GigaScience

The surge in medical imaging has spurred the development of vision-language models (VLMs) to alleviate radiologist workloads. However, clinical deployment is hindered by the lack of meaningful evaluation frameworks. Current metrics - ranging from semantic similarity to large language model (LLM) based judges - often fail to distinguish between clinically trivial and critical discrepancies, poorly reflecting real-world clinical judgment. To address this, we introduce DISCERN (Discordance and Significance-aware Entity-level Radiology Report Comparison). DISCERN is a significance-aware framework that weighs report errors based on their potential impact on patient care. Our results demonstrate that DISCERN powered by closed source LLMs aligns more closely with expert radiologist assessments than traditional metrics or current LLM evaluators, providing a more interpretable and clinically relevant benchmark. By modeling radiologist prioritization and entity-level feedback, DISCERN facilitates targeted model refinement and ensures the safer integration of generative AI into clinical workflows.

Background: Generative AI tools can support data-intensive research by writing code, drafting prose, searching analytical possibilities, and stress-testing claims. They can also produce false citations, drift between statistical specifications, and lose continuity across long investigations. This paper describes a practical workflow for using AI systems in empirical research while keeping discovery, verification, and accountability inspectable. Methods: We developed and applied a three-phase human-AI workflow to a case study of 14 elite Ethiopian distance runners. The dataset contained 22,605 GPS-segments collected across 97 consecutive days in late 2025, supplemented by venue and athlete metadata collected in the field. Phase 1 used an autonomous data-exploration tool to pre-filter the hypothesis space across five seeded research questions. Phase 2 used an AI system under direct human guidance to construct candidate findings into numerical claims, verification scripts, and draft text. Phase 3 used an independent AI system in an adversarial role to stress-test methods, statistics, prose, figures, and citations. The workflow was informed by Pearl's distinction between association, intervention, and counterfactual reasoning, with human judgement retained for research direction, interpretation, and final claims. Results: The workflow produced three empirical analyses and a documented correction process. The analyses estimated an altitude-to-sea-level pace correction of +0.10 min/km per 1,000 m at matched heart rate, showed why pooled altitude-surface regression was not identifiable within this venue system, documented method-dependence in heart-rate-based intensity classification, characterised within-venue route variation as a 64/36 path-fixed-to-trail-variable split with the Sululta label resolving into two functionally distinct sub-venues, and reframed the cohort's training through a 3x3x3 prescription lattice grounded in Ethiopian coaching practice. The adversarial phase identified several hallucinated citations, a terminology error between HC1 and cluster-robust standard errors, and several inconsistencies between prose, figures, and computed results. Verification scripts re-derived nearly all numerical claims from the cleaned lap-level data. Conclusions: The case study shows how researchers can organise AI-assisted empirical work so that candidate discovery, claim construction, independent stress-testing, and final accountability remain separated. The workflow did not remove the need for domain expertise or human judgement. Its value was in making the route from candidate finding to manuscript claim explicit, reproducible, and open to challenge. Trial registration: Not applicable.

Longitudinal evaluations of national electronic health record repositories often track document counts alone, obscuring changes in content size, structure and standards implementation. We decomposed growth in the Estonian Health Information System across document counts, per-document size, section-level structure and version uptake in a 10% random population sample of 4.97 million HL7 Clinical Document Architecture Release 2 documents from 147,819 patients, spanning 2012--2019 and four prespecified document types. Growth patterns differed by document type. Inpatient summaries increased 48.5% in total content volume despite a 2.4% decline in document counts. Section presence and within-section content were highly skewed; 44.6% of 892 data locations carried one fixed value. Code-system diversity increased from 45 to 79, and version uptake took years: inpatient summaries reached 80% organisational uptake after a median 44 months (95% CI 11--78). This decomposition can guide extraction pipelines, secondary use and standards governance in CDA- and FHIR-based repositories.

