J Biomed Opt. 2025 Feb;30(2):026001. doi: 10.1117/1.JBO.30.2.026001. Epub 2025 Feb 17.

ABSTRACT

SIGNIFICANCE: Current methods for complex conjugate removal (CCR) in frequency-domain optical coherence tomography (FD-OCT) often require additional hardware components, which increase system complexity and cost. A software-based solution would provide a more efficient and cost-effective alternative.

AIM: We aim to develop a deep learning approach to effectively remove complex conjugate artifacts (CCAs) from OCT scans without the need for extra hardware components.

APPROACH: We introduce a deep learning method that employs generative adversarial networks to eliminate CCAs from OCT scans. Our model leverages both conventional intensity images and phase images from the OCT scans to enhance the artifact removal process.

RESULTS: Our CCR-generative adversarial network models successfully converted conventional OCT scans with CCAs into artifact-free scans across various samples, including phantoms, human skin, and mouse eyes imaged in vivo with a phase-stable swept source-OCT prototype. The inclusion of phase images significantly improved the performance of the deep learning models in removing CCAs.

CONCLUSIONS: Our method provides a low-cost, data-driven, and software-based solution to enhance FD-OCT imaging capabilities by the removal of CCAs.

PMID:39963188 | PMC:PMC11831228 | DOI:10.1117/1.JBO.30.2.026001

Nanoscale. 2025 Feb 18. doi: 10.1039/d4nr05458j. Online ahead of print.

ABSTRACT

In addition to its excellent photoelectronic properties, the CsPbBr3 perovskite has been reported as a low thermal conductivity (k) material. However, few studies investigated the microscopic mechanisms underlying its low k. Studying its thermal transport behavior is crucial for understanding its thermal properties and thus improving its thermal stability. Here, we train a DFT-level deep-learning potential (DP) of CsPbBr3 and explore its ultra-low k using nonequilibrium molecular dynamics (NEMD). The k calculated using NEMD is 0.43 ± 0.01 W m-1 K-1, which is consistent with experimental results. Furthermore, the Cs vacancy contributes to the decrease in k due to the distortion of the Pb-Br cage, which enhances phonon scattering and reduces the phonon lifetime. Our research reveals the significant potential of machine learning force fields in thermal and phonon behavior research and the valuable insights gained from defect-regulated thermal conductivity.

PMID:39963065 | DOI:10.1039/d4nr05458j

J Pathol Transl Med. 2025 Feb 18. doi: 10.4132/jptm.2024.10.23. Online ahead of print.

ABSTRACT

BACKGROUND: High-grade serous ovarian carcinoma (HGSC) exhibits significant heterogeneity, posing challenges for effective clinical categorization. Understanding the histomorphological diversity within HGSC could lead to improved prognostic stratification and personalized treatment approaches.

METHODS: We applied the Histomic Atlases of Variation Of Cancers model to whole slide images from The Cancer Genome Atlas dataset for ovarian cancer. Histologically distinct tumor clones were grouped into common histomic clusters. Principal component analysis and K-means clustering classified HGSC samples into three groups: highly differentiated (HD), intermediately differentiated (ID), and lowly differentiated (LD).

RESULTS: HD tumors showed diverse patterns, lower densities, and stronger eosin staining. ID tumors had intermediate densities and balanced staining, while LD tumors were dense, patternless, and strongly hematoxylin-stained. RNA sequencing revealed distinct patterns in mitochondrial oxidative phosphorylation and energy metabolism, with upregulation in the HD, downregulation in the LD, and the ID positioned in between. Survival analysis showed significantly lower overall survival for the LD compared to the HD and ID, underscoring the critical role of mitochondrial dynamics and energy metabolism in HGSC progression.

CONCLUSIONS: Deep learning-based histologic analysis effectively stratifies HGSC into clinically relevant prognostic groups, highlighting the role of mitochondrial dynamics and energy metabolism in disease progression. This method offers a novel approach to HGSC categorization.

PMID:39962925 | DOI:10.4132/jptm.2024.10.23

Sci Rep. 2025 Feb 17;15(1):5693. doi: 10.1038/s41598-025-85802-9.

