The advent of endovascular thrombectomy has significantly improved outcomes for stroke patients with intracranial large vessel occlusion, yet individual benefits can vary widely. As demand for thrombectomy rises and geographical disparities in stroke care access persist, there is a growing need for predictive models that quantify individual benefits. However, current imaging methods for estimating outcomes may not fully capture the dynamic nature of cerebral ischaemia and lack a patient-specific assessment of thrombectomy benefits. Our study introduces a deep learning approach to predict individual responses to thrombectomy in acute ischaemic stroke patients. The proposed models provide predictions for both tissue and clinical outcomes under two scenarios: one assuming successful reperfusion and another assuming unsuccessful reperfusion. The resulting simulations of penumbral salvage and difference in National Institutes of Health Stroke Scale (NIHSS) at discharge quantify the potential individual benefits of the intervention. Our models were developed on an extensive dataset from routine stroke care, which included 405 ischaemic stroke patients who underwent thrombectomy. We used acute data for training (n = 304), including multimodal CT imaging and clinical characteristics, along with post hoc markers such as thrombectomy success, final infarct localization and NIHSS at discharge. We benchmarked our tissue outcome predictions under the observed reperfusion scenario against a thresholding-based clinical method and a generalized linear model. Our deep learning model showed significant superiority, with a mean Dice score of 0.48 on internal test data (n = 50) and 0.52 on external test data (n = 51), versus 0.26/0.36 and 0.34/0.35 for the baselines, respectively. The NIHSS sum score prediction achieved median absolute errors of 1.5 NIHSS points on the internal test dataset and 3.0 NIHSS points on the external test dataset, outperforming other machine learning models. By predicting the patient-specific response to thrombectomy for both tissue and clinical outcomes, our approach offers an innovative biomarker that captures the dynamics of cerebral ischaemia. We believe this method holds significant potential to enhance personalized therapeutic strategies and to facilitate efficient resource allocation in acute stroke care.

Artificial intelligence (AI) has significant potential in healthcare applications, but its training and deployment faces challenges due to healthcare's diverse data, complex tasks, and the need to preserve privacy. Foundation models that perform well on medical tasks and require less task-specific tuning data are critical to accelerate the development of healthcare AI applications. We introduce MedGemma, a collection of medical vision-language foundation models based on Gemma 3 4B and 27B. MedGemma demonstrates advanced medical understanding and reasoning on images and text, significantly exceeding the performance of similar-sized generative models and approaching the performance of task-specific models, while maintaining the general capabilities of the Gemma 3 base models. For out-of-distribution tasks, MedGemma achieves 2.6-10% improvement on medical multimodal question answering, 15.5-18.1% improvement on chest X-ray finding classification, and 10.8% improvement on agentic evaluations compared to the base models. Fine-tuning MedGemma further improves performance in subdomains, reducing errors in electronic health record information retrieval by 50% and reaching comparable performance to existing specialized state-of-the-art methods for pneumothorax classification and histopathology patch classification. We additionally introduce MedSigLIP, a medically-tuned vision encoder derived from SigLIP. MedSigLIP powers the visual understanding capabilities of MedGemma and as an encoder achieves comparable or better performance than specialized medical image encoders. Taken together, the MedGemma collection provides a strong foundation of medical image and text capabilities, with potential to significantly accelerate medical research and development of downstream applications. The MedGemma collection, including tutorials and model weights, can be found at https://goo.gle/medgemma.

Brain tumor resection is a complex procedure with significant implications for patient survival and quality of life. Predictions of patient outcomes provide clinicians and patients the opportunity to select the most suitable onco-functional balance. In this study, global features derived from structural magnetic resonance imaging in a clinical dataset of 49 pre- and post-surgery patients identified potential biomarkers associated with survival outcomes. We propose a framework that integrates Explainable AI (XAI) with neuroimaging-based feature engineering for survival assessment, offering guidance for surgical decision-making. In this study, we introduce a global explanation optimizer that refines survival-related feature attribution in deep learning models, enhancing interpretability and reliability. Our findings suggest that survival is influenced by alterations in regions associated with cognitive and sensory functions, indicating the importance of preserving areas involved in decision-making and emotional regulation during surgery to improve outcomes. The global explanation optimizer improves both fidelity and comprehensibility of explanations compared to state-of-the-art XAI methods. It effectively identifies survival-related variability, underscoring its relevance in precision medicine for brain tumor treatment.

