EIBIR 2 - EIBIR stage bonus session 2

Author Block: L. Steinhelfer, F. Jungmann, F. Lohöfer, R. Braren; Munich/DE
Purpose: 177Lu-DOTATATE therapy is an effective treatment for advanced neuroendocrine tumors, despite its dose-limiting hematotoxicity. Herein, the significance of off-target splenic irradiation is unknown. Our study aims to identify predictive markers of peptide receptor radionuclide therapy–induced leukopenia.
Methods or Background: We retrospectively analyzed blood counts and imaging data of 88 patients with histologically confirmed, unresectable
metastatic neuroendocrine tumors who received 177Lu-DOTATATE treatment at our institution fromFebruary 2009 to July 2021. Inclusion criterium was a tumor uptake equivalent to or greater than that in the liver on baseline receptor imaging.We excluded patients with less than 24mo of follow-up and those patients who received fewer than 4 treatment cycles, additional therapies, or blood transfusions during follow-up.
Results or Findings: Our study revealed absolute and relative white blood cell counts and relative spleen volume reduction as independent predictors of
radiation-induced leukopenia at 24mo. However, a 30% decline in spleen volume 12mo after treatment most accurately predicted patients proceeding to leukopenia at 24mo (receiver operating characteristic area under the curve of 0.91, sensitivity of 0.93, and specificity of 0.90), outperforming all other parameters by far.
Conclusion: Automated splenic volume assessments demonstrated superior predictive capabilities forthe development of leukopenia in patients undergoing 177Lu-DOTATATE treatment compared with conventional laboratory parameters. The reduction in spleen size proves to be a valuable, routinely available, and quantitative imaging-based biomarker for predicting radiation-induced leukopenia. This suggests potential clinical applications for risk assessment andmanagement.
Limitations: The retrospective design of this study introduces the potential for bias from unidentified confounding factors, emphasizing the
need for further investigations using different datasets to validate these findings. To avoid investigator bias, blood count and splenic
volumetry were done independently.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: reference 87/18S

Author Block: A. Olesinski, R. Lederman, J. Sosna, L. Joskowicz; Jerusalem/IL
Purpose: Evaluate the performance of a novel method for accurate detection and measurement of mediastinal lymph nodes (LNs) in chest CT by annotation-efficient deep learning.
Methods or Background: The method consists of a 3D nnU-Net voxel classification model for LN detection. It was initially trained on a few annotated scans and used to generate LN pseudolabels for a larger dataset of unannotated scans. The LN pseudolabels were filtered with automatic segmentations of 20 non-LN mediastinal structures. An enhanced 3D nnU-Net model was then trained with the labeled and the selected pseudolabels. LNs with axial short axis >10mm were identified as enlarged.
We obtained 298 chest CECT studies of patients with suspected mediastinal lymphadenopathy from three sources: 108 from our hospital, 100 from the Lnq2023 Challenge dataset, and 90 from an NIH dataset. They included 2,014 annotated LNs, 1,078 normal (short axis of 5-10mm) and 836 enlarged (> 10mm). The training/test set partition was 134/164 scans (1,069/945 LNs). An additional 710 unannotated scans were collected from our hospital for pseudolabel generation. The ratio of annotated/unnanoted scans used for training was 1 to 4.3. Detected LNs and their computed short axes were compared to the annotated ground truth.
Results or Findings: The enhanced 3D nnU-Net yielded a detection precision and recall (std) of 0.85(0.26) and 0.89(0.24) for LNs>10mm and 0.73(0.29) and 0.72(0.28) for LNs 5-10mm. It significantly (p<0.01) improved the recall(std) by 6(14)% and 15(30)% with respect to the initial model, with similar precision. The mean short axis difference(std) was 4.6(4.7)mm and 2.3(3.2)mm, respectively.
Conclusion: Label-efficient automatic detection and measurement of mediastinum lymph nodes in chest CECT yields acurrate results and may help in the evaluation of patients with enlarged mediastinal LNs.
Limitations: Unknown quality of public datasets' annotations.
Funding for this study: None.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Helsinki Committee, Hadassah University Medical Center

