Purpose:This study introduces the first adaptation of RETFound for joint optic disc (OD) and optic cup (OC) segmentation. RETFound is a well-known foundation model developed for fundus camera and optical coherence tomography images, which has shown promising performance in disease diagnosis. Methods:We propose FunduSegmenter, a model integrating a series of novel modules with RETFound, including a Pre-adapter, a Decoder, a Post-adapter, skip connections with a Convolutional Block Attention Module, and a Vision Transformer block adapter. The model was evaluated on a proprietary dataset, GoDARTS, and four public datasets (IDRiD, Drishti-GS, RIM-ONE-r3, and REFUGE) through internal verification, external verification, and domain generalization experiments. Results:An average Dice similarity coefficient of 90.51% was achieved in internal verification, which outperformed all baselines, some substantially (nnU-Net, 82.91%; DUNet, 89.17%; TransUNet, 87.91%). In all external verification experiments, the average results were about 3% higher than those of the best baseline, and our model was also competitive in domain generalization. Conclusions:This study explored the potential of the latent general representations learned by RETFound for OD and OC segmentation in fundus camera images. Our FunduSegmenter model generally outperformed state-of-the-art baseline methods. The proposed modules are general and can be extended to fine-tuning other foundation models. Translational Relevance:The model showed strong stability and generalization on both in-distribution and out-of-distribution data, providing stable OD and OC segmentation. This is an essential step for many automated tasks, from setting the accurate retinal coordinate to biomarker discovery.

Background Early identification of individuals at risk of dementia is essential for preventive care and timely enrolment into disease-modifying interventions. However, most existing prediction approaches rely on invasive, costly, or research-only biomarkers that are not scalable within public healthcare systems. Routinely acquired National Health Service (NHS) brain magnetic resonance imaging (MRI) scans, when linked with electronic health records, represent a widely available and privacy-preserving resource for population-level dementia risk stratification. A key challenge for clinical translation is ensuring that machine-learning predictions are reliable, interpretable, and safe to apply, particularly when models are used years before clinical diagnosis.Methods We conducted a retrospective case-control study entirely within a secure NHS Trusted Research Environment using routine T1-weighted brain MRI scans linked to electronic health records from Tayside and Fife, Scotland. The study included 518 participants: 259 individuals who subsequently developed dementia and 259 age- and sex-matched controls. Structural brain features were derived from MRI data and analysed using a support-vector-machine classifier with nested cross-validation to minimise overfitting. Prediction confidence was quantified using distance-from-hyperplane (DFH) calibration, enabling stratification of model outputs by certainty. Primary outcomes were classification accuracy and area under the receiver-operating-characteristic curve (AUC). Secondary analyses examined DFH-stratified performance and the relationship between prediction accuracy and time from scan to first recorded dementia diagnosis.Results The model predicted future dementia up to five years before first recorded NHS diagnosis with an AUC of 0.71, a performance consistent with real-world clinical imaging rather than research-optimised datasets. Model sensitivity increased for scans acquired closer to diagnosis, indicating stronger predictive signal as disease onset approached. Confidence-based stratification identified a high-confidence subgroup comprising approximately 35% of scans, within which prediction accuracy increased to around 80%. Performance was consistent across heterogeneous routine NHS scanners and imaging protocols, demonstrating robustness and generalisability to real-world clinical data rather than research-optimised acquisitions.Conclusion Routinely collected NHS brain MRI data can be used to predict future dementia several years before clinical diagnosis. Incorporating confidence calibration transforms a conventional machine-learning classifier into a safety-aware and clinically interpretable framework by enabling selective use of high-certainty predictions. This approach supports scalable early detection, population-level risk stratification, and targeted recruitment into preventive or disease-modifying clinical trials, with clear potential for integration into public health systems.

