Table of Content

29 September 2026, Volume 31 Issue 5 Previous Issue   
---
Engineering & Al for Health


For Selected:

Download Citations EndNote Ris BibTeX

Toggle Thumbnails

Select

Cover and Table of Contents 2026, 31 (5):  0.  Abstract ( 38 )   PDF (5713KB) ( 20 )   Related Articles | Metrics

Engineering & Al for Health

Select

When the Interface Becomes the Discipline ZENG Xiaoqin 2026, 31 (5):  1115-1117.  doi: 10.1007/s12204-026-2952-7 Abstract ( 42 )   PDF (98KB) ( 23 )   For most of the past century, the relation between medicine and engineering could be accurately described as one of exchange. Engineering laboratories produced instruments; clinical medicine adopted them; each field returned to its own methods, its own literatures, and its own criteria of judgement. That description is no longer sufficient. What has emerged over the past decade is not a wider channel of exchange between two established disciplines, but a working environment in which the interface itself is where the principal questions are posed. The ten works recognized in this year's list of selected advances are not individually exercises in medicine and engineering. Considered collectively, they indicate how far this reorganization has already proceeded. References | Related Articles | Metrics

Select

Mixture of Expert System for IESS Detection Based on EEG Signal WANG Zong, FU Lijun, YANG Guang, WAN Lin, CHANG Zhijun 2026, 31 (5):  1118-1126.  doi: 10.1007/s12204-024-2771-7 Abstract ( 40 )   PDF (440KB) ( 26 )   Childhood epilepsy presents significant challenges in terms of medical diagnosis and treatment, particularly when it comes to infantile epileptic spasm syndrome (IESS). This condition is characterized by electroencephalogram (EEG) signals with short seizure duration and insignificant waveform, which complicates manual efforts to detect IESS seizures. To address these challenges, we introduce an innovative model called the IESS mixture of experts (MOE) model to assist EEG technologists in their work. The model utilizes a group of sub-data experts operating through a 3D ResNet network to enhance the detection of seizure signals. Specifically, we propose an unsupervised clustering method for EEG signal partitioning into some balanced sub-data domains, addressing the imbalance between IESS seizure and non-seizure data. The IESS MOE model consists of a navigation network that controls input and output, as well as some chosen sub-expert networks that make judgments on sub-expert data domains. We use the PLA IESS dataset for unsupervised learning-based data partitioning among sub-models and subsequent expert modeling. Additionally, we establish a comparison between the performance of the IESS MOE model and six human experts from different levels of medical centers using external validation on 19 cross-individual patient data sets. Finally, test results demonstrate that the sensitivity of the IESS MOE model is 0.539 3, surpassing the average level achieved by human experts at 0.514 3 while maintaining high specificity at 0.905 0 and achieving a high F1 score at 0.675 8, confirming its superior generalization ability. Our proposed model can assist human experts in monitoring episodes related to IESS. References | Related Articles | Metrics

Select

Prediction of Needle Deflection During Insertion into Multilayer Tissues YANG Xu, LIN Yanping 2026, 31 (5):  1127-1133.  doi: 10.1007/s12204-025-2817-5 Abstract ( 52 )   PDF (1145KB) ( 31 )   Percutaneous needle insertion, a minimally invasive procedure widely used in medical applications, requires precise needle targeting for successful results. However, the interaction between the needle and soft tissue can cause needle deflection, resulting in deviations from the target. This study introduces a robust finite element model employing the coupled Eulerian-Lagrangian method to predict needle deflection during insertion into multilayer soft tissues. The model includes multilayer elastic components to capture the nonlinear behavior of soft tissues, providing an accurate representation of needle-tissue interaction. A multilayer phantom mimicking human tissue was fabricated using polyvinyl alcohol. Needle insertion experiments with varying depths and needle types were conducted to validate the simulation results. The prediction model demonstrated an average deflection error of (1.38 ± 0.25) mm. This research enhances understanding of needle-tissue interaction and provides a valuable tool for optimizing percutaneous puncture procedures. The model has potential for integration into real-time puncture navigation systems, improving precision, enhancing safety, and minimizing complications caused by needle deflection. References | Related Articles | Metrics