ABSTRACT

Recently, deep learning-based medical image classification models have made substantial advancements. However, many existing models prioritize performance at the cost of efficiency, limiting their practicality in clinical use. Traditional Convolutional Neural Network (CNN)-based methods, Transformer-based methods, and hybrid approaches combining these two struggle to balance performance and model complexity. To achieve efficient predictions with a low parameter count, we propose DenseAttentionNetwork (DANet), a lightweight model for Alzheimer's disease detection in 3D MRI images. DANet leverages dense connections and a linear attention mechanism to enhance feature extraction and capture long-range dependencies. Its architecture integrates convolutional layers for localized feature extraction with linear attention for global context, enabling efficient multi-scale feature reuse across the network. By replacing traditional self-attention with a parameter-efficient linear attention mechanism, DANet overcomes some limitations of standard self-attention. Extensive experiments across multi-institutional datasets demonstrate that DANet achieves the best performance in area under the receiver operating characteristic curve (AUC), which underscores the model's robustness and effectiveness in capturing relevant features for Alzheimer's disease detection while also attaining a strong accuracy structure with fewer parameters. Visualizations based on activation maps further verify the model's ability to highlight AD-relevant regions in 3D MRI images, providing clinically interpretable insights into disease progression.

PMID:39962113 | DOI:10.1038/s41598-025-85802-9

Sci Rep. 2025 Feb 17;15(1):5815. doi: 10.1038/s41598-025-85703-x.

ABSTRACT

Alzheimer's Disease (AD) is a progressive condition of a neurological brain disorder recognized by symptoms such as dementia, memory loss, alterations in behaviour, and impaired reasoning abilities. Recently, many researchers have been working to develop an effective AD recognition system using deep learning (DL) based convolutional neural network (CNN) model aiming to deploy the automatic medical image diagnosis system. The existing system is still facing difficulties in achieving satisfactory performance in terms of accuracy and efficiency because of the lack of feature ineffectiveness. This study proposes a lightweight Stacked Convolutional Neural Network with a Channel Attention Network (SCCAN) for MRI based on AD classification to overcome the challenges. In the procedure, we sequentially integrate 5 CNN modules, which form a stack CNN aiming to generate a hierarchical understanding of features through multi-level extraction, effectively reducing noise and enhancing the weight's efficacy. This feature is then fed into a channel attention module to select the practical features based on the channel dimension, facilitating the selection of influential features. . Consequently, the model exhibits reduced parameters, making it suitable for training on smaller datasets. Addressing the class imbalance in the Kaggle MRI dataset, a balanced distribution of samples among classes is emphasized. Extensive experiments of the proposed model with the ADNI1 Complete 1Yr 1.5T, Kaggle, and OASIS-1 datasets showed 99.58%, 99.22%, and 99.70% accuracy, respectively. The proposed model's high performance surpassed state-of-the-art (SOTA) models and proved its excellence as a significant advancement in AD classification using MRI images.

PMID:39962097 | DOI:10.1038/s41598-025-85703-x

Insights Imaging. 2025 Feb 17;16(1):44. doi: 10.1186/s13244-025-01911-z.

ABSTRACT

OBJECTIVE: To assess the accuracy of deep learning reconstruction (DLR) technique on synthetic MRI (SyMRI) including T2 measurements and diagnostic performance of DLR synthetic MRI (SyMRIDL) in patients with knee osteoarthritis (KOA) using conventional MRI as standard reference.

MATERIALS AND METHODS: This prospective study recruited 36 volunteers and 70 patients with suspected KOA from May to October 2023. DLR and non-DLR synthetic T2 measurements (T2-SyMRIDL, T2-SyMRI) for phantom and in vivo knee cartilage were compared with multi-echo fast-spin-echo (MESE) sequence acquired standard T2 values (T2MESE). The inter-reader agreement on qualitative evaluation of SyMRIDL and the positive percent agreement (PPA) and negative percentage agreement (NPA) were analyzed using routine images as standard diagnosis.

RESULTS: DLR significantly narrowed the quantitative differences between T2-SyMRIDL and T2MESE for 0.8 ms with 95% LOA [-5.5, 7.1]. The subjective assessment between DLR synthetic MR images and conventional MRI was comparable (all p > 0.05); Inter-reader agreement for SyMRIDL and conventional MRI was substantial to almost perfect with values between 0.62 and 0.88. SyMRIDL MOAKS had substantial inter-reader agreement and high PPA/NPA values (95%/99%) using conventional MRI as standard reference. Moreover, T2-SyMRIDL measurements instead of non-DLR ones significantly differentiated normal-appearing from injury-visible cartilages.

