medRxiv [Preprint]. 2025 Apr 11:2025.04.08.25325262. doi: 10.1101/2025.04.08.25325262.

ABSTRACT

Late-onset Tay-Sachs (LOTS) disease and late-onset Sandhoff disease (LOSD) have long been considered indistinguishable due to similar clinical presentations and shared biochemical deficits. However, recent magnetic resonance imaging (MRI) studies have shown distinct cerebellar atrophy associated with LOTS. In this study, we furthered this investigation to determine if the cerebellar atrophy is globally uniform or preferentially targets certain cerebellar regions. We utilized DeepCERES , a deep learning cerebellar specific segmentation and cortical thickness pipeline to analyze differences between LOTS (n=20), LOSD (n=5), and neurotypical controls (n=1038). LOTS had smaller volumes of the whole cerebellum as well as cerebellar lobules IV, V, VI, VIIB, VIIIA, VIIIB, IX, and both Crus I and II compared to both LOSD and neurotypical controls. LOTS patients also had smaller cortical thickness of cerebellar lobules V, VI, VIIB, VIIIA, VIIIB, and both Crus I and II compared to both LOSD and neurotypical controls. Cerebellar functional and lesion localization studies have implicated lobules V and VI in speech articulation and execution while lobules VI, Crus I, VIIA, among others, have been implicated in a variety of behaviors and neuropsychiatric symptoms. Our observations provide a possible anatomical substrate to the higher prevalence of dysarthria and psychosis in our LOTS but not LOSD patients. Future studies are needed for direct comparisons considering phenotypic aspects such as age of symptom onset, presence and severity of dysarthria and ataxia, full characterization of neuropsychiatric profiles, molecular pathology and biochemical differences to fully understand the dichotomy observed in these two diseases.

PMID:40297453 | PMC:PMC12036421 | DOI:10.1101/2025.04.08.25325262

medRxiv [Preprint]. 2025 Apr 7:2025.04.06.25325337. doi: 10.1101/2025.04.06.25325337.

ABSTRACT

Alzheimer's disease (AD) and AD-related dementias (ADRD) exhibit heterogeneous progression rates, with rapid progression (RP) posing significant challenges for timely intervention and treatment. The increasingly available patient-centered electronic health records (EHRs) have made it possible to develop advanced machine learning models for risk prediction of disease progression by leveraging comprehensive clinical, demographic, and laboratory data. In this study, we propose AutoRADP, an interpretable autoencoder-based framework that predicts rapid AD/ADRD progression using both structured and unstructured EHR data from UFHealth. AutoRADP incorporates a rule-based natural language processing method to extract critical cognitive assessments from clinical notes, combined with feature selection techniques to identify essential structured EHR features. To address the data imbalance issue, we implement a hybrid sampling strategy that combines similarity-based and clustering-based upsampling. Additionally, by utilizing SHapley Additive exPlanations (SHAP) values, we provide interpretable predictions, shedding light on the key factors driving the rapid progression of AD/ADRD. We demonstrate that AutoRADP outperforms existing methods, highlighting the potential of our framework to advance precision medicine by enabling accurate and interpretable predictions of rapid AD/ADRD progression, and thereby supporting improved clinical decision-making and personalized interventions.

PMID:40297450 | PMC:PMC12036374 | DOI:10.1101/2025.04.06.25325337

medRxiv [Preprint]. 2025 Apr 8:2025.04.07.25325386. doi: 10.1101/2025.04.07.25325386.