Select

Novel Digital Assessment System for Upper-Limb Movement in Stroke Patients Using Markless-Sensing Technology and Deep Learning Algorithms SHENG Bo, LEI Ximin, CHENG Jian, XIE Qiurong, TAO Jing, CHEN Yujie 2026, 31 (5):  1134-1146.  doi: 10.1007/s12204-024-2754-8 Abstract ( 35 )   PDF (872KB) ( 26 )   Traditional motor function assessment for stroke patients involves subjective scoring by rehabilitation physicians, a process that is time-consuming, expensive, and subject to variability. By utilizing sensors (markers) and machine learning algorithms, digital assessment systems offer the potential for more objective clinical decision support. Nevertheless, these algorithms often rely on feature extraction, which opens up opportunities for improving diagnostic accuracy and reliability. Therefore, this study proposed a novel assessment approach based on markless-sensing technology and deep learning algorithms, which can perform contactless data measurement and nonfeature-based digital assessment. Specifically, the movement data of stroke patients were collected via the Microsoft Kinect V2 with a customized motion tracking system. The raw dataset was then processed by the Savitzky-Golay filter and long short-term memory-attention-based assessment model. A total of 25 volunteers (15 stroke patients and 10 healthy subjects) were recruited for experimental validation by conducting the commonly used clinical scale (wolf motor function test-functional ability scores, WMFT-FAS). The experiments showed that the proposed novel digital assessment system could achieve favorable results: an average accuracy of 91.7%, average precision of 87.4%, average recall of 87.3%, and average F1-score of 87.3%. In summary, the proposed system can provide objective and reasonably accurate assessment outcomes, which might have the potential to offer essential clinical information for rehabilitation physicians to make reasonable clinical intervention plans. References | Related Articles | Metrics

Select

Flexible Lower Limb Exoskeleton Robot for Rehabilitation Training of Children with Cerebral Palsy JIANG Haoyu, ZHOU Xuan, LI Xin, CHEN Zhengquan, DU Qing, XIE Le 2026, 31 (5):  1147-1161.  doi: 10.1007/s12204-024-2760-x Abstract ( 43 )   PDF (2369KB) ( 26 )   Current research on lower limb exoskeleton robots is relatively mature, but existing studies are mostly applicable to adult patients. The population of pediatric patients is also substantial, but there is still a significant gap in the research and application in pediatric diseases rehabilitation. A flexible lower limb exoskeleton robot for pediatric patients with cerebral palsy around 10 years old who suffer from knee valgus and foot inversion is proposed in this paper. By analyzing the system transmission mode, the relationship between the Bowden cable input and output ends is established. Additionally, a kinematic model of the ankle joint is created based on the characteristics of human ankle joint. The relationship between the Bowden cable length and ankle joint angle is derived. Furthermore, a highly integrated control system is designed and an applicable motion control strategy is proposed. Through simulation experiments, the impact of non-completely fixed components on the expected displacement of the Bowden cable is eliminated. Patient wearing experiments are conducted to demonstrate the effectiveness of the exoskeleton robot in clinical settings. The results confirm that the robot can achieve a favorable assistance effect on children's rehabilitation training. References | Related Articles | Metrics

Select

Psychological Counseling with Integration of Knowledge Graph and Multi-Agent Collaboration Liu Xi, Zheng Xiaohu, Cui Weiwei 2026, 31 (5):  1162-1174.  doi: 10.1007/s12204-024-2785-1 Abstract ( 44 )   PDF (1960KB) ( 23 )   Applying large language models (LLMs) to psychological counseling can reduce costs and lower barriers to access. However, issues inherent to LLMs, such as output hallucinations, knowledge lags, and response biases, limit the quality of counseling services. To address these challenges, we propose a psychological counseling method that integrates a knowledge graph (KG) and multi-agent collaboration (MAC). First, we design an ontology model and use an improved convolution residual network (CRN) to extract relational triples for constructing a psychological medicine KG. Second, we perform low-rank adaptation (LoRA) on Qwen1.5-7B-chat using the Text2Cypher training set. Finally, we design MAC prompts based on the characteristics of the retrieval-based question-answering (QA) process to facilitate the collaborative implementation of psychological medicine knowledge QA and updates. The experiments showed that the improved CRN achieved an extraction accuracy of 81%. The mean squared error of the general-domain capability of the LLM before and after LoRA was only 0.738, with a Text2Cypher score of 99.8 after LoRA. The results generated after prompting received a 70.44% approval rate compared to other methods. Thus, the proposed method can effectively enhance the intelligence, accuracy, and controllability of psychological counseling. References | Related Articles | Metrics

