Citations

Abstract Background aneurysmal subarachnoid hemorrhage (aSAH) is neurological emergency associated with substantial mortality and disability. Current grading systems such as the modified Fisher Scale (mFS) and World Federation of Neurological Societies (WFNS) score, rely on semiquantitative and examination based assessments. Hence, they demonstrate limited predictive precision. The enhanced subarachnoid hemorrhage (eSAH) score is a simplified quantitative model integrating age, Glasgow Coma Scale (GCS), and cisternal subarachnoid hemorrhage volume (SAHV) to predict clinical outcomes after aSAH. Methods We performed a retrospective multicenter cohort study that included 1088 patients across three tertiary-care centers the United States. Predictive performance for unfavorable functional outcome, in-hospital mortality and delayed cerebral ischemia (DCI) was evaluated using receiver operating characteristic (ROC) analysis and area under the curve (AUC). Comparative analyses were performed and compared to the WFNS and mFS grading systems. Results the eSAH score demonstrated excellent discrimination for unfavorable functional outcome at discharge ( AUC 0.89 ) and in-hospital mortality (AUC 0.87). The DCI subscore demonstrated good discriminatory performance for predicting DCI (AUC 0.77). Compared with conventional grading systems, this was superior to both the WFNS (AUC 0.75) and the mFS ( AUC 0.70). increasing eSAH scores were additionally associated with progressively higher rates of mortality and unfavorable functional outcomes. Conclusion the eSAH score demonstrates strong external validity, reproducibility and superior predictive performance compared with conventional grading systems in a large multicenter cohort. These findings support the clinical utility of quantitative hemorrhage burden integration for early risk stratification in patients with aSAH.

Abstract Background & Objectives Cholangiocarcinoma (CCA) is a rare and aggressive malignancy of the bile ducts characterized by complex imaging patterns, high inter-reader variability, and limited subspecialty expertise. Despite the critical importance of early detection for improving patient outcomes, most artificial intelligence (AI) tools developed for radiology have focused on other organ systems, leaving CCA largely underserved. Furthermore, clinicians with relevant domain expertise frequently lack the programming background required to participate meaningfully in AI development. This work aimed to create an accessible, educational AI framework “AI Starter” designed to empower radiologists and clinicians to initiate medical image analysis projects for CCA without requiring extensive coding knowledge. Methods A structured, end-to-end AI pipeline was introduced covering the key stages of image acquisition, preprocessing, segmentation, feature extraction, and predictive modeling. Open-source frameworks, specifically MONAI and PyTorch, were employed to ensure reproducibility and accessibility. Transfer learning strategies were incorporated to address the challenge of limited CCA-specific training datasets. The pipeline was designed with modular components to reduce technical barriers and accommodate users across varying levels of computational experience. Results Radiologists and clinicians were able to successfully engage with and deploy AI tools for CCA imaging analysis using this framework. Transfer learning demonstrably improved model performance under data-limited conditions, and the open-source implementation supported reproducible research workflows. Key challenges identified included scanner variability, segmentation instability, feature instability, and model overfitting, all of which were transparently mapped within the pipeline. Conclusion The AI Starter framework offers a practical and educational entry point for the CCA community including basic scientists and clinicians to meaningfully engage with and understand AI-driven imaging analysis.