ABSTRACT

Alzheimer's disease (AD), particularly late-onset AD, stands as the most prevalent neurodegenerative disorder globally. Owing to its substantial heritability, genetic studies have emerged as indispensable for elucidating genes and biological pathways driving AD onset and progression. However, genetic and molecular mechanisms underlying AD remain poorly defined, largely due to the pronounced heterogeneity of AD and the intricate interactions among AD genetic factors. Notably, approximately 90% of AD-associated genetic variants reside in intronic and intergenic regions, yet their functional significance has remained largely uncharacterized. To address this challenge, we developed a deep learning framework combining bulk and single-cell epigenomic data to evaluate the regulatory potential (i.e., silencing and activating strength) of noncoding AD variants in the dorsolateral prefrontal cortex (DLPFCs) and its major cell types. This model identified 1,457 silencer and 3,084 enhancer AD-associated variants in the DLPFC and binned them into silencer variants only (SL), enhancer variants only (EN), or both variant types (ENSL) classes. Each class exerts distinct cellular and molecular influences on AD pathogenesis. EN loci predominantly regulate housekeeping metabolic processes, whereas SL loci (including the genes MS4A6A , TREM2 , USP6NL , HLA-D ) are selectively linked to immune responses. Notably, 71% of these genes are significantly upregulated in AD and pro-inflammation-stimulated microglia. Furthermore, genes associated with SL loci are, in neuronal cells, often responsive to glutamate receptor antagonists (e.g, NBQX) and anti-inflammatory perturbagens (such as D-64131), the compound classes known for reducing the AD risk. ENSL loci, in contrast, are uniquely implicated in memory maintenance, neurofibrillary tangle assembly, and are also shared by other neurological disorders such as Parkinson's disease and schizophrenia. Key genes in this class of loci, such as MAPT , CR1/2 , and CLU , are frequently upregulated in AD subtypes with hyperphosphorylated tau aggregates. Critically, our model can accurately predict the impact of regulatory variants, with an average Pearson correlation coefficient of 0.54 and a directional concordance rate of 70% between our predictions and experimental outcomes. This model identified rs636317 as a causal AD variant in the MS4A locus, distinguishing it from the 7bp-away allele-neutral variant rs636341. Similarly, rs7922621 was prioritized over its 54-bp-away allele-neutral rs7901634 in the TSPAN14 locus. Additional causal variants include rs6701713 in the CR1 locus, and rs28834970 and rs755951 in the PTK2B locus. Collectively, this work advances our understanding of the regulatory landscape of AD-associated genetic variants, providing a framework to explore their functional roles in the pathogenesis of this complex disease.

PMID:40297423 | PMC:PMC12036408 | DOI:10.1101/2025.04.07.25325386

Digit Health. 2025 Apr 15;11:20552076251335542. doi: 10.1177/20552076251335542. eCollection 2025 Jan-Dec.

ABSTRACT

BACKGROUND: In modern workplaces, emotional well-being significantly impacts productivity, interpersonal relationships, and organizational stability. This study introduced an innovative facial-based emotion recognition system aimed at the real-time monitoring and management of employee emotional states.

METHODS: Utilizing the RetinaFace model for facial detection, the Dlib algorithm for feature extraction, and VGG16 for micro-expression classification, the system constructed a 10-dimensional emotion feature vector. Emotional anomalies were identified using the K-Nearest Neighbors algorithm and assessed with a 3σ-based risk evaluation method.

RESULTS: The system achieved high accuracy in emotion classification, as demonstrated by an empirical analysis, where VGG16 outperformed MobileNet and ResNet50 in key metrics such as accuracy, precision, and recall. Data augmentation techniques were employed to enhance the performance of the micro-expression classification model.

CONCLUSION: These techniques improved the across diverse emotional expressions, resulting in more accurate and robust emotion recognition. When deployed in a corporate environment, the system successfully monitored employees' emotional trends, identified potential risks, and provided actionable insights into early intervention. This study contributes to advancing corporate mental health management and lays the foundation for scalable emotion-based support systems in organizational settings.

PMID:40297378 | PMC:PMC12035250 | DOI:10.1177/20552076251335542

Digit Health. 2025 Apr 22;11:20552076251337183. doi: 10.1177/20552076251337183. eCollection 2025 Jan-Dec.

ABSTRACT

INTRODUCTION: The progression of Alzheimer's disease (AD) has been shown to significantly correlate with changes in brain tissue structure and leads to cognitive decline and dementia. Using radiomic features derived from brain magnetic resonance imaging (MRI) scan, we can get the help of deep learning (DL) model for diagnosing AD.

METHODS: This study proposes the use of the DL model under the framework of MR radiomics for AD diagnosis. Two cross-racial independent cohorts from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (141 AD, 166 Mild Cognitive Impairment (MCI), and 231 normal control (NC) subjects) and Huashan hospital (45 AD, 35 MCI, and 31 NC subjects) were enrolled. We first performed preprocessing of MRI using methods such as spatial normalization and denoizing filtering. Next, we conducted Statistical Parametric Mapping analysis based on a two-sample t-test to identify regions of interest and extracted radiomic features using Radiomics tools. Subsequently, feature selection was carried out using the Least Absolute Shrinkage and Selection Operator model. Finally, the selected radiomic features were used to implement the AD diagnosis task with the TabNet model.

RESULTS: The model was quantitatively evaluated using the average values obtained from five-fold cross-validation. In the three-way classification task, the model achieved classification average area under the curve (AUC) of 0.8728 and average accuracy (ACC) of 0.7111 for AD versus MCI versus NC. For the binary classification task, the average AUC values were 0.8778, 0.8864, and 0.9506 for AD versus MCI, MCI versus NC, and AD versus NC, respectively, with average ACC of 0.8667, 0.8556, and 0.9222 for these comparisons.

CONCLUSIONS: The proposed model exhibited excellent performance in the AD diagnosis task, accurately distinguishing different stages of AD. This confirms the value of MR DL radiomic model for AD diagnosis.