Precision medicine aims to advance our ability from a "one-size-fits-all" approach to personalized and predictive healthcare across diverse populations. It promotes integration of multi-omics and phenotypic data to understand disease mechanisms and discover novel biomarkers and risk factors, which could be used to predict and prevent critical diseases in individual patients across diverse populations. The potential implications of precision medicine approach can accelerate our ability to classify patients at higher risk of developing critical diseases, improve diagnostic capabilities, develop deeper understanding of individual risk, investigate racial differences and demographic characteristics, and find relationships between genetic variants, expressions, and diseases. This study focuses on implementing an innovative and data driven framework of translational bioinformatics and Machine Learning (ML) techniques to analyze multi-omics, including RNA-seq and Whole-Genome Sequencing (WGS) data, generated using blood samples of randomly consented patients. First, we utilized bioinformatics pipelines to identify differentially expressed genes and their pathogenic and likely pathogenic variants for the downstream data analysis, annotation, and visualization. Then, applied a nexus of ML models for multi-omics biomarker discovery, disease prediction, density-based clustering, single-patient profiling, and pathogenicity classification. WGS data analysis supported the exploration of genetic variation and diversity among patients to identify known and novel biomarkers, whereas RNA-seq data analysis improved our understanding of functional and biological pathways that underlying disease states. We classified and clustered pathogenic variants and expressions across various genes and discovered numerous diseases leading risk factors. Our results include gene-disease associations and captured common pathways across the broader population, demonstrating a level of sensitivity and accuracy that has broad clinical implications. We validated our results through clinical records, and state of the science literature. This study delves into the strengths of multi-omics data integration and capabilities of ML application in genetically diverse and complex patient cohorts. Our approach has the potential to elucidate complex gene-disease interactions for genetically diverse populations, which can support earlier diagnoses for patients in many disease realms.

The availability of electronic health records has facilitated data-driven approaches to the understanding of multimorbidity, with clustering becoming a common tool for uncovering relevant groups of associated conditions. Previous studies, however, have found challenges in their reproducibility, with wide disparity in the reported clusters. At the core of this issue lays a vagueness of the definition of a cluster, leading to a lack of standards in their methods and evaluation, while implementation details are often not completely reported or explicit in their assumptions. We present a methodological pipeline that can be adapted to different cluster definitions (e.g. multiple cluster membership or clusters where all nodes are mutually associated) and a set of scores that can be composed into an evaluation metric that explicitly incorporates assumptions that align with the research aims. We apply our pipeline to a healthcare dataset of over 7 million patients in England and show how clusters may drastically differ when varying the parameter choices, exposing the risks of reporting a single clustering realisation. Our methodological pipeline, evaluation framework, and tools for analysis and network visualisation serve as a reference to transparently explore and align methodological decisions to the aims of multimorbidity clustering, contributing to overcome the reproducibility challenges of the field.

One of the prominent methods in neuroimaging data processing is SSM-PCA, which is based on principal component analysis and allows for the identification of diagnostically significant patterns in the form of statistical maps. We developed software, PIE Toolbox, employs SSM-PCA and classification based on the obtained diagnostic patterns revealed from functional and structural tomographic brain imaging. The program supports the entire analysis pipeline including preprocessing of brain images, diagnostic patterns extraction, building classification models, and prediction based on them. The resulting diagnostic patterns are weighted principal components obtained through SSM-PCA, or their linear combinations. PIE Toolbox allows selection of relevant structural and functional brain patterns, computation of their expression values in regions of interest, classification using support vector machines, and evaluation of model performance via cross-validation. This approach enables the use of patterns as features of intergroup differences for individual diagnosis. The software has been validated on both simulated and ADNI datasets.

We present the TopBrain 2025 Challenge, the first benchmark for fine-grained multiclass segmentation of the whole brain vasculature in both computed tomography angiography (CTA) and magnetic resonance angiography (MRA). Building on the TopCoW challenge, TopBrain scales vessel annotation from the Circle of Willis to the entire brain, introducing a dataset of 90 annotated volumes across 48 landmark vessel classes spanning arterial and venous systems, of which 50 training volumes are publicly released. Vessel definitions were consolidated from established neuroanatomical references into a unified annotation scheme, and vessel caliber measurements along the centerline are reported for the first time across the whole brain vascular anatomy. To address the unique challenges of multiclass brain vessel segmentation, we propose an evaluation framework that accounts for detection in segmentation performance, assesses anatomical plausibility, and introduces novel contamination metrics that characterize inter-class prediction errors. Fifteen teams from over 220 registered participants submitted algorithms to the benchmark. The top-performing teams built on nnUNet with principled system design choices, achieving around 80% Dice scores, near-zero invalid neighbor counts, over 60% F1 scores for side-road vessels, and below 18% foreground contamination ratio. Larger vessels are easier to segment, while smaller and more complex vessels remain the true bottleneck. The annotated datasets and podium-finish algorithms are made publicly available on Zenodo.

