Purpose To demonstrate the value of pretraining with millions of radiologic images compared with ImageNet photographic images on downstream medical applications when using transfer learning. Materials and Methods This retrospective study included patients who underwent a radiologic study between 2005 and 2020 at an outpatient imaging facility. Key images and associated labels from the studies were retrospectively extracted from the original study interpretation. These images were used for RadImageNet model training with random weight initiation. The RadImageNet models were compared with ImageNet models using the area under the receiver operating characteristic curve (AUC) for eight classification tasks and using Dice scores for two segmentation problems. Results The RadImageNet database consists of 1.35 million annotated medical images in 131 872 patients who underwent CT, MRI, and US for musculoskeletal, neurologic, oncologic, gastrointestinal, endocrine, abdominal, and pulmonary pathologic conditions. For transfer learning tasks on small datasets—thyroid nodules (US), breast masses (US), anterior cruciate ligament injuries (MRI), and meniscal tears (MRI)—the RadImageNet models demonstrated a significant advantage (P < .001) to ImageNet models (9.4%, 4.0%, 4.8%, and 4.5% AUC improvements, respectively). For larger datasets—pneumonia (chest radiography), COVID-19 (CT), SARS-CoV-2 (CT), and intracranial hemorrhage (CT)—the RadImageNet models also illustrated improved AUC (P < .001) by 1.9%, 6.1%, 1.7%, and 0.9%, respectively. Additionally, lesion localizations of the RadImageNet models were improved by 64.6% and 16.4% on thyroid and breast US datasets, respectively. Conclusion RadImageNet pretrained models demonstrated better interpretability compared with ImageNet models, especially for smaller radiologic datasets. Keywords: CT, MR Imaging, US, Head/Neck, Thorax, Brain/Brain Stem, Evidence-based Medicine, Computer Applications–General (Informatics) Supplemental material is available for this article. Published under a CC BY 4.0 license. See also the commentary by Cadrin-Chênevert in this issue.
Elisabetta Leogrande, Chiara Botrugno, Giulio Trono
et al.
We present a dataset of photoplethysmographic (PPG) signals acquired from 127 seated subjects under controlled conditions. The signals were recorded using the MAX86150EVSYS, a commercial multiwavelength optical sensor capable of simultaneous acquisition at red and infrared wavelengths. Each signal was sampled at 100 Hz, following a standardized protocol that was formally reviewed and approved by the institutional ethics committee of Ospedali Riuniti di Foggia for biomedical data collection. In addition to the PPG waveforms, reference values of systolic and diastolic blood pressure, as well as heart rate, were obtained using a commercial sphygmomanometer. All measurements were performed while minimizing motion to ensure high-quality signals suitable for physiological modeling and algorithm development. This dataset is intended to support research in blood pressure estimation, PPG signal analysis, and the development of robust cardiovascular monitoring techniques for wearable health technologies. All data were collected with informed consent and in accordance with institutional ethical guidelines.
Computer applications to medicine. Medical informatics, Science (General)
Background: The demand for health services among elderly patients with chronic diseases in rural ethnic minority areas of northwest Yunnan is increasing. Yet, service utilization remains imbalanced. Existing studies mainly focus on disease combinations, overlooking temporal and spatial variations in medical behavior. Methods: This study applies graph neural networks to construct a heterogeneous graph integrating patients, medical institutions, and geographic units, modeling dynamic service paths to identify high-frequency and potentially lost-contact patients. Using a heterogeneous graph attention network for feature embedding and a graph attention network classifier, the model captures behavioral similarity and service path patterns. Geographic and social variables such as ethnicity, terrain, and road accessibility further enhance sensitivity to regional disparities.Based on node centrality and path distribution, targeted service optimization strategies—such as mobile medical points and cross-regional collaboration nodes—are proposed for resource allocation. Results: Experimental results reveal marked spatial and structural disparities: Diqing Prefecture shows an accessibility index of 68 min versus 29 min in Dali; multimorbidity (3+) groups have a 68.6% matching rate but a 1.138 utilization rate, indicating resource imbalance; and mountain unit G18’s coverage index is only 0.31. Conclusion: The proposed model achieves a Macro-F1 of 0.83, outperforming XGBoost (0.76), effectively identifying high-risk groups, locating service bottlenecks, and supporting precise health resource optimization.