PMID:40297370 | PMC:PMC12035500 | DOI:10.1177/20552076251337183

Digit Health. 2025 Apr 15;11:20552076251333660. doi: 10.1177/20552076251333660. eCollection 2025 Jan-Dec.

ABSTRACT

OBJECTIVE: This study aims to develop and validate a deep learning model to predict smoking status in patients with chronic obstructive pulmonary disease (COPD) using data from a national survey.

METHODS: Data from the Korea National Health and Nutrition Examination Survey (2007-2018) were used to extract 5466 COPD-eligible cases. The data collection involved demographic, behavioral, and clinical variables, including 21 predictors such as age, sex, and pulmonary function test results. The dependent variable, smoking status, was categorized as smoker or nonsmoker. A residual neural network (ResNN) model was developed and compared with five machine learning algorithms (random forest, decision tree, Gaussian Naive Bayes, K-nearest neighbor, and AdaBoost) and two deep learning models (multilayer perceptron and TabNet). Internal validation was performed using five-fold cross-validation, and model performance was evaluated using the area under the receiver operating characteristic (AUROC) curve, sensitivity, specificity, and F1-score.

RESULTS: The ResNN achieved an AUROC, sensitivity, specificity, and F1-score of 0.73, 70.1%, 75.2%, and 0.67, respectively, outperforming previous machine learning and deep learning models in predicting smoking status in patients with COPD. Explainable artificial intelligence (Shapley additive explanations) identified key predictors, including sex, age, and perceived health status.

CONCLUSION: This deep learning model accurately predicts smoking status in patients with COPD, offering potential as a decision-support tool to detect high-risk persistent smokers for targeted interventions. Future studies should focus on external validation and incorporate additional behavioral and psychological variables to improve its generalizability and performance.

PMID:40297369 | PMC:PMC12035114 | DOI:10.1177/20552076251333660

Digit Health. 2025 Apr 15;11:20552076251334431. doi: 10.1177/20552076251334431. eCollection 2025 Jan-Dec.

ABSTRACT

BACKGROUND: When applying for teleconsultations, medical laboratory reports are usually photographed with a mobile phone, and the photographic results are uploaded as teleconsultation application materials. It is very meaningful to extract the content of the image medical laboratory report and store the content digitally. There are already applications of OCR technology for medical text file recognition, but no researchers have recognized the format of the medical laboratory report and obtained the report content as a serialized process to digitize the image report. This article proposes a serialization method to digitize the medical laboratory report image.

MATERIALS AND METHODS: This article first collects 330 image-based medical laboratory reports, annotates the format of the medical laboratory reports, and forms a training dataset for the layout analysis model. Then, using the pre-trained model, the dataset is trained to obtain a layout analysis model that can correctly recognize the format of the medical laboratory report. Then, the layout of the input image-based medical laboratory report is analyzed, and the layout analysis results are used to call the text detection and text recognition models to obtain the digital content of the image report. Finally, adjusting the layout of the digital content and storing the digital content as a docx file.

RESULTS: After training the layout analysis model, integrating layout analysis, text detection, and text recognition, we have obtained a serialization method that digitizes the content of the image medical laboratory report, restores the report format, shields sensitive and irrelevant content, and digitizes the report content of interest.

CONCLUSIONS: By digitizing the image medical laboratory report through the serialization method, we can correctly display the content of the medical laboratory report for teleconsultation, while removing irrelevant content in the report, such as user names, examination equipment numbers, etc.

PMID:40297365 | PMC:PMC12035204 | DOI:10.1177/20552076251334431

Digit Health. 2025 Apr 15;11:20552076251334040. doi: 10.1177/20552076251334040. eCollection 2025 Jan-Dec.

ABSTRACT

BACKGROUND: Rapid and accurate identification of large vessel occlusion (LVO) is crucial for determining eligibility for endovascular treatment. We aimed to validate whether computed tomography combined with clinical information (CT&CI) or diffusion-weighted imaging (DWI) offers better predictive accuracy for anterior circulation LVO.

METHODS: Computed tomography combined with clinical information and DWI data from patients diagnosed with acute ischemic stroke were collected. Three deep-learning models, convolutional neural network, EfficientNet-B2, and DenseNet121, were used to compare CT&CI and DWI for detecting anterior circulation LVO.

RESULTS: A total of 456 patients, 228 patients with LVO [68.91 ± 12.84 years, 63.60% male; initial National Institutes of Health Stroke Scale (NIHSS) score: median 11 (7-14)] and without LVO [67.06 ± 12.29 years, 64.04% male; initial NIHSS score: median 2 (1-4)] were enrolled. Diffusion-weighted imaging achieved better results than CT&CI did in each performance metric. In DenseNet121, the area under the curves (AUCs) were found to be 0.833 and 0.756, respectively, while in EfficientNet-B2, the AUCs were 0.815 and 0.647, respectively.