CONCLUSION: DLR synthetic knee MRI provided both weighted images for clinical diagnosis and accurate T2 measurements for more confidently identifying early cartilage degeneration from normal-appearing cartilages.

CRITICAL RELEVANCE STATEMENT: One-acquisition synthetic MRI based on deep learning reconstruction provided an accurate quantitative T2 map and morphologic images in relatively short scan time for more confidently identifying early cartilage degeneration from normal-appearing cartilages compared to the conventional morphologic knee sequences.

KEY POINTS: Deep learning reconstruction (DLR) synthetic knee cartilage T2 values showed no difference from conventional ones. DLR synthetic T1-, proton density-, STIR-weighted images had high positive percent agreement and negative percentage agreement using MRI OA Knee Score features. DLR synthetic T2 measurements could identify early cartilage degeneration from normal-appearing ones.

PMID:39961957 | DOI:10.1186/s13244-025-01911-z

Med Biol Eng Comput. 2025 Feb 17. doi: 10.1007/s11517-025-03311-3. Online ahead of print.

ABSTRACT

Aortic aneurysms pose a significant risk of rupture. Previous research has shown that areas exposed to low wall shear stress (WSS) are more prone to rupture. Therefore, precise WSS determination on the aneurysm is crucial for rupture risk assessment. Computational fluid dynamics (CFD) is a powerful approach for WSS calculations, but they are computationally intensive, hindering time-sensitive clinical decision-making. In this study, we propose a deep learning (DL) surrogate, MultiViewUNet, to rapidly predict time-averaged WSS (TAWSS) distributions on abdominal aortic aneurysms (AAA). Our novel approach employs a domain transformation technique to translate complex aortic geometries into representations compatible with state-of-the-art neural networks. MultiViewUNet was trained on 23 real and 230 synthetic AAA geometries, demonstrating an average normalized mean absolute error (NMAE) of just 0.362 % in WSS prediction. This framework has the potential to streamline hemodynamic analysis in AAA and other clinical scenarios where fast and accurate stress quantification is essential.

PMID:39961912 | DOI:10.1007/s11517-025-03311-3

Med Biol Eng Comput. 2025 Feb 17. doi: 10.1007/s11517-025-03318-w. Online ahead of print.

ABSTRACT

Caries segmentation holds significant clinical importance in medical image analysis, particularly in the early detection and treatment of dental caries. However, existing deep learning segmentation methods often struggle with accurately segmenting complex caries boundaries. To address this challenge, this paper proposes a novel network, named AEDD-Net, which combines an attention mechanism with a dual-decoder structure to enhance the performance of boundary segmentation for caries. Unlike traditional methods, AEDD-Net integrates atrous spatial pyramid pooling with cross-coordinate attention mechanisms to effectively fuse global and multi-scale features. Additionally, the network introduces a dedicated boundary generation module that precisely extracts caries boundary information. Moreover, we propose an innovative boundary loss function to further improve the learning of boundary features. Experimental results demonstrate that AEDD-Net significantly outperforms other comparison networks in terms of Dice coefficient, Jaccard similarity, precision, and sensitivity, particularly showing superior performance in boundary segmentation. This study provides an innovative approach for automated caries segmentation, with promising potential for clinical applications.

PMID:39961911 | DOI:10.1007/s11517-025-03318-w

Eur Radiol. 2025 Feb 17. doi: 10.1007/s00330-025-11450-2. Online ahead of print.

ABSTRACT

OBJECTIVES: This study proposes a multimodal deep learning (DL) approach to investigate the impact of tumors and visceral fat on occult peritoneal metastasis in colorectal cancer (CRC) patients.

METHODS: We developed a DL model named Multi-scale Feature Fusion Network (MSFF-Net) based on ResNet18, which extracted features of tumors and visceral fat from the longest diameter tumor section and the third lumbar vertebra level (L3) in preoperative CT scans of CRC patients. Logistic regression analysis was applied to patients' clinical data that integrated with DL features. A random forest (RF) classifier was established to evaluate the MSFF-Net's performance on internal and external test sets and compare it with radiologists' performance.

