Hepatol Int. 2025 Mar 16. doi: 10.1007/s12072-025-10793-8. Online ahead of print.

ABSTRACT

PURPOSE: Existing prognostic staging systems depend on expensive manual extraction by pathologists, potentially overlooking latent patterns critical for prognosis, or use black-box deep learning models, limiting clinical acceptance. This study introduces a novel deep learning-assisted paradigm that complements existing approaches by generating interpretable, multi-view risk scores to stratify prognostic risk in hepatocellular carcinoma (HCC) patients.

METHODS: 510 HCC patients were enrolled in an internal dataset (SYSUCC) as training and validation cohorts to develop the Hybrid Deep Score (HDS). The Attention Activator (ATAT) was designed to heuristically identify tissues with high prognostic risk, and a multi-view risk-scoring system based on ATAT established HDS from microscopic to macroscopic levels. HDS was also validated on an external testing cohort (TCGA-LIHC) with 341 HCC patients. We assessed prognostic significance using Cox regression and the concordance index (c-index).

RESULTS: The ATAT first heuristically identified regions where necrosis, lymphocytes, and tumor tissues converge, particularly focusing on their junctions in high-risk patients. From this, this study developed three independent risk factors: microscopic morphological, co-localization, and deep global indicators, which were concatenated and then input into a neural network to generate the final HDS for each patient. The HDS demonstrated competitive results with hazard ratios (HR) (HR 3.24, 95% confidence interval (CI) 1.91-5.43 in SYSUCC; HR 2.34, 95% CI 1.58-3.47 in TCGA-LIHC) and c-index values (0.751 in SYSUCC; 0.729 in TCGA-LIHC) for Disease-Free Survival (DFS). Furthermore, integrating HDS into existing clinical staging systems allows for more refined stratification, which enables the identification of potential high-risk patients within low-risk groups.

CONCLUSION: This novel paradigm, from identifying high-risk tissues to constructing prognostic risk scores, offers fresh insights into HCC research. Additionally, the integration of HDS complements the existing clinical staging system by facilitating more detailed stratification in DFS and Overall Survival (OS).

PMID:40089963 | DOI:10.1007/s12072-025-10793-8

Ann Nucl Med. 2025 Mar 16. doi: 10.1007/s12149-025-02038-3. Online ahead of print.

ABSTRACT

OBJECTIVE: To provide fully automatic scanner-independent 5-level categorization of the [123I]FP-CIT uptake in striatal subregions in dopamine transporter SPECT.

METHODS: A total of 3500 [123I]FP-CIT SPECT scans from two in house (n = 1740, n = 640) and two external (n = 645, n = 475) datasets were used for this study. A convolutional neural network (CNN) was trained for the categorization of the [123I]FP-CIT uptake in unilateral caudate and putamen in both hemispheres according to 5 levels: normal, borderline, moderate reduction, strong reduction, almost missing. Reference standard labels for the network training were created automatically by fitting a Gaussian mixture model to histograms of the specific [123I]FP-CIT binding ratio, separately for caudate and putamen and separately for each dataset. The CNN was trained on a mixed-scanner subsample (n = 1957) and tested on one independent identically distributed (IID, n = 1068) and one out-of-distribution (OOD, n = 475) test dataset.

RESULTS: The accuracy of the CNN for the 5-level prediction of the [123I]FP-CIT uptake in caudate/putamen was 80.1/78.0% in the IID test dataset and 78.1/76.5% in the OOD test dataset. All 4 regional 5-level predictions were correct in 54.3/52.6% of the cases in the IID/OOD test dataset. A global binary score automatically derived from the regional 5-scores achieved 97.4/96.2% accuracy for automatic classification of the scans as normal or reduced relative to visual expert read as reference standard.

CONCLUSIONS: Automatic scanner-independent 5-level categorization of the [123I]FP-CIT uptake in striatal subregions by a CNN model is feasible with clinically useful accuracy.

PMID:40089953 | DOI:10.1007/s12149-025-02038-3

Adv Sci (Weinh). 2025 Mar 16:e2500859. doi: 10.1002/advs.202500859. Online ahead of print.

