This study performed a bibliometric analysis of the top 100 most cited papers on artificial intelligence (AI) and its applications in health and medicine. Data was retrieved from the Scopus database in December 2024, focusing on papers published between 2001 and 2024. The search string included terms such as "artificial intelligence," "machine learning," "deep learning," and "electronic health records," combined with "health" or "medicine." To ensure specificity, only documents with these terms in their titles were included. The analysis was conducted in two stages: (1) exploring publications to identify a focused dataset and (2) selecting the top 100 most cited papers for in-depth evaluation. A total of 9,219 documents were analyzed, comprising 7,959 articles and 1,260 reviews. D.W. Bates emerged as the top author with 103 publications, Harvard Medical School was the leading institution, with 560 publications, while the United States dominated geographically, contributing 5,413 papers. The publication details, top authors, universities, and countries are presented in supplementary tables. Figures illustrate the dynamics of authors, universities, and countries, as well as collaboration networks for authors, departments, and countries. Additionally, uni-, bi-, and tri-gram analyses and the thematic evolution of the top 100 most cited papers were explored. This study provides a detailed analysis of AI applications in health and medicine, highlighting influential authors, institutions, and countries, as well as funding dynamics and publication trends. The findings offer valuable insights into the evolution of research in this domain and its key contributors, serving as a foundation for guiding future investigations and identifying emerging trends.
Nuclear theranostics involves the use of radioactive materials to image biological processes of expressing specific disease targets, such as membrane transporters or cell surface receptors, and designing these agents to deliver ionizing radiation to tissues that express these targets.With the increasing focus on precision medicine and sustainable healthcare, theranostic approaches have emerged as promising strategies for improving disease diagnosis, chemoraditaion therapy, treatment planning, therapeutic monitoring, and toxicity prediction, thereby enhancing patient outcomes and reducing the healthcare burden. This review summarizes recent advances in targeted theranostic radiopharmaceuticals, including monoclonal antibody-based systems, peptide receptor radionuclide therapy, nanotechnology-assisted theranostics, and pre-targeting strategies. Important considerations related to radionuclide selection, dosimetry, radiolabeling stability, and clinical translation are also discussed. Recent developments in novel radionuclides and radiopharmaceutical platforms have expanded the clinical applications of theranostics to neuroendocrine tumors, prostate cancer, and other malignant and non-malignant diseases. Advances in molecular imaging, personalized dosimetry, and targeted drug delivery have accelerated the clinical translation of theranostic agents while supporting innovation in the nuclear medicine field. However, challenges related to radionuclide production, regulatory approval, accessibility, and large-scale manufacturing continue to limit its widespread clinical implementation. Nuclear theranostics is a rapidly evolving field in modern nuclear medicine that supports precision healthcare through integrated diagnosis and targeted therapy. Continued advancements in radiopharmaceutical development and translational research are expected to improve the safety, effectiveness, and accessibility of theranostic approaches, contributing to sustainable healthcare delivery.
To develop and demonstrate a tool for efficient extraction, management, and visualization of DICOM metadata for facilitating retrospective protocol auditing and longitudinal trend analysis in diagnostic radiology. We developed the DICOM Tags Database as an extension module of IndoQCT platform using Python, PyQt5, and SQLite. The software features a dynamic workflow allowing users to configure custom database schemas by selecting tags directly from a DICOM viewer. The system includes integrated visualization tools for generating trend lines and frequency histograms. To evaluate the tool, anonymized thoracic CT datasets from three different hospitals using diverse scanner vendors (Toshiba, Siemens, and GE) were analyzed. The software successfully extracted and archived key dosimetric and mechanical parameters, including CTDIvol, tube voltage, exposure time, slice thickness, and table height. Longitudinal visualization shows specific protocol trends, thereby identifying outliers in slice thickness settings and highlighting vendor-specific variations in coordinate system definitions. The DICOM Tags Database bridges the gap between manual data recording and complex command-line utilities.
