Pancreatic cancer treatment plans involving surgery and/or chemotherapy are highly dependent on disease stage. However, current staging systems are ineffective and poorly correlated with survival outcomes. We investigate how artificial intelligence (AI) can enhance prognostic accuracy in pancreatic cancer by integrating multiple data sources. Patients with histopathology and/or radiology/follow-up confirmed pancreatic ductal adenocarcinoma (PDAC) from a Dutch center (2004-2023) were included in the development cohort. Two additional PDAC cohorts from a Dutch and Spanish center were used for external validation. Prognostic models including clinical variables, contrast-enhanced CT images, and a combination of both were developed to predict high-risk short-term survival. All models were trained using five-fold cross-validation and assessed by the area under the time-dependent receiver operating characteristic curve (AUC). The models were developed on 401 patients (203 females, 198 males, median survival (OS) = 347 days, IQR: 171-585), with 98 (24.4%) short-term survivors (OS < 230 days) and 303 (75.6%) long-term survivors. The external validation cohorts included 361 patients (165 females, 138 males, median OS = 404 days, IQR: 173-736), with 110 (30.5%) short-term survivors and 251 (69.5%) longer survivors. The best AUC for predicting short vs. long-term survival was achieved with the multi-modal model (AUC = 0.637 (95% CI: 0.500-0.774)) in the internal validation set. External validation showed AUCs of 0.571 (95% CI: 0.453-0.689) and 0.675 (95% CI: 0.593-0.757). Multimodal AI can predict long vs. short-term survival in PDAC patients, showing potential as a prognostic tool in clinical decision-making. Question Prognostic tools for pancreatic ductal adenocarcinoma (PDAC) remain limited, with TNM staging offering suboptimal accuracy in predicting patient survival outcomes. Findings The multimodal AI model demonstrated improved prognostic performance over TNM and unimodal models for predicting short- and long-term survival in PDAC patients. Clinical relevance Multimodal AI provides enhanced prognostic accuracy compared to current staging systems, potentially improving clinical decision-making and personalized management strategies for PDAC patients.

Identification of bony landmarks in medical images is of high importance for 3D planning in orthopaedic surgery. Automated landmark identification has the potential to optimize clinical routines and allows for the scientific analysis of large databases. To the authors' knowledge, no direct comparison of different methods for automated landmark detection on the same dataset has been published to date. We compared 3 methods for automated femoral landmark identification: an artificial neural network, a statistical shape model and a geometric approach. All methods were compared against manual measurements of two raters on the task of identifying 6 femoral landmarks on CT data or derived surface models of 202 femora. The accuracy of the methods was in the range of the manual measurements and comparable to those reported in previous studies. The geometric approach showed a significantly higher average deviation compared to the manually selected reference landmarks, while there was no statistically significant difference for the neural network and the SSM. All fully automated methods show potential for use, depending on the use case. Characteristics of the different methods, such as the input data required (raw CT/segmented bone surface models, amount of training data required) and/or the methods robustness, can be used for method selection in the individual application.

Purpose: To evaluate the impact of harmonization and multi-region CT image feature integration on survival prediction in non-small cell lung cancer (NSCLC) patients, using handcrafted radiomics, pretrained foundation model (FM) features, and clinical data from a multicenter dataset. Methods: We analyzed CT scans and clinical data from 876 NSCLC patients (604 training, 272 test) across five centers. Features were extracted from the whole lung, tumor, mediastinal nodes, coronary arteries, and coronary artery calcium (CAC). Handcrafted radiomics and FM deep features were harmonized using ComBat, reconstruction kernel normalization (RKN), and RKN+ComBat. Regularized Cox models predicted overall survival; performance was assessed using the concordance index (C-index), 5-year time-dependent area under the curve (t-AUC), and hazard ratio (HR). SHapley Additive exPlanations (SHAP) values explained feature contributions. A consensus model used agreement across top region of interest (ROI) models to stratify patient risk. Results: TNM staging showed prognostic utility (C-index = 0.67; HR = 2.70; t-AUC = 0.85). The clinical + tumor radiomics model with ComBat achieved a C-index of 0.7552 and t-AUC of 0.8820. FM features (50-voxel cubes) combined with clinical data yielded the highest performance (C-index = 0.7616; t-AUC = 0.8866). An ensemble of all ROIs and FM features reached a C-index of 0.7142 and t-AUC of 0.7885. The consensus model, covering 78% of valid test cases, achieved a t-AUC of 0.92, sensitivity of 97.6%, and specificity of 66.7%. Conclusion: Harmonization and multi-region feature integration improve survival prediction in multicenter NSCLC data. Combining interpretable radiomics, FM features, and consensus modeling enables robust risk stratification across imaging centers.