Computer applications to medicine. Medical informatics
Abstract Background Understanding the metabolic activities of the gut microbiome is vital for deciphering its impact on human health. While direct measurement of these metabolites through metabolomics is effective, it is often expensive and time-consuming. In contrast, microbial composition data obtained through sequencing is more accessible, making it a promising resource for predicting metabolite profiles. However, current computational models frequently face challenges related to limited prediction accuracy, generalizability, and interpretability. Method Here, we present the Deep Mixture of Variational Gaussian Process Experts (DMoVGPE) model, designed to overcome these issues. DMoVGPE utilizes a dynamic gating mechanism, implemented through a neural network with fully connected layers and dropout for regularization, to select the most relevant Gaussian Process experts. During training, the gating network refines expert selection, dynamically adjusting their contribution based on the input features. The model also incorporates an Automatic Relevance Determination (ARD) mechanism, which assigns relevance scores to microbial features by evaluating their predictive power. Features linked to metabolite profiles are given smaller length scales to increase their influence, while irrelevant features are down-weighted through larger length scales, improving both prediction accuracy and interpretability. Conclusions Through extensive evaluations on various datasets, DMoVGPE consistently achieves higher prediction performance than existing models. Furthermore, our model reveals significant associations between specific microbial taxa and metabolites, aligning well with findings from existing studies. These results highlight DMoVGPE’s potential to provide accurate predictions and to uncover biologically meaningful relationships, paving the way for its application in disease research and personalized healthcare strategies.
Computer applications to medicine. Medical informatics, Biology (General)
Anas Elmahi, Nathalie Doolan, Mohiedin Hezima
et al.
ABSTRACT Background Intermittent claudication (IC) is a hallmark symptom of peripheral arterial disease (PAD), causing pain and discomfort during physical activity caused by reduced blood flow to the lower extremities. The condition significantly impairs mobility and quality of life (QoL) in affected individuals. Treatment options for IC range from conservative approaches, including best medical therapy (BMT) and supervised exercise therapy (SET), to invasive interventions like angioplasty and open re‐vascularization. Aim This meta‐analysis and systematic review seek to assess the long‐term results of invasive procedures concerning Noninvasive treatments for the management of patients with IC. Methods A comprehensive search was conducted in October 2024 across databases containing PubMed, MEDLINE, Cochrane Library, Embase, and Scopus. Randomized controlled trials (RCTs) comparing invasive interventions to Noninvasive treatments were included. Primary outcomes were quality of life (QoL), ankle‐brachial pressure index (ABPI), and maximum walking distance (MWD). Secondary outcomes were major adverse cardiovascular events (MACE), mortality, complications, and re‐intervention rates. Data analysis was conducted using the Cochrane Review Manager 5. Follow‐up duration was between 2 and 7 years, longest available between 2 and 7 years; prioritized 2 years when present. Results A total of 11 RCTs with 1379 patients were included in the analysis. Invasive treatments demonstrated a significant improvement in MWD and ABPI compared to Noninvasive treatments (MWD pooled Mean Difference (MD) = 64.94 [10.77, 115.12] 95% CI, p = .02, 5 studies, and ABPI pooled MD = 0.15 [0.04, 0.26] 95% CI, p = .006, 5 studies). However, invasive interventions were associated with a higher rate of complications, including increased amputation risk (Pooled odds ratio (OR) = 2.46 [0.44, 13.94] 95% CI, p = .31, 3 studies), though this was not statistically significant. Long‐term rates were higher in the Noninvasive treatment group (Pooled OR: 0.56 [0.33, 0.97] 95% CI, p = .04). Conclusions Both invasive and Noninvasive treatments are effective in managing IC. Invasive treatments provide greater improvement in blood flow and walking distance, but the risk of complications and re‐interventions should be considered in treatment decisions. Further research with larger sample sizes and designed for long‐term assessment is needed to assess the cost‐effectiveness and long‐term outcomes of invasive treatments.
