EIBIR 4b - EIBIR Stage bonus session 4

Author Block: H. Rogerson-Bevan, S. M. Simpson, C. Higham, F. Frost, L. Berger, F. J. Wong, C. Barker; Manchester/UK
Purpose: Vertebral fragility fractures (VFFs), indicative of underlying osteoporosis, are common in cancer patients, attributed to the aetiology of the disease and treatment regimes. Frequently mis- and under-diagnosed, VFFs are linked to higher mortality, morbidity, and increased hip fracture risk. We assessed the prevalence and clinical context of VFF detected opportunistically on chest X-rays (CXRs) performed for unrelated indications in cancer patients.
Methods or Background: A retrospective review was conducted of adult cancer patients who underwent CXR within a large, tertiary cancer centre between October and December 2024. Patients with VFFs were identified utilising an AI tool (Annalise Container v2.2, Annalise.ai) on CXR. Indications for imaging, fracture status (new vs. pre-existing), and suspected aetiology were assessed.
Results or Findings: 173 patients with a VFF were identified on CXR, 36/173 (21 male:15 female) representing newly detected vertebral fractures; the remaining 137 (79%) had a pre-existing fracture. All patients underwent CXR for infection screening, baseline pre-treatment evaluation, or post-treatment assessment. Among patients with newly detected fractures, 77% (28/36) were classified as osteoporosis-related based on radiological appearance and clinical context. None of the patients with a newly identified fracture had clinical suspicion or prior imaging for vertebral fracture at the time of CXR.
Conclusion: Opportunistic assessment augmented by AI of CXR revealed a substantial number of newly identified vertebral fractures, most of which were not pathological, but are fragility fractures related to cancer treatment. While early identification of such fractures is an important initial step, meaningful improvement in patient outcomes requires the establishment of robust management strategies and dedicated follow-up pathways to ensure timely and effective intervention.
Limitations: Results reflect findings from one large cancer centre, which may limit generalisability; the study focused on fracture detection and clinical impact was not established.
Funding for this study: Deployment of the AI tool was funded by the NHS AIDF Fund
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: A. Rabasco Meneghetti¹, A. Marcos Morales², L. Riobó Mayo², O. Balacescu³, N. Antone³, J. Calderaro⁴, R. Perez Lopez², V. Moreno Aguado², J. Kather¹; ¹Dresden/DE, ²Barcelona/ES, ³Cluj-Napoca/RO, ⁴Créteil/FR
Purpose: Cancer immunotherapy with immune checkpoint inhibitors (ICIs) is widely used in lung cancer, with proven benefits. However, response is not guaranteed, difficult to predict, and serious toxicity may occur. The TANGERINE study (funded through the EU Joint Call 2021 TRANSCAN-3) aims to develop artificial intelligence (AI)-based histology and radiology-based models for predicting immune features related to ICIs response. Here we present current results for radiological data using the subset of lung cancer patients.
Methods or Background: Pre-ICI treatment computerised tomography (CT) scans from lung cancer patients from 3 hospitals from Spain and Romania were retrospectively identified and included for a lung-multidrug model. Patients had received ICls alongside prior or concomitant chemotherapies. Best overall response was obtained according to RECIST 1.1 criteria. Only CTs closest to the treatment start date were included. Whole-CT embeddings were generated through the MERLIN foundation model (Blankemeier et al. 2024). An attention-based multiple-instance learning (ABMIL) model was then trained to predict disease control rate (DCR) and PFS through 5-fold cross-validation (CV) in 80% of the patients (n=141) and deployed in a holdout test set (n=40).
Results or Findings: 181 patients with lung cancer were included (72% male, Treatments: 53% Pembrolizumab, 17% Nivolumab, 13% Durvalumab, 4% Atezolizumab, 13% others, DCR: 42%, median time to progression: 5.5 months). The lung-multidrug model for DCR classification showed an average CV AUC of 0.75 95% CI (0.57-0.79) in the training set and 0.60 (0.50-0.69) in the test set. Models significantly stratified patients into high and low-risk for PFS (p=0.023) with cause-specific hazard ratio:1.49 (1.05-2.10).