8057 Background: The NCI Adjuvant Lung Cancer Enrichment Marker Identification and Sequencing Trials (ALCHEMIST) study tested resected stage II-IIIB(N2) NSCLC for targetable EGFR and ALK alterations and PD-L1 expression to screen patients for biomarker-driven adjuvant therapy trials (NCT02194738). Local lab results were requested when available for comparison to central lab results. Methods: We identified patients with both central and local PD-L1 results in the ALCHEMIST adjuvant immunotherapy study screening population. Discrepancies in tumor proportion score (TPS) between central and local PD-L1 were defined using prespecified cut-points (granular: <1%, 1-<5%, 5-<10%, 10-<20%, 20-<30%, 30-<40%, 40-<50%, 50-60%, 60-<70%, 70-<80%, 80-<90%, 90-<100%; clinical: <1%, 1-<50%, 50-100%). Large discrepancies were defined as ≥ 50 percentage point differences; for these cases, local PD-L1 and hematoxylin & eosin-stained slides were requested. These and all central slides were digitally scanned at 40X. Three expert pathologists reviewed the scanned slides in a blinded fashion. Differences between patient or tumor characteristics were assessed by Chi-square and Wilcoxon rank sum tests. Percent agreement was computed within and between the reviewers. Results: Central and local PD-L1 results were available for 497 patients. 177 cases (36%) had discrepancies that placed patients in different clinical TPS categories (<1%, 1-<50%, 50-100%). Among these, 43 cases had large TPS discrepancies; 14 had available local testing slides. 46 of the 320 cases without discrepancy were randomly selected as controls. No patient or tumor characteristics were associated with discrepancy, including tumor location, histology, grade or cancer stage (p > 0.05). Pairwise comparisons of TPS according to three pathologists re-reading the same scanned PD-L1 slides showed similar levels of disagreement in clinical TPS category whether there was initial discrepancy between central and local results (24%, 34%, and 23%; mean 27%) or not (33%, 30%, and 20%; mean 28%). Among the 14 cases with large initial discrepancies and local slides available, there was agreement in clinical TPS category when the anonymized central and local scanned slides were re-read by the pathologists in 54%, 38%, and 43% of cases, despite 100% non-agreement in initial reads. There were higher rates of agreement when the tested slides both originated from the same (100%) versus different (35%) blocks, when the same (48%) versus different (40%) PD-L1 antibodies were used, and when the cases were of non-squamous (52%) versus squamous (14%) cell carcinoma. Conclusions: There are substantial differences in PD-L1 TPS read by pathologists that can result in placement of patients in different categories commonly used for clinical management. These differences have clear implications in clinical trials and routine patient care.

4514 Background: Maintenance avelumab is a standard of care in patients (pts) with la/mUC who do not progress after first-line PBC. Cabozantinib (CABO) is an oral inhibitor of MET, VEGFR and TAM, active in multiple solid tumors. We hypothesized that an avelumab (Av)-CABO combination would improve outcomes in pts with mUC with an acceptable safety profile compared to avelumab maintenance alone. Methods: MAIN-CAV (NCT05092958) is a phase III randomized, multicenter, international trial for la/mUC pts who did not progress after 4-6 cycles of any PBC (gemcitabine-cisplatin, gemcitabine-carboplatin, MVAC, ddMVAC). Pts were randomized 1:1 to Av 800 mg IV every 2 weeks alone or with CABO 40 mg daily for ≤2 years. Stratification factors were response to first-line PBC (complete response [CR] vs partial response [PR] vs stable disease [SD]) and visceral metastases. The primary endpoint was overall survival (OS); secondary endpoints included progression-free survival (PFS), safety and quality of life. The target accrual was 654 pts; the trial closed early after enrollment of 68 pts following adoption of enfortumab vedotin plus pembrolizumab as the new preferred first-line standard in the US. Results: From 3/2022 to 3/2024, 68 pts were randomized (Av n = 33; CABO-Av n = 35). Median OS was 24.8 months (mo) with Av and 23.0 mo with Avb-CABO (Hazard ratio 1.41; 95% CI 0.66–3.01; p = 0.186). Median PFS was 4.6 mo with Av vs 6.0 mo with Av-CABO (p = 0.343). Objective response rate among pts with PR or SD after first-line PBC was 20.8% with Av vs 11.5% with Av-CABO (p = 0.424). Grade ≥3 treatment-related adverse events occurred in 24.2% pts on the Av arm vs 39.4% of pts on the Av-CABO arm and CABO dose reductions occurred in 63% of pts. Conclusions: Addition of CABO to Av did not improve OS or PFS compared with Av alone in this underpowered, prematurely closed phase III trial; safety was consistent with known profiles. U10CA180821,U10CA180882; U24CA196171, U10CA180863. (CCTG);U10CA180888 (SWOG); https://acknowledgments.alliancefound.org. Clinical trial information: NCT05092958 .

Abstract Heterogeneous treatment effects (HTE) are key to precision medicine, but most real-world studies lack the scale and diversity needed to detect them. While multi-site analyses offer a potential solution, data-sharing constraints often prevent access to patient-level information across institutions. We introduce POLARIS, a federated framework for time-to-event target trial emulation. POLARIS converts each site’s weighted Cox risk function into a compact tensor shared once with the coordinating center, enabling lossless reproduction of pooled estimates without sharing patient-level data. We applied POLARIS across five U.S. health systems to study risk of gastrointestinal outcomes after GLP-1 receptor agonist (GLP-1RAs) initiation versus sodium-glucose cotransporter 2 inhibitors (SGLT2is) and dipeptidyl peptidase 4 inhibitors (DPP4is). Results showed that GLP-1RAs were consistently associated with higher risks of nausea and vomiting, particularly among men, individuals with higher baseline HbA1c (≥ 8.5%), and lipid therapy. POLARIS provides a scalable solution for distributed target trial emulation and fine-grained assessment of HTE across diverse health systems.

