EIBIR Poster Session

EIBIR 1 - EIBIR stage bonus session 1

February 28, 14:30 - 15:30 CET

6 min
Advancing Neurodegenerative Disease Assessment through Concurrent Multimodal FDG-PET, Perfusion, and Diffusion Evaluation
Joachim Strobel, Ulm / Germany
Author Block: J. Strobel, W. Thaiss, J. Kassubek, A. J. Beer, M. J. Beer, H-P. Müller, N. Sollmann; Ulm/DE
Purpose: This study explores a novel approach to neuroimaging for neurodegenerative diseases (NDD) using a hybrid PET/MRI platform that combines Arterial Spin Labeling (ASL), Diffusion Tensor Imaging (DTI), and Positron Emission Tomography (PET). The goal was to integrate these modalities for a comprehensive analysis of NDDs in patients with conditions like Alzheimer’s disease, frontotemporal dementia, and primary progressive aphasia.
Methods or Background: In this study, 76 patients with NDD were examined using the Siemens Biograph PET/MRI system. The standardized acquisition and analysis protocols allowed comparisons between patients and a control group. Specific brain structures of interest were identified based on prior research, and region-of-interest (ROI) analysis was conducted. An artificial intelligence (AI)-based support vector machine (SVM) was employed to categorize different neurodegenerative pathologies.
Results or Findings: PET imaging achieved 92.0% accuracy in distinguishing NDD patients from controls and 78.1% for differentiating among NDD types. Combined ASL/DTI matched or exceeded PET accuracy for nfvPPA (81.0%) and lgPPA (94.7%), but was lower for AD (76.3%) and svPPA (88.9%). Voxel-wise analysis revealed overlap between PET and ASL clusters, especially in left-lateralized regions for svPPA and lgPPA.
Conclusion: The study reveals significant differences in PET, ASL, and DTI measurements between neurodegenerative disease (NDD) patients and controls. PET achieved 92% accuracy in distinguishing NDDs, while ASL/DTI showed comparable or better performance for specific diseases like nfvPPA and lgPPA. Voxel-wise analysis revealed overlap between PET and ASL clusters, particularly in left-lateralized regions for svPPA and lgPPA, with white matter changes detected by DTI.
Limitations: The use of predefined ROIs may miss subtle or widespread brain changes that whole-brain analyses could detect. Additionally, the voxel-wise analysis focused on lateralized regions, potentially overlooking bilateral or more extensive brain pathologies, underscoring the need for further whole-brain studies.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was conducted in full compliance with ethical standards, and approval was obtained from the appropriate ethics committee. All procedures followed were in accordance with the ethical guidelines for human research, and informed consent was obtained from all participants or their legal guardians prior to the study. The ethics committee raised no objections regarding the study design, data collection, or analysis methods used in this research.
6 min
Development of a radiomic signature to predict overall survival probabilities of patients affected by soft-tissue sarcomas of the extremities and trunk wall
Andrea Vanzulli, Tradate / Italy
Author Block: A. Vanzulli, G. Tinè, A. Messina, C. Morosi, R. Miceli, S. Stacchiotti, A. Gronchi, S. Pasquali, D. Callegaro; Milan/IT
Purpose: Soft-tissue sarcomas of the extremities and trunk wall (ESTS) are characterized by heterogeneous clinical behaviour.
Surgery is the standard-of-care, with peri-operative chemotherapy in patients at high risk for recurrence, namely those with a predicted overall survival (pOS) <60% according to validated nomograms such as Sarculator.
Methods or Background: Tumours were manually annotated on pre-operative MRI of 91 patients with primary ESTS who underwent surgery (2011-2015).
A total of 2144 radiomic features pertaining to First Order and GLCM categories were extracted with a customized tool (Corino, 2018) and a Laplacian filter was applied to retain those explaining ≥ 80% of dataset variance (n = 168).
The selected radiomic features were then combined into a prognostic radiomic signature (RS) applying a customized version of the Bayesian Lasso adapted for survival regression.