Public aspects of medicine, Computer applications to medicine. Medical informatics
Background: Evidence synthesis facilitates evidence-based medicine. This task becomes increasingly difficult to accomplished with applying computational solutions, since the medical literature grows at astonishing rates. Objective: This study evaluates an information retrieval-driven workflow, CASMA, to enhance the efficiency, transparency, and reproducibility of systematic reviews. Endometriosis recurrence serves as the ideal case due to its complex and ambiguous literature. Methods: The hybrid approach integrates PRISMA guidelines with fuzzy matching and regular expression (regex) to facilitate semi-automated deduplication and filtered records before manual screening. The workflow synthesised evidence from randomised controlled trials on the efficacy of a subclass of gonadotropin-releasing hormone agonists (GnRH-a). A modified splitting method addressed unit-of-analysis errors in multi-arm trials. Results: The workflow sharply reduced the screening workload, taking only 11 days to fetch and filter 33,444 records. Seven eligible RCTs were synthesized (841 patients). The pooled random-effects model yielded a Risk Ratio (RR) of $0.64$ ($95\%$ CI $0.48$ to $0.86$), demonstrating a $36\%$ reduction in recurrence, with non-significant heterogeneity ($I^2=0.00\%$, $τ^2=0.00$). The findings were robust and stable, as they were backed by sensitivity analyses. Conclusion: This study demonstrates an application of an information-retrieval-driven workflow for medical evidence synthesis. The approach yields valuable clinical results and a generalisable framework to scale up the evidence synthesis, bridging the gap between clinical research and computer science.
This work presents a novel Bayesian framework for unsupervised domain adaptation (UDA) in medical image segmentation. While prior works have explored this clinically significant task using various strategies of domain alignment, they often lack an explicit and explainable mechanism to ensure that target image features capture meaningful structural information. Besides, these methods are prone to the curse of dimensionality, inevitably leading to challenges in interpretability and computational efficiency. To address these limitations, we propose RemInD, a framework inspired by human adaptation. RemInD learns a domain-agnostic latent manifold, characterized by several anchors, to memorize anatomical variations. By mapping images onto this manifold as weighted anchor averages, our approach ensures realistic and reliable predictions. This design mirrors how humans develop representative components to understand images and then retrieve component combinations from memory to guide segmentation. Notably, model prediction is determined by two explainable factors: a low-dimensional anchor weight vector, and a spatial deformation. This design facilitates computationally efficient and geometry-adherent adaptation by aligning weight vectors between domains on a probability simplex. Experiments on two public datasets, encompassing cardiac and abdominal imaging, demonstrate the superiority of RemInD, which achieves state-of-the-art performance using a single alignment approach, outperforming existing methods that often rely on multiple complex alignment strategies.
Background Electronic health records have a significant impact on nursing practice, particularly in specializations such as labor and delivery, or acute care maternity nursing practice. Although primary studies on the use of electronic health records in labor and delivery have been done, no reviews on this topic exist. Moreover, the topic of labor and delivery nurses’ organizing work in the electronic health record-enabled context has not been addressed. Objective To (a) synthesize research on electronic health record use in labor and delivery nursing and (b) map how labor and delivery nursing organizing work is transformed by the electronic health record (as described in the reviewed studies). Methods The scoping review will be guided by a modified methodology based on selected recommendations from the Joanna Briggs Institute and the Preferred Reporting Items for Systematic reviews and Meta-Analyses Extension for Scoping Reviews. A comprehensive search will be conducted in the following databases: CINAHL Complete, MEDLINE, Academic Search Complete, Web of Science, Scopus and Dissertations and Theses Abstracts and Indexes. Included sources will be primary research, dissertations, or theses that address the use of electronic health records in labor and delivery nursing practice in countries with high levels of electronic health record adoption. Data extracted from included sources will be analyzed thematically. Further analysis will theorize labor and delivery nurses’ organizing work in the context of electronic health record use by utilizing concepts from Davina Allen's Translational Mobilization Theory. Findings will be presented in tabular and descriptive formats. Conclusion The findings of this review will help understand transformations of nursing practice in the electronic health record-enabled labor and delivery context and identify areas of future research. We will propose an extension of the Translational Mobilization Theory and theorize nurses’ organizing work involving the use of the electronic health record.