ABSTRACT

Evolutionary constraints significantly limit the diversity of naturally occurring enzymes, thereby reducing the sequence repertoire available for enzyme discovery and engineering. Recent breakthroughs in protein structure prediction and de novo design, powered by artificial intelligence, now enable to create enzymes with desired functions without solely relying on traditional genome mining. Here, a computational strategy is demonstrated for creating new-to-nature polyethylene terephthalate hydrolases (PET hydrolases) by leveraging the known catalytic mechanisms and implementing multiple deep learning algorithms and molecular computations. This strategy includes the extraction of functional motifs from a template enzyme (here leaf-branch compost cutinase, LCC, is used), regeneration of new protein sequences, computational screening, experimental validation, and sequence refinement. PET hydrolytic activity is successfully replicated with designer enzymes that are at least 30% shorter in sequence length than LCC. Among them, RsPETase1 stands out due to its robust expressibility. It exhibits comparable catalytic efficiency (kcat/Km) to LCC and considerable thermostability with a melting temperature of 56 °C, despite sharing only 34% sequence similarity with LCC. This work suggests that enzyme diversity can be expanded by recapitulating functional motifs with computationally built protein scaffolds, thus generating opportunities to acquire highly active and robust enzymes that do not exist in nature.

PMID:40089854 | DOI:10.1002/advs.202500859

Syst Rev. 2025 Mar 15;14(1):62. doi: 10.1186/s13643-025-02779-2.

ABSTRACT

BACKGROUND: Artificial intelligence (AI) has shown immense potential in the field of medicine, but its actual effectiveness and safety still need to be validated through clinical trials. Currently, the research themes, methodologies, and development trends of AI-related clinical trials remain unclear, and further exploration of these studies will be crucial for uncovering AI's practical application potential and promoting its broader adoption in clinical settings.

OBJECTIVE: To analyze the current status, hotspots, and trends of published clinical research on AI applications.

METHODS: Publications related to AI clinical applications were retrieved from the Web of Science database. Relevant data were extracted using VOSviewer 1.6.17 to generate visual cooperation network maps for countries, organizations, authors, and keywords. Burst citation detection for keywords and citations was performed using CiteSpace 5.8.R3 to identify sudden surges in citation frequency within a short period, and the theme evolution was analyzed using SciMAT to track the development and trends of research topics over time.

RESULTS: A total of 22,583 articles were obtained from the Web of Science database. Seven-hundred and thirty-five AI clinical application research were published by 1764 institutions from 53 countries. The majority of publications were contributed by the United States, China, and the UK. Active collaborations were noted among leading authors, particularly those from developed countries. The publications mainly focused on evaluating the application value of AI technology in the fields of disease diagnosis and classification, disease risk prediction and management, assisted surgery, and rehabilitation. Deep learning and chatbot technologies were identified as emerging research hotspots in recent studies on AI applications.

CONCLUSIONS: A total of 735 articles on AI in clinical research were analyzed, with publication volume and citation counts steadily increasing each year. Institutions and researchers from the United States contributed the most to the research output in this field. Key areas of focus included AI applications in surgery, rehabilitation, disease diagnosis, risk prediction, and health management, with emerging trends in deep learning and chatbots. This study also provides detailed and intuitive information about important articles, journals, core authors, institutions, and topics in the field through visualization maps, which will help researchers quickly understand the current status, hotspots, and trends of artificial intelligence clinical application research. Future clinical trials of artificial intelligence should strengthen scientific design, ethical compliance, and interdisciplinary and international cooperation and pay more attention to its practical clinical value and reliable application in diverse scenarios.

PMID:40089747 | DOI:10.1186/s13643-025-02779-2

Sci Rep. 2025 Mar 15;15(1):9015. doi: 10.1038/s41598-025-93011-7.