Accuracy was evaluated with the 5-year AUC and C-index, while calibration was assessed using the 5-year Brier Score (BS).
Results or Findings: Median tumor diameter was 8 cm (5-10.8). Tumor malignancy grade was G1, G2 and G3 in 22(24,2%) 21(23.1%) G2 and 48 (52.7%) patients, respectively.
The following tumor histologies were identified: 8.8%LMS, 26.4%UPS, 8.8%DDPLPS, 18.7%MLPS, 3.3%MPNST, 22%MFS, 4.4%SS, 7.6%Other.
The median follow-up was 55.8 months (IQR, 44.8-72.9).
19 patients had a pOS <60%.
The RS demonstrated excellent predictive accuracy (5-year AUC: 0.883, 95% CI 0.828-0.927; C-index: 0.840, 95% CI 0.791-0.878) and calibration (5-year BS: 0.376, 95% CI: 0.337-0.412) for stratifying patients’ OS and was significantly associated with Sarculator’s risk groups, with RS values 39.6% higher in the low-pOS compared to the high-pOS group (p=0.001).
Conclusion: Radiomics may be employed to stratify ESTS patients at diagnosis - unlike commonly employed nomograms, which are utilized after resection and definitive histopathological examination - and orient clinical management.
Limitations: External validation pending
Funding for this study: Nothing to disclose
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not applicable
6 min
Automated Psoas Muscle Segmentation; Imaging Features and Surgical Fitness in Spinal Metastatic Lung Cancer
Marco Perez Caceres, Brossard / Canada
Author Block: M. Perez Caceres, O. Ahmed, V. Freire, J. Shen, F. Al-Shakfa, D. Boulé, Z. Wang; Montreal, QC/CA
Purpose: Lung cancer’s propensity for spinal metastasis leads to fractures, dysfunction, pain, and reduced quality of life. Spinal interventions are selectively offered to patients fit for surgery. Sarcopenia, measured by psoas (PM) and whole abdominal muscle (WAM), is proposed as a fitness marker, but lacks consensus on thresholds and segmentation tools.This study aims to validate sarcopenia metrics as imaging biomarkers using both open-source and locally tailored neural networks pertaining to bone-metastatic lung cancer and spinal surgery.
Methods or Background: A retrospective cohort of 63 lung cancer patients (age 64±9, 46% female) with spinal metastasis who underwent surgery between 2010 and 2020 was analyzed. Psoas muscle and lumbar vertebrae segmentation were validated by a musculoskeletal radiologist on CT. Segmentation models were trained using nnUNet, and TotalSegmentator (TS) was applied for PM and WAM segmentation. Sarcopenia metrics (i.e. PMI, PLVI, SMI, TMA) and radiomic features were assessed. Survival analysis was performed based on sarcopenia classification using the Wilcoxon log-rank test with sex and smoking status stratification.
Results or Findings: The locally tailored psoas segmentation model outperformed TS in 7 metrics. PMI and PLVI thresholds showed significant survival differences only when measured with the local model (p<0.05), but not SMI or TMA. Percentile-based classification revealed significant survival differences, especially in local PM metrics (p<0.001). Of 108 radiomic feature clusters, 38 were significant with local models, none with TS WAM segmentation.
Conclusion: The local model demonstrated superior performance compared to TS. Percentile-based thresholds based on PM features were more predictive of survival, suggesting the need for disease-specific cutoffs. Radiomic features warrant further investigation.
Limitations: Statistical power was limited in due to a small number of subjects. Additionally, survival data from a non-surgical group with conservative treatment options namely radiotherapy should be collected.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Authorisation for a retrospective patient analysis was obtained from the research ethics board in October 2021(authorization number 21.222).