Intensity-modulated radiotherapy (IMRT) and volumetric-modulated arc therapy (VMAT) are designed to optimise planning target volume (PTV) coverage while sparing organs at risk (OARs). However, plan selection remains variable due to planner’s experience, patient’s characteristics, and operational constraints. Twenty patients were planned in Monaco 6.1.4. Analysis included dose, volume histogram, monitor units, control points, delivery times, modulation degree and gamma analysis (3
Deep learning may be helpful to differentiate gastrointestinal stromal tumors and risk stratification using endoscopic images. A total of 494 patients with gastrointestinal stromal tumors and 1010 patients with gastric cancer from author’s hospital were trained and validated. Another 99 patients with gastrointestinal stromal tumors and 100 patients with gastric cancer from second hospital were enrolled as external validation. Two deep learning networks, Swin Transformer and ConvNeXt were adapted for the differentiation, tumor size, mitotic index and risk stratification prediction with image level, patient level and combined level. In the internal validation dataset, ConvNeXt achieved area under curve of 0.985 in the differentiation gastrointestinal stromal tumors from gastric cancer. Swin Transformer achieved a best area under curve of 0.927, 0.806 and 0.765 in the prediction of tumor size, mitotic index and risk stratification for gastrointestinal stromal tumors, respectively. In the external validation dataset, ConvNeXt achieved a best area under curve of 0.999 in the differentiation gastrointestinal stromal tumors from gastric cancer. Swin Transformer achieved a best area under curve of 0.872, 0.856 and 0.693 in the prediction of tumor size, mitotic index and risk stratification for gastrointestinal stromal tumors. Endoscopic images-based deep learning models demonstrated excellent performance in the differentiation gastrointestinal stromal tumors from gastric cancer, and in the prediction the tumor size, mitotic index and risk stratification of gastrointestinal stromal tumors. They are promising in the differentiation and risk stratification to improve the management of gastrointestinal stromal tumors.
Healthcare systems are increasingly targeted by cyberattacks, posing risks to patient safety and continuity of care. Although the General Data Protection Regulation (GDPR) mandates technical and organisational safeguards, cybersecurity maturity in healthcare remains uneven. This study analyses major publicly documented cyber incidents in Czech hospitals and benchmarks them against selected European cases to identify recurring vulnerabilities and governance implications. An exploratory qualitative comparative case analysis was conducted of three publicly reported cyber incidents affecting Czech hospitals (2019–2020). Data were derived exclusively from publicly accessible sources, including national cybersecurity authority communications, governmental documents, peer-reviewed literature, and established media reporting. Cases were assessed against GDPR Articles 32–34 and relevant ENISA guidance using a matrix-based analytical framework covering technical safeguards, preparedness, governance and resilience implications. Four high-impact European healthcare incidents (Germany, Ireland, the United Kingdom, and France) were included for comparative benchmarking. Publicly available accounts suggested recurring weaknesses in technical safeguards and organisational preparedness, including limited network segmentation, legacy endpoint protection, insufficient staff training, and constrained incident response capacity. These patterns suggested gaps between regulatory requirements and practical resilience. European cases revealed similar vulnerabilities, in some instances resulting in large-scale service disruption and reported clinical risk. Reactive governance responses and delayed detection were common features. Cybersecurity should be embedded within health technology governance and system resilience rather than treated solely as a compliance obligation. Findings suggest that resilience-oriented governance may complement compliance-based regulation in strengthening preparedness. Strengthening executive accountability and preparedness mechanisms may enhance resilience in digitally dependent healthcare systems.