Author Block: A. Venturi¹, M. Interlenghi¹, C. Salvatore², D. Fazzini¹, G. Di Leo³, F. Sardanelli¹, I. Castiglioni¹; ¹Milan/IT, ²Pavia/IT, ³San Donato Milanese/IT
Purpose: We present a clinical validation of a segmentation foundation model guided by a bounding box and fine-tuned on ultrasound images of breast masses for their segmentation and classification.
Methods or Background: We retrospectively collected ultrasound images of 236 breast masses considered suspicious by four board certified radiologists (BI-RADS≥3) from 235 patients at IRCCS Policlinico San Donato (Italy), 105 malignant and 131 benign according to histopathology.
For each mass, we tested the segmentation accuracy and reliability of the Segment-Anything-Model (SAM) with respect to an expert radiologist (>34y of experience) when ten distinct rectangular bounding boxes were chosen to guide the segmentation of the included breast mass. Such boxes were automatically generated from the segmentation manually performed by the expert radiologist with a random enlargement and/or translation of the region up to 20% of its original size and position.
We then assessed the classification performance of a breast mass radiomic analyzer (TRACE4BUS(™), DeepTrace Technology Srl) in discriminating malignant versus benign breast masses in the ten tests, with respect to expert radiologists.
Results or Findings: DSC and IoU (%, mean±std) of the foundation model segmentations, compared with those of expert radiologist, were 89.1±5.9 and 80.8±9.1, respectively (mean over 10 regions).
The breast mass analyzer achieved sensitivity (%) and specificity (%) of 96.3 and 16.9 respectively on suspicious masses, when considering the radiologist’s segmentations; 95.83 [95.19-96.48] and 15.23 [13.77-16.69], respectively (with 95%CI) when using the foundation model, with a good agreement in timing and accuracy (8 vs 7.3 seconds; Cohen's K: 0.57 [0.54-0.59]).
Conclusion: The SAM model can offer a reliable tool for breast mass automatic segmentation and classification in breast ultrasound, with performance similar to those obtained by expert radiologist manual segmentation.
Limitations: Limited size cohort of patients from single center.
Funding for this study: The authors acknowledge support for this research from the European Union’s HORIZON 2020 research and innovation program (CHAIMELEON project, grant agreement #952172).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Ethics Committee of IRCCS Ospedale San Raffaele (protocol code “SenoRetro”, first approved on 9 November 2017, then amended on 18 July 2019, and on 12 May 2021)

Author Block: A-M. Mascarel, J. Paul, D. Geffroy, C. Labbe-Devilliers, R. Movassaghi, M. Neveu, L. Vilcot, C. Morisseau, I. Doutriaux-Dumoulin; Saint Herblain/FR
Purpose: Artificial intelligence (AI) is increasingly used in mammography. When breast cancer is cared in an expert centre, a review of the initial diagnostic work-up, additional tomosynthesis images, ultrasound, and possibly contrast imaging and biopsies are performed. The performance of AI in the pre-treatment phase has not been evaluated.
Aim : To evaluate the concordance between the results of AI and those of the expert radiologist for the detection of malignant lesions on an additional tomosynthesis work-up in the pre-treatment situation for breast cancer.
Methods or Background: Patients referred for assessment of cancerous breast lesions between November 2023 and February 2024 were included. Bilateral craniocaudal and mediolateral oblique tomosynthesis performed during the pre-therapy work-up were analysed retrospectively using Transpara AI software (V1.7.1 ScreenPoint Medical). These results were compared with a review made by an expert radiologist who had no access to the clinical data or the AI results. The gold standard was the complete pre-therapeutic evaluation (mammography, ultrasound, contrast imaging and biopsies). Agreement between AI results and those of the expert radiologist was assessed using Cohen's Kappa coefficient.
Results or Findings: We included 166 patients with 104 benign and 202 malignant lesions. 178/202 malignant lesions were concordant with 160 visualised and 18 not visualised by AI and radiologist (only seen on ultrasound or contrast imaging). The radiologist detected 21 cancers not identified by the AI, while 3 cancers were detected exclusively by the AI. Kappa coefficient = 0.537.
Conclusion: Our results showed moderate agreement between the AI and the expert radiologist. AI contribution was moderate with only 3 cancers identified exclusively by the AI and 21 false negatives. This also led to a significant number of false positives that would need to be negated by expert radiologists.
Limitations: None
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The ethics committee notification can be found under number MR-004.

