Fracture detection on radiographs can be challenging, particularly for subtle or nondisplaced injuries, while development of artificial intelligence (AI) models often requires programming expertise and specialized computational resources. This proof-of-concept study evaluated whether a commercially available code-free automated machine-learning platform could distinguish fractured from non-fractured bone radiographs. A publicly available heterogeneous dataset containing 420 radiographic images, including 130 labeled as fractured and 290 labeled as non-fractured, was imported into Google Cloud Vertex AI AutoML Image Classification (Google, Inc., Mountain View, CA, USA) and divided into 336 training, 42 validation, and 42 test images. The dataset included radiographs from multiple anatomical regions; however, the publicly available metadata did not provide a sufficiently detailed anatomical distribution to reliably quantify upper-extremity, lower-extremity, spinal, or other fracture categories. Importantly, these counts represent radiographic images rather than confirmed independent patients, and patient-level distribution and independence across the training, validation, and test subsets could not be verified. No additional preprocessing, augmentation, or relabeling was performed by the investigators. Model performance was assessed using average precision, platform-reported aggregate precision and recall, confidence-threshold analysis, and a row-normalized confusion matrix. The trained classifier achieved a micro-averaged area under the precision-recall curve (AuPRC), reported by Vertex AI as average precision, of 0.987. At a confidence threshold of 0.50, model-level precision and recall were each 90.5%. Because these values summarize performance across both classes, they should not be interpreted as fracture-specific sensitivity. Class-specific average precision was 0.994 for non-fractured images and 0.972 for fractured images, while fracture-class recall was 77%. The row-normalized confusion matrix demonstrated correct classification of 77% of fractured radiographs and 97% of non-fractured radiographs. However, the retained Vertex AI output provided normalized percentages rather than raw class counts; therefore, the absolute numbers of correctly and incorrectly classified fractured and non-fractured images could not be verified. Given the small 42-image test set, the absence of raw counts limits the precision, interpretability, and reproducibility of these class-specific performance estimates and should be considered an important methodological limitation. The model was also successfully deployed to an online prediction endpoint and generated high-confidence classifications for representative images. These findings demonstrate that code-free AutoML can support rapid development and deployment of a functional, internally evaluated fracture-classification model without custom programming. However, the 77% fracture-class recall and small 42-image test set indicate that the model's performance should be interpreted cautiously and not as evidence of clinical diagnostic effectiveness. The small test set, class imbalance, limited clinical metadata, uncertain patient-level independence, and absence of external validation limit generalizability. Code-free AutoML may therefore be useful for proof-of-concept medical imaging research and education, but external validation is required before clinical implementation.

Understanding how visual information processing unfolds over time requires models that not only predict neural responses but also expose the representations that support them and generalize beyond sampled stimulus spaces. Here we introduce Img2EEG, a participant-specific image-to-EEG encoding framework that integrates hierarchical visual and semantic representations to generate temporally resolved multichannel EEG responses. Trained on THINGS EEG2, Img2EEG generalized to unseen images while preserving stimulus-specific and participant-specific response structure. Controlled perturbations of internal representations and visual inputs revealed distinct temporally structured contributions of visual and semantic information, and in silico experiments reproduced classic human neural responses, such as the face-sensitive N170, while enabling targeted representational interventions. Scaling Img2EEG to 1.28 million ImageNet images produced over 12 million synthetic EEG responses that supported cross-dataset visual reconstruction and improved the behavioral alignment of an artificial vision model. Img2EEG provides an interpretable and scalable framework for experimentally manipulable modeling of visual neural dynamics.

Sensitive human data underpin discoveries across medicine, biology and the social sciences, yet privacy regulation often prevents sharing them with collaborators or artificial intelligence (AI) systems. We introduce SPHERE, a model-free method that makes sensitive datasets directly usable by AI and shareable for open science as a synthetic twin, while the original records never leave the local environment. Across 33 datasets spanning five scientific domains, SPHERE protects individual privacy against adversarial re-identification attacks while preserving the data's statistical structure: means, variances and correlations are reproduced exactly, effect size and P value in linear statistical analysis are numerically identical, nonlinear machine-learning utility is retained, and each twin is generated in seconds on a laptop. Frontier AI agents running on the twin reach the same scientific conclusions as on the original records. Analyses of the twin reproduce genome- and proteome-wide results at UK Biobank scale and recover the findings of landmark studies across three independent cohorts and consortia. The approach also extends to deep-learning embeddings across language, vision and time-series, with minimal utility loss. We make the Stanford Alzheimer's Disease Research Center cohort openly available for the first time, as a SPHERE twin spanning nine modalities that any registered researcher can analyze without an approval process. We release SPHERE with certification of each twin's privacy and fidelity, and an AI agent that autonomously executes research tasks on sensitive data without ever accessing it. Sensitive datasets that are currently closed to research could thus become routine inputs to open science and AI to enable key discoveries.

