Medical & Biological Engineering & Computing最新文献

Traumatic brain injury (TBI) is one of the most prevalent health conditions, with severity assessment serving as an initial step for management, prognosis, and targeted therapy. Existing studies on automated outcome prediction using machine learning (ML) often overlook the importance of TBI features in decision-making and the challenges posed by limited and imbalanced training data. Furthermore, many attempts have focused on quantitatively evaluating ML algorithms without explaining the decisions, making the outcomes difficult to interpret and apply for less-experienced doctors. This study presents a novel supportive tool, named E-TBI (explainable outcome prediction after TBI), designed with a user-friendly web-based interface to assist doctors in outcome prediction after TBI using machine learning. The tool is developed with the capability to visualize rules applied in the decision-making process. At the tool's core is a feature selection and classification module that receives multimodal data from TBI patients (demographic data, clinical data, laboratory test results, and CT findings). It then infers one of four TBI severity levels. This research investigates various machine learning models and feature selection techniques, ultimately identifying the optimal combination of gradient boosting machine and random forest for the task, which we refer to as GBMRF. This method enabled us to identify a small set of essential features, reducing patient testing costs by 35%, while achieving the highest accuracy rates of 88.82% and 89.78% on two datasets (a public TBI dataset and our self-collected dataset, TBI_MH103). Classification modules are available at https://github.com/auverngo110/Traumatic_Brain_Injury_103 .

Reconfigurable cable-driven parallel robots (RCDPRs) have attracted much attention as a novel type of cable-driven robot that can change their cable anchor position. The reconfigurable cable-driven lower limb rehabilitation robot (RCDLR) employs RCDPRs in lower limb rehabilitation to achieve multiple training modes. This paper investigates the reconfiguration planning and structural parameter design of the RCDLR. The RCDLR aims to fulfill the requirements of early passive rehabilitation training. Therefore, motion capture data are analyzed and mapped to the target trajectory of the RCDLR. Through dynamics modeling, the Wrench-Feasible Anchor-point Space (WFAS) is defined, from which an objective function for optimal reconfiguration planning is derived. The genetic algorithm is used to solve the optimal reconfiguration planning problem. Additionally, we propose the reconfigurability and safety coefficients as components of a structure parameter design method aimed at satisfying multiple target rehabilitation trajectories. Finally, numerical simulations are implemented based on the instance data and target trajectories to compute the specific structure parameters and verify the effectiveness of the reconfiguration planning method.

Echocardiography sequence segmentation is vital in modern cardiology. While the Segment Anything Model (SAM) excels in general segmentation, its direct use in echocardiography faces challenges due to complex cardiac anatomy and subtle ultrasound boundaries. We introduce SAID (Segment Anything with Implicit Decoding), a novel framework integrating implicit neural representations (INR) with SAM to enhance accuracy, adaptability, and robustness. SAID employs a Hiera-based encoder for multi-scale feature extraction and a Mask Unit Attention Decoder for fine detail capture, critical for cardiac delineation. Orthogonalization boosts feature diversity, and I $_{}^2$ Net improves handling of misaligned contextual features. Tested on CAMUS and EchoNet-Dynamics datasets, SAID outperforms state-of-the-art methods, achieving a Dice Similarity Coefficient (DSC) of 93.2% and Hausdorff Distance (HD95) of 5.02 mm on CAMUS, and a DSC of 92.3% and HD95 of 4.05 mm on EchoNet-Dynamics, confirming its efficacy and robustness for echocardiography sequence segmentation.

Bone fractures are a common medical condition requiring accurate simulation for diagnosis and treatment planning. This study introduces a comprehensive method for simulating bone fractures using two-dimensional fracture patterns and real fractured bones applied to three-dimensional bone models. The approach begins with selecting and adjusting a fracture pattern, projecting it onto a 3D bone model and applying triangulation guided by quality metrics to simulate the cortical layer. Perturbation techniques add irregularities to the fracture surface, enhancing realism. Validation involved comparing simulated fragments with real fragments obtained from CT scans to ensure accuracy. Fracture patterns derived from real fragments were applied to non-fractured bone models to generate simulated fragments. A comparison of real and simulated fracture zones verified the minimal deviation in the results. Specifically, the distance between MMAR and MMAS scaled values varies between $-$ 0.36 and 1.44, confirming the accuracy of the simulation. The resulting models have diverse applications, such as accurate surgical planning, enhanced training, and medical simulation. These models also support personalized medicine by improving patient-specific surgical interventions. This advancement has the potential to significantly enhance fracture treatment strategies and elevate overall patient care.

Electroencephalography (EEG) usage in emotion recognition has garnered significant interest in brain-computer interface (BCI) research. Nevertheless, in order to develop an effective model for emotion identification, features need to be extracted from EEG data in terms of multi-view. In order to tackle the problems of multi-feature interaction and domain adaptation, we suggest an innovative network, IF-MMCL, which leverages multi-modal data in multi-view representation and integrates an individual focused network. In our approach, we build an individual focused network with multi-view that utilizes individual focused contrastive learning to improve model generalization. The network employs different structures for multi-view feature extraction and uses multi-feature relationship computation to identify the relationships between features from various views and modalities. Our model is validated using four public emotion datasets, each containing various emotion classification tasks. In leave-one-subject-out experiments, IF-MMCL performs better than the previous methods in model generalization with limited data.