This study aimed to comprehensively determine the radiation dose and image quality of rapid kVp-switching dual-energy CT (DECT) protocol for whole-abdominal emergency and compare its performance with a standard single-energy CT (SECT) protocol. Retrospective data were collected from 130 patients who underwent contrast-enhanced whole-abdominal CT using either a fast kVp-switching DECT protocol (n = 65) or a routine SECT protocol (n = 65). Radiation dose in each protocol was determined using the volume CT dose index (CTDIvol). Quantitative image quality was assessed by measuring the mean attenuation (HU) and image noise (standard deviation of HU) to calculate the signal-to-noise ratio (SNR) in the aorta, main portal vein (MPV), liver, spleen, and psoas muscle. Subjective image quality including diagnostic acceptability and image noise was assessed independently by two experienced emergency radiologists. The Mann–Whitney U test was used to compare differences between the DECT and SECT protocols. The average CTDIvol obtained from SECT was higher than that from DECT; however, the difference was not statistically significant (10.7 ± 2.3 vs. 10.3 ± 2.8 mGy, p = 0.579). Although DECT demonstrated significantly higher signal and noise levels across all measured structures (p < 0.05), SNR values were not significantly different between protocols, except for the MPV (p = 0.016). Subjective assessments showed no significant differences in diagnostic acceptability or image noise (p > 0.05). Fast kVp-switching DECT demonstrates comparable radiation dose and image quality to SECT in patients undergoing CT for acute abdominal conditions in the emergency department. These findings suggest that DECT may be considered as an alternative CT imaging technique in the emergency abdominal setting.
This study presents an automated quality assurance (QA) method for LINAC (linear accelerator) performance verification using EPID-based imaging (electronic portal imaging device), combined with a custom image processing algorithm integrated into the QAtrack+ platform. An in-house Python algorithm, developed and validated for this study using the Pylinac library, was implemented on an Elekta Versa HD LINAC equipped with an iViewGT EPID system. The algorithm automatically analyzes EPID images to extract key dosimetric metrics, field size, flatness, symmetry and dose linearity, and stores results within QAtrack + for centralized tracking and long-term trend evaluation. Calibration followed IAEA TRS-398 standards, and image quality was verified according to AAPM TG-58 recommendations. The algorithm demonstrated high repeatability and reproducibility (coefficient of variation < 2
The purpose of this feasibility study was to test whether changes in the indoor environment and electricity usage can be used to detect changes in occupant activities that reflect changes in health. Temperature, relative humidity, carbon dioxide and electrical current were measured over 14+ months in 36 single-occupant homes. Data regarding health and household activities were collected via frequent optional surveys. Characteristics were extracted from the sensor data, including descriptive statistics, timings of peaks and troughs, frequency, deviations from average and modelled values (long short-term memory neural network). Machine learning was used to map these characteristics to occupant responses about health and activities, using a multilayer regressor neural network. Models were trained using all participant data together and separately on data from individual participants. Detection of an existing health condition being worse than normal gave a balanced accuracy of 62
Brain–computer interface (BCI) technology is an emerging technology that makes communication possible and supports motor functions by analyzing brain activity and transmitting control signals to external devices. It has potential as a therapeutic or assistive tool for patients with severe conditions, such as paralysis. Although the number of BCI-related patents has been increasing, clinical applications remain limited. The purpose of this study was to clarify the status of BCI patent applications and clinical trials, analyze their characteristics, and identify the challenges in clinical implementation. Patent data were obtained from Espacenet using the terms “brain–computer interface” and “brain–machine interface” and categorized as invasive or noninvasive based on international patent classifications. Applications were analyzed based on applicant nationality, type, collaboration status, and invasiveness. Clinical trial data were collected from ClinicalTrials.gov and the Chinese Clinical Trial Registry. Trials were similarly classified and analyzed by country, target disease, treatment content, sample size, trial status, and publication status. A total of 4216 BCI-related patent applications were identified, of which 2366 (56.1
Artificial Intelligence (AI) has transformed various sectors, including healthcare, where it holds promises for improving early detection, personalized care, and decision-making. However, despite its potential, AI systems often exacerbate health inequalities, particularly among underrepresented communities, due to biased algorithms, limited accessibility, and data exclusion. This current study aims to examine inclusive AI in healthcare grounded on a patient-centered approach. We conducted a literature review following PRISMA 2020 guidelines, searching across online databases for publications from diverse geographic and cultural contexts. Articles were assessed for inclusion based on specific criteria, and data on healthcare, AI type, inclusivity frameworks, implementation methods, and SWOT analysis were extracted. The quality and risk of bias of included articles were assessed with the JBI Checklist for Expert Opinion; and the PROBAST tool. Findings from the study suggest that various AI types were discussed, including machine learning, chatbots, and convolutional neural networks. Several key frameworks for ensuring inclusive AI were reported in the included articles, such as ACCEPT-AI, Bias Mitigation, and human rights values frameworks. However, some articles did not mention specific inclusivity frameworks. Our findings suggest that AI in healthcare has the potential to improve health outcomes but must be implemented inclusively and transparently to avoid exacerbating inequalities. This review emphasizes the need for ethical, regulated, and collaborative approaches to mitigate bias, enhance accessibility, and ensure patient-centered solutions. Future efforts should focus on stakeholder-inclusive design and standardized frameworks for equitable AI in healthcare. This literature review was not registered.
