In medical image segmentation, convolutional neural networks (CNNs) and transformers are dominant. For CNNs, given the local receptive fields of convolutional layers, long-range spatial correlations are captured through consecutive convolutions and pooling. However, as the computational cost and memory footprint can be prohibitively large, 3D models can only afford fewer layers than 2D models with reduced receptive fields and abstract levels. For transformers, although long-range correlations can be captured by multi-head attention, its quadratic complexity with respect to input size is computationally demanding. Therefore, either model may require input size reduction to allow more filters and layers for better segmentation. Nevertheless, given their discrete nature, models trained with patch-wise training or image downsampling may produce suboptimal results when applied on higher resolutions. To address this issue, here we propose the resolution-robust HNOSeg-XS architecture. We model image segmentation by learnable partial differential equations through the Fourier neural operator which has the zero-shot super-resolution property. By replacing the Fourier transform by the Hartley transform and reformulating the problem in the frequency domain, we created the HNOSeg-XS model, which is resolution robust, fast, memory efficient, and extremely parameter efficient. When tested on the BraTS'23, KiTS'23, and MVSeg'23 datasets with a Tesla V100 GPU, HNOSeg-XS showed its superior resolution robustness with fewer than 34.7k model parameters. It also achieved the overall best inference time (< 0.24 s) and memory efficiency (< 1.8 GiB) compared to the tested CNN and transformer models¹.

Large language models (LLMs) can generate outputs understandable by humans, such as answers to medical questions and radiology reports. With the rapid development of LLMs, clinicians face a growing challenge in determining the most suitable algorithms to support their work. We aimed to provide clinicians and other health care practitioners with systematic guidance in selecting an LLM that is relevant and appropriate to their needs and facilitate the integration process of LLMs in health care. We conducted a literature search of full-text publications in English on clinical applications of LLMs published between January 1, 2022, and March 31, 2025, on PubMed, ScienceDirect, Scopus, and IEEE Xplore. We excluded papers from journals below a set citation threshold, as well as papers that did not focus on LLMs, were not research based, or did not involve clinical applications. We also conducted a literature search on arXiv within the same investigated period and included papers on the clinical applications of innovative multimodal LLMs. This led to a total of 270 studies. We collected 330 LLMs and recorded their application frequency in clinical tasks and frequency of best performance in their context. On the basis of a 5-stage clinical workflow, we found that stages 2, 3, and 4 are key stages in the clinical workflow, involving numerous clinical subtasks and LLMs. However, the diversity of LLMs that may perform optimally in each context remains limited. GPT-3.5 and GPT-4 were the most versatile models in the 5-stage clinical workflow, applied to 52% (29/56) and 71% (40/56) of the clinical subtasks, respectively, and they performed best in 29% (16/56) and 54% (30/56) of the clinical subtasks, respectively. General-purpose LLMs may not perform well in specialized areas as they often require lightweight prompt engineering methods or fine-tuning techniques based on specific datasets to improve model performance. Most LLMs with multimodal abilities are closed-source models and, therefore, lack of transparency, model customization, and fine-tuning for specific clinical tasks and may also pose challenges regarding data protection and privacy, which are common requirements in clinical settings. In this review, we found that LLMs may help clinicians in a variety of clinical tasks. However, we did not find evidence of generalist clinical LLMs successfully applicable to a wide range of clinical tasks. Therefore, their clinical deployment remains challenging. On the basis of this review, we propose an interactive online guideline for clinicians to select suitable LLMs by clinical task. With a clinical perspective and free of unnecessary technical jargon, this guideline may be used as a reference to successfully apply LLMs in clinical settings.

Predicting distant metastases in soft tissue sarcomas (STS) is vital for guiding clinical decision-making. Recent advancements in radiomics and deep learning (DL) models have shown promise, but their diagnostic accuracy remains unclear. This meta-analysis aims to assess the performance of radiomics and DL-based models in predicting metastases in STS by analyzing pooled sensitivity and specificity. Following PRISMA guidelines, a thorough search was conducted in PubMed, Web of Science, and Embase. A random-effects model was used to estimate the pooled area under the curve (AUC), sensitivity, and specificity. Subgroup analyses were performed based on imaging modality (MRI, PET, PET/CT), feature extraction method (DL radiomics [DLR] vs. handcrafted radiomics [HCR]), incorporation of clinical features, and dataset used. Heterogeneity by I² statistic, leave-one-out sensitivity analyses, and publication bias by Egger's test assessed model robustness and potential biases. Ninetheen studies involving 1712 patients were included. The pooled AUC for predicting metastasis was 0.88 (95% CI: 0.80-0.92). The pooled AUC values were 88% (95% CI: 77-89%) for MRI-based models, 80% (95% CI: 76-92%) for PET-based models, and 91% (95% CI: 78-93%) for PET/CT-based models, with no significant differences (p = 0.75). DL-based models showed significantly higher sensitivity than HCR models (p < 0.01). Including clinical features did not significantly improve model performance (AUC: 0.90 vs. 0.88, p = 0.99). Significant heterogeneity was noted (I² > 25%), and Egger's test suggested potential publication bias (p < 0.001). Radiomics models showed promising potential for predicting metastases in STSs, with DL approaches outperforming traditional HCR. While integrating this approach into routine clinical practice is still evolving, it can aid physicians in identifying high-risk patients and implementing targeted monitoring strategies to reduce the risk of severe complications associated with metastasis. However, challenges such as heterogeneity, limited external validation, and potential publication bias persist. Future research should concentrate on standardizing imaging protocols and conducting multi-center validation studies to improve the clinical applicability of radiomics predictive models.

