This data article presents an experimental dataset of scattered images, obtained using a low-cost, open-source, Raspberry Pi-based optical system. Each data sample includes two grayscale images of 256 × 256 resolution: the (i) scattered image, and (ii) original projected pattern as ground truth. The system projects diverse patterns using various optical diffusers with different scattering coefficients and physical thicknesses. The dataset includes geometric shapes, digits, and textures to increase variability and generalization. This variety allows the analysis of distinct scattering regimes and evaluation of image recovery models under varying optical complexities. The dataset supports deep learning research focused on inverse problems in optics. It is particularly useful for training and benchmarking image restoration models in scattering environments.

Visual discomfort is a subjective experience, like many attributes of interest in the field of psychology. Measuring subjective phenomena can be difficult, as there is no ground truth against which to calibrate judgements. There is also a trade-off between the quality of the data and the time and effort of the participant - greater time investment should result in better data. However, whilst long, complex experiments might be possible in controlled lab settings with few observers, it becomes a barrier when attempting to estimate visual discomfort in less controlled but more ecologically valid spaces, and when investigating individual differences, for example young people and clinical populations. It is also difficult to calibrate judgements between participants due to individual variation in criterion - the idiosyncratic mapping of discomfort onto responses. We propose an intuitive method for participants to reduce criterion effects. This method maximises the amount of information gathered in a short space of time, and limits the risk of apparently estimating "discomfort" when the individual does not experience it. We apply this method to test two theoretical contributions to visual discomfort - cortical hyperexcitability (from spatial frequency (f), corresponding to stripe thickness) and ambiguous motion signals (from phase modulation wavelength (μ) corresponding to stripe waviness). Participants gave binary estimations that were used to scale their magnitude estimations. Using Bayesian methods, both these factors were found to affect discomfort judgements in both controlled lab environments (34 observers) and real-world estimations (47 observers).

Osteoporosis (OP) is characterised by loss of bone mineral density (BMD) and deterioration of trabecular microarchitecture, yet routine clinical imaging techniques remain limited in their ability to fully characterise bone microarchitecture. As new imaging technologies are developed, the potential for point of care bone assessment with both density and microarchitectural parameters of bone becomes a reality. Although advanced imaging modalities such as high-resolution peripheral quantitative computed tomography (HR-pQCT) offers improved sensitivity to bone structure, this is primarily focused on research settings. The development of higher resolution digital tomosynthesis (DT), required rethinking of phantoms, otherwise development and pre-clinical validation are constrained by the lack of reproducible, structure-controlled reference standards. In this study, we present a novel anthropomorphic bone phantom designed as a preclinical platform for calibration, benchmarking, and validation of bone imaging systems and quantitative analysis methods. The phantom integrates digital-twin trabecular models derived from micro-computed tomography (μ-CT), enabling parametric control of trabecular thickness and bone volume fraction to represent healthy and osteoporotic conditions. BMD is independently controlled using calibrated contrast agent (PVP-BaSO4), while moulded lean and adipose soft-tissue equivalents are incorporated to provide realistic X-ray attenuation for projection-based imaging. The phantoms were evaluated using multiple imaging modalities, including X-ray, DXA, pQCT, DT, and μCT, to verify their fidelity in reproducing both BMD and trabecular microstructural features. Imaging-derived parameters showed strong correlations with controlled variations in trabecular architecture and BMD, demonstrating the utility of the phantom as a source of controlled ground truth for cross-modality comparison. This reproducible platform enables systematic evaluation of imaging systems and facilitates early osteoporosis detection by bridging structure-density relationships. Our phantom serves as a valuable tool for preclinical diagnostic validation, imaging quality assurance, and the development of bone health biomarkers, thereby reducing reliance on animal or cadaveric studies.

scGImpute: A hybrid BiLayer multi-head graph attention-based imputation framework for zero dropout in single-cell sequencing datasets.