Abstract Solitary plasmacytoma of bone (SPB) is a localized plasma cell neoplasm associated with a substantial risk of progression to multiple myeloma (MM), yet the role of adjuvant systemic therapy following definitive radiation therapy (RT) remains undefined. We conducted a randomized phase III trial (Alliance A061402) to evaluate whether the addition of systemic therapy with ixazomib, lenalidomide, and dexamethasone (IXA-LEN-DEX) to zoledronic acid (ZA) improves outcomes compared with ZA alone in patients with SPB and minimal (< 10%) bone marrow plasma cell involvement following RT. Patients were randomized to receive six 28-day cycles of IXA-LEN-DEX-ZA or ZA alone. The primary endpoint was time to progression (TTP) to MM. The study was closed early due to poor accrual after enrolling 11 patients (IXA-LEN-DEX-ZA, n = 5; ZA, n = 6). All patients completed RT and planned treatment without dose reductions. The combination regimen was well tolerated, with limited grade ≥ 3 toxicities, including hypophosphatemia and lymphopenia. At follow-up extending beyond 4 years, the 1- and 4-year progression rates to MM in the IXA-LEN-DEX-ZA arm were 0% and 33%, respectively, compared with 37.5% and 58.3% at 1 and 2 years in the ZA-alone arm. Although underpowered, these findings demonstrate the feasibility and tolerability of an all-oral adjuvant systemic regimen in SPB and provide prospective data supporting further investigation of systemic strategies to delay progression to MM.

Abstract Background Reference sets are needed to evaluate performance of multi-cancer detection (MCD) assays. The National Cancer Institute funded the Alliance reference set study to assess MCDs for use in future trials. Methods Individuals with cancer and controls were recruited; blood specimens were collected prior to cancer treatment. A performance evaluation study was designed utilizing reference set samples. Companies (n = 6) were selected to participate based on review of performance data and ability to utilize the blood collection tube. Companies received samples from cancer types their assay was designed to detect (“targeted”), plus additional “non-targeted” and control samples. Companies reported positive/negative calls, risk scores, and tissue-of-origin (TOO) predictions. Sensitivity was computed for early (I-II) and late (III-IV) stage cases, based on positive/negative calls (SEPN) and at fixed 98% specificity (SE98). Specificity and TOO accuracy were computed. Results Five hundred and forty nine cases (encompassing 13 cancer types) and 413 controls from the reference set were included in the study. Companies assessed samples from median 6 (range 5-9) targeted cancer types and median 8 (range: 7-11) overall cancer types. Median (range) specificity was 92.3% (76.5%-98.5%). Median (range) SEPN was 32% (25%-42%) for early stage and 73% (48%-89%) for late stage; while median (range) SE98 was 19% (8%-35%) for early stage and 66% (13%-79%) for late stage. Median sensitivity for non-targeted types was 40% (early stage) and 52% (late stage). Median (range) TOO accuracy (primary predicted site) was 75% (64%-78%). Conclusions Sensitivity and specificity varied widely across assays with early-stage sensitivity substantially lower than late-stage sensitivity. Clinical Trial Registration Number ClinicalTrials.gov identifier NCT05334069.

This review explores how multimodal foundation models (MFMs) are transforming biliary tract cancer (BTC) research. BTCs are aggressive malignancies with poor prognosis, presenting unique challenges due to difficult diagnostic methods, molecular complexity, and rarity. Importantly, intrahepatic cholangiocarcinoma (iCCA), perihilar cholangiocarcinoma (pCCA), and distal bile duct cholangiocarcinoma (dCCA) represent fundamentally distinct clinical entities, with iCCA presenting as mass-forming lesions amenable to biopsy and targeted therapies, while pCCA manifests as infiltrative bile duct lesions with challenging diagnosis and primarily palliative management approaches. MFMs offer potential to advance research by integrating radiological images, histopathology, multi-omics profiles, and clinical data into unified computational frameworks, with applications tailored to these distinct BTC subtypes. Key applications include enhanced biomarker discovery that identifies previously unrecognizable cross-modal patterns, potential for improving currently limited diagnostic accuracy—though validation in BTC-specific cohorts remains essential—accelerated drug repurposing, and advanced patient stratification for personalized treatment. Despite promising results, challenges such as data scarcity, high computational demands, and clinical workflow integration remain to be addressed. Future research should focus on standardized data protocols, architectural innovations, and prospective validation studies. The integration of artificial intelligence (AI)-based methodologies offers new solutions for these historically challenging malignancies. However, current evidence for BTC-specific applications remains largely theoretical, with most studies limited to proof-of-concept designs or related cancer types. Comprehensive clinical validation studies and prospective trials demonstrating patient benefit are essential prerequisites for clinical implementation. The timeline for evidence-based clinical adoption likely extends 7–10 years, contingent on successful completion of validation studies addressing current evidence gaps.