Firearm identification plays a crucial role in criminal justice globally. The capability to link firearms to specific crimes is invaluable for investigations and court cases. Each firearm leaves distinctive markings on bullets and cartridge cases, creating a "mechanical fingerprint" that can be used for the comparison of bullets and cartridge cases and underpins this area of forensic science. Cartridge cases fired from the same firearm exhibit similar markings on their bases. These traces can be used for investigation purposes as a means to potentially provide a link between more than one scene where cartridge cases have been recovered, or to provide a potential evidential link between a firearm and a cartridge case. These applications involve comparing the markings on the base of two or more cartridge cases, consisting of the headstamp, breech face and firing pin areas. The headstamp area usually contains information about the manufacturer and the calibre. Once this is considered, the remaining task is to compare the breech and firing pin areas of the two cartridges. Currently, some automated methods exist for this comparison, all of which involve the removal of the headstamp area to minimize bias. Some semi-automated methods for headstamp removal are available, and recently, an automated deep learning method that can be applied to 256 × 256 pixel resolution images has been introduced. In this article, we also propose a deep learning method addressing a more computationally demanding task of removing the head stamp area in higher-resolution images, 512 × 512 and 2592 × 1944 pixels, which will permit the automated extraction of finer features at a higher resolution. We also (a) introduce a post-processing method that improves the performance of our method, (b) provide the labelled data that we have produced so it can be used, together with the NIST database of cartridge case images, as a benchmark for future research, and (c) provide the estimated weights and models of the convolutional neural networks that can either be used directly or as initial values for further research. This article contributes to the emerging body of research on deep learning applications in forensic science.

Abstract Background Prior studies have demonstrated an association between retinal vascular features and cardiovascular disease (CVD), however most studies have only evaluated a few simple parameters at a time. Our aim was to determine whether a deep-learning artificial intelligence (AI) model could be used to predict CVD outcomes from routinely obtained diabetic retinal screening photographs and to compare its performance to a traditional clinical CVD risk score. Methods We included 6127 individuals with type 2 diabetes without myocardial infarction or stroke prior to study entry. The cohort was divided into training (70%), validation (10%) and testing (20%) cohorts. Clinical 10-year CVD risk was calculated using the pooled cohort equation (PCE) risk score. A polygenic risk score (PRS) for coronary heart disease was also obtained. Retinal images were analysed using an EfficientNet-B2 network to predict 10-year CVD risk. The primary outcome was time to first major adverse CV event (MACE) including CV death, myocardial infarction or stroke. Results 1241 individuals were included in the test cohort (mean PCE 10-year CVD risk 35%). There was a strong correlation between retinal predicted CVD risk and the PCE risk score (r = 0.66) but not the polygenic risk score (r = 0.05). There were 288 MACE events. Higher retina-predicted risk was significantly associated with increased 10-year risk of MACE (HR 1.05 per 1% increase; 95% CI 1.04–1.06, p < 0.001) and remained so after adjustment for the PCE and polygenic risk score (HR 1.03; 95% CI 1.02–1.04, p < 0.001). The retinal risk score had similar performance to the PCE (both AUC 0.697) and when combined with the PCE and polygenic risk score had significantly improved performance compared to the PCE alone (AUC 0.728). An increase in retinal-predicted risk within 3 years was associated with subsequent increased MACE likelihood. Conclusions A deep-learning AI model could accurately predict MACE from routine retinal screening photographs with a comparable performance to traditional clinical risk assessment in a diabetic cohort. Combining the AI-derived retinal risk prediction with a coronary heart disease polygenic risk score improved risk prediction. AI retinal assessment might allow a one-stop CVD risk assessment at routine retinal screening. Graphical abstract

Importance: Early identification of dementia risk is essential for preventive care and timely enrolment into disease-modifying interventions. Current approaches rely on costly, invasive, or research-only methods not feasible at scale within public health systems. Objective: To test whether routinely acquired NHS brain MRI scans can be used to predict future dementia diagnosis and whether confidence-based stratification improves prediction reliability and clinical interpretability. Design, Setting, and Participants: Retrospective case-control study conducted entirely within a secure NHS Trusted Research Environment. Routine T1-weighted MRI brain scans were linked with electronic health records for participants from Tayside and Fife, Scotland. The study included 259 individuals who subsequently developed dementia and 259 age- and sex-matched controls. Data were processed and modelled between January and June 2025. Exposure: MRI-derived structural brain features analysed using a support-vector-machine model with nested cross-validation and distance-from-hyperplane (DFH) confidence calibration. Main Outcomes and Measures: Primary outcomes were prediction accuracy and area under the receiver-operating-characteristic curve (AUC). Secondary analyses assessed DFH-stratified performance and time from scan to clinical diagnosis. Results: Confidence-based filtering identified a high-confidence subgroup (≈35% of scans) with ≈80% accuracy. Overall, the model predicted future dementia up to five years before first recorded NHS diagnosis, achieving 66.8% accuracy (AUC = 0.71). Model sensitivity increased for shorter time-to-diagnosis intervals. Analyses were generalisable across heterogeneous routine NHS scanners and datasets. Conclusions and Relevance: This study provides, to our knowledge, the first demonstration that routinely collected NHS MRI data can predict future dementia years before clinical diagnosis. Incorporating confidence calibration transforms a standard classifier into a safety-aware, clinically interpretable framework, supporting scalable early detection, risk stratification, and recruitment to preventive or disease-modifying trials across population health systems. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement This work was funded by the Medical Research Council (MRC) MICA Programme Grant: InterdisciPlInary Collaboration for efficienT and effective Use of clinical images in big data health care RESearch: PICTURES (MR/S010351/1). ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The use of data by this project was covered by University of Dundee, Health Informatics Centre's ethical approval 18/ES/0126, in accordance with conditions set by the East of Scotland Research Ethics Committee, and the approval of the NHS Caldicott Guardians. All SHARE participants consent to the access and use of their electronic health records for eligible research projects such as this one, scrutinised and approved by the SHARE access committee. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes The data related to the results presented in this article can be accessed within the HIC TRE subject to ethical and governance approvals. The details of dementia cohort identification can be accessed online on the HDR UK Phenotype library (https://phenotypes.healthdatagateway.org/phenotypes/PH1717/version/3973/detail/). The ML codebase used in this study is available online at GitHub (https://github.com/HicResearch/PICTURES-DementiaClassifier).