CONCLUSIONS: In detecting the presence of anterior circulation LVO, DWI showed better results in each performance metric than CT&CI did, and the best-performing deep-learning model was DenseNet121.

PMID:40297357 | PMC:PMC12035260 | DOI:10.1177/20552076251334040

Sci Rep. 2025 Apr 28;15(1):14903. doi: 10.1038/s41598-025-98399-w.

ABSTRACT

Alzheimer's disease (AD) is characterized by a progressive spread of neurofibrillary tangles (NFT), beginning in the medial perirhinal cortex (mPRC), advancing to the entorhinal cortex (ERC), and subsequently involving the hippocampus, lateral perirhinal cortex (lPRC), and the rest of the brain. Given the close relationship between NFT accumulation and neuronal loss, the mPRC reflects a promising structural marker for early diagnosis of AD. However, only limited tools that automatically measure the cortical thickness of the mPRC are currently available. Utilizing the nnU-Net framework, we trained models on structural MRI of 126 adults, with manually segmented labels as ground truth. These models were then applied to an independent dataset of 103 adults (comprising patients with Alzheimer's dementia, amnestic mild cognitive impairment (aMCI), and healthy controls). High agreement was observed between manual and automated measurements of cortical thickness. Furthermore, we found significant atrophy in the Alzheimer's dementia group in the mPRC, ERC, and lPRC compared to healthy controls. Comparison of the aMCI group and healthy controls revealed significant differences in the ERC only. The results underscore the utility of our automated segmentation tool in advancing Alzheimer's research.

PMID:40295570 | DOI:10.1038/s41598-025-98399-w

Sci Rep. 2025 Apr 28;15(1):14876. doi: 10.1038/s41598-025-95803-3.

ABSTRACT

A brain tumor is a serious medical condition characterized by the abnormal growth of cells within the brain. It can cause a range of symptoms, including headaches, seizures, cognitive impairment, and changes in behavior. Brain tumors pose a significant health concern, imposing a substantial burden on patients. Timely diagnosis is crucial for effective treatment and patient health. Brain tumors can be either benign or malignant, and their symptoms often overlap with those of other neurological conditions, leading to delays in diagnosis. Early detection and diagnosis allow for timely intervention, potentially preventing the tumor from reaching an advanced stage. This reduces the risk of complications and increases the rate of recovery. Early detection is also significant in the selection of the most suitable treatment. In recent years, Smart IoT devices and deep learning techniques have brought remarkable success in various medical imaging applications. This study proposes a smart monitoring system for the early and timely detection, classification, and prediction of brain tumors. The proposed research employs a custom CNN model and two pre-trained models, specifically Inception-v4 and EfficientNet-B4, for classification of brain tumor cases into ten categories: Meningioma, Pituitary, No tumor, Astrocytoma, Ependymoma, Glioblastoma, Oligodendroglioma, Medulloblastoma, Germinoma, and Schwannoma. The custom CNN model is designed specifically to focus on computational efficiency and adaptability to address the unique challenges of brain tumor classification. Its adaptability to new challenges makes it a key component in the proposed smart monitoring system for brain tumor detection. Extensive experimentation is conducted to study a diverse set of brain MRI datasets and to evaluate the performance of the developed model. The model's precision, sensitivity, accuracy, f1-score, error rate, specificity, Y-index, balanced accuracy, geometric mean, and ROC are considered as performance metrics. The average classification accuracy for CNN, Inception-v4, and EfficientNet-B4 is 97.58%, 99.56%, and 99.76%, respectively. The results demonstrate the excellent accuracy and performance of the previous proposed approaches. Furthermore, the trained models maintain accurate performance after deployment. The method predicts accuracy of 96.5% for CNN, 99.3% for Inception-v4, and 99.7% for EfficientNet-B4 on a test dataset of 1000 brain tumor images.

PMID:40295548 | DOI:10.1038/s41598-025-95803-3

J Neurooncol. 2025 Apr 28. doi: 10.1007/s11060-025-05063-4. Online ahead of print.

ABSTRACT

PURPOSE: Chordomas and chondrosarcomas are rare, aggressive spinal bone tumors with distinct origins, biological behavior, and treatment challenges, primarily due to their resistance to conventional chemotherapy and radiation. This study aimed to compare clinical characteristics, treatment strategies, and long-term outcomes between spinal chordoma and chondrosarcoma, and to develop a robust machine learning-based model for individualized survival prediction.