6 min
Predicting clinically significant versus clinically insignificant prostate cancer with a combined MpMRI clinical model based on machine learning: a multicentric study
Alessandro Venturi, Florence / Italy
Author Block: R. Girometti1, D. Fazzini2, A. Colarieti3, L. Cereser1, G. D'Imperio1, M. Interlenghi2, A. Venturi2, F. Sardanelli2, I. Castiglioni2; 1Udine/IT, 2Milan/IT, 3San Donato Milanese, Milan/IT
Purpose: To build and test an artificial intelligence-based model to predict whether ADC map-visible prostate lesions are clinically significant (csPCa) versus clinically insignificant prostate cancer (ciPCa).
Methods or Background: We retrospectively included men who, between 2016-2023, underwent mpMRI on a 1.5T/3.0T magnets in three different centers, with subsequent prostate biopsy (systematic and target cores) because of PI-RADS (v2.1) ≥3 lesions and/or high clinical suspicion. csPCa and ciPCA were defined as grading group (GG) ≥2 and GG1, respectively. On 193 T2-weighted images from the first center, a U-Net based DL algorithm to automatically extract prostate volume, thus obtaining Prostate-Specific-Antigen-Density [PSAD], was trained and internally tested, evaluating the Dice-Similarity-Coefficient (DSC) in comparison with three board-certified radiologists. It was then externally tested on 86 images from the second center. Separately, a radiomic machine-learning model to classify csPCa versus ciPCa from segmented prostate lesions on ADC map, was trained and internally tested on 85 images (10-fold nested cross-validation) and externally tested on 95 images, all from the third center.
Results or Findings: A total of 459 patients, aged 66.2±7.7 yo were included. The DL algorithm automatically computed PSAD in a range of 0.02−2.36 ng/cc2 (median 0.10 ng/cc2), achieving a DSC of 0.86 and 100% repeatability. Optimal cutoff on age was 67yo. A PSAD cutoff of 0.11ng/cc2, combined with age cutoff achieved 0.84 sensitivity and 0.64 specificity on the external dataset. The radiomic ML classifier combined with the age cutoff achieved 0.93 sensitivity and 0.80 specificity on the external dataset.
Conclusion: Our results showed how cutoffs for PSAD, age and radiomic predictors can be obtained by machine learning applied on mpMRI to distinguish csPCa from ciPCa.
Limitations: All centers were italian; retrospective study.
Funding for this study: No fundings were received.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics committee approved this study.
6 min
Prediction of treatment response and tumor progression in rectal cancer on staging MRI using radiomics
David Luengo Gómez, Granada / Spain
Author Block: D. Luengo Gómez, M. García Cerezo, Á. Salmerón Ruiz, D. López Cornejo, E. González Flores, A. J. Láinez Ramos-Bossini; Granada/ES
Purpose: Radiomics has shown promise in predicting rectal cancer (RC) characteristics, therapeutic responses, and outcomes. The primary objective of this study is to identify reliable radiomic biomarkers on MRI for predicting histologic staging and tumour progression in RC.
Methods or Background: A retrospective cohort of 129 patients who underwent MRI for RC staging was analysed. Sixty-six patients received neoadjuvant therapy (NAT), while 63 did not. Following manual segmentation of T2 sequences from two independent abdominal radiologists, Py-Radiomics was used to extract a total of 1710 radiomic features for each patient. Several Machine Learning models were trained to predict tumour and lymph node status, progression, and complete pathological response (CPR).
Results or Findings: Among patients who did not receive NAT, the Nearest Centroid (NC) model was the most effective in predicting both tumour stage and CPR. For lymph node status, Bagging LDA and Bagging Gaussian Naive Bayes (NB) was employed for progression prediction. The corresponding sensitivity values were 0.68 for pT, 0.82 for lymph node status, 0.48 for progression, and 0.68 for CPR. In the NAT group, NC predicted tumour stage, while LDA was used for lymph node status. Bagging LDA predicted progression, and Gaussian NB was utilised for CPR prediction. Sensitivity values for the NAT group were: 0.51 for tumour prediction, 0.85 for lymph node status, 0.77 for progression, and 0.52 for CPR. Explainability was provided using the SHAP algorithm, which highlighted that the most relevant features were texture-based characteristics.