Conclusion: Pre-treatment CT scans show predictive value for ICI outcomes in lung cancer, specifically for PFS prognosis.
Limitations: Sample size was limited. Further study iterations will include more patients, and fully-external validation cohorts
Funding for this study: This study was financed through the EU Joint Call 2021 TRANSCAN-3 grant.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Protocol and informed consent for the TANGERINE project were approved by the IDIBELL´s Research Ethics Committee (coordinating centre for the TANGERINE study).

Author Block: S. Fouladi, F. Darvizeh, R. Di Meo, I. Bossi Zanetti, G. Gianini, E. Damiani, A. Maiocchi, D. Fazzini, M. Alì; Milan/IT
Purpose: Prostate cancer (PCa) is the second most commonly diagnosed cancer in men, with an estimated 288,300 new cases and over 34,700 deaths annually in the United States. Early detection and accurate lesion localization are crucial for improving outcomes; however, manual segmentation of multiparametric MRI (mpMRI), including T2-weighted (T2W), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) sequences, is labor-intensive and prone to interobserver variability. This challenge has motivated the development of automated deep learning solutions.
Methods or Background: We evaluated two datasets: PI-RADS 4–5 (220 training, 33 test) and PI-RADS 3–5 (270 training, 41 test). In the first stage, U-Net, Dense U-Net, and Attention U-Net were trained separately on T2W, DWI, and ADC to benchmark the contribution of each sequence. In the second stage, we implemented a multi-input U-Net with three parallel encoders, each dedicated to one sequence (T2W, DWI, ADC), enabling joint learning while preserving modality-specific features. Finally, building on the strong performance of ADC, we proposed a novel LSTM U-Net with a Bi-ConvLSTM bottleneck to capture temporal dependencies and improve lesion boundary delineation.
Results or Findings: ADC achieved the highest Dice scores (69% for PI-RADS 4–5 and 68% for PI-RADS 3, 4, and 5). The LSTM U-Net on ADC provided competitive accuracy and improved delineation of challenging lesions, highlighting the benefit of temporal modeling.
Conclusion: Segmentation depends on dataset composition and network design. Multi-input sequences improve accuracy, while temporal modeling refines lesion boundaries, supporting AI-assisted prostate cancer diagnosis.
Limitations: The number of images was limited due to the time-consuming process of manual mask creation. Despite this constraint, the results are promising, and performance is expected to further improve with the inclusion of larger datasets.
Funding for this study: Funding The work was partially supported by the MUSA-Multilayered Urban
Sustainability Action project, funded by the European Union-NextGenerationEU,
under the Mission 4 Component 2 Investment Line of the National Recovery and
Resilience Plan (NRRP) Mission 4 Component 2 Investment Line 1.5: Strengthening
of research structures and creation of R&D ”innovation ecosystems”, set up
of ”territorial leaders in R&D” (CUP G43C22001370007, Code ECS00000037);
Program ”piano sostegno alla ricerca” PSR and the PSR-GSA-Linea 6; Project
ReGAInS (code 2023-NAZ-0207/DIP-ECC-DISCO-23), funded by the Italian
University and Research Ministry, within the Excellence Departments program
2023-2027 (law 232/2016).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approval on September 11, 2024 by CET Lombardia 3 Ethical Committee (Study ID: 5105)

Author Block: V. Chernina¹, V. Gombolevskiy², A. Meldo³, E. Petrash², M. Valkov¹; ¹Arkhangelsk/RU, ²Moscow/RU, ³St. Peterburg/RU
Purpose: Define the prognostic value of AI-derived CT biomarkers of COVID-19 pneumonia and assess the impact of vaccination on survival in oncology patients.
Methods or Background: Adults with cancer (April 2020–December 2021) underwent chest CT, linked with survival and vaccination data. AI extracted COVID-19 pneumonia and cardiopulmonary CT biomarkers. Endpoints: overall and cancer-specific survival; multivariable Cox regression adjusted for demographics, stage, AI detection, and vaccination status.