The operating room is evolving into an information-rich environment where surgical care is increasingly shaped by integrated digital architectures. The proliferation of minimally invasive and robotic platforms has generated a massive influx of high-resolution video, shifting the field toward real-time computational analysis. Artificial intelligence, particularly computer vision, is moving beyond retrospective analysis to extract structured intelligence from the operative field, redefining how procedures are analyzed and supported. This review focuses on the clinical applications of computer vision with the greatest translational relevance: Computer-aided detection, instrument tracking, procedural phase recognition, and video-based skill assessment. However, clinical translation remains constrained by practice variability, a lack of standardized open datasets, ethical and regulatory considerations, and the difficulty of ensuring model generalizability across diverse institutions. This review examines the current evidence for computer-aided detection, instrument tracking, phase recognition, and video-based skill assessment in surgery, and discusses their clinical readiness and key limitations.

Patients with advanced melanoma are treated with immune checkpoint inhibitors (ICIs), yet <50% of patients achieve a durable response while all patients are exposed to the risk of severe side effects. Tumor-infiltrating lymphocytes (TILs) in pathology images are associated with ICI outcomes, but manual assessment is subjective. In addition, the predictive value of other immune cell subsets, including plasma cells, neutrophils, histiocytes, and melanophages, remains unclear. We organized the Panoptic segmentation of nUclei and tissue in advanced MelanomA (PUMA) challenge to evaluate whether the spatial localization of TILs and other immune cell subsets on melanoma H&E slides collected before start of treatment was associated with treatment outcomes. Algorithm performance was evaluated on a hidden test set, after which top-ranked algorithms were applied to pre-treatment metastatic whole-slide images from a large, multicenter cohort of patients with advanced melanoma treated with first-line ICIs (n = 1102). Automatically quantified tissue features and immune cell subsets were then associated with clinical outcomes. Top-performing algorithms improved detection of immune cell subsets through pathology foundation model features, multi-stage tissue-context learning, and class-specific architectures, although accuracy for rare classes remained limited. Across challenge participants, TIL density showed the most consistent association with treatment response and survival. Associations for stromal TILs were weaker, while plasma cells, histiocytes, melanophages, neutrophils, necrosis and blood vessels did not show independent associations with outcomes. Overall, the PUMA challenge advanced immune cell detection in melanoma histopathology and showed that intratumoral lymphocytes are most consistently associated with treatment response and survival.

Manual Small-Incision Cataract Surgery (SICS) is a prevalent technique in low- and middle-income countries (LMICs). Automated analysis of SICS videos based on artificial intelligence (AI) would benefit self-evaluation, training, monitoring, and ultimately facilitate computer aided surgical assistance. However, this procedure remains understudied in terms of automated surgical analysis due to a lack of publicly available data. To advance this field, we organized the SICS-155 phase recognition challenge at MICCAI 2025. It was based on a dataset with 155 videos recorded at Sankara Eye Hospital in Bengaluru, India with an average duration of 13:05 minutes, annotated with 19 distinct surgical phases. In this work, we first present the results and findings of that challenge in accordance with the BIAS guidelines. Second, we integrated key technical innovations from the participating teams to develop a new, state-of-the-art approach for phase recognition in a post-operative, offline setting. This approach uses a boundary-aware FACT architecture, which we evaluated using videos of both SICS and phacoemulsification. Our approach achieved an accuracy of 87.98% [95%-CI: ±2.06], an edit score of 90.80% [95%-CI: ±1.47], and a segmental F1-score of 89.56% [95%-CI: ±1.92] on the SICS-155 challenge dataset, as well as an accuracy of 97.61%/ 93.36%, a precision of 98.42%/ 95.27%, and a recall of 97.97%/ 93.32% on the Cataract-101 phacoemulsification public dataset (depending on the data split). Building on these promising results, our future work will investigate potential real-time quality assessment and complication recognition in SICS and facilitate the development of clinical training and surgical quality monitoring software for ophthalmologists in LMICs.

The human brain's capacity to imagine visual scenes from memory is thought to rely on the medial temporal subsystem of the default mode network (MT-DMN), yet the neural codes supporting this ability remain poorly understood. We combined functional magnetic resonance imaging (fMRI) with vision and language artificial intelligence models to characterize neural codes during autobiographical imagination. Fifty participants imagined reexperiencing 20 natural scenarios while undergoing fMRI when cued by generic text prompts (e.g., wedding, exercising, and driving). Individual scenes were modeled using Stable Diffusion to generate personalized synthetic images from verbal descriptions of the scenarios imagined, collected beforehand. These depictions were then transformed into image recognition network hidden state representations. Representational similarity analysis revealed that the MT-DMN encoded the participant-specific representational structure of image representations, even when controlling for semantic features derived from a large language model. This effect was absent in other brain networks and during reading without active imagination, identifying the MT-DMN as a core substrate for the visual reconstruction of autobiographical experiences.