METHODS: We conducted a retrospective analysis using the National Cancer Database (NCDB) to identify patients diagnosed with spinal chordoma or chondrosarcoma from 2004 to 2017. Demographics, tumor characteristics, comorbidity indices, treatment modalities (surgery, radiation, chemotherapy), and outcomes were extracted. Kaplan-Meier and weighted log-rank analyses assessed overall survival (OS) at predefined intervals (30-day, 90-day, 1-year, 5-year, 10-year). Twelve machine learning and deep learning models were trained to predict 10-year OS. Model performance was evaluated using AUC, Brier Score, and Concordance Index (C-index). A web-based risk calculator was developed using the best-performing ensemble model.

RESULTS: A total of 3175 patients were included (chordoma: n = 1204; chondrosarcoma: n = 1971). Chordoma patients were significantly older, travelled farther for treatment, and had smaller tumors with lower rates of metastatic disease at presentation. Chondrosarcoma patients more frequently underwent gross total resection, while chordoma patients received more radiation therapy, often with higher doses and more frequent use of proton therapy. Kaplan-Meier analysis revealed that chordoma patients had superior 10-year OS compared to chondrosarcoma patients (p < 0.0001). Among those receiving radiation, chondrosarcoma patients treated with radiation alone had the poorest survival. DeepSurv achieved the highest C-index (0.83) and lowest Brier Score (0.14), while ensemble models integrating Gradient Boosting and CatBoost also demonstrated strong performance (AUC > 0.80). Age, tumor type, and radiation therapy were identified as the most influential predictors using SHAP analysis. A publicly accessible, web-based calculator was developed for individualized survival prediction.

CONCLUSION: Spinal chordoma and chondrosarcoma differ significantly in clinical features and outcomes, with chordoma showing more favorable long-term survival. The findings highlight the importance of GTR and individualized radiation therapy in optimizing outcomes. The predictive model employing complicated machine learning models provides a valuable tool for estimating long-term survival and guiding personalized treatment strategies, though external validation is needed to strengthen its generalizability and clinical utility.

PMID:40295452 | DOI:10.1007/s11060-025-05063-4

J Comput Biol. 2025 Apr 28. doi: 10.1089/cmb.2024.0884. Online ahead of print.

ABSTRACT

Predicting the survival outcomes and assessing the risk of patients play a pivotal role in comprehending the microbial composition across various stages of cancer. With the ongoing advancements in deep learning, it has been substantiated that deep learning holds the potential to analyze patient survival risks based on microbial data. However, confronting a common challenge in individual cancer datasets involves the limited sample size and the high dimensionality of the feature space. This predicament often leads to overfitting issues in deep learning models, hindering their ability to effectively extract profound data representations and resulting in suboptimal model performance. To overcome these challenges, we advocate the utilization of pretraining and fine-tuning strategies, which have proven effective in addressing the constraint of having a smaller sample size in individual cancer datasets. In this study, we propose a deep learning model that amalgamates Transformer encoder and variational autoencoder (VAE), VTrans, employing both pre-training and fine-tuning strategies to predict the survival risk of cancer patients using microbial data. Furthermore, we highlight the potential of extending VTrans to integrate microbial multi-omics data. Our method is assessed on three distinct cancer datasets from The Cancer Genome Atlas Program, and the research findings demonstrated that (1) VTrans excels in terms of performance compared to conventional machine learning and other deep learning models. (2) The utilization of pretraning significantly enhances its performance. (3) In contrast to positional encoding, employing VAE encoding proves to be more effective in enriching data representation. (4) Using the idea of saliency map, it is possible to observe which microbes have a high contribution to the classification results. These results demonstrate the effectiveness of VTrans in prediting patient survival risk. Source code and all datasets used in this paper are available at https://github.com/wenwenmin/VTrans and https://doi.org/10.5281/zenodo.14166580.

PMID:40295093 | DOI:10.1089/cmb.2024.0884

Anticancer Res. 2025 May;45(5):1953-1964. doi: 10.21873/anticanres.17572.

ABSTRACT

BACKGROUND/AIM: Bladder cancer (BCa) is associated with high recurrence rates, emphasizing the importance of early and accurate detection. This study aimed to develop a lightweight and fast deep learning model, Light-Bladder-Net (LBN), for non-invasive BCa detection using conventional urine data.

MATERIALS AND METHODS: We improved LBN's generalization by applying data transformations, adding uniform noise, and employing feature selection methods (mRMR, PCA, SVD, t-SNE) to extract key vectors from its fully connected layer. These vectors were integrated into the original dataset, and multiple machine learning models were trained to enhance classification accuracy. Lastly, weighted voting was used to assign importance across these models.

RESULTS: Our approach achieved an accuracy of 0.83, a sensitivity of 0.85, a specificity of 0.80, and a precision of 0.81, indicating robust performance in detecting BCa from urine data.