With the development of deep learning (DL) techniques, there has been a successful application of this approach to determine biological age from latent information contained in retinal images. Retinal age gap (RAG) defined as the difference between chronological age and predicted retinal age has been established previously to predict the age-related disease. In this study, we performed discovery genome-wide association analysis (GWAS) on the RAG using the 31,271 UK Biobank participants and replicated our findings in 8034 GoDARTS participants. The genetic correlation between RAGs predicted from the two cohorts was 0.67 (P = 0.021). After meta-analysis, we found 13 RAG loci which might be related to retinal vessel density and other aging processes. The SNP-wide heritability (h2) of RAG was 0.15. Meanwhile, by performing Mendelian randomization analysis, we found that glycated hemoglobin, inflammation hemocytes, and anemia might be associated with accelerated retinal aging. Our study explored the biological implications and molecular-level mechanism of RAG, which might enable causal inference of the aging process as well as provide potential pharmaceutical intervention targets for further treatment.

Ballistics (the linkage of bullets and cartridge cases to weapons) is a common type of evidence encountered in criminal cases around the world. The interest lies in determining whether two bullets were fired using the same firearm. This paper proposes an automated method to classify bullets from surface topography and Land Engraved Area (LEA) images of the fired pellets using machine and deep learning methods. The curvature of the surface topography was removed using loess fit and features were extracted using Empirical Mode Decomposition (EMD) followed by various entropy measures. The informative features were identified using minimum Redundancy Maximum Relevance (mRMR), finally the classification was performed using Support Vector Machines (SVM), Decision Tree (DT) and Random Forest (RF) classifiers. The results revealed a good predictive performance. In addition, the deep learning model DenseNet121 was used to classify the LEA images. DenseNet121 provided a higher predictive performance than SVM, DT and RF classifiers. Moreover, the Grad-CAM technique was used to visualise the discriminative regions in the LEA images. These results suggest that the proposed deep learning method can be used to expedite the linkage of projectiles to firearms and assist in ballistic examinations. In this work, the bullets that were compared were air pellets fired from both air rifles and a high velocity air pistol. Air guns were used to collect the data because they were more accessible than other firearms and could be used as a proxy, delivering comparable LEAs. The methods developed here can be used as a proof-of-concept and are easily expandable to bullet and cartridge case identification from any weapon.

There is increasing evidence that the complexity of the retinal vasculature measured as fractal dimension, D f , might offer earlier insights into the progression of coronary artery disease (CAD) before traditional biomarkers can be detected. This association could be partly explained by a common genetic basis; however, the genetic component of D f is poorly understood. We present a genome-wide association study (GWAS) of 38,000 individuals with white British ancestry from the UK Biobank aimed to comprehensively study the genetic component of D f and analyse its relationship with CAD. We replicated 5 D f loci and found 4 additional loci with suggestive significance ( P < 1e−05) to contribute to D f variation, which previously were reported in retinal tortuosity and complexity, hypertension, and CAD studies. Significant negative genetic correlation estimates support the inverse relationship between D f and CAD, and between D f and myocardial infarction (MI), one of CAD’s fatal outcomes. Fine-mapping of D f loci revealed Notch signalling regulatory variants supporting a shared mechanism with MI outcomes. We developed a predictive model for MI incident cases, recorded over a 10-year period following clinical and ophthalmic evaluation, combining clinical information, D f , and a CAD polygenic risk score. Internal cross-validation demonstrated a considerable improvement in the area under the curve (AUC) of our predictive model (AUC = 0.770 ± 0.001) when comparing with an established risk model, SCORE, (AUC = 0.741 ± 0.002) and extensions thereof leveraging the PRS (AUC = 0.728 ± 0.001). This evidences that D f provides risk information beyond demographic, lifestyle, and genetic risk factors. Our findings shed new light on the genetic basis of D f , unveiling a common control with MI, and highlighting the benefits of its application in individualised MI risk prediction.