ABSTRACT

Physicians evaluate a patient's respiratory health during a physical examination by visual assessment of the work of breathing (WoB) to determine respiratory stability, and by detecting abnormal lung sounds via lung auscultation using a stethoscope to identify common pathological lung diseases, such as chronic obstructive pulmonary disease (COPD) and pneumonia. Since these assessment methods are subjective, a low-profile device used for an accurate and quantitative monitoring approach could provide valuable preemptive insights into respiratory health, proving to be clinically beneficial. To achieve this goal, we have developed a miniature patch consisting of a sensitive wideband multi-axis seismometer that can be placed on the anatomical areas of a patient's lungs to enable an effective quantification of a patient's WoB and lung sounds. When used on a patch, the seismometer captures chest wall vibrations due to respiratory muscle effort, known as high-frequency mechanomyogram (MMG), during tidal breathing as well as seismic pulmonary-induced vibrations (PIVs) during deep breathing due to normal and/or adventitious lung sounds like crackles, while simultaneously recording respiration rate and phase. A system comprised of multiple patches was evaluated on 124 patients in the hospital setting and shown to accurately assess and quantify a patent's physical signs of WoB by measuring the average respiratory effort extracted from high-frequency MMG signals, demonstrating statistical significance of this method in comparison to clinical bedside observation of WoB and respiration rate. A data fusion deep learning model was developed which combined the inputs of PIVs lung sounds and the corresponding respiration phase to detect crackle, wheeze and normal breath sound features. The model exhibited high accuracy, sensitivity, specificity, precision and F1 score of 93%, 93%, 97%, 93% and 93% respectively, with area under the curve (AUC) of precision recall (PR) of 0.97 on the test set. Additionally, the PIVs with corresponding respiration phase captured from each auscultation point generated an acoustic map of the patient's lung, which correlated with traditional lung radiographic findings.

PMID:40089574 | DOI:10.1038/s41598-025-93011-7

Sci Rep. 2025 Mar 15;15(1):9005. doi: 10.1038/s41598-025-88687-w.

ABSTRACT

Social media has become a powerful tool for public discourse, shaping opinions and the emotional landscape of communities. The extensive use of social media has led to a massive influx of online content. This content includes instances where negativity is amplified through hateful speech but also a significant number of posts that provide support and encouragement, commonly known as hope speech. In recent years, researchers have focused on the automatic detection of hope speech in languages such as Russian, English, Hindi, Spanish, and Bengali. However, to the best of our knowledge, detection of hope speech in Urdu and English, particularly using translation-based techniques, remains unexplored. To contribute to this area we have created a multilingual dataset in English and Urdu and applied a translation-based approach to handle multilingual challenges and utilized several state-of-the-art machine learning, deep learning, and transfer learning based methods to benchmark our dataset. Our observations indicate that a rigorous process for annotator selection, along with detailed annotation guidelines, significantly improved the quality of the dataset. Through extensive experimentation, our proposed methodology, based on the Bert transformer model, achieved benchmark performance, surpassing traditional machine learning models with accuracies of 87% for English and 79% for Urdu. These results show improvements of 8.75% in English and 1.87% in Urdu over baseline models (SVM 80% English and 78% in Urdu).

PMID:40089522 | DOI:10.1038/s41598-025-88687-w

NPJ Sci Food. 2025 Mar 15;9(1):31. doi: 10.1038/s41538-025-00393-z.

ABSTRACT

Veterinary drug residues in poultry and livestock products present persistent challenges to food safety, necessitating precise and efficient detection methods. Surface-enhanced Raman scattering (SERS) has been identified as a powerful tool for veterinary drug residue analysis due to its high sensitivity and specificity. However, the development of reliable SERS substrates and the interpretation of complex spectral data remain significant obstacles. This review summarizes the development process of SERS substrates, categorizing them into metal-based, rigid, and flexible substrates, and highlighting the emerging trend of multifunctional substrates. The diverse application scenarios and detection requirements for these substrates are also discussed, with a focus on their use in veterinary drug detection. Furthermore, the integration of deep learning techniques into SERS-based detection is explored, including substrate structure design optimization, optical property prediction, spectral preprocessing, and both qualitative and quantitative spectral analyses. Finally, key limitations are briefly outlined, such as challenges in selecting reporter molecules, data imbalance, and computational demands. Future trends and directions for improving SERS-based veterinary drug detection are proposed.

PMID:40089516 | DOI:10.1038/s41538-025-00393-z

Sci Rep. 2025 Mar 15;15(1):8938. doi: 10.1038/s41598-025-92994-7.

