Foundation models (FMs) have shown great promise in medical image analysis by improving generalization across diverse downstream tasks. In ophthalmology, several FMs have recently emerged, but there is still no clear answer to fundamental questions: Which FM performs the best? Are they equally good across different tasks? What if we combine all FMs together? To our knowledge, this is the first study to systematically evaluate both single and fused ophthalmic FMs. To address these questions, we propose FusionFM, a comprehensive evaluation suite, along with two fusion approaches to integrate different ophthalmic FMs. Our framework covers both ophthalmic disease detection (glaucoma, diabetic retinopathy, and age-related macular degeneration) and systemic disease prediction (diabetes and hypertension) based on retinal imaging. We benchmarked four state-of-the-art FMs (RETFound, VisionFM, RetiZero, and DINORET) using standardized datasets from multiple countries and evaluated their performance using AUC and F1 metrics. Our results show that DINORET and RetiZero achieve superior performance in both ophthalmic and systemic disease tasks, with RetiZero exhibiting stronger generalization on external datasets. Regarding fusion strategies, the Gating-based approach provides modest improvements in predicting glaucoma, AMD, and hypertension. Despite these advances, predicting systemic diseases, especially hypertension in external cohort remains challenging. These findings provide an evidence-based evaluation of ophthalmic FMs, highlight the benefits of model fusion, and point to strategies for enhancing their clinical applicability.

Lung cancer is the leading cause of cancer-related death worldwide. Deep learning-based computer-aided diagnosis (CAD) systems in screening programs enhance malignancy prediction, assist radiologists in decision-making, and reduce inter-reader variability. However, limited research has explored the analysis of repeated annual exams of indeterminate lung nodules to improve accuracy. We introduced a novel spatio-temporal deep learning framework, the global attention convolutional recurrent neural network (globAttCRNN), to predict indeterminate lung nodule malignancy using serial screening computed tomography (CT) images from the National Lung Screening Trial (NLST) dataset. The model comprises a lightweight 2D convolutional neural network for spatial feature extraction and a recurrent neural network with a global attention module to capture the temporal evolution of lung nodules. Additionally, we proposed new strategies to handle missing data in the temporal dimension to mitigate potential biases arising from missing time steps, including temporal augmentation and temporal dropout. Our model achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.954 in an independent test set of 175 lung nodules, each detected in multiple CT scans over patient follow-up, outperforming baseline single-time and multiple-time architectures. The temporal global attention module prioritizes informative time points, enabling the model to capture key spatial and temporal features while ignoring irrelevant or redundant information. Our evaluation emphasizes its potential as a valuable tool for the diagnosis and stratification of patients at risk of lung cancer.

This study introduces the Structural MRI-based Alzheimer's Disease Score (SMAS), a novel index intended to quantify Alzheimer's Disease (AD)-related morphometric patterns using a deep learning Bayesian-supervised Variational Autoencoder (Bayesian-SVAE). The SMAS index was constructed using baseline structural MRI data from the DELCODE study and evaluated longitudinally in two independent cohorts: DELCODE (n=415) and ADNI (n=190). Our findings indicate that SMAS has strong associations with cognitive performance (DELCODE: r=-0.83; ADNI: r=-0.62), age (DELCODE: r=0.50; ADNI: r=0.28), hippocampal volume (DELCODE: r=-0.44; ADNI: r=-0.66), and total gray matter volume (DELCODE: r=-0.42; ADNI: r=-0.47), suggesting its potential as a biomarker for AD-related brain atrophy. Moreover, our longitudinal studies indicated that SMAS may be useful for the early identification and tracking of AD. The model demonstrated significant predictive accuracy in distinguishing cognitively healthy individuals from those with AD (DELCODE: AUC=0.971 at baseline, 0.833 at 36 months; ADNI: AUC=0.817 at baseline, improving to 0.903 at 24 months). Notably, over 36 months, the SMAS index outperformed existing measures such as SPARE-AD and hippocampal volume. The relevance map analysis revealed significant morphological changes in key AD-related brain regions, including the hippocampus, posterior cingulate cortex, precuneus, and lateral parietal cortex, highlighting that SMAS is a sensitive and interpretable biomarker of brain atrophy, suitable for early AD detection and longitudinal monitoring of disease progression.