Deep learning (DL)-based methods show promise in detecting pulmonary embolism (PE) on CT pulmonary angiography (CTPA), potentially improving diagnostic accuracy and workflow efficiency. This meta-analysis aimed to (1) determine pooled performance estimates of DL algorithms for PE detection; and (2) compare the diagnostic efficacy of convolutional neural network (CNN)- versus U-Net-based architectures. Following PRISMA guidelines, we searched PubMed and EMBASE through April 15, 2025 for English-language studies (2010-2025) reporting DL models for PE detection with extractable 2 × 2 data or performance metrics. True/false positives and negatives were reconstructed when necessary under an assumed 50 % PE prevalence (with 0.5 continuity correction). We approximated AUROC as the mean of sensitivity and specificity if not directly reported. Sensitivity, specificity, accuracy, PPV and NPV were pooled using a DerSimonian-Laird random-effects model with Freeman-Tukey transformation; AUROC values were combined via a fixed-effect inverse-variance approach. Heterogeneity was assessed by Cochran's Q and I². Subgroup analyses contrasted CNN versus U-Net models. Twenty-four studies (n = 22,984 patients) met inclusion criteria. Pooled estimates were: AUROC 0.895 (95 % CI: 0.874-0.917), sensitivity 0.894 (0.856-0.923), specificity 0.871 (0.831-0.903), accuracy 0.857 (0.833-0.882), PPV 0.832 (0.794-0.869) and NPV 0.902 (0.874-0.929). Between-study heterogeneity was high (I² ≈ 97 % for sensitivity/specificity). U-Net models exhibited higher sensitivity (0.899 vs 0.893) and CNN models higher specificity (0.926 vs 0.900); subgroup Q-tests confirmed significant differences for both sensitivity (p = 0.0002) and specificity (p < 0.001). DL algorithms demonstrate high diagnostic accuracy for PE detection on CTPA, with complementary strengths: U-Net architectures excel in true-positive identification, whereas CNNs yield fewer false positives. However, marked heterogeneity underscores the need for standardized, prospective validation before routine clinical implementation.

Stroke is among the top three causes of death worldwide, and accurate identification of stroke lesion boundaries is critical for diagnosis and treatment. Supervised deep learning methods have emerged as the leading solution for stroke lesion segmentation but require large, diverse, and annotated datasets. The ISLES'24 challenge addresses this need by providing longitudinal stroke imaging data, including CT scans taken on arrival to the hospital and follow-up MRI taken 2-9 days from initial arrival, with annotations derived from follow-up MRI. Importantly, models submitted to the ISLES'24 challenge are evaluated using only CT inputs, requiring prediction of lesion progression that may not be visible in CT scans for segmentation. Our winning solution shows that a carefully designed preprocessing pipeline including deep-learning-based skull stripping and custom intensity windowing is beneficial for accurate segmentation. Combined with a standard large residual nnU-Net architecture for segmentation, this approach achieves a mean test Dice of 28.5 with a standard deviation of 21.27.

This paper presents a novel method for monocular patient-to-image intraoperative registration, specifically designed to operate without any external hardware tracking equipment or fiducial point markers. Leveraging a synthetic microscopy surgical scene dataset with a wide range of transformations, our approach directly maps preoperative CT scans to 2D intraoperative surgical frames through a lightweight neural network for real-time cochlear implant surgery guidance via a zero-shot learning approach. Unlike traditional methods, our framework seamlessly integrates with monocular surgical microscopes, making it highly practical for clinical use without additional hardware dependencies and requirements. Our method estimates camera poses, which include a rotation matrix and a translation vector, by learning from the synthetic dataset, enabling accurate and efficient intraoperative registration. The proposed framework was evaluated on nine clinical cases using a patient-specific and cross-patient validation strategy. Our results suggest that our approach achieves clinically relevant accuracy in predicting 6D camera poses for registering 3D preoperative CT scans to 2D surgical scenes with an angular error within 10 degrees in most cases, while also addressing limitations of traditional methods, such as reliance on external tracking systems or fiducial markers.

This study is aimed to clarify the effectiveness of the image-rotation technique and zooming augmentation to improve the accuracy of the deep learning (DL)-based dose conversion from pencil beam (PB) to Monte Carlo (MC) in proton beam therapy (PBT). We adapted 85 patients with head and neck cancers. The patient dataset was randomly divided into 101 plans (334 beams) for training/validation and 11 plans (34 beams) for testing. Further, we trained a DL model that inputs a computed tomography (CT) image and the PB dose in a single-proton field and outputs the MC dose, applying the image-rotation technique and zooming augmentation. We evaluated the DL-based dose conversion accuracy in a single-proton field. The average γ-passing rates (a criterion of 3%/3 mm) were 80.6 ± 6.6% for the PB dose, 87.6 ± 6.0% for the baseline model, 92.1 ± 4.7% for the image-rotation model, and 93.0 ± 5.2% for the data-augmentation model, respectively. Moreover, the average range differences for R90 were - 1.5 ± 3.6% in the PB dose, 0.2 ± 2.3% in the baseline model, -0.5 ± 1.2% in the image-rotation model, and - 0.5 ± 1.1% in the data-augmentation model, respectively. The doses as well as ranges were improved by the image-rotation technique and zooming augmentation. The image-rotation technique and zooming augmentation greatly improved the DL-based dose conversion accuracy from the PB to the MC. These techniques can be powerful tools for improving the DL-based dose calculation accuracy in PBT.