ABSTRACT

Cracks are common defects in physical structures, and if not detected and addressed in a timely manner, they can pose a severe threat to the overall safety of the structure. In recent years, with advancements in deep learning, particularly the widespread use of Convolutional Neural Networks (CNNs) and Transformers, significant breakthroughs have been made in the field of crack detection. However, CNNs still face limitations in capturing global information due to their local receptive fields when processing images. On the other hand, while Transformers are powerful in handling long-range dependencies, their high computational cost remains a significant challenge. To effectively address these issues, this paper proposes an innovative modification to the VM-UNet model. This modified model strategically integrates the strengths of the Mamba architecture and UNet to significantly improve the accuracy of crack segmentation. In this study, we optimized the original VM-UNet architecture to better meet the practical needs of crack segmentation tasks. Through comparative experiments on the Crack500 and Ozgenel public datasets, the results clearly demonstrate that the improved VM-UNet achieves significant advancements in segmentation accuracy. Compared to the original VM-UNet and other state-of-the-art models, VM-UNet++ shows a 3% improvement in mDS and a 4.6-6.2% increase in mIoU. These results fully validate the effectiveness of our improvement strategy. Additionally, VM-UNet++ demonstrates lower parameter count and floating-point operations, while maintaining a relatively satisfactory inference speed. These improvements make VM-UNet++ advantageous for practical applications.

PMID:40089495 | DOI:10.1038/s41598-025-92994-7

Magn Reson Imaging. 2025 Mar 13:110370. doi: 10.1016/j.mri.2025.110370. Online ahead of print.

ABSTRACT

OBJECTIVE: To evaluate whether deep learning analysis (DL) of intratumor subregion based on dynamic contrast-enhanced MRI (DCE-MRI) can help predict Ki-67 expression level in breast cancer.

MATERIALS AND METHODS: A total of 290 breast cancer patients from two hospitals were retrospectively collected. A k-means clustering algorithm confirmed subregions of tumor. DL features of whole tumor and subregions were extracted from DCE-MRI images based on 3D ResNet18 pre-trained model. The logistic regression model was constructed after dimension reduction. Model performance was assessed using the area under the curve (AUC), and clinical value was demonstrated through decision curve analysis (DCA).

RESULTS: The k-means clustering method clustered the tumor into two subregions (habitat 1 and habitat 2) based on voxel values. Both the habitat 1 model (validation set: AUC = 0.771, 95 %CI: 0.642-0.900 and external test set: AUC = 0.794, 95 %CI: 0.696-0.891) and the habitat 2 model (AUC = 0.734, 95 %CI: 0.605-0.862 and AUC = 0.756, 95 %CI: 0.646-0.866) showed better predictive capabilities for Ki-67 expression level than the whole tumor model (AUC = 0.686, 95 %CI: 0.550-0.823 and AUC = 0.680, 95 %CI: 0.555-0.804). The combined model based on the two subregions further enhanced the predictive capability (AUC = 0.808, 95 %CI: 0.696-0.921 and AUC = 0.842, 95 %CI: 0.758-0.926), and it demonstrated higher clinical value than other models in DCA.

CONCLUSIONS: The deep learning model derived from subregion of tumor showed better performance for predicting Ki-67 expression level in breast cancer patients. Additionally, the model that integrated two subregions further enhanced the predictive performance.

PMID:40089082 | DOI:10.1016/j.mri.2025.110370

Int J Radiat Oncol Biol Phys. 2025 Mar 13:S0360-3016(25)00243-3. doi: 10.1016/j.ijrobp.2025.03.012. Online ahead of print.

ABSTRACT

BACKGROUND: Patients with identical stages or similar tumor volumes can vary significantly in their responses to radiotherapy (RT) due to individual characteristics, making personalized RT for non-small cell lung cancer (NSCLC) challenging. This study aimed to develop a deep learning (DL) model by integrating pre- and mid-treatment computed tomography (CT) to predict the treatment response in NSCLC patients.

METHODS AND MATERIAL: We retrospectively collected data from 168 NSCLC patients across three hospitals. Data from A (35 patients) and B (93 patients) were used for model training and internal validation, while data from C (40 patients) was used for external validation. DL, radiomics, and clinical features were extracted to establish a varying time-interval long short-term memory network (VTI-LSTM) for response prediction. Furthermore, we derived a model-deduced personalize dose escalation (DE) for patients predicted to have suboptimal gross tumor volume (GTV) regression. The area under the receiver operating characteristic curve (AUC) and predicted absolute error (PAE) were used to evaluate the predictive Response Evaluation Criteria in Solid Tumors (RECIST) classification and proportion of GTV residual. DE was calculated as biological equivalent dose (BED) using an α/β ratio of 10 Gy.