Synthetic images are an option for augmenting limited medical imaging datasets to improve the performance of various machine learning models. A common metric for evaluating synthetic image quality is the Fr\'echet Inception Distance (FID) which measures the similarity of two image datasets. In this study we evaluate the relationship between this metric and the improvement which synthetic images, generated by a Progressively Growing Generative Adversarial Network (PGGAN), grant when augmenting Diabetes-related Macular Edema (DME) intraretinal fluid segmentation performed by a U-Net model with limited amounts of training data. We find that the behaviour of augmenting with standard and synthetic images agrees with previously conducted experiments. Additionally, we show that dissimilar (high FID) datasets do not improve segmentation significantly. As FID between the training and augmenting datasets decreases, the augmentation datasets are shown to contribute to significant and robust improvements in image segmentation. Finally, we find that there is significant evidence to suggest that synthetic and standard augmentations follow separate log-normal trends between FID and improvements in model performance, with synthetic data proving more effective than standard augmentation techniques. Our findings show that more similar datasets (lower FID) will be more effective at improving U-Net performance, however, the results also suggest that this improvement may only occur when images are sufficiently dissimilar.

X-ray imaging is a rapid and cost-effective tool for visualizing internal human anatomy. While multi-view X-ray imaging provides complementary information that enhances diagnosis, intervention, and education, acquiring images from multiple angles increases radiation exposure and complicates clinical workflows. To address these challenges, we propose a novel view-conditioned diffusion model for synthesizing multi-view X-ray images from a single view. Unlike prior methods, which are limited in angular range, resolution, and image quality, our approach leverages the Diffusion Transformer to preserve fine details and employs a weak-to-strong training strategy for stable high-resolution image generation. Experimental results demonstrate that our method generates higher-resolution outputs with improved control over viewing angles. This capability has significant implications not only for clinical applications but also for medical education and data extension, enabling the creation of diverse, high-quality datasets for training and analysis. Our code is available at GitHub.

Deep neural networks are increasingly applied for automated histopathology. Yet, whole-slide images (WSIs) are often acquired at gigapixel sizes, rendering it computationally infeasible to analyze them entirely at high resolution. Diagnostic labels are largely available only at the slide-level, because expert annotation of images at a finer (patch) level is both laborious and expensive. Moreover, regions with diagnostic information typically occupy only a small fraction of the WSI, making it inefficient to examine the entire slide at full resolution. Here, we propose SASHA -- {\it S}equential {\it A}ttention-based {\it S}ampling for {\it H}istopathological {\it A}nalysis -- a deep reinforcement learning approach for efficient analysis of histopathological images. First, SASHA learns informative features with a lightweight hierarchical, attention-based multiple instance learning (MIL) model. Second, SASHA samples intelligently and zooms selectively into a small fraction (10-20\%) of high-resolution patches, to achieve reliable diagnosis. We show that SASHA matches state-of-the-art methods that analyze the WSI fully at high-resolution, albeit at a fraction of their computational and memory costs. In addition, it significantly outperforms competing, sparse sampling methods. We propose SASHA as an intelligent sampling model for medical imaging challenges that involve automated diagnosis with exceptionally large images containing sparsely informative features.

Colorectal cancer (CRC) is closely linked to the malignant transformation of colorectal polyps, making early detection essential. However, current models struggle with detecting small lesions, accurately localizing boundaries, and providing interpretable decisions. To address these issues, we propose HGNet, which integrates High-Order Spatial Awareness Hypergraph and Multi-Scale Context Attention. Key innovations include: (1) an Efficient Multi-Scale Context Attention (EMCA) module to enhance lesion feature representation and boundary modeling; (2) the deployment of a spatial hypergraph convolution module before the detection head to capture higher-order spatial relationships between nodes; (3) the application of transfer learning to address the scarcity of medical image data; and (4) Eigen Class Activation Map (Eigen-CAM) for decision visualization. Experimental results show that HGNet achieves 94% accuracy, 90.6% recall, and 90% [email protected], significantly improving small lesion differentiation and clinical interpretability. The source code will be made publicly available upon publication of this paper.