RESULTS: The model incorporating fat features outperformed the single tumor modality in the internal test set. Combining clinical information with DL provided the best diagnostic performance for predicting peritoneal metastasis in CRC patients. The AUCs were 0.941 (95% CI: [0.891, 0.986], p = 0.03) for the internal test set and 0.911 (95% CI: [0.857, 0.971], p = 0.013) for the external test set. CRC patients with peritoneal metastasis had a higher visceral adipose tissue index (VATI) compared to those without. Maximum tumor diameter and VATI were identified as independent prognostic factors for peritoneal metastasis. Grad-CAM decision regions corresponded with the independent prognostic factors identified by logistic regression analysis.

CONCLUSION: The study confirms the network features of tumors and visceral fat significantly enhance predictive performance for peritoneal metastasis in CRC. Visceral fat is a meaningful imaging biomarker for peritoneal metastasis's early detection in CRC patients.

KEY POINTS: Question Current research on predicting colorectal cancer with peritoneal metastasis mainly focuses on single-modality analysis, while studies based on multimodal imaging information are relatively scarce. Findings The Multi-scale Feature Fusion Network, constructed based on ResNet18, can utilize CT images of tumors and visceral fat to detect occult peritoneal metastasis in colorectal cancer. Clinical relevance This study identified independent prognostic factors for colorectal cancer peritoneal metastasis and combines them with tumor and visceral fat network features, aiding early diagnosis and accurate prognostic assessment.

PMID:39961863 | DOI:10.1007/s00330-025-11450-2

Sci Rep. 2025 Feb 17;15(1):5759. doi: 10.1038/s41598-025-90430-4.

ABSTRACT

The concept of a smart city has spread as a solution ensuring wider availability of data and services to citizens, apart from as a means to lower the environmental footprint of cities. Crowd density monitoring is a cutting-edge technology that enables smart cities to monitor and effectively manage crowd movements in real time. By utilizing advanced artificial intelligence and video analytics, valuable insights are accumulated from crowd behaviour, assisting cities in improving operational efficiency, improving public safety, and urban planning. This technology also significantly contributes to resource allocation and emergency response, contributing to smarter, safer urban environments. Crowd density classification in smart cities using deep learning (DL) employs cutting-edge NN models to interpret and analyze information from sensors such as IoT devices and CCTV cameras. This technique trains DL models on large datasets to accurately count people in a region, assisting traffic management, safety, and urban planning. By utilizing recurrent neural networks (RNNs) for time-series data and convolutional neural networks (CNNs) for image processing, the model adapts to varying crowd scenarios, lighting, and angles. This manuscript presents a Deep Convolutional Neural Network-based Crowd Density Monitoring for Intelligent Urban Planning (DCNNCDM-IUP) technique on smart cities. The proposed DCNNCDM-IUP technique utilizes DL methods to detect crowd densities, which can significantly assist in urban planning for smart cities. Initially, the DCNNCDM-IUP technique performs image preprocessing using Gaussian filtering (GF). The DCNNCDM-IUP technique utilizes the SE-DenseNet approach, which effectually learns complex feature patterns for feature extraction. Moreover, the hyperparameter selection of the SE-DenseNet approach is accomplished by using the red fox optimization (RFO) methodology. Finally, the convolutional long short-term memory (ConvLSTM) methodology recognizes varied crowd densities. A comprehensive simulation analysis is conducted to demonstrate the improved performance of the DCNNCDM-IUP technique. The experimental validation of the DCNNCDM-IUP technique portrayed a superior accuracy value of 98.40% compared to existing DL models.

PMID:39962323 | DOI:10.1038/s41598-025-90430-4

Nat Methods. 2025 Feb 17. doi: 10.1038/s41592-024-02581-3. Online ahead of print.

NO ABSTRACT

PMID:39962313 | DOI:10.1038/s41592-024-02581-3

Nat Methods. 2025 Feb 17. doi: 10.1038/s41592-024-02582-2. Online ahead of print.

ABSTRACT

The dynamics of neuron populations commonly evolve on low-dimensional manifolds. Thus, we need methods that learn the dynamical processes over neural manifolds to infer interpretable and consistent latent representations. We introduce a representation learning method, MARBLE, which decomposes on-manifold dynamics into local flow fields and maps them into a common latent space using unsupervised geometric deep learning. In simulated nonlinear dynamical systems, recurrent neural networks and experimental single-neuron recordings from primates and rodents, we discover emergent low-dimensional latent representations that parametrize high-dimensional neural dynamics during gain modulation, decision-making and changes in the internal state. These representations are consistent across neural networks and animals, enabling the robust comparison of cognitive computations. Extensive benchmarking demonstrates state-of-the-art within- and across-animal decoding accuracy of MARBLE compared to current representation learning approaches, with minimal user input. Our results suggest that a manifold structure provides a powerful inductive bias to develop decoding algorithms and assimilate data across experiments.