RESULTS: The model using only pre-treatment CT achieved the highest AUC of 0.762 and 0.687 in internal and external validation respectively, while the model integrating both pre- and mid-treatment CT achieved AUC of 0.869 and 0.798, with PAE of 0.137 and 0.185. We performed personalized DE for 29 patients. Their original BED was approximately 72 Gy, within the range of 71.6 Gy to 75 Gy. DE ranged from 77.7 to 120 Gy for 29 patients, with 17 patients exceeding 100 Gy and eight patients reaching the model's preset upper limit of 120 Gy.

CONCLUSIONS: Combining pre- and mid-treatment CT enhances prediction performance for RT response and offers a promising approach for personalized DE in NSCLC.

PMID:40089073 | DOI:10.1016/j.ijrobp.2025.03.012

Neural Netw. 2025 Mar 10;187:107337. doi: 10.1016/j.neunet.2025.107337. Online ahead of print.

ABSTRACT

Deep learning for Electroencephalography (EEG) has become dominant in the tasks of discrimination and evaluation of brain disorders. However, despite its significant successes, this approach has long been facing challenges due to the limited availability of labeled samples and the individuality of subjects, particularly in complex scenarios such as Autism Spectrum Disorders (ASD). To facilitate the efficient optimization of EEG discrimination models in the face of these limitations, this study has developed a framework called STEM (Self-Training EEG Model). STEM accomplishes this by self-training the model, which involves initializing it with limited labeled samples and optimizing it with self-constructed samples. (1) Model initialization with multi-task learning: A multi-task model (MAC) comprising an AutoEncoder and a classifier offers guidance for subsequent pseudo-labeling. This guidance includes task-related latent EEG representations and prediction probabilities of unlabeled samples. The AutoEncoder, which consists of depth-separable convolutions and BiGRUs, is responsible for learning comprehensive EEG representations through the EEG reconstruction task. Meanwhile, the classifier, trained using limited labeled samples through supervised learning, directs the model's attention towards capturing task-related features. (2) Model optimization aided by pseudo-labeled samples construction: Next, trustworthy pseudo-labels are assigned to the unlabeled samples, and this approach (PLASC) combines the sample's distance relationship in the feature space mapped by the encoder with the sample's predicted probability, using the initial MAC model as a reference. The constructed pseudo-labeled samples then support the self-training of MAC to learn individual information from new subjects, potentially enhancing the adaptation of the optimized model to samples from new subjects. The STEM framework has undergone an extensive evaluation, comparing it to state-of-the-art counterparts, using resting-state EEG data collected from 175 ASD-suspicious children spanning different age groups. The observed results indicate the following: (1) STEM achieves the best performance, with an accuracy of 88.33% and an F1-score of 87.24%, and (2) STEM's multi-task learning capability outperforms supervised methods when labeled data is limited. More importantly, the use of PLASC improves the model's performance in ASD discrimination across different age groups, resulting in an increase in accuracy (3%-8%) and F1-scores (4%-10%). These increments are approximately 6% higher than those achieved by the comparison methods.

PMID:40088831 | DOI:10.1016/j.neunet.2025.107337

Waste Manag. 2025 Mar 14;200:114742. doi: 10.1016/j.wasman.2025.114742. Online ahead of print.

ABSTRACT

The surge in medical waste has highlighted the urgent need for cost-effective and advanced management solutions. In this paper, a novel medical waste management approach, "MedBin," is proposed for automated sorting, reusing, and recycling. A comprehensive medical waste dataset, "MedBin-Dataset" is established, comprising 2,119 original images spanning 36 categories, with samples captured in various backgrounds. The lightweight "MedBin-Net" model is introduced to enable detection and instance segmentation of medical waste, enhancing waste recognition capabilities. Experimental results demonstrate the effectiveness of the proposed approach, achieving an average precision of 0.91, recall of 0.97, and F1-score of 0.94 across all categories with just 2.51 M parameters (where M stands for million, i.e., 2.51 million parameters), 5.20G FLOPs (where G stands for billion, i.e., 5.20 billion floating-point operations per second), and 0.60 ms inference time. Additionally, the proposed method includes a World Health Organization (WHO) Guideline-Based Classifier that categorizes detected waste into 5 types, each with a corresponding disposal method, following WHO medical waste classification standards. The proposed method, along with the dedicated dataset, offers a promising solution that supports sustainable medical waste management and other related applications. To access the MedBin-Dataset samples, please visit https://universe.roboflow.com/uob-ylti8/medbin_dataset. The source code for MedBin-Net can be found at https://github.com/Wayne3918/MedbinNet.