This dataset presents the first open-access collection of near-infrared hyperspectral imaging (NIR-HSI) data for the optical identification of textiles, with a focus on supporting research in sensor-based textile sorting and recycling. The dataset comprises hyperspectral images, RGB photographs, and detailed metadata, including fibre composition and colour, for 71 post-industrial textile samples, collected in Finland. Over 11 million spectra are included in the hyperspectral images, with more than 6 million annotated, providing a robust foundation for machine learning and data analysis. In addition, we provide a single representative NIR spectra and RGB value for each sample in order to accommodate classic spectroscopic analysis. Used garments were sourced from a partner company specializing in end-of-life textile management, with ground truth information on fibre composition obtained from suppliers. Small pieces of each garment were measured using Specim SWIR 3 hyperspectral camera and photographed with high-resolution mobile phone camera (Samsung Galaxy A52). The dataset is organized into folders containing raw and processed data, including ENVI-format hyperspectral images, RGB images, as well as CSV files with mean spectra, mean RGB values, and sample metadata. An example Python script is provided to facilitate data access and processing. Potential reuse scenarios include classification of textiles by material or colour, prediction of natural fibre content, image segmentation, algorithm development for spectral classification, and use as a reference spectral library. The dataset's comprehensive structure and open availability address the limitations of previous research, which often relied on small or non-public datasets, and is intended to accelerate advances in optical identification technologies for textile recycling.

This work introduces a novel deep learning-based architecture, termed the Deep Belief Markov Model (DBMM), which provides efficient, model-formulation agnostic inference in Partially Observable Markov Decision Process (POMDP) problems. The POMDP framework allows for modeling and solving sequential decision-making problems under observation uncertainty. In complex, high-dimensional, partially observable environments, existing methods for inference based on exact computations (e.g., via Bayes' theorem) or sampling algorithms do not always scale well. Furthermore, ground truth states may not be available for learning the exact transition dynamics. DBMMs extend deep Markov models into the partially observable decision-making framework and allow efficient belief inference entirely based on available observation data via variational inference methods. By leveraging the potency of neural networks, DBMMs can infer and simulate non-linear relationships in the system dynamics and naturally scale to problems with high dimensionality and discrete or continuous variables. In addition, neural network parameters can be dynamically updated efficiently based on data availability. DBMMs can thus be used to infer a belief variable, thus enabling the derivation of POMDP solutions over the belief space. We evaluate the efficacy of the proposed methodology by evaluating the capability of model-formulation agnostic inference of DBMMs in benchmark problems that include discrete and continuous variables. Finally, we demonstrate the practical utility of the inferred beliefs in a downstream decision-making task, showing that an RL agent guided by DBMMs beliefs significantly outperforms powerful model-free baselines and achieves near-optimal performance.1.

Image fusion aims to blend complementary information from diverse sensing modalities, yet most current methods lack robustness in complex fusion scenarios and cannot flexibly accommodate user intent. We present DiTFuse, the first Diffusion-Transformer (DiT) framework for instruction-driven, dynamic fusion control. Guided by natural-language instructions, DiTFuse flexibly blends multimodal content to enable hierarchical and fine-grained control over fusion dynamics. The training phase employs a multi-degrade-mask-image-modeling (M3) strategy, so the network jointly learns cross-modal alignment, modality-invariant restoration, and task-aware feature selection without relying on ideal reference images. A curated, multi-granularity instruction dataset further equips the model with interactive fusion capabilities. DiTFuse unifies infrared-visible, multi-focus, and multi-exposure fusion-as well as text-controlled refinement and downstream tasks-within a single architecture. Experiments on public IVIF, MFF, and MEF benchmarks confirm superior quantitative and qualitative performance, sharper textures, and better semantic retention. The model also supports multi-level user control and zero-shot generalization to other multiimage fusion scenarios, including instruction-conditioned segmentation.