PMID:39962310 | DOI:10.1038/s41592-024-02582-2

Sci Rep. 2025 Feb 17;15(1):5707. doi: 10.1038/s41598-025-90164-3.

ABSTRACT

In order to detect epileptic spikes, this paper suggests a deep learning architecture that blends 1D residual convolutional neural networks (1D-ResCNN) with a hybrid optimization strategy. The Layer-wise Adaptive Moments (LAMB) and AdamW algorithms have been used in the model's optimization to improve efficiency and accelerate convergence while extracting features from time and frequency domain EEG data. The framework has been considered on two public epilepsy datasets CHB-MIT and Siena. In the CHB-MIT dataset, comprising 24-channel EEG recordings from 12 patients, the model achieved an accuracy of 99.71%, a sensitivity of 99.60%, and a specificity of 99.61% for detecting epileptic spikes. Similarly, in the Siena dataset, which includes EEG data from 14 adult patients, the model demonstrated an average accuracy of 99.75%. Sensitivity averaged 99.94%, while specificity averaged 99.95%. The false positive rate (FPR) remained low at 0.0011, and the model obtained an average F1-score of 99.74%. For real-time hardware validation, the 1D-ResCNN model was deployed within the Typhoon HIL simulator, utilizing embedded C2000 microcontrollers. This hardware configuration allowed for immediate spike detection with minimal latency, ensuring reliable performance in real-time clinical applications. The findings imply that the suggested approach provides suitable for identifying epileptic spikes in real time for medical settings.

PMID:39962290 | DOI:10.1038/s41598-025-90164-3

Sci Rep. 2025 Feb 17;15(1):5766. doi: 10.1038/s41598-025-89672-z.

ABSTRACT

Accurately discerning lead-zinc open pit mining areas using traditional remote sensing methods is challenging due to spectral signature class mixing. However, machine learning (ML) algorithms have been implemented to classify satellite images, achieving better accuracy in discriminating complex landcover features. This study aims to characterise various ML models for detecting and classifying lead-zinc open pit mining areas amidst surrounding landcover features based on Sentinel 2 image analysis. Various associated band ratios and spectral indices were integrated with processed Sentinel 2 reflectance bands to enhance detection accuracy. Suitable bands highlighting lead and zinc mine areas were identified based on optimal index factor (OIF) analysis and various deep learning-based stacked encoders. Furthermore, 15 different ML classifiers were tested to identify optimised algorithms for accurately discriminating complex mining areas and associated landcover features. After detailed evaluation and comparison of their accuracies, the extra tree classifier (et) was the most effective, achieving an overall accuracy of 0.94 and a kappa coefficient of 0.93. The light gradient boosting machine classifier (lightgbm) and random forest classifier (rf) models also performed well, with overall accuracies of 0.937 and 0.936 and kappa coefficients of 0.925 and 0.925, respectively.

PMID:39962260 | DOI:10.1038/s41598-025-89672-z

Sci Rep. 2025 Feb 17;15(1):5734. doi: 10.1038/s41598-025-90034-y.

ABSTRACT

This study aims to address the diagnostic challenges in distinguishing gastric polyps from protrusions, emphasizing the need for accurate and cost-effective diagnosis strategies. It explores the application of Convolutional Neural Networks (CNNs) to improve diagnostic accuracy. This research introduces MultiAttentiveScopeNet, a deep learning model that incorporates multi-layer feature ensemble and attention mechanisms to enhance gastroscopy image analysis accuracy. A weakly supervised labeling strategy was employed to construct a large multi-class gastroscopy image dataset for training and validation. MultiAttentiveScopeNet demonstrates significant improvements in prediction accuracy and interpretability. The integrated attention mechanism effectively identifies critical areas in images to aid clinical decisions. Its multi-layer feature ensemble enables robust analysis of complex gastroscopy images. Comparative testing against human experts shows exceptional diagnostic performance, with accuracy, micro and macro precision, micro and macro recall, and micro and macro AUC reaching 0.9308, 0.9312, 0.9325, 0.9283, 0.9308, 0.9847 and 0.9853 respectively. This highlights its potential as an effective tool for primary healthcare settings. This study provides a comprehensive solution to address diagnostic challenges differentiating gastric polyps and protrusions. MultiAttentiveScopeNet improves accuracy and interpretability, demonstrating the potential of deep learning for gastroscopy image analysis. The constructed dataset facilitates continued model optimization and validation. The model shows promise in enhancing diagnostic outcomes in primary care.