We explored automated classification models of lung nodules into three morphological types [solid, part-solid, and pure ground-glass nodules (pGGN)] based on computed tomography (CT) using features related to the three-dimensional (3D) consolidation-to-tumor ratio (CTR), which is the maximum diameter ratio of consolidation (= solid component) to a whole tumor. This retrospective study included an internal dataset including 275 patients with non-small cell lung cancer (solid: 179; part-solid: 69; pGGN: 27) from our university hospital and an external test dataset comprising 60 nodules (20 nodules for each type) from The Cancer Imaging Archive. Twelve tumor contours for each patient were predicted by using three deep learning networks (U-net, V-net, and dense V-net) and nine fusion models. The CTR, consolidation-to-tumor-volume ratio (CTVR), and average voxel value of non-solid components, which directly associated with the solid and ground-glass components of nodules, were calculated as the interpretable features from 12 predicted contours. Six machine-learning and two ensemble-learning models (stacking and bagging) were constructed using CTR-related features to classify the nodules into the three morphological types. The classification performance was evaluated using the area under receiver operating characteristic curves (AUCs). The proposed interpretable model achieved the classification accuracies of 0.869 for the internal test dataset and 0.733 for the external test dataset. The sensitivity on the external test dataset for pGGNs was improved to 0.900 compared with conventional models. This study suggests that the proposed interpretable classification model using CTR-related features could robustly predict morphological nodule types.
Evaluating the effectiveness and limitations of ChatGPT, Gemini, and DeepSeek in benchmarking the execution of deterministic eligibility rules for lung cancer screening from a structured patient list. 1,000 simulated patient cases (500 smokers and 500 non-smokers) were included. Three large language models (LLMs) were evaluated: ChatGPT-4.0, Gemini Advanced, and DeepSeek-V3. Various personal and smoking-associated characteristics, including age, gender, weight, height, smoking status (smoker versus non-smoker), packs per day, years smoked, pack years, and years since quitting smoking, were included in a file and presented to the LLMs. Eligibility criteria from three countries were included: the USA, South Korea, and Germany. The LLMs were evaluated for robustness, repeatability, factual correctness, and diagnostic accuracy after simple and advanced prompting. ChatGPT-4.0 achieved a specificity of 94.6
Published end-to-end validation data for 10 MV photon beams in mixed-vendor radiotherapy environments remain limited. A 10 MV photon beam from an Elekta Synergy Agility linear accelerator was commissioned in Eclipse (version 17.1), and the dosimetric performance of AAA and Acuros XB was evaluated using end-to-end testing. Beam data were measured in a PTW Beamscan water phantom and then imported into Eclipse to generate beam models for AAA and AXB. Eclipse was integrated with the Mosaiq oncology information system to enable better workflow. Following IAEA TECDOC-1583 guidelines, end-to-end test plans were created in Eclipse and delivered to a CIRS thorax phantom (model 002LFC) containing soft-tissue, lung, and bone-equivalent inserts. Absolute dose measurements were compared for AAA, AXB dose-to-water (Dw), and AXB dose-to-medium (Dm), with stopping-power corrections applied for lung and bone regions. End-to-end testing demonstrated good overall agreement for both algorithms. AAA showed consistent performance in soft tissue and closer agreement in bone-equivalent regions, while AXB performed better in lung-equivalent regions. Larger deviations were observed for AXB in posterior beam arrangements where the beam passed through bone before reaching the target. The 10 MV photon beam was successfully commissioned and verified for clinical use in the Eclipse system. Both AAA and AXB demonstrated clinically acceptable accuracy for 3DCRT. Although algorithm-dependent differences were observed in heterogeneous conditions, AAA remains suitable for 3DCRT planning when AXB is unavailable. This work contributes new end-to-end validation data for a 10 MV Elekta-Eclipse configuration, addressing a current gap in the literature for mixed-vendor radiotherapy systems.