Author Block: D. A. Motoc; Oradea/RO
Purpose: The aim of this study was to develop an efficient data augmentation pipeline to address the challenges of small medical imaging datasets, specifically for brain tumour segmentation. By focusing on coding efficiency and optimisation, the objective was to create a scalable process capable of generating diverse training samples, enhancing future model development and diagnostic performance.
Methods or Background: This retrospective study used 500 training images and 500 corresponding segmentation masks from the Decathlon brain tumour dataset. The dataset consisted of 4D MRI volumes of patients with confirmed brain tumours. The pipeline was developed in Python, applying both 3D and 2D augmentations. NIfTI images were loaded with NiBabel, converted into PyTorch tensors, and processed. The 3D augmentations included affine transformations, intensity rescaling, and random flipping, using TorchIO. For 2D data, slices were extracted and further augmented with rotations, cropping, and intensity normalisation. A custom visualisation function (plot_tensor) and NiLearn were used to verify anatomical accuracy.
Results or Findings: The pipeline generated diverse variations of 4D MRI volumes by acquiring and augmenting 3D, 2D slices. A key focus was the alignment between augmented images and segmentation masks, which remained accurate throughout transformations. Visualisation confirmed that the dataset maintained clinical relevance while enhancing variability, crucial for deep learning model generalisation. Diagnostic accuracy metrics, such as specificity or sensitivity, were not calculated but will be evaluated in future model comparisons.
Conclusion: The pipeline efficiently enhances small medical imaging datasets, creating diverse and clinically relevant data for future deep learning models.
Limitations: The limitations of the study are that, due to limited computational resources, no model training or benchmarking was conducted. Future studies will focus on training models using data augmentation through the provided pipeline framework, which is designed for building robust Dataloaders.
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 study is retrospective and educational.

Author Block: F. J. Pérez García, P. Montosa Rodenas, J. P. Martinez Barbero, D. López Cornejo, J. M. Benítez, A. J. Láinez Ramos-Bossini; Granada/ES
Purpose: The aim of this study was to develop a machine learning model to differentiate between radionecrosis and tumor progression in glioma patients using MRI perfusion parameters and imaging features.
Methods or Background: We conducted a retrospective observational study at a tertiary neuro-traumatological hospital. Patients with previously treated brain glioma on follow-up and confirmed radionecrosis or tumor progression were included. T2*-Dynamic Susceptibility Contrast (DSC) perfusion images were used to manually segment the suspicious areas. Radiomics features were extracted using PyRadiomics, yielding 829 features. These features, along with intensity-time curve (TIC) characteristics, including relative cerebral blood volume (rCBV) and percentage signal recovery (PSR), were integrated into a single dataset. A machine learning model was built using PyCaret, employing GroupKFold cross-validation to prevent data leakage. A total of 15 models were evaluated, and performance was assessed through ROC analysis and confusion matrices.
Results or Findings: A total of 46 patients (mean age, 54.5 years; 51% women; 50% radionecrosis) were included in the study. The dataset was divided into a training and a test set of 40 and 6 patients, respectively. The area under the curve (AUC) of the best model, ExtraTreesClassifier, was 0.94. The model achieved an accuracy of 72.8% for radionecrosis and 92.2% for progression. Additionally, the model demonstrated a sensitivity of 91.7% for radionecrosis and 74.4% for progression, along with F1 scores of 0.812 for radionecrosis and 0.824 for progression.
Conclusion: This study demonstrated the potential of a combined radiomics and perfusion-based machine learning approach in accurately differentiating radionecrosis from tumor progression in glioma patients.
Limitations: The study was limited by the small sample size of 46 patients, which may affect the generalizability of the results. Future studies with larger cohorts are needed to validate our findings.
Funding for this study: This study was funded by MICIN/AEI/10.13039/501100011033 (project reference) PID2020-118224RB-I00
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: IANeuro24)