PMID:40088805 | DOI:10.1016/j.wasman.2025.114742

Ecohealth. 2025 Mar 15. doi: 10.1007/s10393-025-01705-1. Online ahead of print.

ABSTRACT

The World Health Organization (WHO) aims to ensure the highest level of health for all populations. Despite progress, increased life expectancy has not translated into a proportional rise in healthy life years, as chronic diseases are on the rise. In this context, health ecology emerges as a new scientific discipline focused on preserving health rather than curing diseases. It seeks to calculate healthy life expectancy by analyzing individual, social, and systemic choices, offering a proactive and rigorous approach to making informed decisions and improving long-term well-being.

PMID:40088354 | DOI:10.1007/s10393-025-01705-1

Interdiscip Sci. 2025 Mar 15. doi: 10.1007/s12539-025-00695-6. Online ahead of print.

ABSTRACT

Drug combination therapy has shown improved efficacy and decreased adverse effects, making it a practical approach for conditions like cancer. However, discovering all potential synergistic drug combinations requires extensive experimentation, which can be challenging. Recent research utilizing deep learning techniques has shown promise in reducing the number of experiments and overall workload by predicting synergistic drug combinations. Therefore, developing reliable and effective computational methods for predicting these combinations is essential. This paper proposed a novel method called Drug-molecule Connect Cell (DconnC) for predicting synergistic drug combinations. DconnC leverages cellular features as nodes to establish connections between drug molecular structures, allowing the extraction of pertinent features. These features are then optimized through self-augmented contrastive learning using bidirectional recurrent neural networks (Bi-RNN) and long short-term memory (LSTM) models, ultimately predicting the drug synergy. By integrating information about the molecular structure of drugs for the extraction of cell features, DconnC uncovers the inherent connection between drug molecular structures and cellular characteristics, thus improving the accuracy of predictions. The performance of our method is evaluated using a five-fold cross validation approach, demonstrating a 35 % reduction in the mean square error (MSE) compared to the next-best method. Moreover, our method significantly outperformed alternative approaches in various evaluation criteria, particularly in predicting different cell lines and Loewe synergy score intervals.

PMID:40088336 | DOI:10.1007/s12539-025-00695-6

J Mol Neurosci. 2025 Mar 15;75(1):36. doi: 10.1007/s12031-025-02329-4.

ABSTRACT

Parkinson's disease recognition (PDR) involves identifying Parkinson's disease using clinical evaluations, imaging studies, and biomarkers, focusing on early symptoms like tremors, rigidity, and bradykinesia to facilitate timely treatment. However, due to noise, variability, and the non-stationary nature of EEG signals, distinguishing PD remains a challenge. Traditional deep learning methods struggle to capture the intricate temporal and spatial dependencies in EEG data, limiting their precision. To address this, a novel fusion framework called graph embedding class-based convolutional recurrent attention network with Brown Bear Optimization Algorithm (GECCR2ANet + BBOA) is introduced for EEG-based PD recognition. Preprocessing is conducted using numerical operations and noise removal with weighted guided image filtering and entropy evaluation weighting (WGIF-EEW). Feature extraction is performed via the improved VGG19 with graph triple attention network (IVGG19-GTAN), which captures spatial and temporal dependencies in EEG data. The extracted features are classified using the graph embedding class-based convolutional recurrent attention network (GECCR2ANet), further optimized through the Brown Bear Optimization Algorithm (BBOA) to enhance classification accuracy. The model achieves 99.9% accuracy, 99.4% sensitivity, and a 99.3% F1-score on the UNM dataset, and 99.8% accuracy, 99.1% sensitivity, and 99.2% F1-score on the UC San Diego dataset, significantly outperforming existing methods. Additionally, it records an error rate of 0.5% and a computing time of 0.25 s. Previous models like 2D-MDAGTS, A-TQWT, and CWCNN achieved below 95% accuracy, while the proposed model's 99.9% accuracy underscores its superior performance in real-world clinical applications, enhancing early PD detection and improving diagnostic efficiency.