Conclusion: The evaluated radiomic models demonstrated potential for improving the management of RC. The ability to accurately predict tumour response and progression risk post-treatment may improve neoadjuvant treatment or even dispense with surgery in selected cases.
Limitations: The main limitations of the study are an imbalanced and limited dataset.
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: The study was approved by the Provincial Ethics Committee of Granada (code CRIOMA2).
6 min
Artificial Intelligence accurately differentiates radionecrosis from true progression in brain metastasis treated with stereotactic radiosurgery: analysis of 106 histologically assessed lesions
Gaia Ressa, Milan / Italy
Author Block: G. Ressa1, R. Levi1, G. Savini1, L. Raspagliesi1, E. Clerici2, F. Pessina2, M. Scorsetti2, L. S. Politi2; 1pieve emanuele/IT, 2Rozzano/IT
Purpose: To develop automated articial intelligence models able to discriminate radionecrosis from disease progression in brain metastases treated with stereotactic radiosurgery.
Methods or Background: This single-centre retrospective study included all patients who underwent neurosurgery for brain metastasis after stereotactic radiosurgery, between 2012 to 2022.
Pre-neurosurgical FLAIR and post-contrast T1 (T1ce) aquisitions were segmented using a convolutional neural network (CNN) into non-enhancing, enhancing, and edema volumes of interest (VOI). Radiomics features (RFs, n=321) were extracted from each sequence and VOI, and their significance to radionecrosis was assessed through univariate and multivariate analyses.
A Random Forest machine learning model was trained on 70% of the sample and evaluated on the remaining 30% using Bayesian optimization and 10-fold cross-validation. A 3D ResNet-based CNN was trained on the same split dataset.
Occlusion sensitivity (OS) algorithm (Monai) was employed on the trained CNN networks as explainable AI model.
Post-surgical histology was available for all cases.
Results or Findings: This study included 106 patients (41 males, mean age 56.4).
Histological analysis revealed exclusive radionecrosis in 30 cases (28%), while others showed varying percentages of neoplastic cells.
Univariate analysis identified 131 significant RFs between groups, including GLDM_DNUN and GLDM_SDE (p<0.001) within the enhancing area of T1ce. Multivariate analysis confirmed GLDM_DNUN (OR 0.39) and GLDM_SDE (OR 0.28) as significant for radionecrosis.
The Random Forest model achieved 0.79 accuracy, 0.81 AUROC, and 0.87 sensitivity, while the CNN ResNet model performed better (86.9% accuracy, 0.889 AUROC, 87.8% sensitivity).
ResNet yielded similar AUROC on segmented (0.889) and unsegmented (0.890) brain images, highlighting it can automatically focuses on lesions (demostrated by OS Maps).
Conclusion: AI models could be employed to differentiate between radionecrosis and disease progression in brain metastasis following SRS, potentially reducing unnecessary brain surgery interventions.
Limitations: Single-centre retrospective study.
Funding for this study: No funding has been received.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: IRB approval was waived considering the retrospective non-interventional study design.
6 min
AI-Assessed Sarcopenia as a Predictor of Neoadjuvant Chemotherapy Outcomes in Muscle-Invasive Bladder Cancer
Antonella Borrelli, Rome / Italy
Author Block: A. Borrelli, S. Novelli, F. Mezzapesa, E. Messina, C. Catalano, V. Panebianco; Rome/IT
Purpose: Sarcopenia has already been widely investigated as a potential biomarker of negative outcomes in oncological patients. Our aim was to evaluate the potential predictive role of sarcopenia, assessed using an Artificial Intelligence-powered software, in response to neoadjuvant chemotherapy (NAC) in patients with muscle-invasive bladder cancer (MIBC).