Computer applications to medicine. Medical informatics
Background and Objective Current tongue segmentation methods often struggle with extracting global features and performing selective filtering, particularly in complex environments where background objects resemble the tongue. These challenges significantly reduce segmentation efficiency. To address these issues, this article proposes a novel model for tongue segmentation in complex environments, combining Mamba and U-Net. By leveraging Mamba’s global feature selection capabilities, this model assists U-Net in accurately excluding tongue-like objects from the background, thereby enhancing segmentation accuracy and efficiency. Methods To improved the segmentation accuracy of the U-Net backbone model, we incorporated the Mamba attention module along with a multi-stage feature fusion module. The Mamba attention module serially connects spatial and channel attention mechanisms at the U-Net ’s skip connections, selectively filtering the feature maps passed into the deep network. Additionally, the multi-stage feature fusion module integrates feature maps from different stages, further improving segmentation performance. Results Compared with state-of-the-art semantic segmentation and tongue segmentation models, our model improved the mean intersection over union by 1.17%. Ablation experiments further demonstrated that each module proposed in this study contributes to enhancing the model’s segmentation efficiency. Conclusion This study constructs a T ongue segmentation model based on U -Net and Mamba (TUMamba). The model effectively extracted global spatial and channel features using the Mamba attention module, captured local detail features through U-Net, and enhanced image features via multi-stage feature fusion. The results demonstrate that the model performs exceptionally well in tongue segmentation tasks, proving its value in handling complex environments.
Computer applications to medicine. Medical informatics
Since the release of ChatGPT and GPT-4, large language models (LLMs) and multimodal large language models (MLLMs) have attracted widespread attention for their exceptional capabilities in understanding, reasoning, and generation, introducing transformative paradigms for integrating artificial intelligence into medicine. This survey provides a comprehensive overview of the development, principles, application scenarios, challenges, and future directions of LLMs and MLLMs in medicine. Specifically, it begins by examining the paradigm shift, tracing the transition from traditional models to LLMs and MLLMs, and highlighting the unique advantages of these LLMs and MLLMs in medical applications. Next, the survey reviews existing medical LLMs and MLLMs, providing detailed guidance on their construction and evaluation in a clear and systematic manner. Subsequently, to underscore the substantial value of LLMs and MLLMs in healthcare, the survey explores five promising applications in the field. Finally, the survey addresses the challenges confronting medical LLMs and MLLMs and proposes practical strategies and future directions for their integration into medicine. In summary, this survey offers a comprehensive analysis of the technical methodologies and practical clinical applications of medical LLMs and MLLMs, with the goal of bridging the gap between these advanced technologies and clinical practice, thereby fostering the evolution of the next generation of intelligent healthcare systems.
The mixing problem is classically encountered in the study of differential equations applied to fluid dynamics. An understanding of fluid movement under constraints is particularly important in the field of medicine as many therapeutics and biologic molecules are dissolved in bodily fluids. Many areas of biomedical research and diagnostics also rely on fluid sampling to obtain accurate measurements of biologic markers. We present in this manuscript the general solution to the mixing problem in the context of studying physiological phenomena based on the movement of fluid acting as a carrier for medically relevant molecules/solutes. We also expanded the general solution to become more compatible with areas of biomedical research and diagnostics that seek to characterize bodily fluids located in areas that are difficult to sample.