Although GLP1/GIP receptor agonists demonstrate unprecedented weight loss efficacy, their rapid clinical adoption has revealed significant real-world tolerability challenges. To evaluate their dynamic safety profiles, we developed a macro to micro pharmacovigilance framework by combining global FAERS reports with local UT Physician EHR. Macroscopically, we distilled 17 shared adverse events across the drug class from FAERS with disproportionality analysis. Microscopically, local EHR data (289,655 longitudinal treatment sessions across 71,316 patients) revealed 51.6% of GLP1 sessions terminated within 90 days. Furthermore, temporal stratified logistic regression demonstrated that initial exposure (0 to 30 days) correlated strongly with nausea and vomiting, which attenuated in extended sessions, whereas extended exposure (>2 years) uncovered late onset risks, notably incident hepatic steatosis. Ultimately, this time aware framework reveals that GLP1 safety profiles are profoundly duration dependent, providing critical insights into both acute intolerances and long-term medication safety.

Objective. Biomedical knowledge graphs (KGs) such as PrimeKG, Hetionet, UMLS, and PharmGKB are increasingly used as the substrate for downstream machine-learning, retrieval-augmented generation, drug-repurposing, and electronic health record (EHR) augmentation pipelines. The dominant assumption in published work is that integrating two or more such KGs is a tractable engineering step solved by identifier (ID) matching. This paper interrogates that assumption empirically. We quantify how much concept overlap survives realistic alignment, and we characterize the new failure modes introduced by the methods that practitioners reach for when ID matching is insufficient. Materials and Methods. We compared four widely used biomedical KGs (PrimeKG, Hetionet v1.0, the full UMLS Metathesaurus, and PharmGKB) across eleven node types using a tiered alignment pipeline: (1) direct ID matching for nodes sharing a primary vocabulary; (2) cross-ontology bridging using standard mappings (e.g., MONDO-DOID, HPO-UMLS, HPO-UMLS-MeSH for side effects, NCBI Gene-HGNC-UMLS, UBERON-FMA/SNOMEDCT_US/NCI/MeSH for anatomy); (3) ClinicalBERT cosine-similarity grouping at threshold >= 0.98 for over-segmented disease nodes, with a deterministic suffix-stripping canonicalizer; (4) exact name matching for ontology-poor types (anatomy, REACTOME pathways); and (5) embedding-based fuzzy matching with UMLS lookup (SapBERT and ClinicalBERT) for free-text microbiome concepts. We applied the pipeline to a 698-concept gut-microbiome benchmark spanning taxa, pathways, and disease labels, validated grouping decisions against the curated SSSOM mappings released by the MONDO project, and audited the ClinicalBERT consolidation against five clinical-genetics case studies drawn from the literature. Results. Per-type pairwise coverage was strikingly asymmetric. Genes/proteins and the three Gene Ontology categories aligned cleanly across PrimeKG and Hetionet (mutual coverage 94-99%), but disease overlap was sparse: only 0.7% of PrimeKG individual disease nodes mapped to Hetionet, rising to 2.0% after MONDO grouping (versus 78.7% and 18.4% from the Hetionet side). PrimeKG-to-UMLS coverage spanned 100% (effect/phenotype via HPO) down to 20.8% (REACTOME pathways), with drugs at 73.7% and anatomy at 58.8%. PrimeKG-to-PharmGKB drug coverage required up to two bridging hops (DrugBank -> UMLS -> RxNorm/ATC/MeSH). Bigger was not uniformly more complete: on a 698-concept microbiome drug benchmark, Hetionet missed 0 concepts while PrimeKG missed 16. ClinicalBERT-based grouping consolidated 22,205 raw MONDO disease nodes into 17,080 groups but introduced three reproducible failure modes documented in case studies: (i) peer over-merging: for example, all 22 osteogenesis imperfecta subtypes collapsed into a single node despite distinct severity classes; (ii) parent-child collapse: e.g. acute myeloid leukemia merged with myeloid leukemia, erasing the acute/chronic distinction that drives clinical management; and (iii) lexical false positives: neurofibromatosis and schwannomatosis grouped together despite cellular-pathology differences. Discussion. Identifier matching alone is a weak baseline for biomedical KG integration. Cross-ontology bridges and embedding-based consolidation expand coverage but do so at the cost of clinically meaningful resolution, and the resulting failures are systematic rather than random. Reporting only aggregate coverage statistics obscures these losses, which propagate silently into downstream tasks. Conclusion. We provide reusable per-type coverage tables, a taxonomy of three integration failure modes, and concrete recommendations for downstream studies that depend on a unified biomedical KG. We argue that future KG integration work should report per-type coverage and per-cluster confidence rather than aggregate match rates.