PMID:39962226 | DOI:10.1038/s41598-025-90034-y

Sci Rep. 2025 Feb 17;15(1):5787. doi: 10.1038/s41598-025-90115-y.

ABSTRACT

Current deep learning methods for diagnosing Alzheimer's disease (AD) typically rely on analyzing all or parts of high-resolution 3D volumetric features, which demand expensive computational resources and powerful GPUs, particularly when using multimodal data. In contrast, lightweight cortical surface representations offer a more efficient approach for quantifying AD-related changes across different cortical regions, such as alterations in cortical structures, impaired glucose metabolism, and the deposition of pathological biomarkers like amyloid-β and tau. Despite these advantages, few studies have focused on diagnosing AD using multimodal surface-based data. This study pioneers a novel method that leverages multimodal, lightweight cortical surface features extracted from MRI and PET scans, providing an alternative to computationally intensive 3D volumetric features. Our model employs a middle-fusion approach with a cross-attention mechanism to efficiently integrate features from different modalities. Experimental evaluations on the ADNI series dataset, using T1-weighted MRI and [Formula: see text]Fluorodeoxyglucose PET, demonstrate that the proposed model outperforms volume-based methods in both early AD diagnosis accuracy and computational efficiency. The effectiveness of our model is further validated with the combination of T1-weighted MRI, Aβ PET, and Tau PET scans, yielding favorable results. Our findings highlight the potential of surface-based transformer models as a superior alternative to conventional volume-based approaches.

PMID:39962212 | DOI:10.1038/s41598-025-90115-y

Commun Eng. 2025 Feb 17;4(1):24. doi: 10.1038/s44172-025-00367-9.

ABSTRACT

Regular monitoring of marine life is essential for preserving the stability of marine ecosystems. However, underwater target detection presents several challenges, particularly in balancing accuracy with model efficiency and real-time performance. To address these issues, we propose an innovative approach that combines the Structured Space Model (SSM) with feature enhancement, specifically designed for small target detection in underwater environments. We developed a high-accuracy, lightweight detection model-UWNet. The results demonstrate that UWNet excels in detection accuracy, particularly in identifying difficult-to-detect organisms like starfish and scallops. Compared to other models, UWNet reduces the number of model parameters by 5% to 390%, substantially improving computational efficiency while maintaining top detection accuracy. Its lightweight design enhances the model's applicability for deployment on underwater robots.

PMID:39962196 | DOI:10.1038/s44172-025-00367-9

Sci Rep. 2025 Feb 17;15(1):5805. doi: 10.1038/s41598-024-72539-0.

ABSTRACT

Cervical spinal cord injury is often catastrophic, frequently leading to irreversible impairment. MRI has become the gold standard for evaluating spinal cord injuries (SCI). Our study aimed to assess the accuracy of a radiomics approach, based on machine learning and utilizing conventional MRI, in predicting the prognosis of patients with SCI. In a retrospective analysis of 82 SCI patients from three hospitals, we categorized them into good (n = 49) and poor (n = 33) prognosis groups. Preoperative T2-weighted MRI images were segmented using 3D-Region of Interest (ROI) techniques, and both radiomic and deep transfer learning features were extracted. These features were normalized using Z-score and harmonized via ComBat. Feature selection was performed using a greedy algorithm and Least absolute shrinkage and selection operator (LASSO), and others, followed by the calculation of radiomics scores through linear regression. Machine learning was then used to identify the most predictive radiomic features. Model performance was evaluated by analyzing the area under the curve (AUC) and other indicators.Univariate analysis indicated that the demographic characteristics of cervical spinal cord injury were not statistically significant. In the test dataset, the random forest (RF) combined with radiomics and ResNet34 demonstrated better performance, with an accuracy of 0.800 and an AUC of 0.893.Using MRI, deep learning-based radiomics signals show promise in evaluating and predicting the postoperative prognosis of these patients.