The TraX Engine is an advanced data processing tool developed by ADVACAM in collaboration with the European Space Agency (ESA), specifically designed for analyzing data from Timepix detectors equipped with various sensor materials (Si, CdTe, GaAs, SiC). TraX Engine can process large datasets across various scientific and medical applications, including space radiation monitoring, particle therapy, and imaging. In space applications, the TraX Engine has been used to process data from satellites like OneWeb JoeySat deployed in LEO orbit, where it continuously monitors space radiation environments measuring flux, dose, and dose rate in real-time. In medical applications, particularly in particle therapy, the TraX Engine is used to process data to characterize radiation fields in terms of particle flux, Linear Energy Transfer, and spatial distribution of the radiation dose. The TraX Engine can identify and classify scattered particles, such as secondary protons and electrons, and estimate their contribution to out-of-field doses. In imaging applications, the TraX Engine is integrated into Compton cameras, where it supports photon source localization through directional reconstruction of photons. The system ability to identify gamma radiation source with high precision makes it suitable for medical imaging tasks, such as tracking I-131 used in thyroid cancer treatment or localizing radiation sources. This paper presents the architecture and capabilities of the newly developed software TraX Engine, alongside results from various applications, demonstrating its role in particle tracking, radiation monitoring, imaging, and others. With its modular architecture, the TraX Engine offers multiple interfaces, including a command-line tool, an API, a web portal, and a graphical user interface, ensuring usability across different fields and user expertise levels.
Recent advancements in deep learning have shown transformative potential in medical imaging, yet concerns about fairness persist due to performance disparities across demographic subgroups. Existing methods aim to address these biases by mitigating sensitive attributes in image data; however, these attributes often carry clinically relevant information, and their removal can compromise model performance-a highly undesirable outcome. To address this challenge, we propose Fair Re-fusion After Disentanglement (FairREAD), a novel, simple, and efficient framework that mitigates unfairness by re-integrating sensitive demographic attributes into fair image representations. FairREAD employs orthogonality constraints and adversarial training to disentangle demographic information while using a controlled re-fusion mechanism to preserve clinically relevant details. Additionally, subgroup-specific threshold adjustments ensure equitable performance across demographic groups. Comprehensive evaluations on a large-scale clinical X-ray dataset demonstrate that FairREAD significantly reduces unfairness metrics while maintaining diagnostic accuracy, establishing a new benchmark for fairness and performance in medical image classification.
Anvar Kurmukov, Bogdan Zavolovich, Aleksandra Dalechina
et al.
Image compression is a critical tool in decreasing the cost of storage and improving the speed of transmission over the internet. While deep learning applications for natural images widely adopts the usage of lossy compression techniques, it is not widespread for 3D medical images. Using three CT datasets (17 tasks) and one MRI dataset (3 tasks) we demonstrate that lossy compression up to 20 times have no negative impact on segmentation quality with deep neural networks (DNN). In addition, we demonstrate the ability of DNN models trained on compressed data to predict on uncompressed data and vice versa with no quality deterioration.
Large Language Models (LLMs) have exhibited remarkable proficiency in natural language understanding, prompting extensive exploration of their potential applications across diverse domains. In the medical domain, open-source LLMs have demonstrated moderate efficacy following domain-specific fine-tuning; however, they remain substantially inferior to proprietary models such as GPT-4 and GPT-3.5. These open-source models encounter limitations in the comprehensiveness of domain-specific knowledge and exhibit a propensity for 'hallucinations' during text generation. To mitigate these issues, researchers have implemented the Retrieval-Augmented Generation (RAG) approach, which augments LLMs with background information from external knowledge bases while preserving the model's internal parameters. However, document noise can adversely affect performance, and the application of RAG in the medical field remains in its nascent stages. This study presents the Bailicai framework: a novel integration of retrieval-augmented generation with large language models optimized for the medical domain. The Bailicai framework augments the performance of LLMs in medicine through the implementation of four sub-modules. Experimental results demonstrate that the Bailicai approach surpasses existing medical domain LLMs across multiple medical benchmarks and exceeds the performance of GPT-3.5. Furthermore, the Bailicai method effectively attenuates the prevalent issue of hallucinations in medical applications of LLMs and ameliorates the noise-related challenges associated with traditional RAG techniques when processing irrelevant or pseudo-relevant documents.