Background: Three-dimensional visualization and quantitative analysis of cerebral arteries on 3DRA are central to endovascular treatment planning, device selection, and cerebrovascular research. Manual segmentation is time-consuming and operator-dependent, yet no open-source deep learning model has been prospectively validated for this task on 3DRA. Methods: A nnUNet v2 model was trained for binary cerebral artery segmentation on 400 consecutive 3DRA acquisitions from three angiographic systems, comparing four configurations across architectures and loss functions. The best-performing configurations were prospectively validated on 40 patients using a dual approach: quantitative metrics (DSC, clDice, HD95, ASD, Precision, Recall), and blinded expert qualitative evaluation by two interventional neuroradiologists assessing 12 arterial segments, a global quality score, and clinical usability across 40 test cases. Results: The ensemble model achieved median DSC 0.917, clDice 0.932, and HD95 1.494 mm. Global quality scores were significantly lower for nnUNet v2 than for expert segmentations (median 4 vs 5, p<0.001), but nnUNet v2 segmentations were rated clinically usable in 88-90% of cases versus 95-98% for expert segmentations, without significant difference on the binary usability criterion. A consistent proximal-to-distal quality gradient was identified, with comparable scores at proximal arteries and the largest differences at distal arterial segments. Conclusion: nnUNet v2 with topology-aware training provides clinically usable cerebral artery segmentations on 3DRA, prospectively validated through both quantitative metrics and structured expert qualitative assessment, and represents a reproducible open-source foundation for endovascular and research applications.

Pediatric patients with cardiac implantable electronic devices (CIEDs) face limited MRI access due to RF-induced heating, and computational modeling is increasingly used to characterize this risk. The validity of these simulations, however, depends on pairing body models with clinically realistic lead configurations, guidance that is currently lacking. We retrospectively analyzed 302 CIED surgeries in 281 pediatric patients to derive weight-based constraints for simulation design. Weight alone discriminated epicardial from endocardial lead implantation with AUC = 0.90, and adding age and height yielded no improvement, supporting weight as a sufficient single-parameter selection metric. The probabilistic crossover between approaches occurred at 44~kg, substantially higher than the 10 to 15~kg threshold commonly cited in the literature, with a broad transition zone of 21 to 66~kg in which both lead types were routinely used. Lead length was likewise weight-constrained: only 25~cm leads were observed in patients below 6~kg, and leads of 45~cm or longer were uncommon below 50~kg. These findings yield a three-tier framework, with epicardial-only configurations below 21~kg, dual configurations within 21 to 66~kg, and weight-thresholded lead lengths throughout, enabling MRI safety simulations to focus on clinically realizable anatomy and device combinations.