Spontaneous intracranial hemorrhages have a high disease burden. Due to increasing medical imaging, new technological solutions for assisting in image interpretation are warranted. We developed a deep learning (DL) solution for spontaneous intracranial hemorrhage detection from head CT scans. The DL solution included four base convolutional neural networks (CNNs), which were trained using 300 head CT scans. A metamodel was trained on top of the four base CNNs, and simple post processing steps were applied to improve the solution's accuracy. The solution performance was evaluated using a retrospective dataset of consecutive emergency head CTs imaged in ten different emergency rooms. 7797 head CT scans were included in the validation dataset and 118 CT scans presented with spontaneous intracranial hemorrhage. The trained metamodel together with a simple rule-based post-processing step showed 89.8% sensitivity and 89.5% specificity for hemorrhage detection at the case-level. The solution detected all 78 spontaneous hemorrhage cases imaged presumably or confirmedly within 12 h from the symptom onset and identified five hemorrhages missed in the initial on-call reports. Although the success of DL algorithms depends on multiple factors, including training data versatility and quality of annotations, using the proposed ensemble-learning approach and rule-based post-processing may help clinicians to develop highly accurate DL solutions for clinical imaging diagnostics.

Focal cortical dysplasia (FCD) is the leading cause of drug-resistant epilepsy (DRE) in pediatric patients. Accurate detection of FCDs is crucial for successful surgical outcomes, yet remains challenging due to frequently subtle MRI findings, especially in children, whose brain morphology undergoes significant developmental changes. Automated detection algorithms have the potential to improve diagnostic precision, particularly in cases, where standard visual assessment fails. This study aimed to evaluate the performance of automated algorithms in detecting FCD type II in pediatric patients and to examine the impact of adult versus pediatric templates on detection accuracy. MRI data from 23 surgical pediatric patients with histologically confirmed FCD type II were retrospectively analyzed. Three imaging-based detection algorithms were applied to T1-weighted images, each targeting key structural features: cortical thickness, gray matter intensity (extension), and gray-white matter junction blurring. Their performance was assessed using adult and pediatric healthy controls templates, with validation against both predictive radiological ROIs (PRR) and post-resection cavities (PRC). The junction algorithm achieved the highest median dice score (0.028, IQR 0.038, p < 0.01 when compared with other algorithms) and detected relevant clusters even in MRI-negative cases. The adult template (median dice score 0.013, IQR 0.027) significantly outperformed the pediatric template (0.0032, IQR 0.023) (p < 0.001), highlighting the importance of template consistency. Despite superior performance of the adult template, its use in pediatric populations may introduce bias, as it does not account for age-specific morphological features such as cortical maturation and incomplete myelination. Automated algorithms, especially those targeting junction blurring, enhance FCD detection in pediatric populations. These algorithms may serve as valuable decision-support tools, particularly in settings where neuroradiological expertise is limited.

Radiomics has emerged as a promising non-invasive imaging approach in pediatric neuro-oncology, offering the ability to extract high-dimensional quantitative features from routine MRI to support diagnosis, risk stratification, molecular characterization, and outcome prediction. Pediatric brain tumors, which differ significantly from adult tumors in biology and imaging appearance, present unique diagnostic and prognostic challenges. By integrating radiomics with machine learning algorithms, studies have demonstrated strong performance in classifying tumor types such as medulloblastoma, ependymoma, and gliomas, and predicting molecular subgroups and mutations such as H3K27M and BRAF. Recent studies combining radiomics with machine learning algorithms - including support vector machines, random forests, and deep learning CNNs - have demonstrated promising performance, with AUCs ranging from 0.75 to 0.98 for tumor classification and 0.77 to 0.88 for molecular subgroup prediction, across cohorts from 50 to over 450 patients, with internal cross-validation and external validation in some cases. In resource-limited settings or regions with limited radiologist manpower, radiomics-based tools could help augment diagnostic accuracy and consistency, serving as decision support to prioritize patients for further evaluation or biopsy. Emerging applications such as radio-immunomics and radio-pathomics may further enhance understanding of tumor biology but remain investigational. Despite its potential, clinical translation faces notable barriers, including limited pediatric-specific datasets, variable imaging protocols, and the lack of standardized, reproducible workflows. Multi-institutional collaboration, harmonized pipelines, and prospective validation are essential next steps. Radiomics should be viewed as a supplementary tool that complements existing clinical and pathological frameworks, supporting more informed and equitable care in pediatric brain tumor management.

This study developed an end-to-end deep learning (DL) model using non-enhanced MRI to diagnose benign and malignant pelvic and sacral tumors (PSTs). Retrospective data from 835 patients across four hospitals were employed to train, validate, and test the models. Six diagnostic models with varied input sources were compared. Performance (AUC, accuracy/ACC) and reading times of three radiologists were compared. The proposed Model SEG-CL-NC achieved AUC/ACC of 0.823/0.776 (Internal Test Set 1) and 0.836/0.781 (Internal Test Set 2). In External Dataset Centers 2, 3, and 4, its ACC was 0.714, 0.740, and 0.756, comparable to contrast-enhanced models and radiologists (P > 0.05), while its diagnosis time was significantly shorter than radiologists (P < 0.01). Our results suggested that the proposed Model SEG-CL-NC could achieve comparable performance to contrast-enhanced models and radiologists in diagnosing benign and malignant PSTs, offering an accurate, efficient, and cost-effective tool for clinical practice.