Accurate prediction of recurrence in clear cell renal cell carcinoma (ccRCC) remains a major clinical challenge due to the disease complex molecular, pathological, and clinical heterogeneity. Traditional prognostic models, which rely on single data modalities such as radiology, histopathology, or genomics, often fail to capture the full spectrum of disease complexity, resulting in suboptimal predictive accuracy. This study aims to overcome these limitations by proposing a deep learning (DL) framework that integrates multimodal data, including CT, MRI, histopathology whole slide images (WSI), clinical data, and genomic profiles, to improve the prediction of ccRCC recurrence and enhance clinical decision-making. The proposed framework utilizes a comprehensive dataset curated from multiple publicly available sources, including TCGA, TCIA, and CPTAC. To process the diverse modalities, domain-specific models are employed: CLAM, a ResNet50-based model, is used for histopathology WSIs, while MeD-3D, a pre-trained 3D-ResNet18 model, processes CT and MRI images. For structured clinical and genomic data, a multi-layer perceptron (MLP) is used. These models are designed to extract deep feature embeddings from each modality, which are then fused through an early and late integration architecture. This fusion strategy enables the model to combine complementary information from multiple sources. Additionally, the framework is designed to handle incomplete data, a common challenge in clinical settings, by enabling inference even when certain modalities are missing.

Neurogenetic disorders (NeD) are a group of neurological conditions resulting from inherited genetic defects. By affecting the normal functioning of the nervous system, these diseases lead to serious problems in movement, cognition, and other body functions. In recent years, machine learning (ML) approaches have proven highly effective, enabling the analysis and processing of vast amounts of medical data. By analyzing genetic data, medical imaging, and other clinical data, these techniques can contribute to early diagnosis and more effective treatment of NeD. However, using these approaches is challenged by issues including data variability, model explainability, and the requirement for interdisciplinary collaboration. This paper investigates the impact of ML on healthcare diagnosis and care of common NeD, such as Alzheimer's disease (AD), Parkinson's disease (PD), Huntington's disease (HD), and Multiple Sclerosis disease (MSD). The purpose of this research is to determine the opportunities and challenges of using these techniques in the field of neurogenetic medicine. Our findings show that using ML can increase the detection accuracy by 85 % and reduce the detection time by 60 %. Additionally, the use of these techniques in predicting patient prognosis has been 70 % more accurate than traditional methods. Ultimately, this research will enable medical professionals and researchers to leverage ML approaches in advancing the diagnostic and therapeutic processes of NeD by identifying the opportunities and challenges.

In medical image segmentation, convolutional neural networks (CNNs) and transformers are dominant. For CNNs, given the local receptive fields of convolutional layers, long-range spatial correlations are captured through consecutive convolutions and pooling. However, as the computational cost and memory footprint can be prohibitively large, 3D models can only afford fewer layers than 2D models with reduced receptive fields and abstract levels. For transformers, although long-range correlations can be captured by multi-head attention, its quadratic complexity with respect to input size is computationally demanding. Therefore, either model may require input size reduction to allow more filters and layers for better segmentation. Nevertheless, given their discrete nature, models trained with patch-wise training or image downsampling may produce suboptimal results when applied on higher resolutions. To address this issue, here we propose the resolution-robust HNOSeg-XS architecture. We model image segmentation by learnable partial differential equations through the Fourier neural operator which has the zero-shot super-resolution property. By replacing the Fourier transform by the Hartley transform and reformulating the problem in the frequency domain, we created the HNOSeg-XS model, which is resolution robust, fast, memory efficient, and extremely parameter efficient. When tested on the BraTS'23, KiTS'23, and MVSeg'23 datasets with a Tesla V100 GPU, HNOSeg-XS showed its superior resolution robustness with fewer than 34.7k model parameters. It also achieved the overall best inference time (< 0.24 s) and memory efficiency (< 1.8 GiB) compared to the tested CNN and transformer models.