CONCLUSION: This non-invasive diagnostic method offers rapid, cost-effective predictions. A free online tool is available for clinicians and patients to conveniently detect BCa using standard urine samples at http://merlin.nchu.edu.tw/LBN/.

PMID:40295062 | DOI:10.21873/anticanres.17572

Med Phys. 2025 Apr 28. doi: 10.1002/mp.17855. Online ahead of print.

ABSTRACT

BACKGROUND: The accurate automated extraction of brain vessel centerlines from Computed tomographic angiography (CTA) images plays an important role in diagnosing and treating cerebrovascular diseases such as stroke. Despite its significance, this task is complicated by the complex cerebrovascular structure and heterogeneous imaging quality.

PURPOSE: This study aims to develop and validate a segmentation-assisted framework designed to improve the accuracy and efficiency of brain vessel centerline extraction from CTA images. We streamline the process of lumen segmentation generation without additional annotation effort from physicians, enhancing the effectiveness of centerline extraction.

METHODS: The framework integrates four modules: (1) pre-processing techniques that register CTA images with a CT atlas and divide these images into input patches, (2) lumen segmentation generation from annotated vessel centerlines using graph cuts and robust kernel regression, (3) a dual-branch topology-aware UNet (DTUNet) that optimizes the use of the annotated vessel centerlines and the generated lumen segmentation via a topology-aware loss (TAL) and its dual-branch structure, and (4) post-processing methods that skeletonize and refine the lumen segmentation predicted by the DTUNet.

RESULTS: An in-house dataset derived from a subset of the MR CLEAN Registry is used to evaluate the proposed framework. The dataset comprises 10 intracranial CTA images, and 40 cube CTA sub-images with a resolution of 128 × 128 × 128 $128 \times 128 \times 128$ voxels. Via five-fold cross-validation on this dataset, we demonstrate that the proposed framework consistently outperforms state-of-the-art methods in terms of average symmetric centerline distance (ASCD) and overlap (OV). Specifically, it achieves an ASCD of 0.84, an OV 1.0 $\textrm {OV}_{1.0}$ of 0.839, and an OV 1.5 $\textrm {OV}_{1.5}$ of 0.885 for intracranial CTA images, and obtains an ASCD of 1.26, an OV 1.0 $\textrm {OV}_{1.0}$ of 0.779, and an OV 1.5 $\textrm {OV}_{1.5}$ of 0.824 for cube CTA sub-images. Subgroup analyses further suggest that the proposed framework holds promise in clinical applications for stroke diagnosis and treatment.

CONCLUSIONS: By automating the process of lumen segmentation generation and optimizing the network design of vessel centerline extraction, DTUnet achieves high performance without introducing additional annotation demands. This solution promises to be beneficial in various clinical applications in cerebrovascular disease management.

PMID:40296200 | DOI:10.1002/mp.17855

Med Phys. 2025 Apr 28. doi: 10.1002/mp.17845. Online ahead of print.

ABSTRACT

BACKGROUND: The main task of deep learning (DL) based brain tumor segmentation is to get accurate projection from learned image features to their corresponding semantic labels (i.e., brain tumor sub-regions). To achieve this goal, segmentation networks are required to learn image features with high intra-class consistency. However, brain tumor are known to be heterogeneous, and it often causes high diversity in image gray values which further influences the learned image features. Therefore, projecting such diverse image features (i.e., low intra-class consistency) to the same semantic label is often difficult and inefficient.

PURPOSE: The purpose of this study is to address the issue of low intra-class consistency of image features learned from heterogeneous brain tumor regions and ease the projection of image features to their corresponding semantic labels. In this way, accurate segmentation of brain tumor can be achieved.

METHODS: We propose a new DL-based method for brain tumor segmentation, where a semantic feature module (SFM) is introduced to consolidate image features with meaningful semantic information and enhance their intra-class consistency. Specifically, in the SFM, deep semantic vectors are derived and used as prototypes to re-encode image features learned in the segmentation network. Since the relatively consistent deep semantic vectors, diversity of the resulting image features can be reduced; moreover, semantic information in the resulting image features can also be enriched, both facilitating accurate projection to the final semantic labels.

RESULTS: In the experiment, a public brain tumor dataset, BraTS2022 containing, multi-sequence MR images of 1251 patients is used to evaluate our method in the task of brain tumor sub-region segmentation, and the experimental results demonstrate that, benefiting from the SFM, our method outperforms the state-of-the-art methods with statistical significance ( p < 0.05 $p<0.05$ using the Wilcoxon signed rank test). Further ablation study shows that the proposed SFM can yield an improvement in segmentation accuracy (Dice index) of up to 11% comparing with that without the SFM.