There is increasing evidence that the complexity of the retinal vasculature (measured as fractal dimension, Df) might offer earlier insights into the progression of coronary artery disease (CAD) before traditional biomarkers can be detected. This association could be partly explained by a common genetic basis; however, the genetic component of Df is poorly understood. We present here a genome-wide association study (GWAS) aimed to elucidate the genetic component of Df and to analyse its relationship with CAD. To this end, we obtained Df from retinal fundus images and genotyping information from ~38,000 white-British participants in the UK Biobank. We discovered 9 loci associated with Df, previously reported in pigmentation, retinal width and tortuosity, hypertension, and CAD studies. Significant negative genetic correlation estimates endorse the inverse relationship between Df and CAD, and between Df and myocardial infarction (MI), one of CAD fatal outcomes. This strong association motivated us to developing a MI predictive model combining clinical information, Df, a CAD polygenic risk score and using a random forest algorithm. Internal cross validation evidenced a considerable improvement in the area under the curve (AUC) of our predictive model (AUC=0.770) when comparing with an established risk model, SCORE, (AUC=0.719). Our findings shed new light on the genetic basis of Df, unveiling a common control with CAD, and highlights the benefits of its application in individualised MI risk prediction.

Importance:The potential association of schizophrenia with distinct retinal changes is of clinical interest but has been challenging to investigate because of a lack of sufficiently large and detailed cohorts. Objective:To investigate the association between retinal biomarkers from multimodal imaging (oculomics) and schizophrenia in a large real-world population. Design, Setting, and Participants:This cross-sectional analysis used data from a retrospective cohort of 154 830 patients 40 years and older from the AlzEye study, which linked ophthalmic data with hospital admission data across England. Patients attended Moorfields Eye Hospital, a secondary care ophthalmic hospital with a principal central site, 4 district hubs, and 5 satellite clinics in and around London, United Kingdom, and had retinal imaging during the study period (January 2008 and April 2018). Data were analyzed from January 2022 to July 2022. Main Outcomes and Measures:Retinovascular and optic nerve indices were computed from color fundus photography. Macular retinal nerve fiber layer (RNFL) and ganglion cell-inner plexiform layer (mGC-IPL) thicknesses were extracted from optical coherence tomography. Linear mixed-effects models were used to examine the association between schizophrenia and retinal biomarkers. Results:A total of 485 individuals (747 eyes) with schizophrenia (mean [SD] age, 64.9 years [12.2]; 258 [53.2%] female) and 100 931 individuals (165 400 eyes) without schizophrenia (mean age, 65.9 years [13.7]; 53 253 [52.8%] female) were included after images underwent quality control and potentially confounding conditions were excluded. Individuals with schizophrenia were more likely to have hypertension (407 [83.9%] vs 49 971 [48.0%]) and diabetes (364 [75.1%] vs 28 762 [27.6%]). The schizophrenia group had thinner mGC-IPL (-4.05 μm, 95% CI, -5.40 to -2.69; P = 5.4 × 10-9), which persisted when investigating only patients without diabetes (-3.99 μm; 95% CI, -6.67 to -1.30; P = .004) or just those 55 years and younger (-2.90 μm; 95% CI, -5.55 to -0.24; P = .03). On adjusted analysis, retinal fractal dimension among vascular variables was reduced in individuals with schizophrenia (-0.14 units; 95% CI, -0.22 to -0.05; P = .001), although this was not present when excluding patients with diabetes. Conclusions and Relevance:In this study, patients with schizophrenia had measurable differences in neural and vascular integrity of the retina. Differences in retinal vasculature were mostly secondary to the higher prevalence of diabetes and hypertension in patients with schizophrenia. The role of retinal features as adjunct outcomes in patients with schizophrenia warrants further investigation.