A closed-loop quality system deployed across thirteen US hospital sites resolved physician complaints with zero regressions on 42 tracked cases across 1,089 optimization iterations, while a deterministic assembly-agent replacement cut H+P trace latency from 19.6 s to 10.8 s (-8.8 s, 95% CI [-10.5, -7.1] s; n = 100 pre, n = 100 post). We report four observations and an architectural follow-through. First, the same binary-check instrument produces opposite outcomes depending on the question asked: "maximize this score" produces structurally-correct notes that physicians reject (Spearman rho = -0.077, 95% CI [-0.40, 0.26], n = 36); "did this specific fabrication stop?" produces rater-invariant deployment decisions. Second, in our pipeline, assembly-stage agents did not respond to prompt optimization the way reasoning agents did: four consecutive optimization attempts produced 18-28 point regressions. Third, physician preference is rater-fragile at typical clinical-AI calibration sample sizes (Cohen's kappa = 0.028 between two board-certified physicians, 95% CI [-0.30, 0.36] on n = 35 overlapping pairs). Fourth, the architectural punchline: six weeks after the prediction, the LLM call at the chart-assembly step was replaced with a deterministic renderer (sub-500-character template plus sandboxed scripting), lifting the defect-free rate on a 51-case holdout from 49% to 84%. We introduce a Pareto-with-absolute-floors acceptance rule (multi-axis commit with severity-class categorical vetoes) as a methodological contribution distinct from scalar-reward acceptance in standard prompt-optimization frameworks. Cross-iteration rejection memory prevents the loop from re-proposing edits already rejected three or more times. A reproducibility bundle (anonymized ablation per-case counts, bootstrap-CI data, analysis scripts) is released under CC BY 4.0 at github.com/sayvant/SQS-Auditor-paper-data.

Background: Store-and-forward teledermatology commonly relies on several patient-submitted photographs of the same concern, but most dermatology artificial intelligence models classify single images independently. Objective: To develop and internally validate a case-level diagnostic-support model that aggregates multiple patient-submitted photographs for common dermatologic conditions. Methods: We conducted a retrospective diagnostic-modeling study using the Skin Condition Image Network, a public dataset of deidentified self-taken dermatology images from US adults. We curated 2,336 cases comprising 5,041 images across 10 common inflammatory, allergic, and infectious conditions. Cases were split at the submission level into training, validation, and held-out test sets. Frozen general-purpose and dermatology-specific encoders were compared with image-level classifiers and a gated-attention multiple instance learning model that generated one case-level output from 1-3 images. Results: The strongest image-level baseline, dermatology-specific embeddings with random forest classification, achieved macro/micro ROC-AUCs of 0.797/0.854. Case-level aggregation improved discrimination, with dermatology-specific embeddings plus multiple instance learning achieving mean macro/micro ROC-AUCs of 0.819/0.863 across repeated stratified experiments. The locked final model achieved macro/micro ROC-AUCs of 0.800/0.849 on the held-out test set. Balanced-threshold sensitivity/specificity examples were 0.702/0.688 for eczema and 0.818/0.826 for urticaria. Limitations: Internal validation used a 10-condition subset from a US volunteer dataset; external validation, calibration, subgroup performance analysis, and prospective workflow studies are required. Conclusion: Modeling the teledermatology submission as a multi-image case better reflects asynchronous dermatology workflow than single-image classification. The model is preliminary clinician-facing support for structured review and triage, not autonomous diagnosis.