To construct an intelligent segmentation scheme for precise localization of central serous chorioretinopathy (CSC) leakage points, thereby enabling ophthalmologists to deliver accurate laser treatment without navigational laser equipment. A dataset with dual labels (point-level and pixel-level) was first established based on fundus fluorescein angiography (FFA) images of CSC and subsequently divided into training (102 images), validation (40 images), and test (40 images) datasets. An intelligent segmentation method was then developed, based on the You Only Look Once version 8 Pose Estimation (YOLOv8-Pose) model and segment anything model (SAM), to segment CSC leakage points. Next, the YOLOv8-Pose model was trained for 200 epochs, and the best-performing model was selected to form the optimal combination with SAM. Additionally, the classic five types of U-Net series models [i.e., U-Net, recurrent residual U-Net (R2U-Net), attention U-Net (AttU-Net), recurrent residual attention U-Net (R2AttU-Net), and nested U-Net (UNet++)] were initialized with three random seeds and trained for 200 epochs, resulting in a total of 15 baseline models for comparison. Finally, based on the metrics including Dice similarity coefficient (DICE), intersection over union (IoU), precision, recall, precision-recall (PR) curve, and receiver operating characteristic (ROC) curve, the proposed method was compared with baseline models through quantitative and qualitative experiments for leakage point segmentation, thereby demonstrating its effectiveness. With the increase of training epochs, the mAP50-95, Recall, and precision of the YOLOv8-Pose model showed a significant increase and tended to stabilize, and it achieved a preliminary localization success rate of 90% (i.e., 36 images) for CSC leakage points in 40 test images. Using manually expert-annotated pixel-level labels as the ground truth, the proposed method achieved outcomes with a DICE of 57.13%, an IoU of 45.31%, a precision of 45.91%, a recall of 93.57%, an area under the PR curve (AUC-PR) of 0.78 and an area under the ROC curve (AUC-ROC) of 0.97, which enables more accurate segmentation of CSC leakage points. By combining the precise localization capability of the YOLOv8-Pose model with the robust and flexible segmentation ability of SAM, the proposed method not only demonstrates the effectiveness of the YOLOv8-Pose model in detecting keypoint coordinates of CSC leakage points from the perspective of application innovation but also establishes a novel approach for accurate segmentation of CSC leakage points through the "detect-then-segment" strategy, thereby providing a potential auxiliary means for the automatic and precise real-time localization of leakage points during traditional laser photocoagulation for CSC.

Accurately modelling physical perturbations in optical systems is critical for photonic device design, yet existing characterization methods are often computationally prohibitive. We introduce a data-efficient machine learning framework that learns the perturbation-dependent transmission matrix of a multimode fiber. To circumvent the spectral bias that prevents standard neural networks from resolving high-frequency phase changes, we explicitly encode perturbations into a Fourier Feature basis. This approach enables a compact multi-layer perceptron to learn the mapping from sparse training data with high fidelity. Using experimental data from a mechanically deformed fiber, our model achieves a 0.996 complex correlation with the ground truth, improving phase accuracy by an order of magnitude over standard networks while using significantly fewer parameters. This framework transforms the transmission matrix into a continuous, differentiable "digital twin" of the system, providing a robust tool for characterizing complex media in rapidly evolving environments.