In mechanical thrombectomy (MT), extracranial vascular tortuosity is among the main determinants of procedure duration and success. Currently, no rapid and reliable method exists to identify the anatomical features precluding fast and stable access to the cervical vessels. A retrospective sample of 513 patients were included in this study. Patients underwent first-line transfemoral MT following anterior circulation large vessel occlusion stroke. Difficult transfemoral access (DTFA) was defined as impossible common carotid catheterization or time from groin puncture to first carotid angiogram >30 min. A machine learning model based on 29 anatomical features automatically extracted from head-and-neck computed tomography angiography (CTA) was developed to predict DTFA. Three experienced raters independently assessed the likelihood of DTFA on a reduced cohort of 116 cases using a Likert scale as benchmark for the model, using preprocedural CTA as well as automatic 3D vascular segmentation separately. Among the study population, 11.5% of procedures (59/513) presented DTFA. Six different features from the aortic, supra-aortic, and cervical regions were included in the model. Cross-validation resulted in an area under the receiver operating characteristic (AUROC) curve of 0.76 (95% CI 0.75 to 0.76) for DTFA prediction, with high sensitivity for impossible access identification (0.90, 95% CI 0.81 to 0.94). The model outperformed human assessment in the reduced cohort [F1-score (95% CI) by experts with CTA: 0.43 (0.37 to 0.50); experts with 3D segmentation: 0.50 (0.46 to 0.54); and model: 0.70 (0.65 to 0.75)]. A fully automatic model for DTFA prediction was developed and validated. The presented method improved expert assessment of difficult access prediction in stroke MT. Derived information could be used to guide decisions regarding arterial access for MT.

The sphenoid sinus is an important access point for trans-sphenoidal surgeries, but variations in its pneumatization may complicate surgical safety. Deep learning can be used to identify these anatomical variations. We developed a convolutional neural network (CNN) model for the automated prediction of sphenoid sinus pneumatization patterns in computed tomography (CT) scans. This model was tested on mid-sagittal CT images. Two radiologists labeled all CT images into four pneumatization patterns: Conchal (type I), presellar (type II), sellar (type III), and postsellar (type IV). We then augmented the training set to address the limited size and imbalanced nature of the data. The initial dataset included 249 CT images, divided into training (n = 174) and test (n = 75) datasets. The training dataset was augmented to 378 images. Following augmentation, the overall diagnostic accuracy of the model improved from 76.71% to 84%, with an area under the curve (AUC) of 0.84, indicating very good diagnostic performance. Subgroup analysis showed excellent results for type IV, with the highest AUC of 0.93, perfect sensitivity (100%), and an F1-score of 0.94. The model also performed robustly for type I, achieving an accuracy of 97.33% and high specificity (99%). These metrics highlight the model's potential for reliable clinical application. The proposed CNN model demonstrates very good diagnostic accuracy in identifying various sphenoid sinus pneumatization patterns, particularly excelling in type IV, which is crucial for endoscopic sinus surgery due to its higher risk of surgical complications. By assisting radiologists and surgeons, this model enhances the safety of transsphenoidal surgery, highlighting its value, novelty, and applicability in clinical settings.

To evaluate the variability of fully automated airway quantitative CT (QCT) measures caused by different kernels and the effect of kernel conversion. This retrospective study included 96 patients who underwent non-enhanced chest CT at two centers. CT scans were reconstructed using four kernels (medium soft, medium sharp, sharp, very sharp) from three vendors. Kernel conversion targeting the medium soft kernel as reference was applied to sharp kernel images. Fully automated airway quantification was performed before and after conversion. The effects of kernel type and conversion on airway quantification were evaluated using analysis of variance, paired t-tests, and concordance correlation coefficient (CCC). Airway QCT measures (e.g., Pi10, wall thickness, wall area percentage, lumen diameter) decreased with sharper kernels (all, p < 0.001), with varying degrees of variability across variables and vendors. Kernel conversion substantially reduced variability between medium soft and sharp kernel images for vendors A (pooled CCC: 0.59 vs. 0.92) and B (0.40 vs. 0.91) and lung-dedicated sharp kernels of vendor C (0.26 vs. 0.71). However, it was ineffective for non-lung-dedicated sharp kernels of vendor C (0.81 vs. 0.43) and showed limited improvement in variability of QCT measures at the subsegmental level. Consistent airway segmentation and identical anatomic labeling improved subsegmental airway variability in theoretical tests. Deep learning-based kernel conversion reduced the measurement variability of airway QCT across various kernels and vendors but was less effective for non-lung-dedicated kernels and subsegmental airways. Consistent airway segmentation and precise anatomic labeling can further enhance reproducibility for reliable automated quantification. Question How do different CT reconstruction kernels affect the measurement variability of automated airway measurements, and can deep learning-based kernel conversion reduce this variability? Findings Kernel conversion improved measurement consistency across vendors for lung-dedicated kernels, but showed limited effectiveness for non-lung-dedicated kernels and subsegmental airways. Clinical relevance Understanding kernel-related variability in airway quantification and mitigating it through deep learning enables standardized analysis, but further refinements are needed for robust airway segmentation, particularly for improving measurement variability in subsegmental airways and specific kernels.