Select

Medical Image Segmentation Using Multi-Head Self-Attention-Based Residual Double U-Net Pandu J., Ravi Shankar Reddy G., Ashok Babu Ch. 2026, 31 (5):  1175-1186.  doi: 10.1007/s12204-024-2756-6 Abstract ( 33 )   PDF (714KB) ( 25 )   Medical image segmentation plays a crucial role in facilitating clinical diagnosis and treatment, yet it poses numerous challenges due to variations in object appearances and sizes with indistinct boundaries. This paper introduces the MHSAttResDU-Net architecture, a novel approach to automatic medical image segmentation. Drawing inspiration from the double U-Net, multi-head self-attention (MHSA) model, and residual connections, the proposed model is trained on images pre-processed by the innovative ranking-based color constancy approach (RCC). The MHSAttResDU-Net includes the integration of RCC to control model complexity and enhance generalization across diverse lighting conditions. Additionally, the incorporation of the sparse salient region pooling (SSRP) unit in the encoder-decoder blocks reduces the dimension of feature maps, capturing essential local and global channel descriptors without introducing learnable parameters. MHSA gates are strategically employed in both down-sampling and up-sampling paths, allowing the recollection of additional relevant dimensional data. This effectively addresses dissimilar feature representations, minimizing unfocused noise and artifacts while reducing computational costs. Furthermore, Leaky ReLU-based residual connections between the encoder and decoder enhance the model's capability to recognize complex shapes and structures, ensuring improved gradient flow and faster convergence. Experimental results demonstrate the superiority of the MHSAttResDU-Net architecture across diverse datasets, including COVID-19, ISIC 2018, CVC-ClinicDB, and the 2018 Data Science Bowl. The model achieves an accuracy of up to 99.6% on the CVC-ClinicDB dataset and accuracies ranging from 97.0% to 99.6% across the four datasets, representing a promising advancement in automated medical image analysis with potential implications for improving patient care and diagnostic accuracy. References | Related Articles | Metrics

Select

Multi-Layer Propagation Neural Network for Diffraction Tomography Peng Jiageng, Miao Yubin 2026, 31 (5):  1187-1196.  doi: 10.1007/s12204-025-2809-5 Abstract ( 26 )   PDF (1785KB) ( 24 )   Intensity diffraction tomography is a label-free and non-invasive three-dimensional microscopy imaging technique, widely used in the analysis of cell morphology and study of physicochemical properties. However, the existing intensity diffraction tomography techniques often require complex computations and converge slowly, particularly when applied to samples with multiple scattering. To address the above issues, this paper presents a multi-layer propagation neural network model that incorporates the obliquity factor into forward propagation. In this model, the refractive index distribution of biological samples is represented as learnable parameters. Using the propagation equations, the predicted intensity at the measured plane is obtained, and the reconstruction process is fully automated, simplifying gradient computation and enhancing reconstruction efficiency. The experimental results demonstrate that, compared to existing multi-layer propagation models, our approach increases the average reconstruction speed by 5% to 40%, significantly reducing reconstruction time and facilitating faster cell analysis. References | Related Articles | Metrics

Select

Structural Design of Implant Fusion Based on Triple Period Minimal Surface and Its Performance Analysis LI Haijie, HAI Jizhe, SHAN Chunlong, XU Qingyu, XU Zhiguo, LEI Yueheng, JING Lei 2026, 31 (5):  1197-1212.  doi: 10.1007/s12204-025-2822-8 Abstract ( 36 )   PDF (3616KB) ( 23 )   Porous structures of triple periodic minimal surface (TPMS) based on diamond (D), gyroid (G), and primitive (P) have become a leading choice in dental implant design because of their excellent mechanical and biological characteristics. However, the natural mandibular tissue is inherently heterogeneous, and using homogeneous porous implants might cause pathological bone stress concentrations. To address this issue, this study constructs a TPMS implant with an axially fused structure based on the Sigmoid function and parametric design. The fused structure axially integrates two types of TPMS units. A simulation analysis for 18 different structures was conducted to analyze the differences between fused structures (D-G, D-P, and G-P) and homogeneous structures (D, G, P) under constant porosities of 50%, 60%, and 70% in terms of equivalent elastic modulus, equivalent yield strength, permeability, fluid wall shear stress, and micromotion at the implant-bone interface. The results show that compared with the homogeneous scaffolds, the fusion structure (D-G) has a low and wide range of elastic modulus, which decreases from 12.23 GPa to 4.58 GPa in a porosity range of 50%—70%, and its elastic modulus range increases by 28% on average. The yield strength was similar to that of cortical bone. The permeability was lower than that of the homogeneous scaffold (G, P), but it met the requirements of natural bone. In terms of fluid shear stress, the average wall shear stress of the fusion structure (D-G) was greater than that of the homogeneous structure, and its suitable growth zone accounted for 98.9%. In the simulation of implant-bone interface fretting under the effect of the occluding force, the axial fusion structure (D-G) showed a stress distribution more similar to that of the mandible; that is, a larger stress distribution was in the upper D structure region in contact with cortical bone, and a smaller stress was evenly distributed in the lower G structure region in contact with cancellous bone. This stress distribution effectively reduces the difference in stress and microstrain between cortical bone and cancellous bone, avoids the phenomenon of stress concentration in cancellous bone, reduces the risk of bone resorption, and improves the overall biomechanical stability of the implant. References | Supplementary Material | Related Articles | Metrics