Early identification of internal carotid artery (ICA) system diseases is critical for preventing stroke and other cerebrovascular events. Traditional diagnostic methods rely heavily on clinician expertise and costly imaging, limiting accessibility. This study aims to develop an interpretable machine learning (ML) model using common carotid artery (CCA) features to predict ICA disease risk, enabling efficient screening. Clinical data from 1612 patients (806 high-risk vs. 806 low-risk ICA disease) were analyzed. CCA features-blood flow, intima-media thickness, internal diameter, age, and gender-were used to train five ML models. Model performance was evaluated via accuracy, sensitivity, specificity, AUC-ROC, and F1 score. SHAP analysis identified key predictors. The support vector machine (SVM) achieved optimal performance (accuracy, 84.9%; AUC, 92.6%), outperforming neural networks (accuracy, 81.4%; AUC, 89.8%). SHAP analysis revealed CCA blood flow (negative correlation) and intima-media thickness (positive correlation) as dominant predictors. This study demonstrates that CCA hemodynamic and structural features, combined with interpretable ML models, can effectively predict ICA disease risk. The SVM-based framework offers a cost-effective screening tool for early intervention, particularly in resource-limited settings. Future work will validate these findings in multi-center cohorts.

Trauma has become a major cause of increased morbidity and mortality worldwide. In emergency response, the classification of injuries is crucial as it helps to quickly determine the criticality of the injured, allocate rescue resources rationally, and decide the priority order of treatment. However, emergency scenes are often chaotic environments, making it difficult for rescue personnel to collect complete and accurate information about the injured in a short period. The combination of artificial intelligence and emergency rescue is gradually changing the rescue model, improving the efficiency of rescue operations. We selected data from 26,810 trauma patients admitted to Chongqing Daping Hospital between 2013 and 2024. We propose a fast tiered medical treatment method with a two-layer structure under emergency limited data conditions, which integrates natural language processing (NLP) and machine learning (ML) techniques. The tiered medical treatment model utilizes NLP to capture semantic features of unstructured text data, while utilizing four ML algorithms to process structured numerical data. Additionally, we conducted external validation using 245 data entries from the Chongqing Emergency Center. The experimental results show that gradient boosting and logistic regression have the best performance in the two-layer ML algorithms. Based on these two algorithms, our model outperformed the multilayer perceptron (MLP) model on the test dataset, achieving an accuracy of 91.17%, which is 4.33% higher than that of the MLP model. The specificity, F1-score, and AUC of our model were 97.06%, 86.85%, and 0.949, respectively. For the external dataset, the model achieved accuracy, specificity, F1-score, and AUC of 87.35%, 95.78%, 80.37%, and 0.848, respectively. These results demonstrate the model's high generalizability and prediction accuracy. A model integrating NLP and ML techniques enables rapid tiered medical treatment based on limited data from the emergency scene, with significant advantages in terms of prediction accuracy.

Multiple sclerosis (MS) is a neurodegenerative disease that impacts retinal layer thickness. Thus, several works proposed to diagnose MS from the retinal optical coherence tomography (OCT) images. Recent clinical studies affirmed that thinning occurs on the four top layers, explicitly in the macular region. However, existing MS detection methods have not considered all MS symptoms, which may impact the MS detection performance. In this research, we propose a new automated method to detect MS from the retinal OCT images. The main principle is based on extracting the relevant retinal layers and figuring out the layer thicknesses, which are investigated to deduce the MS disease. The main challenge is to guarantee a higher performance biomarker extraction within an efficient exploration of OCT cuts. Our contribution consists of the following: (1) employing two DL architectures to segment separately sub-images based on their morphology, in order to enhance segmentation quality; (2) extracting thickness features from the four top layers; (3) dedicating a classifier for each OCT cut that is selected based on its position with respect to the macula center; and (4) merging the classifier knowledge through an ensemble learning approach. Our suggested method achieved 97% accuracy, 100% sensitivity, and 94% precision and specificity, which outperforms several state-of-the-art methods.

Accurate segmentation of lung adenocarcinoma nodules in computed tomography (CT) images is critical for clinical staging and diagnosis. However, irregular nodule shapes and ambiguous boundaries pose significant challenges for existing methods. This study introduces S³TU-Net, a hybrid CNN-Transformer architecture designed to enhance feature extraction, fusion, and global context modeling. The model integrates three key innovations: (1) structured convolution blocks (DWF-Conv/D²BR-Conv) for multi-scale feature extraction and overfitting mitigation; (2) S²-MLP Link, a spatial-shift-enhanced skip-connection module to improve multi-level feature fusion; and 3) residual-based superpixel vision transformer (RM-SViT) to capture long-range dependencies efficiently. Evaluated on the LIDC-IDRI dataset, S³TU-Net achieves a Dice score of 89.04%, precision of 90.73%, and IoU of 90.70%, outperforming recent methods by 4.52% in Dice. Validation on the EPDB dataset further confirms its generalizability (Dice, 86.40%). This work contributes to bridging the gap between local feature sensitivity and global context awareness by integrating structured convolutions and superpixel-based transformers, offering a robust tool for clinical decision support.

Accurate segmentation of hard exudate in fundus images is crucial for early diagnosis of retinal diseases. However, hard exudate segmentation is still a challenge task for accurately detecting small lesions and precisely locating the boundaries of ambiguous lesions. In this paper, the longitudinal multi-scale fusion network (LMSF-Net) is proposed for accurate hard exudate segmentation in fundus images. In this network, an adjacent complementary correction module (ACCM) is proposed on the encoding path for complementary fusion between adjacent encoding features, and a progressive iterative fusion module (PIFM) is designed on the decoding path for fusion between adjacent decoding features. Furthermore, a spatial awareness fusion module (SAFM) is proposed at the end of the decoding path for calibration and aggregation of the two decoding outputs. The proposed method can improve segmentation results of hard exudates with different scales and shapes. The experimental results confirm the superiority of the proposed method for hard exudate segmentation with AUPR of 0.6954, 0.9017, and 0.6745 on the DDR, IDRID, and E-Ophtha EX datasets, respectively.