While large language models (LLMs) have shown promise in medical text analysis, their application in automated medical billing code extraction remains underexplored, particularly for the German medical fee schedule system (GOÄ). Therefore, an LLM was fine-tuned to perform multi-label classification of GOÄ codes from radiology reports automatically, and its performance was compared with state-of-the-art commercial and open-source LLMs. Following ethics committee approval, we analyzed 499,601 radiology reports from 124,497 patients, containing 1,799,971 manually identified GOÄ codes as ground truth. The MediPhi-Instruct 4B model was fine-tuned using five-fold cross-validation. Performance was evaluated on the hold-out test set and compared against GPT-5, GPT-4.1, GPT-oss, Kimi-K2, Deepseek-R1, Deepseek-V3, Gemini 2.5, Llama-70B, and Qwen-3 LLMs on a subset of 500 anonymized and 350 cleaned reports using zero-shot and few-shot prompting techniques. The fine-tuned model achieved an accuracy of 77.15% ± 0.47% and a micro-average F1-score of 87.79% ± 0.31% on the hold-out test set. On a subset of 500 real-world samples, our models outperformed the best-performing LLM, Gemini 2.5 Flash, with an F1-score of 70.32% ± 1.54% compared to 58.22% ± 1.50% (p < 0.001). For the cleaned dataset of 350 samples, GPT-5 achieved the best F1-score of 89.51 ± 1.52% and outperformed the fine-tuned models (p < 0.001). Fine-tuned LLMs can effectively automate GOÄ code classification from radiology reports, with the potential of outperforming commercial LLMs. This approach shows promise for improving billing efficiency and accuracy in healthcare settings, though manual verification is still recommended. Question LLMs with high parameters possess medical knowledge, but how effective are they at predicting billing codes from radiology reports compared to smaller, fine-tuned models? Finidngs A fine-tuned ensemble model achieved competitive results and can outperform larger, proprietary LLMs. Clinical relevance Smaller, fine-tuned models offer an efficient alternative to proprietary LLMs in generating billing codes and can be integrated to assist clinical coding. This technology has the potential to transform clinical billing procedures, but its use should be overseen by qualified professional personnel.

Cerebral aneurysm is a life-threatening condition characterized by the formation of a saccular bulge in brain blood vessels, which can rupture and lead to severe complications. One treatment involves inserting a soft, flexible wire (coil) into the aneurysm to promote clotting and sealing. Mediators are often used to simulate tissue ingrowth within the sac to stabilize healing and prevent recurrence. However, quantitative assessment of tissue ingrowth in preclinical models remains labor-intensive, subjective, and poorly standardized, limiting the ability to compare therapeutic strategies and healing mechanisms. We developed a robust machine learning (ML) pipeline based on a Unet + + convolutional neural network (CNN), optimized for segmenting and quantifying tissue ingrowth in a preclinical carotid aneurysm mouse model. The model was trained and validated on 64 high-resolution histological images using 10-fold cross-validation. Image preprocessing included resizing, normalization, and augmentation, while post-processing applied thresholding techniques to CNN-generated heatmaps. Our method achieved Dice coefficients of 94.58% for sac segmentation and 95.23% for tissue ingrowth detection, with AUCs of 99.24% and 96.78%, respectively. The model's predictions showed strong agreement with ground truth ([Formula: see text]), supporting its potential for assessing biological stability and informing clinical decisions. In a blinded evaluation against expert annotations, our AI model achieved the highest agreement (Cohen's κ) among all raters, demonstrating its potential to provide consistent and expert-level tissue ingrowth assessments. A user-friendly graphical interface was developed, to enable non-technical users to perform segmentation and quantify tissue ingrowth. By providing objective, reproducible metrics of intra-aneurysmal healing, this approach supports mechanistic studies of therapeutic efficacy in preclinical aneurysm models and establishes a foundation for standardized evaluation of pro-healing interventions.

Statistical methods to identify Neanderthal ancestry in modern human genomes rest on varying assumptions and inputs. Nonetheless, most studies of introgression use only a single method to define Neanderthal ancestry. Due to a lack of "ground truth," we have a limited understanding of the accuracy, comparative strengths and weaknesses, and the sensitivity of downstream conclusions for these methods. Here, we performed large-scale comparisons of 14 genome-wide introgression maps computed by 11 representative Neanderthal introgression detection algorithms: admixfrog, ArchaicSeeker2, ArchIE, ARGWeaver-D, CRF, DICAL-ADMIX, hmmix, IBDmix, SARGE, Sprime, and S*. These algorithms span statistical approaches based on summary statistics, probabilistic modeling, and machine learning, and vary in their use of archaic, modern, and simulated genomes as input. Our results highlight a core set of regions predicted by nearly all methods, as well as substantial heterogeneity in commonly used Neanderthal introgression maps, especially at the individual genome level. Furthermore, we find that downstream analyses may result in different conclusions depending on the map used. Thus, we recommend careful consideration of map(s) chosen for downstream analysis and support the use of multiple maps to ensure robustness of conclusions. We make integrated prediction sets available, enabling further understanding of Neanderthal introgression's legacy on modern humans.