Methods or Background: Single-center retrospective study; patients with non-metastatic MIBC and addressed to NAC who had a mpMRI and a total-body CT scan available prior to platinum-based neoadjuvant therapy were enrolled. The follow-up MRI assessment was performed using the NacVI-RADS algorithm for evaluation of response to therapy. An AI-powered software was applied to the CT images to calculate the Skeletal Muscle Index (SMI-L3), based on the psoas, long spine, and abdominal muscle areas at the L3 level on axial images. Multivariate logistic regression analysis was performed to assess predictors of response defined as the proportion of patients who achieved objective clinical response rate (CR).
Results or Findings: This study included 45 bladder cancer patients with an average age of 63.5 years, 31 (68.89%) of whom formed the non-sarcopenic group (NSG) and 14 (31.11%) the sarcopenic group (SG). Multivariate logistic regression analysis revealed sarcopenia (odds ratio [OR]: 0.26; 95% confidence interval[CI]: 0.08–0.87; p= 0.029), pre-treatment IRC (<35 g/L) (OR: 0.28; 95% CI: 0.09–0.94; p= 0.040) and Nac-VI-RADS ([OR]: 0.26; 95%[CI]: 0.08–0.87; p= 0.029) as negative risk factors for achieving CR. The CR was significantly higher in NSG compared with the SG (p < 0.001).
Conclusion: Sarcopenia, assessed by an AI-powered software, were negatively associated with tumor response following NAC in MIBC patients. Our findings, supported by further evidence, suggest to integrate sarcopenia AI-analysis with radiological staging in order guide patients toward personalized nutritional therapies that potentially could improve outcomes.
Limitations: Small sample size
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: None
6 min
Utilization of Machine Learning on preoperative CT for the prediction of oesophagus cancer recurrence - A pilot study
Emmanouil Koltsakis, Stockholm / Sweden
Author Block: E. Koltsakis1, M. E. Klontzas2, I. Rouvelas1, A. Tzortzakakis1; 1Stockholm/SE, 2Herakleion/GR
Purpose: Oesophagus cancer is the 6th leading cause of cancer worldwide and the number of cases is predicted to increase in the near future due to lifestyle changes. Although total resection following neoadjuvant chemo/radiotherapy can have promising results, five-year survival ranges between 6-49% according to the American Cancer Society. With half of the patients readmitting with distant metastasis within 5 years an accurate prediction model based on preoperative and preneoadjuvant imaging can strongly affect the treatment decision which could consecutively decrease mortality. Thus, the purpose of this pilot study is to predict metastasising or recurrence of oesophagus tumours based on the preoperative CT scan.
Methods or Background: A total of 30 patients with oesophageal cancer who underwent a pretreatment CT scan were included in this study. The tumours were segmented using 3D Slicer on thin slices on venous phase. Non-robust features were excluded with Boruta analysis, and the rest were used on an XGBoost model.
Results or Findings: The XGBoost model showed an AUC on 77.5% (42.1%-100.0%) with a sensitivity at 0.75 and a specificity at 0.80 on predicting recurrence on the 30-case series.
Conclusion: As recurrence is one of the main concerns on oesophageal cancer follow up, clinical practise can benefit by additional tools which can identify such high-risk cases. Our model showed promising results on the identification of such cases.
Limitations: The main limitation of the study is the small sample size. Furthermore, this is a single centre study without an external validation set. Lastly, addition of other factors i.e. clinical data may improve the prediction model.
Funding for this study: Medicinsk Diagnostik Karolinska Kliniknära FoU 2024 Projektnr: 929511, Kostnadsställe: 25013
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Regional Ethical Review Authority in Stockholm (Regionalaetikprövningsnämnden i Stockholm)
6 min
CT-based radiomics model in differentiating benign ovarian masses from early-stage ovarian carcinoma
Anna Preda, Pavia / Italy
Author Block: A. Preda1, J. Zhang2, E. Man Fung Wong2, L. Han3, G. Ho2, W. H. K. Chiu2, A. Kwun To Leung2, R. Singh2, E. Y. P. Lee2; 1Pavia/IT, 2Hong Kong/HK, 3Guangzhou/CN
Purpose: Incidental adnexal masses are common with the exponential growth of CT utilization. Differentiating benign ovarian masses (BOM) from early-stage ovarian carcinoma (ES-OC) is challenging as the imaging features overlap resulting in diagnostic dilemma. This multicentre study aimed to build a CT-based radiomic model to differentiate BOM from ES-OC.
