integrating radiomic, clinical, and immunological features to predict overall survival (OS).
Methods or Background: We retrospectively analyzed 298 training and 123 validation GBM patients. A total of 6,792 radiomic features were extracted from tumor core and peritumoral edema regions on T1CE, T2WI, and T2-FLAIR MRI. Clinical variables (cortical, ependymal, and contralateral invasion) and immune markers (PD-1, CD68+, CD86+, CD163+, HIF-1α+, etc.) were evaluated. Tumor and peritumoral radiomic scores were constructed via LASSO-Cox regression. A Cox model integrated these scores with clinical and immune scores to predict OS.
Results or Findings: Tumor-core radiomic scores: 2 wavelet-based features were selected via LASSO-Cox regression to construct the tumor core score (HR=3.02, 95%CI 1.38–6.64, p=0.006). Peritumoral radiomic scores: 5 skewness-based features were identified for the peritumoral score (HR=3.18, p=0.049). Clinical score​​: Multivariate Cox analysis confirmed cortical invasion, ependymal invasion, and contralateral invasion as independent prognostic factors (all p<0.05). The clinical score stratified patients into high- and low-risk groups with distinct survival trajectories (p<0.05). ​​Immune stratification​​: Based on densities of nine immune markers (PD1+, CD68+, etc.), patients were classified into immune-cold, -intermediate, and -hot subgroups. Survival analysis revealed significant OS differences among subgroups (p<0.05). Multimodal integration​​: The integrated nomogram combining Tumor-core radiomic scores, peritumoral radiomic scores, clinical scores, and immune scores demonstrated superior predictive accuracy compared to unimodal models.It achieved an AUC of 0.86. High-risk patients had significantly shorter OS (p<0.05).
Conclusion: The nomogram synergizes radiological, pathological, and immunological data to refine GBM prognostication. Peritumoral skewness features and specific invasion patterns signal aggressive biology, while immune-hot phenotypes correlate with poorer outcomes. This framework aids personalized therapy decisions.
Limitations: This study was a ​​single-center design​​.
Funding for this study: No.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: M. Nemy, E. Stoklasa, D. Bojko, V. Sedlák, A. Kavková, L. Polášková, I. Jacečková, M. Majovsky; Prague/CZ
Purpose: This study presents a hospital-integrated real-time system for multi-shell diffusion MRI (dMRI) analysis in gliomas, including its architecture, secure PACS-connected workflow, and embedded machine-learning models for molecular prediction and automatic segmentation.
Methods or Background: A cohort of 210 patients with histologically confirmed intra-axial gliomas underwent a 3T MRI protocol including high-angular, multi-shell dMRI (134 directions, 7 b-values). Advanced diffusion models (DKI, RSI, VERDICT, NODDI) were derived to characterize microscopic tissue composition. Cellularity maps were used to train classical machine-learning models (XGBoost, random forest) and deep-learning models (CNNs, ResNet50), with data augmentation and transfer learning to address sample size. Automatic tumor and peritumoral segmentation was performed using CNNs and benchmarked against semi-automatic annotations. A secure hospital-integrated pipeline was implemented to enable real-time use: scans are anonymized at acquisition, transferred via encrypted channels to a high-performance cluster, processed under strict time limits, and results returned directly into PACS for radiologist review.
Results or Findings: Advanced dMRI models outperformed conventional ADC in glioma characterization. For IDH mutation prediction, AUC improved from 0.82 (ADC) to 0.91 with advanced models, and for glioma grading from 0.88 to 0.94. Deep-learning achieved the best performance with IDH classification (AUC >0.95, sensitivity >0.93). Automatic segmentation achieved mean Dice coefficients of >0.83 for enhancing tumor and >0.78 for peritumoral zones. The hospital-integrated system processed cases end-to-end robustly within clinically acceptable time limits.
Conclusion: This work demonstrates a clinically viable hospital-integrated system that combines advanced multi-shell dMRI and machine learning for accurate molecular prediction and automatic glioma segmentation. The architecture enables secure, real-time analysis with seamless PACS integration, provides advanced imaging biomarkers directly to radiologists, and may reduce manual workload in routine practice.
Limitations: The limitations of the study are its single-center design and relatively limited sample size.
Funding for this study: Funding was provided by the Ministry of Health of the Czech Republic, grant no. NW25J-08-00023, and by Charles University, project GA UK no. 222623.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the Ethics Committee of the Military University Hospital Prague.

Author Block: Z. Zhu¹, H. Yang¹, H. Wang², Y. Song¹, M. Xu¹, X. Zhang¹, W-W. Lin², T. Li², B. Zhang¹; ¹Nanjing/CN, ²Shanghai/CN
Purpose: Our study focused on tumor regions segmentation and predicting the World Health Organization (WHO) grade, isocitrate dehydrogenase (IDH) mutation, and 1p/19q codeletion status using deep learning models on preoperative Magnetic Resonance Imaging (MRI).
Methods or Background: Glioma is the most common primary malignant brain tumor and preoperative genetic profiling is essential for management of glioma patients. We developed an optimal mass transport (OMT) approach to transform irregular MRI brain images into tensors. In addition, we proposed an algebraic pre-classification (APC) model utilizing multi-mode OMT tensor singular value decomposition (SVD) to estimate pre-classification probabilities. The fully automated deep learning model named OMT-APC was used for multitask classification. Our study incorporated preoperative brain MRI data from 3,565 glioma patients across 16 datasets spanning Asia, Europe, and America. Among these, 2,551 patients from 5 datasets were used for training and internal validation. In comparison, 1,014 patients from 11 datasets, including 242 patients from The Cancer Genome Atlas (TCGA), were used as independent external test.
Results or Findings: OMT segmentation model achieved mean lesion-wise Dice scores of 0.880. The OMT-APC model was evaluated on the TCGA dataset, achieving accuracies of 0.855, 0.917 and 0.809, with AUC scores of 0.845, 0.908 and 0.769 for WHO grade, IDH mutation, and 1p/19q codeletion, respectively, which outperformed the four radiologists in all tasks.
Conclusion: These results highlighted the effectiveness of our OMT and tensor SVD-based methods in brain tumor genetic profiling, suggesting promising applications for algebraic and geometric methods in medical image analysis.
Limitations: Lacking prospective validation cohort.
Funding for this study: National Science and Technology Innovation 2030 -- Major program of "Brain Science and Brain-Like Research" (2022ZD0211800), National Natural Science Foundation of China (82271965, 82330059, 12371377)
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Medical Ethical committee of Nanjing Drum Tower Hospital