Artificial Intelligence (AI) and Machine Learning (ML) are increasingly deployed as Software as a Medical Device (SaMD), including adaptive systems whose performance evolves through continuous learning. While these technologies offer clinical and operational benefits, their dynamic nature challenges traditional regulatory frameworks designed for static medical products. This article aims to critically examine the regulatory implications of adaptive AI/ML-based medical devices and to identify key governance gaps within existing oversight models. A critical-integrative analysis of the current regulatory landscape was conducted, focusing on major frameworks and guidance documents governing AI-enabled medical devices. The analysis examines regulatory approaches such as the United States Food and Drug Administration’s Total Product Lifecycle framework, Predetermined Change Control Plans, and emerging Good Machine Learning Practice principles, alongside international harmonization initiatives. Key regulatory challenges—including continuous learning, post-market monitoring, algorithmic bias, transparency, cybersecurity, and accountability—were systematically evaluated. The analysis identifies operational and governance gaps in existing regulatory approaches, in regulating continuously learning systems across their lifecycle. While lifecycle-oriented and risk-based frameworks represent important advances, current models remain insufficiently operationalized for adaptive AI systems. Challenges related to post-market performance surveillance, real-world data integration, transparency, and international regulatory alignment remain unresolved. Rather than proposing a wholly new regulatory framework, this study outlines concrete operational mechanisms for strengthening governance of adaptive AI/ML medical devices. These include risk-tiered, lifecycle-oriented oversight, enhanced post-market surveillance mechanisms, AI-specific technical standards, and sustained multi-stakeholder collaboration. Addressing these gaps is essential to ensure the safe and trustworthy integration of adaptive AI into healthcare.
Extended reality (XR) technologies, incorporating virtual and augmented reality, are increasingly being explored as potential adjuncts in the management of complex regional pain syndrome (CRPS). However, the appropriateness of these interventions in terms of their alignment with CRPS symptomatology, patient experiences, and rehabilitation principles remains unclear. This scoping review aimed to critically examine the current approaches used in XR-based rehabilitation for CRPS and assess their suitability for managing this condition. A systematic search was conducted in several health databases for articles describing the use of XR for management of people with CRPS. Data were extracted on study characteristics and intervention details, and the intervention was critiqued using a custom framework designed to evaluate the incorporation of rehabilitation principles. Sixteen studies met the inclusion criteria. There was considerable diversity in study methodologies, participant characteristics, and technological approaches. All studies provided appropriate theoretical justifications of their intervention and were suitable for clinical use, while most showed potential for CRPS symptom reduction. However, there were significant gaps in the interventions related to fostering independence, feasibility of home use, potential for progression, and cultural considerations. XR interventions show promise in certain aspects of CRPS management but opportunities exist for more comprehensive intervention delivery formats that address key rehabilitation principles. Future development and evaluation studies should place greater emphasis on fostering independent use, integration of user feedback, and overtly incorporate cultural considerations.