Results or Findings: The cohort included 1148 patients (66.2 ± 12.5 years; 52% female) (Fig. 1); 6.6% were vaccinated. AI detected COVID-19 pneumonia in 27.2%, emphysema in 36.4%, aorta aneurysm in 1.6%, main pulmonary artery enlargement in 13.6%, CAC ≥1 in 24.7%, and epicardial fat ≥125 mL in 17.8% (Fig. 2). Median follow-up was 38.4 months.
OS at 1/3/5 years was 70.2%, 48.3%, and 41.7%; CSS was 74.2%, 49.0%, and 47.8%, respectively. Patients without vs. with AI-detected pneumonia had 1-year OS of 80.8% vs. 65.9% and 3-year OS of 55.2% vs. 36.4% (p < 0.001); CSS was 84.1% vs. 63.2% (p < 0.001) (Fig. 3). Fig. 4 illustrates AI output.
Multivariable Cox analysis identified independent mortality predictors: COVID-19 pneumonia (HR 1.31; 95% CI 1.09–1.58; p=0.004), stage III–IV, male sex, and pulmonary/GI tumor site. Vaccination reduced mortality by 61% (HR 0.39; 95% CI 0.24–0.64; p<0.001), raising 1-/3-year OS to 99% and 80% vs. 68% and 45% in unvaccinated (Fig. 5).
Conclusion: AI-detected COVID-19 pneumonia was the sole independent CT biomarker, increasing mortality risk by 31% (HR 1.31) across all subgroups. Vaccination emerged as the only modifiable factor, reducing mortality by 61% (HR 0.39) in oncology patients.
Limitations: Single-region retrospective design limits causal inference. Excluding >70% of CTs and protocol variability may bias selection, but strict filtering and a certified, heterogeneity-trained AI improve robustness.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Permission to conduct the study was obtained from the Local Ethics Committee of Northern State Medical University (No. 07/10-238, 2023).

Author Block: M. Jaisankar¹, A. Ostreika¹, B. García-Zapirain Soto²; ¹Kaunas/LT, ²Bilbao/ES
Purpose: This systematic review synthesizes literature on attention mechanisms in Self-Explainable AI (SXAI) for the interpretable analysis of cancer using multimodal data. The primary objective is to evaluate how attention mechanisms enhance both model performance and explainability for clinical tasks (diagnosis, prognosis, treatment prediction), thereby improving clinical trust and facilitating adoption.
Methods or Background: Following PRISMA guidelines, the review analyzes studies using AI to integrate heterogeneous data, including histopathology whole-slide images (WSIs), MRI, and multi-omics. Combining these modalities offers complementary insights, but the "black box" nature of deep learning limits clinical utility. Attention mechanisms, core to architectures like Transformers, are widely explored to create inherently transparent SXAI models by enabling them to focus on and highlight the most relevant data features used in model's predictions.
Results or Findings: The literature consistently demonstrates that multimodal models significantly outperform unimodal approaches in cancer survival prediction and risk stratification, with some showing improvements of over 4.5% in the Concordance Index (C-index). These AI models accurately predict critical molecular features, such as TP53 mutations, directly from standard medical images. Attention mechanisms are central to this success, serving as a powerful tool for effective data fusion and model interpretability. Attention-based visualizations reveal biologically relevant morpho-molecular correlates, like identifying areas with high densities of tumor-infiltrating lymphocytes (TILs).
Conclusion: Attention-based SXAI models represent a robust and effective strategy for multimodal cancer analysis. By providing transparent, biologically relevant insights into their decision-making, these models enhance predictive accuracy, foster clinical trust, and serve as a promising pathway toward developing rapid, cost-effective biomarkers that can advance personalized oncology and improve patient outcomes.
Limitations: Common limitations identified include the retrospective nature of studies and heavy reliance on public datasets, which may lack demographic diversity and completeness.
Funding for this study: Erasmus+, Lithuanian State fund for PhD.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study has been approved by Ethics committee.