Methods or Background: In this retrospective study, 529 lesions were detected in 483 patients with their CT collected from six centres, all were histologically confirmed. The cohort was divided into training and test sets in a stratified ratio of 80:20. Radiomic features were extracted from the primary adnexal mass using PyRadiomics after image normalization. Highly correlated features (Spearman correlation coefficient >0.85) were removed. Feature selection was performed using elastic-net regression with 5-fold stratified cross-validation (SCV), repeated across 100 iterations. Features retained for more than 475 times were considered for subsequent analyses. Mann-Whitney U-test was applied to identify statistically significant radiomic features. Radscores were calculated for both groups to evaluate the discriminative capability of 10 selected features. Support vector machine optimized using hyper-parameter tuning and validated through 10-fold SCV was employed to construct the prediction model.
Results or Findings: 291 patients (40±15 years old) presented with BOM (n=337 lesions) and 192 patients (49±7 years old) had ES-OC (FIGO stage I-II, n=192). The proposed radiomics model achieved a mean area under the curve (AUC) of 0.95, sensitivity of 0.84, specificity of 0.91 and accuracy of 0.89 in distinguishing BOM from ES-OC in the test set.
Conclusion: CT-based radiomic analysis could serve as a valuable, non-invasive tool for improving diagnostic accuracy of adnexal masses, reducing further examination for characterisation.
Limitations: The limitations of the study are the small cohort, that may result in model overfitting, and selection bias given the retrospective design.
Funding for this study: Funding was provided by the Hong Kong Medical Research Fund 08192016.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This retrospective study was approved by local review boards (UW20-251, UW24-148, YB2018-52, HKECREC-2020-040).
6 min
Early clinical evaluation of an artificial intelligence (AI) decision-supporting tool to identify high confidence unremarkable chest radiographs in a single NHS Trust
William Pettit, London / United Kingdom
Author Block: W. Pettit1, M. Ryan2, A. Raginis-Zborowska2, E. Compton2, A. Kumar1; 1Berkshire and Surrey/UK, 2Sydney/AU
Purpose: The purpose of this retrospective evaluation was to evaluate the performance and generalisability of an artificial intelligence (AI) tool (Annalise Enterprise v3.8) for identification of “clinically remarkable” (study contained findings indicative of urgent suspected lung cancer and/or clinically acute findings) chest radiographs in a clinical setting prior to clinical use. The project is a part of the national AI Diagnostic Fund where it was deployed on a regional level as part of 5 hospitals
Methods or Background: A cohort of General Practice and Outpatient referred frontal chest radiographs (CXRs) collected from a National Health Service Trust was collected retrospectively from examinations conducted during November 2022 - January 2023. The reference index was established by consensus between at least two radiologists.

The analysis included basic performance metrics: sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV).
Results or Findings: 538 patients were included alongside patient demographic data on age, sex and ethnicity to confirm generalisability. The AI device correctly identified 40/64 (62.5%) of the remarkable CXRs as remarkable and 409/474 (85.74%) of the unremarkable CXRs as unremarkable. The sensitivity, specificity, Positive Predictive Value and Negative Predictive Value were 62.5%, 86.3%, 38.10%, and 94.5% respectively.
Conclusion: The results demonstrate the ability of an AI device to prioritise studies for clinical review based on AI-determined presence of findings. The highly configurable design of the AI device allows the customisation of findings that warrant the classification of urgent and tuning of thresholds for the subject population. The device performance will be monitored ongoing through a limited user roll-out to confirm usability and governance, before proceeding to rollout across the entire NHS Trust, transitioning the system into prospective real-world operations.
Limitations: Population size due to time and workforce capacity constraints
Funding for this study: Funding from UK Department of Health AI diagnostic fund
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approved by the local AI governance committee at the trust.