PMID:40088329 | DOI:10.1007/s12031-025-02329-4

Abdom Radiol (NY). 2025 Mar 15. doi: 10.1007/s00261-025-04883-2. Online ahead of print.

ABSTRACT

PURPOSE: Magnetic resonance elastography (MRE) measures liver stiffness for fibrosis staging, but its utility can be hindered by quality control (QC) challenges and measurement variability. The objective of the study was to fully automate liver MRE QC and liver stiffness measurement (LSM) using a deep learning (DL) method.

METHODS: In this retrospective, single center, IRB-approved human study, a curated dataset involved 897 MRE magnitude slices from 146 2D MRE scans [1.5 T and 3 T MRI, 2D Gradient Echo (GRE), and 2D Spin Echo-Echo Planar Imaging (SE-EPI)] of 69 patients (37 males, mean age 51.6 years). A SqueezeNet-based binary QC model was trained using combined and individual inputs of MRE magnitude slices and their 2D Fast-Fourier transforms to detect artifacts from patient motion, aliasing, and blurring. Three independent observers labeled MRE magnitude images as 0 (non-diagnostic quality) or 1 (diagnostic quality) to create a reference standard. A 2D U-Net segmentation model was trained on diagnostic slices with liver masks to support LSM. Intersection over union between the predicted segmentation and confidence masks identified measurable areas for LSM on elastograms. Cohen's unweighted Kappa coefficient, mean LSM error (%), and intra-class correlation coefficient were calculated to compare the DL-assisted approach with the observers' annotations. An efficiency analysis compared the DL-assisted vs manual LSM durations.

RESULTS: The top QC ensemble model (using MRE magnitude alone) achieved accuracy, precision, and recall of 0.958, 0.982, and 0.886, respectively. The mean LSM error between the DL-assisted approach and the reference standard was 1.9% ± 4.6%. DL-assisted approach completed LSM for 29 diagnostic slices in under 1 s, compared to 20 min manually.

CONCLUSION: An automated DL-based classification of liver MRE diagnostic quality, liver segmentation, and LSM approach demonstrates a promising high performance, with potential for clinical adoption.

PMID:40088296 | DOI:10.1007/s00261-025-04883-2

Med Biol Eng Comput. 2025 Mar 15. doi: 10.1007/s11517-025-03334-w. Online ahead of print.

ABSTRACT

Neurodegenerative disorders (NDD) represent a broad spectrum of diseases that progressively impact neurological function, yet available therapeutics remain conspicuously limited. They lead to altered rhythms and dynamics of walking, which are evident in the sequential footfall contact times measured from one stride to the next. Early detection of aberrant walking patterns can prevent the progression of risks associated with neurodegenerative diseases, enabling timely intervention and management. In this study, we propose a new methodology based on a parallel-path ConvMixer neural network for neurodegenerative disease classification from gait analysis. Earlier research in this field depended on either gait parameter-derived features or the ground reaction force signal. This study has emerged to combine both ground reaction force signals and extracted features to improve gait pattern analysis. The study is being carried out on the gait dynamics in the NDD database, i.e., on the benchmark dataset Physionet gaitndd. Leave one out cross-validation is carried out. The proposed model achieved the best average rates of accuracy, precision, recall, and an F1-score of 97.77 % , 96.37 % , 96.5 % , and 96.25 % , respectively. The experimental findings demonstrate that our approach outperforms the best results achieved by other state-of-the-art methods.

PMID:40088256 | DOI:10.1007/s11517-025-03334-w

J Am Chem Soc. 2025 Mar 15. doi: 10.1021/jacs.5c02046. Online ahead of print.

ABSTRACT

Determining hydrogen positions in metal hydride clusters remains a formidable challenge, which relies heavily on unaffordable neutron diffraction. While machine learning has shown promise, only one deep learning-based method has been proposed so far, which relies heavily on neutron diffraction data for training, limiting its general applicability. In this work, we present an innovative strategy─SSW-NN (stochastic surface walking with neural network)─a robust, non-neutron diffraction-dependent technique that accurately predicts hydrogen positions. Validated against neutron diffraction data for copper hydride clusters, SSW-NN proved effective for clusters where only X-ray diffraction data or DFT predictions are available. It offers superior accuracy, efficiency, and versatility across different metal hydrides, including silver and alloy hydride systems, currently without any neutron diffraction references. This approach not only establishes a new research paradigm for metal hydride clusters but also provides a universal solution for hydrogen localization in other research fields constrained by neutron sources.