Lameness in cattle is a significant welfare and economic concern. To address this, we developed an end-to-end deep learning framework for 24/7 lameness monitoring using top-down depth images of cattle. The framework integrates three key stages: instance segmentation for detection, a custom multi-object tracking algorithm for identity preservation, and a spatio-temporal model for classification. We compared multiple instance segmentation models (Mask R-CNN, YOLOv8m-seg, YOLOv11m-seg) and evaluated three proposed tracking algorithms version1, 2 and 3 (PTAV1, PTAV2, and PTAV3). For classification, we tested multiple configurations integrating various pre-processing conditions (no filter, Gaussian, median), seven EfficientNet backbones (B1-B7), two temporal sequence lengths (5 and 7 frames), and a Long Short-Term Memory (LSTM) network to assign a lameness score from 1 (healthy) to 4 (lame) based on expert ground truth. In the detection model comparison, the YOLOv11m-seg model emerged as the top performer for detection, achieving a BBox AP@50 of 99.38%, Mask AP@50 of 99.26%, at 75.49 FPS. Our proposed tracking algorithm, PTAV3, which leverages location and direction prediction, achieved an exceptional overall accuracy of 99.94% (95% CI: 99.7-100%). For classification, the best model-an EfficientNet-B7 + LSTM architecture-yielded an accuracy of 95.95% (95% CI: 94.8-97.1%) and an F1-score of 96.06% (95% CI: 94.8-97.1%) on unseen test data, using a 5-frame sequence with no pre-processing filter. This integrated system provides a robust, automated, and objective solution for lameness scoring, showcasing the potential for real-time animal welfare monitoring in agricultural settings.

To review recent advances in the use of artificial intelligence (AI) to address shortcomings in assessing and improving surgical performance/training by automating surgical skills assessment and feedback. We searched PubMed for studies published between 2015 and 2025 pertaining to AI for surgical training. Search terms included 'artificial intelligence or 'machine learning' or 'deep learning' and 'surgical feedback' or 'surgical training' or 'surgical skill'. Articles were identified with special attention given to those published in the last 5 years with a focus on AI for surgical skill assessment or feedback. Artificial intelligence has been used to successfully automate surgical skill assessment across a variety of surgical disciplines via approaches such as kinematics, sabermetrics, computer vision, and gesture analysis. Many of these studies have developed AI models capable of a binary classification of skill (novice vs expert), which demonstrate concordance when verified against ground truths from human raters. Based on these skills assessments, AI approaches may be further leveraged to generate automatic feedback, which has proven effective in improving surgeon performance metrics, particularly for underperformers. AI has also shown utility in categorising and analysing the content and impact of live surgical feedback, enabling more efficient analysis of how feedback can be best delivered to trainees. Artificial intelligence is a promising tool for augmenting surgical training and improving the objectivity and scalability of surgical skill assessment and feedback. To date, AI models are adept at detecting relatively large differences in surgical performance and providing rudimentary feedback. Further work is required to create models capable of doing more fine-tuned skill assessments and generating more detailed, constructive feedback.

Deciphering disease-specific progression from low sample size, high-dimensional omic profiles remains challenging. Traditional biomarker discovery methods are costly and limited, while Nonnegative Matrix Factorization (NMF), though popular, suffers from instability and lack of biologically relevant solutions. This study aims to overcome these limitations by introducing a more robust framework. This article proposes TopConNMF, a topology-constrained extension of NMF which incorporates structural constraints, ensures stability, accuracy, and faster performance while maintaining biological interpretability. The method was evaluated on two publicly available time-varying omic datasets with established ground truths and compared against other state-of-the-art approaches. The TopConNMF consistently demonstrated stable performance across both the datasets, delivering superior accuracy and biologically relevant factorization compared to conventional NMF and other benchmark methods. The exhaustive evaluation confirmed its robustness in capturing disease-specific profiles and its efficiency in handling complex, high-dimensional data. Thus, TopConNMF provides a deeper understanding of complex biological systems by producing stable and interpretable factorization. Its broad applicability across multiple disease manifestations highlights its potential as a valuable tool for advancing omic data analysis and biomarker discovery. Clinical Impact: TopConNMF enables reliable biomarker discovery from limited omic data, supporting early diagnosis, patient stratification, and personalized treatment, thereby bridging computational findings with clinical applications.