The integration of artificial intelligence (AI) in medical imaging has revolutionized diagnostic capabilities, yet the black-box nature of deep learning models poses significant challenges for clinical adoption. Current explainable AI (XAI) approaches, including SHAP, LIME, and Grad-CAM, predominantly focus on post hoc explanations that may inadvertently undermine clinical decision-making by providing misleading confidence in AI outputs. This paper presents a systematic review and meta-analysis of 67 studies (covering 23 radiology, 19 pathology, and 25 ophthalmology applications) evaluating XAI fidelity, stability, and performance trade-offs across medical imaging modalities. Our meta-analysis of 847 initially identified studies reveals that LIME achieves superior fidelity (0.81, 95% CI: 0.78–0.84) compared to SHAP (0.38, 95% CI: 0.35–0.41) and Grad-CAM (0.54, 95% CI: 0.51–0.57) across all modalities. Post hoc explanations demonstrated poor stability under noise perturbation, with SHAP showing 53% degradation in ophthalmology applications (ρ = 0.42 at 10% noise) compared to 11% in radiology (ρ = 0.89). We demonstrate a consistent 5–7% AUC performance penalty for interpretable models but identify modality-specific stability patterns suggesting that tailored XAI approaches are necessary. Based on these empirical findings, we propose a comprehensive three-pillar accountability framework that prioritizes transparency in model development, interpretability in architecture design, and a cautious deployment of post hoc explanations with explicit uncertainty quantification. This approach offers a pathway toward genuinely accountable AI systems that enhance rather than compromise clinical decision-making quality and patient safety.

Background We automated subarachnoid hemorrhage volume (SAHV) calculation with artificial intelligence (SAHVAI) and created 3‐dimensional volumetric images (SAHVAI‐3D) using noncontrast head computed tomography data for patients with aneurysmal SAH. We also defined 4‐dimensional SAHVAI (SAHVAI‐4D), representing SAHV over time. We aimed to compare automated SAHVAI values and computational times to manual SAHV measurement methods, explore the potential of imaging biomarkers to identify at‐risk brain regions for delayed cerebral ischemia and explore potential insights in future neurotherapeutic interventions for patient recovery after SAH. Methods A training set of 10 consecutive patients with aneurysmal SAH was used to manually compute SAHV, SAHVAI‐3D, and SAHVAI‐4D, including 92 noncontrast computed tomography scans (182 slices each). The SAHVAI deep learning algorithm generated automated SAHV values in cubic centimeters. A 3‐dimensional SAH brain map was created for each patient for the SAHVAI and manual evaluations. Blood volumetric outputs were analyzed and compared to neurologic outcomes at discharge, including delayed cerebral ischemia events, symptomatic vasospasm, and areas with the thickest SAH blood concentration. Results SAHVAI quantified SAHV in a mean of 6.7 seconds per scan, significantly faster than the manual method, which took >60 minutes per scan (Fisher exact test, P <0.001). SAHVAI demonstrated an accuracy of 99.8%, Dice score of 0.701, false‐positive rate of 0.0005, and negative predictive value of 0.999. The mean absolute error between SAHVAI and manual methods was 5.67 mL. SAHVAI‐3D brain map and total SAHV at admission were inversely associated with Glasgow Coma Scale (R 2 = 0.23, P = 0.017) and directly associated with length of hospital stay (R 2 = 0.175, P = 0.004), especially in regions with dense blood concentration. Conclusion SAHVAI‐3D and SAHVAI‐4D brain mapping techniques represent innovative imaging biomarkers for SAH. These advancements enable rapid evaluation and targeted interventions, potentially improving patient care in SAH management.