Radiology, radiation oncology, and medical physics require decision-making that integrates medical images, textual reports, and quantitative data under high-stakes conditions. With the introduction of GPT-5, it is critical to assess whether recent advances in large multimodal models translate into measurable gains in these safety-critical domains. We present a targeted zero-shot evaluation of GPT-5 and its smaller variants (GPT-5-mini, GPT-5-nano) against GPT-4o across three representative tasks. We present a targeted zero-shot evaluation of GPT-5 and its smaller variants (GPT-5-mini, GPT-5-nano) against GPT-4o across three representative tasks: (1) VQA-RAD, a benchmark for visual question answering in radiology; (2) SLAKE, a semantically annotated, multilingual VQA dataset testing cross-modal grounding; and (3) a curated Medical Physics Board Examination-style dataset of 150 multiple-choice questions spanning treatment planning, dosimetry, imaging, and quality assurance. Across all datasets, GPT-5 achieved the highest accuracy, with substantial gains over GPT-4o up to +20.00% in challenging anatomical regions such as the chest-mediastinal, +13.60% in lung-focused questions, and +11.44% in brain-tissue interpretation. On the board-style physics questions, GPT-5 attained 90.7% accuracy (136/150), exceeding the estimated human passing threshold, while GPT-4o trailed at 78.0%. These results demonstrate that GPT-5 delivers consistent and often pronounced performance improvements over GPT-4o in both image-grounded reasoning and domain-specific numerical problem-solving, highlighting its potential to augment expert workflows in medical imaging and therapeutic physics.

BackgroundThe inbuilt ability to adapt existing models to new applications has been one of the key drivers of the success of deep learning models. Thereby, sharing trained models is crucial for their adaptation to different populations and domains. Not sharing models prohibits validation and potentially following translation into clinical practice, and hinders scientific progress. In this paper we examine the current state of data and model sharing in the medical field using cervical cancer staging on colposcopy images as a case example. MethodsWe conducted a comprehensive literature search in PubMed to identify studies employing machine learning techniques in the analysis of colposcopy images. For studies where raw data was not directly accessible, we systematically inquired about accessing the pre-trained model weights and/or raw colposcopy image data by contacting the authors using various channels. ResultsWe included 46 studies and one publicly available dataset in our study. We retrieved data of the latter and inquired about data access for the 46 studies by contacting a total of 92 authors. We received 15 responses related to 14 studies (30%). The remaining 32 studies remained unresponsive (70%). Of the 15 responses received, two responses redirected our inquiry to other authors, two responses were initially pending, and 11 declined data sharing. Despite our follow-up efforts on all responses received, none of the inquiries led to actual data sharing (0%). The only available data source remained the publicly available dataset. ConclusionsDespite the long-standing demands for reproducible research and efforts to incentivize data sharing, such as the requirement of data availability statements, our case study reveals a persistent lack of data sharing culture. Reasons identified in this case study include a lack of resources to provide the data, data privacy concerns, ongoing trial registrations and low response rates to inquiries. Potential routes for improvement could include comprehensive data availability statements required by journals, data preparation and deposition in a repository as part of the publication process, an automatic maximal embargo time after which data will become openly accessible and data sharing rules set by funders.

The claustrum is a band-like gray matter structure located between putamen and insula whose exact functions are still actively researched. Its sheet-like structure makes it barely visible in in vivo magnetic resonance imaging (MRI) scans at typical resolutions, and neuroimaging tools for its study, including methods for automatic segmentation, are currently very limited. In this paper, we propose a contrast- and resolution-agnostic method for claustrum segmentation at ultra-high resolution (0.35 mm isotropic); the method is based on the SynthSeg segmentation framework, which leverages the use of synthetic training intensity images to achieve excellent generalization. In particular, SynthSeg requires only label maps to be trained, since corresponding intensity images are synthesized on the fly with random contrast and resolution. We trained a deep learning network for automatic claustrum segmentation, using claustrum manual labels obtained from 18 ultra-high resolution MRI scans (mostly ex vivo). We demonstrated the method to work on these 18 high resolution cases (Dice score = 0.632, mean surface distance = 0.458 mm, and volumetric similarity = 0.867 using 6-fold cross validation (CV)), and also on in vivo T1-weighted MRI scans at typical resolutions (≈1 mm isotropic). We also demonstrated that the method is robust in a test-retest setting and when applied to multimodal imaging (T2-weighted, proton density, and quantitative T1 scans). To the best of our knowledge this is the first accurate method for automatic ultra-high resolution claustrum segmentation, which is robust against changes in contrast and resolution. The method is released at https://github.com/chiara-mauri/claustrum_segmentation and as part of the neuroimaging package FreeSurfer.