This study examines the ability of YOLO (You Only Look Once) 11x, a widely used and state of the art object detection model, trained on publicly available datasets, to identify and count neutrophils in tissue samples taken at prosthetic joint revision surgery, with the objective of automating a laborious but necessary part of the diagnostic workup for periprosthetic joint infection. Three datasets containing blood film microscopic slides with neutrophils were downloaded, combined, and labelled. The resulting dataset of 3,923 images was augmented with ten additional histological slides from periprosthetic tissue, taken at the time of revision surgery (5 infected, 5 sterile), and split into training (70%), validation (20%), and test (10%) sets. The dataset was used to train YOLO 11x object detection model optimized for a mean average precision above 50%. The trained network was tested on a ground truth specimen and histological whole slide images from 19 additional cases, previously unseen by the model, for validation. The threshold for diagnosis of infection on histological sections was set at more than five neutrophils per 0.2 mm2 (equivalent to one high-powered microscope field). The model performed well as ground truth image returned precision at 82%, recall (sensitivity) 79%, and F1 harmonic mean 80%. When assessed against formal histopathological, microbiological, and multidisciplinary team (MDT) diagnosis, precision was 78%, 80%, and 90%; recall 78%, 89%, and 82%; and F1 score 78%, 84%, and 86%, respectively. Against the definitive MDT diagnosis, our model identified nine out of the ten infected cases and excluded seven out of nine cases that were not infected. This study demonstrates ability of the trained model to identify neutrophils in tissue taken at revision surgery and could assist in diagnosis of periprosthetic infection. Further work is needed to improve confidence in the identifications and diagnostic accuracy of periprosthetic infection.

An offshore oil spill near Karawang, West Java, in July 2019 caused a considerable impact on the Pulau Rambut Wildlife Sanctuary, a protected area known for its mangrove ecosystems, which provide crucial habitat for waterbirds in Jakarta Bay. The purpose of this study was to map and quantify land cover types on Rambut Island, as well as examine land cover changes three years after the spill, with a focus on mangrove dynamics. Land cover categorization was performed using the Maximum Likelihood (ML) method applied to remote sensing data from SPOT-6 and SPOT-7 satellite photos for 2019, 2020, and 2021. Ground truthing and drone imagery were used to validate categorization results, and accuracy was determined using the Kappa statistics. All classes had strong levels of agreement, with Kappa values of 85.66%, 81.40%, and 82% in 2019, 2020, and 2021, respectively. Four types of land cover were identified: mangroves, non-mangrove forests, water bodies, and open spaces. Rambut Island has an expected mangrove covering of 18.80 ha in 2019, which increased to 21.15 ha in 2020 before significantly declining to 18.84 ha in 2021. These findings are consistent with field data, in which 12 of 13 MHI (Mangrove Health Index) plots were classed as moderate. This data implies that the 2019 oil spill did not result in a significant or long-term loss in mangrove area on Rambut Island.Keywords: mangrove, maximum likelihood, oil spill, Rambut Island, SPOT-6/7