Hypoxia-targeted BOLD MRI is a novel technique, which probes oxygenation physiology in response to a controlled transient hypoxia stimulus. In glioblastoma, the signal response is spatially and temporally heterogeneous. We developed a voxel-wise temporal decomposition framework for hypoxia-targeted BOLD MRI that separates the arrival of responses, transition phases, and steady state during controlled isocapnic hypoxia. Twenty healthy controls underwent 3-T BOLD MRI during a double hypoxic step challenge to establish a normative reference. Three patients with newly diagnosed glioblastoma were included as proof-of-concept cases. For each voxel, we estimated response arrival delay (Delaycorr), delay to plateau, delay to return and an O2-normalized steady-state response (HypoxiaSS). Healthy-control maps were used to construct a voxel-wise normative atlas and, for HypoxiaSS, a global-response-adjusted model for patient deviation mapping. In healthy controls, HypoxiaSS showed lower supratentorial between-subject variabilitythan both whole-stimulus comparators (coefficient of variation: 1.77 versus 2.36 for Hypoxiaavg) and higher voxel-level step-to-step agreement (ICC(2,1): median 0.951 versus 0.792 for Hypoxiaavg). Whole-stimulus averaging exhibited a systematic step-2 signal amplification present in 19 of 20 subjects, which was absent from HypoxiaSS. Asingle global response scalar explained a median 72.5% of voxel-wise between-subject variance in HypoxiaSS. In proof-of-concept patient analyses, G-adjusted HypoxiaSS deviation maps and timing maps identified spatially coherentabnormalities that were partly complementary and extended beyond conventional MRI-defined lesion margins.Temporal decomposition improves the stability and interpretability of hypoxia-targeted BOLD MRI and provides a practical framework for population-referenced physiological mapping and atlas-based deviation mapping in glioblastoma.

Background: Traditional diagnostic models lack explainability, while multimodal language models prone to hallucination remain unsafe for medical education. An interactive, risk-free artificial intelligence framework is required to serve as a reliable clinical mentor for radiology trainees. Methods: We propose a multi-agent architecture decoupling deterministic image analysis from generative consultation. Specialized computer vision models perform anatomical localization and pathological segmentation. These quantitative outputs are synthesized into a structured payload, which grounds a locally hosted large language model (LLaVA 7B) using strict prompt guardrails and prerequisite protocols. Results: The system effectively eliminates visual hallucinations by intercepting unanchored queries. The artificial intelligence tutor successfully contextualizes spatial anomalies and baseline metrics, generating accurate conversational explanations and formally structured radiology reports while strictly enforcing medical safety disclaimers. Discussion and Conclusion: By anchoring language generation exclusively to verified algorithmic realities, this framework transforms opaque diagnostic models into safe, interactive educational simulators. This establishes a highly reliable paradigm for integrating explainable artificial intelligence into medical training.

To evaluate the international interoperability of food composition databases, we assessed the compatibility of seven national food composition tables with USDA FoodData Central (FDC) using the LLM-based matching method reported previously (Nakagawa and Yamamoto, 2026). Databases from four English-speaking countries (Canada, United Kingdom, Australia, and New Zealand), South Korea, and Japan were compared with 8,158 USDA FDC entries (SR Legacy and Foundation Foods, excluding Survey/FNDDS). Match rates varied by country (62.0-89.7%) and food category. After excluding six USDA categories unsuitable for cross-national comparison, 45.2% of the remaining 6,290 entries were not matched by any country. Canada showed the highest concordance, reflecting shared North American food supply. Japan and South Korea showed similar low coverage for vegetables and spices. These findings suggest that while USDA FDC represents a practical foundation for a globally comprehensive food composition database given its breadth, systematic incorporation of country-specific foods and classification schemes will be necessary to achieve true international interoperability.

Background and Objective: Access to real-world electronic health records (EHRs) remains limited by privacy, governance and annotation constraints, hindering the development of clinical natural language processing models. Realistic synthetic progress notes may provide EHR-like corpora that preserve clinically rigorous information on diagnoses, treatments, symptoms, imaging, laboratory findings and therapeutic trajectories without relying directly on sensitive patient records. This study evaluates whether large language models (LLMs) can generate realistic Spanish prostate cancer progress notes from published case reports, preserving clinical content, temporality and hospital-style conventions.