Accurate detection of anatomic landmarks is essential for assessing alveolar bone and root conditions, thereby optimizing clinical outcomes in orthodontics, periodontics, and implant dentistry. Manual annotation of landmarks on cone-beam computed tomography (CBCT) by dentists is time-consuming, labor-intensive, and subject to inter-observer variability. Deep learning-based automated methods present a promising approach to streamline this process efficiently. However, the scarcity of training data and the high cost of expert annotations hinder the adoption of conventional deep learning techniques. To overcome these challenges, we introduce GeoSapiens, a novel few-shot learning framework designed for robust dental landmark detection using limited annotated CBCT of anterior teeth. Our GeoSapiens framework comprises two key components: (1) a robust baseline adapted from Sapiens, a foundational model that has achieved state-of-the-art performance in human-centric vision tasks, and (2) a novel geometric loss function that improves the model's capacity to capture critical geometric relationships among anatomical structures. Experiments conducted on our collected dataset of anterior teeth landmarks revealed that GeoSapiens surpassed existing landmark detection methods, outperforming the leading approach by an 8.18% higher success detection rate at a strict 0.5 mm threshold-a standard widely recognized in dental diagnostics. Code is available at: https://github.com/xmed-lab/GeoSapiens.

Quantitative imaging (QI) is demonstrating strong promise across multiple clinical applications. For clinical translation of QI methods, objective evaluation on clinically relevant tasks is essential. To address this need, multiple evaluation strategies are being developed. In this paper, based on previous literature, we outline four emerging frameworks to perform evaluation studies of QI methods. We first discuss the use of virtual imaging trials (VITs) to evaluate QI methods. Next, we outline a no-gold-standard evaluation framework to clinically evaluate QI methods without ground truth. Third, a framework to evaluate QI methods for joint detection and quantification tasks is outlined. Finally, we outline a framework to evaluate QI methods that output multi-dimensional parameters, such as radiomic features. We review these frameworks, discussing their utilities and limitations. Further, we examine future research areas in evaluation of QI methods. Given the recent advancements in PET, including long axial field-of-view scanners and the development of artificial-intelligence algorithms, we present these frameworks in the context of PET.

Large language models (LLMs) have shown promising potential in analyzing complex textual data, including radiological reports. These models can assist clinicians, particularly those with limited experience, by integrating and presenting diagnostic criteria within radiological classifications. However, before clinical adoption, LLMs must be rigorously validated by medical professionals to ensure accuracy, especially in the context of advanced radiological classification systems. This study evaluates the performance of four LLMs-ChatGPT-4o, AmbossGPT, Claude 3.5 Sonnet, and Gemini 2.0 Flash-in classifying fractures based on the AO classification system using CT reports. A dataset of 292 fictitious physician-generated CT reports, representing 310 fractures, was used to assess the accuracy of each LLM in AO fracture classification retrospectively. Performance was evaluated by comparing the models' classifications to ground truth labels, with accuracy rates analyzed across different fracture types and subtypes. ChatGPT-4o and AmbossGPT achieved the highest overall accuracy (74.6 and 74.3%, respectively), outperforming Claude 3.5 Sonnet (69.5%) and Gemini 2.0 Flash (62.7%). Statistically significant differences were observed in fracture type classification, particularly between ChatGPT-4o and Gemini 2.0 Flash (Δ12%, p < 0.001). While all models demonstrated strong bone recognition rates (90-99%), their accuracy in fracture subtype classification remained lower (71-77%), indicating limitations in nuanced diagnostic categorization. LLMs show potential in assisting radiologists with initial fracture classification, particularly in high-volume or resource-limited settings. However, their performance remains inconsistent for detailed subtype classification, highlighting the need for further refinement and validation before clinical integration in advanced diagnostic workflows.