Select

Computer-Aided Diagnosis of Endometrial Cancer Histopathologic Images Based on an Improved InceptionNeXt Model PAN Zhonghan, NAN Zhuojiang, GONG Chenbo, TAO Wei 2026, 31 (5):  1213-1225.  doi: 10.1007/s12204-024-2780-6 Abstract ( 35 )   PDF (1695KB) ( 23 )   Endometrial cancer is a common female cancer that poses a threat to the health of women worldwide. Several researchers have developed deep learning-based computer-aided diagnosis (CAD) systems for endometrial cancer with good results. However, when finer diagnostic results are needed, the results of existing models are not satisfactory. Therefore, we propose a CAD method for endometrial cancer based on an improved InceptionNeXt network. We designed an efficient coordinate attention module to be added to the network structure, which further enhances the model's ability to extract location information and channel relationships. We improved the training strategy to effectively improve the generalization ability and robustness of the model. Our model performs well in both four-classification and two-classification tasks, with diagnostic accuracy of 70.74% and 95.43%, respectively. We also performed a visualization analysis of the model to validate its ability to identify morphological features of images and provide pathological interpretations. Our method has a reliable potential to assist physicians in making clinically graded diagnoses. References | Related Articles | Metrics

Select

Dual-Branch Knowledge Fusion Network for Differential Diagnosis of Lung Cancer Cytopathology Images TANG Songqiao, GENG Chen, GONG Wei, HU Fuyuan, DAI Yakang 2026, 31 (5):  1226-1236.  doi: 10.1007/s12204-024-2783-3 Abstract ( 47 )   PDF (1627KB) ( 23 )   Adenocarcinoma and squamous cell carcinoma constitute a significant proportion of lung cancer cases, with distinct treatment strategies and prognoses. Accurate differentiation is crucial for clinical treatment. Advanced-stage lung cancer patients often cannot undergo tissue pathology, making cytopathology based on rapid on-site evaluation a feasible approach. However, this method faces challenges such as a shortage of digital pathology experts and diagnostic subjectivity. Current computer-aided digital cytopathology studies often overlook the spatial information between cells. This study proposes a knowledge-fusion cross-attention classification network, utilizing a graph convolutional network branch and a convolutional neural network branch to respectively introduce spatial information knowledge and texture knowledge. The knowledge is fused and classified using a multilayer perceptron, with a cross-attention mechanism proposed for node optimization during the pooling stage. Experiments on a pulmonary cell pathology dataset containing 209 cases showed that the proposed method achieved an accuracy of 0.925 in distinguishing adenocarcinoma from squamous cell carcinoma, with both recall and precision at 0.950, an F1-score of 0.920, and AUC of 0.980. This performance surpasses all comparison methods. In summary, the proposed method can improve the accuracy of diagnosing lung adenocarcinoma and squamous cell carcinoma, providing assistance in rapid intraoperative pathology for lung cancer patients. References | Related Articles | Metrics

Select

MCFF-Net: Multi-Scale Contextual Feature Fusion Network for Polyp Segmentation MA Lei, YAN Jiangkai, SHAO Dangguo, LI Jingtao, WANG Jiawei, YAN Yukun 2026, 31 (5):  1237-1249.  doi: 10.1007/s12204-024-2791-3 Abstract ( 39 )   PDF (1378KB) ( 23 )   In clinical practice, technologies for automatically segmenting polyps might greatly improve the speed and accuracy of medical diagnosis, significantly reducing the chances of patients developing colorectal cancer. Nevertheless, the segmentation of polyps presents a considerable challenge due to their low contrast, presence of occlusion, and intricate background disturbances. In this paper, we propose an innovative approach for segmenting polyps, called multi-scale contextual feature fusion network for polyp segmentation (MCFF-Net). Specifically, MCFF-Net includes three modules: feature enhancement module (FEM), global-local feature fusion module (GLFF), and cross-scale feature fusion module (CFFM). Among them, FEM is used to improve the features at each scale. Building on this, the CFFM enables comprehensive interaction among the information across various scales, enhancing the diversity of the feature scales extracted. Moreover, the GLFF fuses global-local features to generate a rough object localization map. This map further modulates the fusion features from the CFFM, facilitating more precise segmentation. On five publicly available datasets, our approach not only segments various polyp types but also surpasses existing methods in terms of accuracy and generalizability. References | Related Articles | Metrics