Introduction: The use of artificial intelligence (AI) by clinicians has increased rapidly in recent years, with large language models (LLMs) emerging as tools that can equal clinician diagnostic performance in simulated settings. However, limited data exist regarding physicians use of LLMs in real-world clinical practice. This study aimed to evaluate the frequency of LLM use among practicing hospitalists, identify which LLMs are most commonly utilized, and assess hospitalists' perceptions of the benefits and limitations of LLM use in clinical care. Methods: We conducted a cross-sectional survey study of academic hospital medicine faculty across 8 institutions within the Hospital Medicine Reengineering Network (HOMERuN), a collaborative research consortium. Eligible participants included hospitalists practicing within participating HOMERuN sites during the study period. The survey assessed the frequency of LLM use, types of LLMs used, clinical applications, and physician perceptions regarding usefulness, efficiency, and concerns associated with LLM adoption. Results: 170 respondents (67.1%) reported ever using an LLM in clinical practice. Among LLM users, OpenEvidence was the most used tool (88.9%), followed by ChatGPT (58.5%), Google Gemini (26.9%), and Microsoft Copilot (20.5%). Only a minority of hospitalists reported using LLMs daily while seeing patients. The most common use cases of LLMs were answering diagnostic (77.1%) and management (77.6%) questions. A majority also reported using LLMs to identify or summarize primary literature (60.0%). Lack of trust in outputs (49.8%), uncertainty around institutional policies (48.6%), and lack of access to secure applications (43.1%) were cited as the most frequent barriers to using LLMs in practice. Discussion: The use of LLMs in clinical practice is already widespread, though regular or daily use is not yet typical. Concerns regarding reliability, patient privacy, and safe integration into clinical workflows remain significant barriers to broader adoption. The responsible implementation of LLMs in hospital medicine will require addressing these barriers.

We trained a self-configuring nnU-Net model for CMB segmentation in a heterogeneous multicenter sample (n=264), including 1.5T and 3T field strengths, SWI and T2*-GRE sequences, and community and clinical cohorts. Model performance was evaluated using 5-fold cross-validation with a focus on object-level detection metrics. Real-world performance was evaluated on scans from an unseen dataset of people with cerebrovascular disease (n=20). The model achieved 0.82 cluster Dice, 0.88 precision, and 0.77 sensitivity on hold-out test data. Notably, the model demonstrated a low false-positive rate, averaging 0.58 false positives (FPs) per scan, an improvement on existing publicly available models. The model achieved high performance in dataset of those with Alzheimer's disease and mild cognitive impairment (0.89 cluster Dice, 0.94 sensitivity), supporting its utility in clinical settings where ARIA-H monitoring is critical. In external validation, the model maintained high robustness with 0.79 sensitivity and 0.95 FPs per scan. By leveraging a heterogenous training strategy and a self-adapting architecture, we demonstrate that deep learning can achieve high-precision CMB detection that is robust to domain shifts. The low FP rate suggests this publicly available pipeline is suitable for automated screening and lesion counting in heterogenous large-scale clinical trials, reducing the burden of manual quantification.

Teledermatology expands access to dermatologic expertise in rural settings, yet diagnostic uncertainty persists in low-resource primary care. This retrospective study evaluated MedGemma-4B-IT, a compact multimodal vision-language model, as adjunctive clinical decision support for challenging diagnostic cases. We analyzed 77 zero-concordance cases (360 clinical photographs) from a Dermatology Extension for Community Healthcare Outcomes (ECHO) tele-mentoring program (2016-2021). Zero-concordance cases showed no overlap between primary clinician provisional diagnosis and dermatologist-confirmed diagnosis. The model was prompted using dermatologist-style format to generate ranked differential diagnoses. Performance was assessed using strict case-level top-k exact-match accuracy and relaxed matching criteria based on fuzzy string similarity. MedGemma achieved 0.0% strict top-1 accuracy, 1.3% top-3 accuracy, 3.9% top-5 accuracy, and 3.9% top-10 accuracy. Relaxed concept-level matching achieved 28.6% top-1, 63.6% top-5, and 67.5% top-10 accuracy. Image-level accuracy was 44.2% (159/360, 95% CI 39.0-49.5%). The model surfaced the correct diagnosis within differential lists in 45.5% of cases despite no exact top-1 matches, suggesting utility for differential expansion rather than definitive diagnosis. Performance varied across diagnostic categories, with highest accuracy in Other categories (54.5%) and lowest in neoplastic conditions (0.0%). Common errors included confusion between inflammatory and other diagnostic groupings. These findings characterize MedGemma performance on real-world teledermatology cases and inform safe, clinician-in-the-loop integration into teledermatology workflows where specialist oversight remains essential.