PMID:39962172 | DOI:10.1038/s41598-024-72539-0

Sci Rep. 2025 Feb 17;15(1):5761. doi: 10.1038/s41598-025-86673-w.

ABSTRACT

Software requirement smell detection is an important part of establishing high-quality software specifications. These smells, which frequently indicate difficulties like ambiguity, vagueness, or incompleteness, can lead to misunderstandings and mistakes in the latter phases of software development. Traditionally, identifying requirement smells was a manual process, time-consuming, prone to inconsistency, and human mistakes. Moreover, the previous machine learning and deep learning research was insufficient for detecting multiple smells in a single requirement statement. To address this problem, we developed a multi-label software requirement smell model to detect multiple software requirement smells in a single requirement. Therefore, this study explores a deep learning-based approach to multi-label classification of software requirement smells, incorporating advanced neural network architectures such as LSTM, Bi-LSTM, and GRU with combined word embedding like ELMo and Word2Vec. We collected and prepared an 8120 requirements dataset from different sources categorized into 11 linguistic aspects and we used a binary relevance multi-label classification strategy in which each category was treated independently and used the F1-macro average of each label of the smell. Next, we built models that can classify software requirement smell in a multi-label manner using deep learning algorithms. After executing numerous experiments with different parameters in the Bi-LSTM, LSTM, and GRU models, we obtained 90.3%, 89%, and 88.7% of F1-score macro averages with ELMo, respectively. Therefore, Bi-LSTM achieved a greater F1-score macro average than the other algorithms.

PMID:39962114 | DOI:10.1038/s41598-025-86673-w

Invest Radiol. 2025 Feb 18. doi: 10.1097/RLI.0000000000001162. Online ahead of print.

ABSTRACT

OBJECTIVES: Deep learning for body composition analysis (BCA) is gaining traction in clinical research, offering rapid and automated ways to measure body features like muscle or fat volume. However, most current methods prioritize computed tomography (CT) over magnetic resonance imaging (MRI). This study presents a deep learning approach for automatic BCA using MR T2-weighted sequences.

METHODS: Initial BCA segmentations (10 body regions and 4 body parts) were generated by mapping CT segmentations from body and organ analysis (BOA) model to synthetic MR images created using an in-house trained CycleGAN. In total, 30 synthetic data pairs were used to train an initial nnU-Net V2 in 3D, and this preliminary model was then applied to segment 120 real T2-weighted MRI sequences from 120 patients (46% female) with a median age of 56 (interquartile range, 17.75), generating early segmentation proposals. These proposals were refined by human annotators, and nnU-Net V2 2D and 3D models were trained using 5-fold cross-validation on this optimized dataset of real MR images. Performance was evaluated using Sørensen-Dice, Surface Dice, and Hausdorff Distance metrics including 95% confidence intervals for cross-validation and ensemble models.

RESULTS: The 3D ensemble segmentation model achieved the highest Dice scores for the body region classes: bone 0.926 (95% confidence interval [CI], 0.914-0.937), muscle 0.968 (95% CI, 0.961-0.975), subcutaneous fat 0.98 (95% CI, 0.971-0.986), nervous system 0.973 (95% CI, 0.965-0.98), thoracic cavity 0.978 (95% CI, 0.969-0.984), abdominal cavity 0.989 (95% CI, 0.986-0.991), mediastinum 0.92 (95% CI, 0.901-0.936), pericardium 0.945 (95% CI, 0.924-0.96), brain 0.966 (95% CI, 0.927-0.989), and glands 0.905 (95% CI, 0.886-0.921). Furthermore, body part 2D ensemble model reached the highest Dice scores for all labels: arms 0.952 (95% CI, 0.937-0.965), head + neck 0.965 (95% CI, 0.953-0.976), legs 0.978 (95% CI, 0.968-0.988), and torso 0.99 (95% CI, 0.988-0.991). The overall average Dice across body parts (2D = 0.971, 3D = 0.969, P = ns) and body regions (2D = 0.935, 3D = 0.955, P < 0.001) ensemble models indicates stable performance across all classes.

CONCLUSIONS: The presented approach facilitates efficient and automated extraction of BCA parameters from T2-weighted MRI sequences, providing precise and detailed body composition information across various regions and body parts.

PMID:39961134 | DOI:10.1097/RLI.0000000000001162