This review examines the safety of diagnostic ultrasound in Ghana and other African countries by linking established bioeffects evidence to real-world practice. It asks how international guidance on ultrasound safety (justification, ALARA, and attention to Thermal Index [TI] and Mechanical Index [MI]) can be implemented effectively in typical African service conditions, particularly in obstetrics, pediatric care, Doppler applications, and emerging techniques such as contrast-enhanced ultrasound (CEUS) and elastography. A narrative review was conducted using published literature on ultrasound bioeffects, safety indices, and exposure-related risk modifiers, alongside key international guidance from major professional and regulatory bodies (e.g., WHO, WFUMB, ISUOG, EFSUMB, FDA). Additional sources addressing implementation challenges in Ghana and Africa were considered, including training pathways, maintenance constraints, quality assurance/quality control (QA/QC), infrastructure limitations, and patient communication practices. Across the reviewed evidence and guidance, diagnostic ultrasound has a strong safety profile when used appropriately, but the likelihood of thermal and mechanical bioeffects increases with higher output settings, Doppler modes, prolonged dwell time, scanning near bone, and the use of microbubble contrast. Reported practice challenges in Ghana and similar contexts include variable operator training, limited preventive maintenance and quality control, power instability, weak governance, and inconsistent patient counselling. These factors can contribute to suboptimal optimization and unnecessary repeat scanning. Ultrasound safety in Ghana and Africa depends less on lack of guidance than on consistent implementation. Practical, locally feasible measures include justified scanning, ALARA-based presets, short focused protocols, routine TI/MI awareness, competency-based training, strengthened maintenance and basic QC led with medical physics support, and standardized patient education.
Multiple sclerosis (MS) is an autoimmune, demyelinating chronic disease that affects the central nervous system (CNS). The sensitivity of magnetic resonance imaging (MRI) to MS lesions is under study and improvement. MRI technologies like 3D T2-SPACE-Dark fluid sequences offer enhanced lesion visibility by cerebrospinal fluid (CSF) suppression and isotropic resolution, yet they are rarely compared directly with routinely used 2D T2-TSE sequences, especially in clinical MS imaging. This study aims to analyze the outperformance of 3D T2-SPACE-Dark fluid sequence over 2D T2-TSE sequence in detecting MS lesions, by calculating contrast to noise ratio (CNR), signal to noise ratio (SNR) in deep white matter region of the brain, and evaluating cerebral lesions counts across 3 brain anatomical regions, juxtacortical, periventricular and deep white matter. The clinical protocol 2D T2-TSE sequence was compared to an additional 3D T2-SPACE-Dark fluid sequence in the axial plane in 32 patients with multiple sclerosis. To obtain qualitative and quantitative analysis of the lesions, the acquired images were post-processed using the SYNGOVIA MRI workstation. The resulting data were statistically analyzed using the Wilcoxon signed-rank test. Demographic data such as the subtype and disease duration were collected and evaluated. Quantitative analysis demonstrated a significantly higher number of total lesion detections in all brain regions in the 3D T2-SPACE-Dark fluid sequence 24.03 (5.64) than in the 2D T2-TSE sequence 17.90 (4.36), W < 0.0001. SNR and CNR were also superior in the 3D sequence (SNR: 332.47 ± 125.09, CNR: 117.75 ± 53.01) relative to 2D acquisition (SNR: 195.26 ± 61.43, CNR: 79.37 ± 50.69). No significant correlation was observed between lesion load and expanded disability scale score (EDSS) for either 2D (r = -0.17, p = 0.33) or 3D (r = 0.12, p = 0.50) sequences; similarly, lesion load did not correlate significantly with disease duration (3D: r = 0.041, p = 0.82; 2D: r = -0.093, p = 0.61). However, a statistically significant positive correlation was indicated between disease duration and EDSS with a value of r = 0.60, p = 0.00023, suggesting progressive impairment over time. The 3D T2-SPACE-Dark fluid sequence can provide more accurate estimation of the lesion load for visualizing brain lesions and is a useful sequence that has advantages over 2D T2-TSE, though lesion burden alone may not be a reliable indicator for clinical disability or disease progression.