Author Block: M. Wodrich, F. Sahlin, J. Karlsson, I. Arvidsson, K. Lang; Lund/SE
Purpose: Breast cancer diagnosis using high-end ultrasound is costly in terms of time and machines, and requires highly trained breast radiologists. The use of point-of-care ultrasound (POCUS) combined with artificial intelligence (AI) could be a cost-effective solution for limited-resource settings.
Methods or Background: A multi-reader multi-case study involving 40 women of which 70 POCUS and 70 case-matched standard breast ultrasound (BUS) images (11 malignant, 21 benign, and 38 normal cases each) were acquired. Four breast radiologists read each set of ultrasound images with a washout period of at least 4 weeks in between. The reader task was to score the images on a 5-level scale (≥ 3 was considered positive). The images were also analysed by an in-house developed deep learning based AI algorithm. The AI includes an uncertainty quantification method that can measure predictive uncertainties to flag unsuitable images. The performance of radiologists and AI, respectively, on POCUS and BUS were assessed using receiver operating characteristic (ROC) curve, the area under the ROC curve (AUC), sensitivity, and specificity.
Results or Findings: For radiologists reading POCUS/BUS, the sensitivity was 100.0%/100.0%, and the specificity was 76.7%/78.0% respectively. AI performance on POCUS/BUS was sensitivity 90.9%/100.0%, specificity 88.1%/86.4%. AI performance when excluding cases where it had high uncertainty about the prediction for POCUS/BUS was sensitivity 83.3%/100.0%, specificity 94.3%/94.3%.
Conclusion: Radiologists performance on POCUS and BUS was similar. The performance of AI on BUS was slightly better than on POCUS, but both showing higher specificity than average breast radiologists. Our results indicate the potential of using AI to automatically analyze POCUS breast images in limited-resource settings.
Limitations: Small sample size. While the lesions are case-matched, we cannot guarantee that the images are of the same angle and quality.
Funding for this study: This work was supported by strategic research area eSSENCE and Analytic Imaging Diagnostics Arena (AIDA).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approval was granted by the Swedish Ethical Review Authority of Region Skåne (2019-04607).

Author Block: R. J. Jóźwiak¹, M. Gonet¹, J. Mycka¹, I. Mykhalevych¹, P. Sobecki¹, T. Lorenc¹, A. Zacharzewska-Gondek², J. Dołowy², K. Tupikowski²; ¹Warsaw/PL, ²Wrocław/PL
Purpose: To serve as a user-friendly educational platform tailored for radiologists who want to better understand the idea of structured reporting and AI support.
Methods or Background: In radiology education, acquiring new knowledge and skills, as well as understanding standards and the potential and limitations of new technologies like AI, is crucial. The European Society of Radiology (ESR) highlights the importance of structured reporting to enhance service quality for patients and physicians. ESR also identifies opportunities for integrating AI tools with radiology structured reporting, which could lead to improved synergy and natural integration.
Results or Findings: AI4AR is an upcoming platform for radiology education that combines modern forms of structured reporting, AI models, and rich sets of reference cases with multi-reader annotations. The platform allows users to track their learning progress and compare the reporting agreement. The course can be completed with and without AI support, allowing the assessment of the impact of AI on the reporting process. By integrating with a DICOM browser, the platform offers a seamless interaction with AI model suggestions, allowing users to verify indications by rejecting, correcting, or accepting them. This approach respects the human primacy over AI algorithms and adheres to ethical guidelines. The available module dedicated to prostate cancer diagnosis is standardized according to the PI-RADS guidelines. Developed AI models support the analysis of mpMRI by segmenting prostate gland anatomy and identifying potentially suspicious areas.
Conclusion: AI4AR gives the unique opportunity to learn by interpreting real cases using standardized forms of structured reporting with AI augmentation. The conscious use of AI supports the idea of augmented radiology - using AI technology to improve human decision-making, providing added value to patients and the radiological society.
Limitations: Currently, only the prostate module is available.
Funding for this study: This work has been funded by the Polish National Centre for Research and Development under the program INFOSTRATEG I, project INFOSTRATEG-I/0036/2021 “AI-augmented radiology - detection, reporting and clinical decision making in prostate cancer diagnosis”.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Our study uses a single-center retrospective database collected in Lower-Silesian Center of Oncology,
Pulmonary and Hematology (DCOHiP) in Wroclaw, collected as part of an ongoing research project
INFOSTRATEG-I/0036/2021 AI-augmented radiology - detection, reporting and clinical decision making
in prostate cancer diagnosis (AI4AR), for which ethical approval with the need for informed consent for
data analysis was obtained from the Bioethics Committee of the Hirszfeld Institute of Immunology and
Experimental Therapy (KB – 4/2022).