Medical image segmentation is vital for accurate diagnosis. While U-Net-based models are effective, they struggle to capture long-range dependencies in complex anatomy. We propose GH-UNet, a Group-wise Hybrid Convolution-ViT model within the U-Net framework, to address this limitation. GH-UNet integrates a hybrid convolution-Transformer encoder for both local detail and global context modeling, a Group-wise Dynamic Gating (GDG) module for adaptive feature weighting, and a cascaded decoder for multi-scale integration. Both the encoder and GDG are modular, enabling compatibility with various CNN or ViT backbones. Extensive experiments on five public and one private dataset show GH-UNet consistently achieves superior performance. On ISIC2016, it surpasses H2Former with 1.37% and 1.94% gains in DICE and IOU, respectively, using only 38% of the parameters and 49.61% of the FLOPs. The code is freely accessible via: https://github.com/xiachashuanghua/GH-UNet .

Manual curation of radiographic features in pancreatic cyst registries for data abstraction and longitudinal evaluation is time consuming and limits widespread implementation. We examined the feasibility and accuracy of using large language models (LLMs) to extract clinical variables from radiology reports. A single center retrospective study included patients under surveillance for pancreatic cysts. Nine radiographic elements used to monitor cyst progression were included: cyst size, main pancreatic duct (MPD) size (continuous variable), number of lesions, MPD dilation ≥5mm (categorical), branch duct dilation, presence of solid component, calcific lesion, pancreatic atrophy, and pancreatitis. LLMs (GPT) on the OpenAI GPT-4 platform were employed to extract elements of interest with a zero-shot learning approach using prompting to facilitate annotation without any training data. A manually annotated institutional cyst database was used as the ground truth (GT) for comparison. Overall, 3198 longitudinal scans from 991 patients were included. GPT successfully extracted the selected radiographic elements with high accuracy. Among categorical variables, accuracy ranged from 97% for solid component to 99% for calcific lesions. In the continuous variables, accuracy varied from 92% for cyst size to 97% for MPD size. However, Cohen's Kappa was higher for cyst size (0.92) compared to MPD size (0.82). Lowest accuracy (81%) was noted in the multi-class variable for number of cysts. LLM can accurately extract and curate data from radiology reports for pancreatic cyst surveillance and can be reliably used to assemble longitudinal databases. Future application of this work may potentiate the development of artificial intelligence-based surveillance models.

The rapid evolution of deepfake technology poses critical challenges to healthcare systems, particularly in safeguarding the integrity of medical imaging, electronic health records (EHR), and telemedicine platforms. As autonomous AI becomes increasingly integrated into smart healthcare, the potential misuse of deepfakes to manipulate sensitive healthcare data or impersonate medical professionals highlights the urgent need for robust and adaptive detection mechanisms. In this work, we propose DProm, a dynamic deepfake detection framework leveraging visual prompt tuning (VPT) with a pre-trained Swin Transformer. Unlike traditional static detection models, which struggle to adapt to rapidly evolving deepfake techniques, DProm fine-tunes a small set of visual prompts to efficiently adapt to new data distributions with minimal computational and storage requirements. Comprehensive experiments demonstrate that DProm achieves state-of-the-art performance in both static cross-dataset evaluations and dynamic scenarios, ensuring robust detection across diverse data distributions. By addressing the challenges of scalability, adaptability, and resource efficiency, DProm offers a transformative solution for enhancing the security and trustworthiness of autonomous AI systems in healthcare, paving the way for safer and more reliable smart healthcare applications.

Stroke is a leading cause of mortality and disability worldwide, with approximately 15 million people suffering strokes annually. Machine learning (ML) techniques have emerged as powerful tools for stroke prediction, enabling early identification of risk factors through data-driven approaches. However, the clinical utility and performance characteristics of these approaches require systematic evaluation. To systematically review and analyze ML techniques used for stroke prediction, systematically synthesize performance metrics across different prediction targets and data sources, evaluate their clinical applicability, and identify research trends focusing on patient population characteristics and stroke prevalence patterns. A systematic review was conducted following PRISMA guidelines. Five databases (Google Scholar, Lens, PubMed, ResearchGate, and Semantic Scholar) were searched for open-access publications on ML-based stroke prediction published between January 2013 and December 2024. Data were extracted on publication characteristics, datasets, ML methodologies, evaluation metrics, prediction targets (stroke occurrence vs. outcomes), data sources (EHR, imaging, biosignals), patient demographics, and stroke prevalence. Descriptive synthesis was performed due to substantial heterogeneity precluding quantitative meta-analysis. Fifty-eight studies were included, with peak publication output in 2021 (21 articles). Studies targeted three main prediction objectives: stroke occurrence prediction (n = 52, 62.7 %), stroke outcome prediction (n = 19, 22.9 %), and stroke type classification (n = 12, 14.4 %). Data sources included electronic health records (n = 48, 57.8 %), medical imaging (n = 21, 25.3 %), and biosignals (n = 14, 16.9 %). Systematic analysis revealed ensemble methods consistently achieved highest accuracies for stroke occurrence prediction (range: 90.4-97.8 %), while deep learning excelled in imaging-based applications. African populations, despite highest stroke mortality rates globally, were represented in fewer than 4 studies. ML techniques show promising results for stroke prediction. However, significant gaps exist in representation of high-risk populations and real-world clinical validation. Future research should prioritize population-specific model development and clinical implementation frameworks.