CONCLUSIONS: In DL-based segmentation, low intra-class consistency of learned image features degrades segmentation performance. The proposed SFM can effectively enhance the intra-class consistency with high-level semantic information, making the projection of image features to their corresponding semantic labels more accurate.

PMID:40296197 | DOI:10.1002/mp.17845

Med Phys. 2025 Apr 28. doi: 10.1002/mp.17860. Online ahead of print.

ABSTRACT

BACKGROUND: Magnetic resonance imaging (MRI) is a crucial medical imaging technique that can determine the structural and functional status of body tissues and organs. However, the prolonged MRI acquisition time increases the scanning cost and limits its use in less developed areas.

PURPOSE: The objective of this study is to design a lightweight, data-driven under-sampling pattern for fastMRI to achieve a balance between MRI reconstruction quality and sampling time while also being able to be integrated with deep learning to further improve reconstruction quality.

METHODS: In this study, we attempted to establish a connection between k-space and the corresponding MRI through singular value decomposition(SVD). Specifically, we apply SVD to MRI to decouple it into multiple components, which are sorted by energy contribution. Then, the sampling points that match the energy contribution in the k-space, which correspond to each component are selected sequentially. Finally, the sampling points obtained from all components are merged to obtain a mask. This mask can be used directly as a sampler or integrated into deep learning as an initial or fixed sampling points.

RESULTS: The experiments were conducted on two public datasets, and the results demonstrate that when the mask generated based on our method is directly used as the sampler, the MRI reconstruction quality surpasses that of state-of-the-art heuristic samplers. In addition, when integrated into the deep learning models, the models converge faster and the sampler performance is significantly improved.

CONCLUSIONS: The proposed lightweight data-driven sampling approach avoids time-consuming parameter tuning and the establishment of complex mathematical models, achieving a balance between reconstruction quality and sampling time.

PMID:40296184 | DOI:10.1002/mp.17860

BMC Sports Sci Med Rehabil. 2025 Apr 28;17(1):104. doi: 10.1186/s13102-025-01147-w.

ABSTRACT

BACKGROUND: This study aims to define and cluster tennis match types based on how they are played.

METHODS: The research data selected for this study were from the 100th round of 32 matches of the five finals of the 2023 International Tennis Open Tournament. Based on expert knowledge and sports expertise, 27 variables were included across seven areas. Three models were applied and the silhouette coefficient was calculated to identify the optimal number of clusters. A difference test was conducted on the game record variables based on the cluster results.

RESULTS: Calculation of the silhouette coefficients for the three models showed that Model 3 (silhouette coefficient: 0.406) had the highest performance. The clustering results for the tennis match types are as follows. First, the NEt Rusher Defensive type, which is defensive and induces net play. Second, the ALl Courter Defensive type, which is either defensive or all-round. Third, the STroke Placement Offensive type, which is aggressive and has strengths in stroke. Fourth, the SErve Placement Offensive type, which is aggressive and has strengths in sub courses.

CONCLUSION: This study's findings are not only provide basic data to cluster game types in tennis matches but also to contribute to establishing game strategies for each game type, thereby further improving performance.

PMID:40296175 | DOI:10.1186/s13102-025-01147-w

BMC Nephrol. 2025 Apr 28;26(1):214. doi: 10.1186/s12882-025-04133-z.

ABSTRACT

BACKGROUND: Maintenance hemodialysis patients experience high morbidity and mortality, primarily from cardiovascular and infectious diseases. It was discovered recently that low arterial oxygen saturation (SaO2) is associated with a pro-inflammatory phenotype and poor patient outcomes. Sleep apnea is highly prevalent in maintenance hemodialysis patients and may contribute to intradialytic hypoxemia. In sleep apnea, normal respiration patterns are disrupted by episodes of apnea because of either disturbed respiratory control (i.e., central sleep apnea) or upper airway obstruction (i.e., obstructive sleep apnea). Intermittent SaO2 saw-tooth patterns are a hallmark of sleep apnea. Continuous intradialytic measurements of SaO2 provide an opportunity to follow the temporal evolution of SaO2 during hemodialysis. Using artificial intelligence, we aimed to automatically identify patients with repetitive episodes of intermittent SaO2 saw-tooth patterns.

METHODS: The analysis utilized intradialytic SaO2 measurements by the Crit-Line device (Fresenius Medical Care, Waltham, MA). In patients with an arterio-venous fistula as vascular access, this FDA approved device records 150 SaO2 measurements per second in the extracorporeal blood circuit of the hemodialysis system. The average SaO2 of a 10-second segment is computed and streamed to the cloud. Periods comprising thirty 10-second segments (i.e., 300 s or five minutes) were independently adjudicated by two researchers for the presence or absence of SaO2 saw-tooth pattern. We built one-dimensional convolutional neural networks (1D-CNN), a state-of-the-art deep learning method, for SaO2 pattern classification and randomly assigned SaO2 time series segments to either a training (80%) or a test (20%) set.