Author Block: I. Capaldo¹, M. Signorini²; ¹Salerno/IT, ²Rovigo/IT
Purpose: Today, many patients search for healthcare information online also through Artificial Intelligence (AI). This study aimed to evaluate the ability of three AI platforms to answer frequently asked breast cancer-related questions.
Methods or Background: Three AI platforms (ChatGPT 4.0, Gemini and Bing), were asked to answer 8 questions related to breast radiology. Responses were assessed for accuracy and clarity by a breast radiologist resident with 2 years of experience and a board certified breast radiologist with 10 years of experience using 5-point Scale. A one-way ANOVA test was performed to determine any statistical difference.
Results or Findings: While there were no statistically significant differences in the overall accuracy of the three AI models, two responses from Bing were deemed insufficient. ChatGPT's responses were significantly longer than those of Gemini and Bing (mean word count: 264.75 vs. 130.375 and 94.5 respectively; p<0.05). Although ChatGPT's clarity was good (mean score 4.125), it was significantly lower (p<0.05) compared to Gemini (4.75).
Conclusion: The discrepancies in AI responses pose a risk of misleading patients seeking information about breast cancer. Scientific societies, such as EUSOBI, should collaborate with AI managers to review the sources used to generate answers and ensure that patients receive unambiguous and reliable information.
Limitations: None
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: None

Author Block: M. Spector¹, N. Bogot¹, V. Miskovic², E. Oberstien³, A. Prelaj², S. G. Armato⁴, N. Peled³, O. Benjaminov¹, L. C. Roisman³; ¹Israel/IL, ²Milano/IT, ³Jerusalem/IL, ⁴Chicago, IL/US
Purpose: Lung cancer is a global health challenge. Radiomics, extracts quantitative features from medical images, shows promise in optimizing tumor assessment, treatment planning and outcomes. Implementing radiomics challenges due to varied acquisition protocols, image parameters, and tumor characteristics. This study focuses on overcoming segmentation challenges in an international multicenter setting.
Methods or Background: This retrospective multicenter study enrolled patients with confirmed NSCLC diagnosed between 2012 and 2022. Chest CTs were acquired according to local standard-of-care protocols, collected, and pseudo-anonymized in DICOM format. After quality assessment, images underwent harmonization to address inter-institutional variability in acquisition parameters, and slice thickness and reconstruction kernels were standardized. Tumor segmentation focused on delineating regions of interest using automated and semi-automated methods, with manual corrections as needed.
Results or Findings: From 1,492 patients initially enrolled, 828 patients had chest CTs. One hundred (12%) post-surgical images or images without pulmonary lesions were excluded. Segmentation challenges were categorized into technical issues (77 patients, 9.3%) – including image dimensions or image acquisition - and tumor-related factors (84 patients, 10%), including complex anatomy, diffuse disease, or poorly defined borders. AI-based lung segmentation with ROI-based fast matching, followed by manual correction, was applied to 31 patients (3.7%). Ultimately, 597 patients (72%) were successfully segmented for radiomics analysis.
Conclusion: This study highlights the challenges in segmentation to developing a robust radiomic model for NSCLC in multicenter settings. By addressing the limitations of fully automated methods, segmentation accuracy improved, increasing the number of successfully segmented cases for radiomics analysis.
Limitations: The study's limitations include variability in acquisition protocols, potential selection bias from excluded cases, and manual segmentation corrections. Challenges with poorly defined tumors limit generalizability to advanced NSCLC, while incomplete follow-up data may impact the correlation between radiomic features and patient outcomes.
Funding for this study: Horizon 2020 Framework Program (EU Framework Program for Research and Innovation H2020), 101057695
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was conducted in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments. Ethical approval for the study was granted by the Ethics Committee of Shaare Zedek Medical Center (Approval number: 0240-22-SZMC).