PMID:40088162 | DOI:10.1021/jacs.5c02046

J Appl Clin Med Phys. 2025 Mar 14:e70012. doi: 10.1002/acm2.70012. Online ahead of print.

ABSTRACT

PURPOSE: Training deep learning dose prediction models for the latest cutting-edge radiotherapy techniques, such as AI-based nodal radiotherapy (AINRT) and Daily Adaptive AI-based nodal radiotherapy (DA-AINRT), is challenging due to limited data. This study aims to investigate the impact of transfer learning on the predictive performance of an existing clinical dose prediction model and its potential to enhance emerging radiotherapy approaches for head and neck cancer patients.

METHOD: We evaluated the impact and benefits of transfer learning by fine-tuning a Hierarchically Densely Connected U-net on both AINRT and DA-AINRT patient datasets, creating ModelAINRT (Study 1) and ModelDA-AINRT (Study 2). These models were compared against pretrained and baseline models trained from scratch. In Study 3, both fine-tuned models were tested using DA-AINRT patients' final adaptive sessions to assess ModelAINRT 's effectiveness on DA-AINRT patients, given that the primary difference is planning target volume (PTV) sizes between AINRT and DA-AINRT.

RESULT: Studies 1 and 2 revealed that the transfer learning model accurately predicted the mean dose within 0.71% and 0.86% of the prescription dose on the test data. This outperformed the pretrained and baseline models, which showed PTV mean dose prediction errors of 2.29% and 1.1% in Study 1, and 2.38% and 2.86% in Study 2 (P < 0.05). Additionally, Study 3 demonstrated significant improvements in PTV dose prediction error with ModelDA-AINRT, with a mean dose difference of 0.86% ± 0.73% versus 2.26% ± 1.65% (P < 0.05). This emphasizes the importance of training models for specific patient cohorts to achieve optimal outcomes.

CONCLUSION: Applying transfer learning to dose prediction models significantly improves prediction accuracy for PTV while maintaining similar dose performance in predicting organ-at-risk (OAR) dose compared to pretrained and baseline models. This approach enhances dose prediction models for novel radiotherapy methods with limited training data.

PMID:40087841 | DOI:10.1002/acm2.70012

Magn Reson Med. 2025 Mar 14. doi: 10.1002/mrm.30488. Online ahead of print.

ABSTRACT

PURPOSE: To develop a quantitative multislice chemical exchange saturation transfer (CEST) schedule optimization and pulse sequence that reduces the loss of sensitivity inherent to multislice sequences.

METHODS: A deep learning framework was developed for simultaneous optimization of scan parameters and slice order. The optimized sequence was tested in numerical simulations against a random schedule and an optimized single-slice schedule. The scan efficiency of each schedule was quantified. Three healthy subjects were scanned with the proposed sequence. Regions of interest in white matter (WM) and gray matter (GM) were defined. The sequence was compared with the single-slice sequence in vivo and differences quantified using Bland-Altman plots. Test-retest reproducibility was assessed, and the Lin's concordance correlation coefficient (CCC) was calculated for WM and GM. Intersubject variability was also measured with the CCC. Feasibility of whole-brain clinical imaging was tested using a multislab acquisition in 1 subject.

RESULTS: The optimized multislice sequence yielded a lower mean error than the random schedule for all tissue parameters and a lower error than the optimized single-slice schedule for four of six parameters. The optimized multislice sequence provided the highest scan efficiency. In vivo tissue-parameter values obtained with the proposed sequence agreed well with those of the optimized single-slice sequence and prior studies. The average WM/GM CCC was 0.8151/0.7779 for the test-retest scans and 0.7792/0.7191 for the intersubject variability experiment.

CONCLUSION: A multislice schedule optimization framework and pulse sequence were demonstrated for quantitative CEST. The proposed approach enables accurate and reproducible whole-brain quantitative CEST imaging in clinically relevant scan times.

PMID:40087839 | DOI:10.1002/mrm.30488