Deforestation and forest degradation are the main threats to biodiversity and carbon stocks in tropical forests. Advances in optical and SAR satellite sensors have enabled the development of real-time monitoring of deforestation on a global scale. SAR is particularly appealing in tropical areas due to its insensitivity to cloud cover. However, the automatic detection of small disturbed areas (such as individual tree felling gaps) remains a major challenge. Thanks to a unique dataset consisting of 23,759 locations of individual tree felling gaps and multi-date drone lidar acquisitions, we evaluated the potential of Sentinel-1 dense time series for monitoring small-sized forest disturbances substantially smaller than 0.1 ha on both FSC-certified and artisanal logging sites in the Congo Basin. We designed a new method for forest monitoring using the fused-lasso technique optimized to detect abrupt changes of at least 0.02 ha in Sentinel-1 time series using the fused-lasso technique (Fused-Lasso Change Detection, FLCD). We assessed our new method along with the Cumulative Sum (CuSum) that also proved promising for detecting small impacts, referring for the first time to precise disturbance dates over large areas. Both approaches reached similar rates of confirmed felling gaps that were similarly increasing with gap size, and similar rates of unconfirmed detected gaps. The FLCD method estimates the dates of tree felling more accurately in FSC-certified areas (−2 days difference for FLCD and −19 for CuSum on average). The effective resolution of the S-1 images limits detection for the smallest gaps, yet the approach can help detect and monitor degradation fronts. Fused lasso regression is relevant for modeling the temporal trajectories of the radar signal, which will allow taking advantage of both the increasing availability of UAV-borne data and the lengthening of the S-1 image series.

Reliable identification of fibrotic regions is essential for targeted catheter ablation therapy, as current imaging modalities such as cardiac magnetic resonance imaging face technical and clinical limitations, particularly in resolution and compatibility with implanted devices. This work presents the quantitative assessment of optical coherence tomography (OCT) images to classify myocardium into fibro-elastic versus normal. We acquired ultrahigh resolution OCT images from a sheep model with chronic myocardial infarction and performed pixelwise depth-resolved analysis to generate attenuation coefficient maps. In addition, we extracted radiomic features from three dimensional subvolumes to train a XGBoost classifier and validated our results against histological ground truth using Masson's trichrome staining histology to assess diagnostic accuracy. Attenuation and prediction probabilities effectively highlighted fibro-elastic regions. Widefield en face representations offered fast three dimensional screening of cardiac fibrosis. The radiomics-based XGBoost classifier achieved an area under the curve of 0.97 for binary classification. Combining ultrahigh resolution OCT with a straightforward attenuation coefficient and a robust radiomics pipeline for optical property extraction and high throughput radiomic feature analysis enables label-free assessment of fibrotic microstructures in the myocardium. The proposed quantitative framework enhances the detection and characterization of fibrotic myocardial tissue, offering potential for improved diagnostic precision and clinical integration of OCT in cardiology workflows towards data-driven catheter therapy guidance.

Continuous non-invasive blood pressure monitoring holds significant promise for cardiovascular disease management, yet cuff-based methods remain limited by their intermittent nature. Machine learning approaches leveraging photoplethysmography (PPG) and electrocardiography (ECG) signals present compelling alternatives, though questions persist about which signal type contributes more predictive value. This study compares traditional machine learning models, ensemble methods, and deep learning architectures for estimating systolic blood pressure from physiological waveforms. We extracted 55 features from PPG and ECG recordings of 100 subjects in the MIMIC-III Waveform Database, yielding 3000 segments with invasive arterial blood pressure as ground truth. Data splitting was performed at the subject level (70/15/15 train/validation/test) to prevent data leakage. Evaluation included regression metrics, British Hypertension Society grading, SHAP-based explainability, and ablation studies. Among all models, LightGBM achieved the best performance with mean absolute error of 15.97 mmHg, placing it at BHS Grade D. While SHAP analysis showed ECG features contributing 54.7% of importance versus 45.3% for PPG, our ablation study revealed that PPG-only models achieved comparable performance (MAE 15.97 vs. 16.23 mmHg), with the difference not statistically significant (p = 0.226). These results suggest that PPG-only wearable devices are viable for blood pressure estimation, as adding ECG features provides no statistically significant improvement. However, all configurations achieved only BHS Grade D, indicating that personalized calibration may be necessary for clinical acceptability.