RESULTS: We analyzed 4,075 consecutive 5-minute segments from 89 hemodialysis treatments in 22 hemodialysis patients. While 891 (21.9%) segments showed saw-tooth pattern, 3,184 (78.1%) did not. In the test data set, the rate of correct SaO2 pattern classification was 96% with an area under the receiver operating curve of 0.995 (95% CI: 0.993 to 0.998).

CONCLUSION: Our 1D-CNN algorithm accurately classifies SaO2 saw-tooth pattern. The SaO2 pattern classification can be performed in real time during an ongoing hemodialysis treatment, provide timely alert in the event of respiratory instability or sleep apnea, and trigger further diagnostic and therapeutic interventions.

PMID:40295983 | DOI:10.1186/s12882-025-04133-z

BMC Med Imaging. 2025 Apr 28;25(1):138. doi: 10.1186/s12880-025-01684-3.

ABSTRACT

BACKGROUND: To develop and validate deep learning (DL) and traditional clinical-metabolic (CM) models based on 18 F-FDG PET/CT images for noninvasively predicting high-grade patterns (HGPs) of invasive lung adenocarcinoma (LUAD).

METHODS: A total of 303 patients with invasive LUAD were enrolled in this retrospective study; these patients were randomly divided into training, validation and test sets at a ratio of 7:1:2. DL models were trained and optimized on PET, CT and PET/CT fusion images, respectively. CM model was built from clinical and PET/CT metabolic parameters via backwards stepwise logistic regression and visualized via a nomogram. The prediction performance of the models was evaluated mainly by the area under the curve (AUC). We also compared the AUCs of different models for the test set.

RESULTS: CM model was established upon clinical stage (OR: 7.30; 95% CI: 2.46-26.37), cytokeratin 19 fragment 21 - 1 (CYFRA 21-1, OR: 1.18; 95% CI: 0.96-1.57), mean standardized uptake value (SUVmean, OR: 1.31; 95% CI: 1.17-1.49), total lesion glycolysis (TLG, OR: 0.994; 95% CI: 0.990-1.000) and size (OR: 1.37; 95% CI: 0.95-2.02). Both the DL and CM models exhibited good prediction efficacy in the three cohorts, with AUCs ranging from 0.817 to 0.977. For the test set, the highest AUC was yielded by the CT-DL model (0.895), followed by the PET/CT-DL model (0.882), CM model (0.879) and PET-DL model (0.817), but no significant difference was revealed between any two models.

CONCLUSIONS: Deep learning and clinical-metabolic models based on the 18F-FDG PET/CT model could effectively identify LUAD patients with HGP. These models could aid in treatment planning and precision medicine.

CLINICAL TRIAL NUMBER: Not applicable.

PMID:40295979 | DOI:10.1186/s12880-025-01684-3

BMC Cardiovasc Disord. 2025 Apr 28;25(1):335. doi: 10.1186/s12872-025-04728-2.

ABSTRACT

BACKGROUND: The risk stratification and prognosis of cardiac arrhythmia depend on the individual condition of patients, while invasive diagnostic methods may be risky to patient health, and current non-invasive diagnostic methods are applicable to few disease types without sensitivity and specificity. Cardiac electrophysiologic imaging (ECGI) technology reflects cardiac activities accurately and non-invasively, which is of great significance for the diagnosis and treatment of cardiac diseases. This paper aims to provide a new solution for the realization of ECGI by combining simulation model and deep learning methods.

METHODS: A complete three-dimensional bidomain cardiac electrophysiologic activity model was constructed, and simulated electrocardiogram data were obtained as training samples. Particle swarm optimization-back propagation neural network, convolutional neural network, and long short-term memory network were used respectively to reconstruct the cardiac surface potential.

RESULTS: The correlation coefficients between the simulation results and the clinical data range from 75.76 to 84.61%. The P waves, PR intervals, QRS complex, and T waves in the simulated waveforms were within the normal clinical range, and the distribution trend of the simulated body surface potential mapping was consistent with the clinical data. The coefficient of determination R2 between the reconstruction results of all the algorithms and the true value is above 0.80, and the mean absolute error is below 2.1 mV, among which the R2 of long short-term memory network is about 0.99 and the mean absolute error about 0.5 mV.

CONCLUSIONS: The electrophysiologic model constructed in this study can reflect cardiac electrical activity, and contains the mapping relationship between the cardiac potential and the body surface potential. In cardiac potential reconstruction, long short-term memory network has significant advantages over other algorithms.

CLINICAL TRIAL NUMBER: Not applicable.

PMID:40295939 | DOI:10.1186/s12872-025-04728-2