Ocular trauma represents a threat to crew safety and mission performance in space. Microgravity, confined environments, and exposure to particulate matter, chemicals, and mechanical hazards place astronauts at risk for corneal abrasions, open-globe injuries, chemical burns, lens dislocation, retinal detachment, orbital fractures, and barotrauma. Diagnostic capabilities during spaceflight remain limited by resources, lack of specialist expertise, and communication delays with Earth. Artificial intelligence, particularly convolutional neural networks and multimodal models, may help address these gaps through image interpretation, risk stratification, and longitudinal monitoring. Convolutional neural networks can extract hierarchical features from imaging data to identify subtle structural abnormalities, while multimodal models integrate imaging with clinical and environmental parameters to generate more comprehensive assessments. Terrestrial ophthalmology studies demonstrate the potential of these approaches across optical coherence tomography, ultrasound, fundus photography, and anterior-segment imaging. This review examines how these capabilities can be matched to ocular injuries during spaceflight, compares the suitability of different approaches across injury types, and identifies pathways toward autonomous care. Particular emphasis is given to spaceflight-related imaging and physiologic changes, constrained onboard hardware, and integration into workflows that support non-expert crew members. Collectively, these applications may expand diagnostic capabilities and enable earlier, more informed management during long-duration missions.

We propose CardioTransfer-X, a cross-cohort transfer learning framework within a related clinical benchmark family for tabular CVD risk prediction that yields performance comparable to training from scratch while preserving transparency. Predictive models are pre-trained on a composite multi-hospital heart disease dataset to learn generalized risk patterns, then fine-tuned on the smaller, distributionally distinct Cleveland Heart Disease dataset. To mitigate dataset shift and class imbalance, we introduce a stability-aware, correlation-based feature selection strategy and apply SMOTE only during training. CardioTransfer-X is instantiated with XGBoost, TabNet, and a MLP, enabling systematic comparison across tree-based and neural models. SHAP-based explanations provide global interpretability across architectures, with consistent feature importance rankings before and after fine-tuning. Fine-tuned XGBoost achieves 80.65±4.08% accuracy and 81.74±3.71% F1-score on the Cleveland dataset while maintaining stable source-domain performance, showing performance competitive with training from scratch, while shifting the decision boundary toward higher sensitivity and maintaining clinically meaningful feature attributions.

GeoSimp is a Spanish-language dataset for automatic text simplification (ATS), comprising approximately 3200 discourse segments derived from peer-reviewed geological scientific articles. The dataset was constructed to address two documented gaps in existing ATS resources: the limited availability of Spanish-language corpora for text simplification and text complexity classification, and the scarcity of domain-specific datasets focused on specialized scientific writing. Source texts were drawn exclusively from articles published in the Revista Geológica de América Central, a peer-reviewed, open-access journal, all of which are distributed under a single Creative Commons Attribution-NonCommercial-ShareAlike 3.0 (CC BY-NC-SA 3.0) license. Text extraction was performed manually to preserve structural coherence. Paragraphs were copied into Microsoft Excel spreadsheets, assigned unique identifiers, and subsequently segmented into discourse segments by one of the authors, a researcher with expertise in linguistics. Four annotators with university degrees in Spanish philology and professional experience in text editing performed attribute identification and simplification, using a predefined scheme of 21 linguistic complexity attributes. Prior to annotation, these annotators completed a structured training and calibration phase. The dataset is distributed as two Excel files. The training file (GeoSimp_train.xlsx) contains 2956 records, each comprising the original discourse segment, one simplified version, the annotator identifier, the identified complexity attributes, the paragraph identifier, and the identifier of the source article. The test file (GeoSimp_test.xlsx) contains 301 records, each comprising the original discourse segment and four independently produced simplified versions with their corresponding attribute annotations. A simplification guidelines manual used during annotation is also included in the repository. GeoSimp is suitable for training and evaluating ATS models, for developing text complexity classification models, and for research on paragraph-level simplification. The dataset is specifically oriented toward simplification for blind and low-vision users who rely on screen readers, a population underrepresented in the ATS literature. The explicit annotation of linguistic complexity attributes per segment supports reproducibility and fine-grained error analysis. The dataset is openly available in a public GitHub repository.

The rapid advancement of Assisted Reproductive Technology (ART) demands equally innovative approaches to professional training. Traditional educational models in reproductive medicine are often limited by inconsistent quality, variable clinical exposure, and prolonged learning curves. This paper proposes a comprehensive, AI-enhanced training framework designed to standardize and accelerate the education of clinicians, surgeons, and embryologists. The framework integrates six core domains: adaptive learning for theoretical knowledge, AI-driven simulations for clinical decision-making, virtual reality for surgical and embryology skills, large language model-based interactions for patient communication training, and automated tools for objective competency assessment. By leveraging technologies such as computer vision, machine learning, and immersive simulation, this platform is intended to deliberate practice in a safe, repeatable environment while potentially reducing training costs and ensuring uniform competency standards. It should be emphasized that this framework remains conceptual at this stage; it has not yet been implemented or empirically validated, and the described benefits are prospective rather than demonstrated. Future directions include the integration of digital twin technology and the development of ethical guidelines to support widespread implementation in reproductive medicine education.