Background and Objectives: This paper reports a quantitative analysis of the effects of joint photographic experts group (JPEG) image compression of retinal fundus camera images on automatic vessel segmentation and on morphometric vascular measurements derived from it, including vessel width, tortuosity and fractal dimension. Methods: Measurements are computed with vascular assessment and measurement platform for images of the retina (VAMPIRE), a specialized software application adopted in many international studies on retinal biomarkers. For reproducibility, we use three public archives of fundus images (digital retinal images for vessel extraction (DRIVE), automated retinal image analyzer (ARIA), high-resolution fundus (HRF)). We generate compressed versions of original images in a range of representative levels. Results: We compare the resulting vessel segmentations with ground truth maps and morphological measurements of the vascular network with those obtained from the original (uncompressed) images. We assess the segmentation quality with sensitivity, specificity, accuracy, area under the curve and Dice coefficient. We assess the agreement between VAMPIRE measurements from compressed and uncompressed images with correlation, intra-class correlation and Bland-Altman analysis. Conclusions: Results suggest that VAMPIRE width-related measurements (central retinal artery equivalent (CRAE), central retinal vein equivalent (CRVE), arteriolar-venular width ratio (AVR)), the fractal dimension (FD) and arteriolar tortuosity have excellent agreement with those from the original images, remaining substantially stable even for strong loss of quality (20% of the original), suggesting the suitability of VAMPIRE in association studies with compressed images. ? 2021 Elsevier B.V. All rights reserved. Background and Objectives: This paper reports a quantitative analysis of the effects of joint photographic experts group (JPEG) image compression of retinal fundus camera images on automatic vessel segmentation and on morphometric vascular measurements derived from it, including vessel width, tortuosity and fractal dimension. Methods: Measurements are computed with vascular assessment and measurement platform for images of the retina (VAMPIRE), a specialized software application adopted in many international studies on retinal biomarkers. For reproducibility, we use three public archives of fundus images (digital retinal images for vessel extraction (DRIVE), automated retinal image analyzer (ARIA), high-resolution fundus (HRF)). We generate compressed versions of original images in a range of representative levels. Results: We compare the resulting vessel segmentations with ground truth maps and morphological measurements of the vascular network with those obtained from the original (uncompressed) images. We assess the segmentation quality with sensitivity, specificity, accuracy, area under the curve and Dice coefficient. We assess the agreement between VAMPIRE measurements from compressed and uncompressed images with correlation, intra-class correlation and Bland-Altman analysis. Conclusions: Results suggest that VAMPIRE width-related measurements (central retinal artery equivalent (CRAE), central retinal vein equivalent (CRVE), arteriolar-venular width ratio (AVR)), the fractal dimension (FD) and arteriolar tortuosity have excellent agreement with those from the original images, remaining substantially stable even for strong loss of quality (20% of the original), suggesting the suitability of VAMPIRE in association studies with compressed images.

Texture analysis (TA) has shown promise as a surrogate marker for tissue structure, based on conventional and quantitative MRI sequences. Chemical-shift-encoding-based MRI (CSE-MRI)-derived proton density fat fraction (PDFF) of paraspinal muscles has been associated with various medical conditions including lumbar back pain (LBP) and neuromuscular diseases (NMD). Its application has been shown to improve the prediction of paraspinal muscle strength beyond muscle volume. Since mean PDFF values do not fully reflect muscle tissue structure, the purpose of our study was to investigate PDFF-based TA of paraspinal muscles as a predictor of muscle strength, as compared to mean PDFF. We performed 3T-MRI of the lumbar spine in 26 healthy subjects (age = 30 ± 6 years; 15 females) using a six-echo 3D spoiled gradient echo sequence for chemical-shift-encoding-based water–fat separation. Erector spinae (ES) and psoas (PS) muscles were segmented bilaterally from level L2–L5 to extract mean PDFF and texture features. Muscle flexion and extension strength was measured with an isokinetic dynamometer. Out of the eleven texture features extracted for each muscle, Kurtosis(global) of ES showed the highest significant correlation (r = 0.59, p = 0.001) with extension strength and Variance(global) of PS showed the highest significant correlation (r = 0.63, p = 0.001) with flexion strength. Using multivariate linear regression models, Kurtosis(global) of ES and BMI were identified as significant predictors of extension strength (R2adj = 0.42; p < 0.001), and Variance(global) and Skewness(global) of PS were identified as significant predictors of flexion strength (R2adj = 0.59; p = 0.001), while mean PDFF was not identified as a significant predictor. TA of CSE-MRI-based PDFF maps improves the prediction of paraspinal muscle strength beyond mean PDFF, potentially reflecting the ability to quantify the pattern of muscular fat infiltration. In the future, this may help to improve the pathophysiological understanding, diagnosis, monitoring and treatment evaluation of diseases with paraspinal muscle involvement, e.g., NMD and LBP.