Select

Spatio-temporal Context Fusion Network for Endoscopic Ultrasound Video Segmentation ZHENG Hanyao, ZHENG Hangbin, WAN Xinjian, DONG Zhixia, LI Baiyan, BAO Jinsong 2026, 31 (5):  1250-1259.  doi: 10.1007/s12204-024-2786-0 Abstract ( 36 )   PDF (1352KB) ( 27 )   The accurate and automatic segmentation of endoscopic ultrasound (EUS) videos is crucial for assessing the progression of submucosal tumors (SMTs) in the digestive tract and developing treatment plans. However, the dynamic nature of these videos and the rich temporal information that they contain render traditional static image processing techniques insufficient for fully extracting potential diagnostic information. To address this issue, a spatio-temporal aggregation network is proposed for the precise segmentation of EUS videos. This network comprises a spatio-temporal attention module (STAM) and a spatio-temporal context fusion module (STFM). The STAM extracts spatio-temporal features using attention mechanisms in both the temporal and spatial dimensions. The STFM synchronously integrates the spatial context from static images with the temporal context across different frames to fully explore the potential diagnostic information. The experimental results demonstrate that this method significantly improves the semantic segmentation performance of SMT videos, such as those of gastrointestinal stromal tumors and leiomyomas. References | Related Articles | Metrics

Select

Macro-Meso-Micro Biomechanical Analysis of the Lumbar Spine After Pedicle Subtraction Osteotomy for Idiopathic Scoliosis ZHANG Huaiyue, FU Rongchang 2026, 31 (5):  1260-1276.  doi: 10.1007/s12204-024-2788-y Abstract ( 38 )   PDF (3592KB) ( 14 )   To investigate the multiscale biomechanical properties and bone tissue osteogenicity at the bone contact surface after pedicle subtraction osteotomy in idiopathic scoliosis, this study constructs a postoperative macroscopic finite element model, a mesoscopic bone unit, and a microscopic lacunar-canalicular system's finite element model by applying the sub-model approach and based on microstrain. With these models, normal physiological working conditions were simulated, the osteogenicity of bone tissues at the vertebral bone contact surfaces at different scales was analyzed, and the differences in mechanical properties between the mesoscopic and macroscopic models were compared. The results indicated that the bone unit stress and osteocyte synaptic flow velocity and pressure in the end face of the upper vertebra were greater than those in the end face of the lower half of the vertebral body, and the highest values were obtained under the lateral bending condition; the average stress in the end face of the upper half of the vertebral body was ranked from large to small as the right side, the left side, and the rear, while in the lower half of the vertebral body, it was ranked as the left side, the front, and the right side; the stress at the mesoscopic scale in this case was about 1.442—5.217 times higher than that in the macroscopic scale. It can be inferred that there is a difference in osteogenicity of bone tissue at the different end surfaces and around the end surfaces at the bone contact surface; frequent postoperative lateral bending exercises can help the growth of bone tissues, and the stress amplification relationship between the mesoscopic bone unit and the macroscopic cortical bone is revealed. This study demonstrates the importance of multiscale studies for the fusion of bone contact surfaces and provides a theoretical basis for designing and optimizing surgical protocols for idiopathic scoliosis. References | Related Articles | Metrics

Select

DSFNet: Deep Fusion Parallel MRI Reconstruction Based on SPIRiT and Complex Convolution DUAN Jizhong, REN Xinmin, CHEN Shengyi 2026, 31 (5):  1277-1289.  doi: 10.1007/s12204-024-2778-0 Abstract ( 29 )   PDF (1784KB) ( 13 )   Long data acquisition time is an inherent disadvantage of magnetic resonance imaging (MRI). To accelerate the data acquisition speed of MRI, undersampling is required, which results in low imaging quality. Based on an iterative self-consistent parallel imaging reconstruction (SPIRiT) and improved complex convolutional neural networks, a deep SPIRiT fusion network (DSFNet) is proposed to improve the quality of reconstructed images. The DSFNet model first uses the SPIRiT model for the reconstruction of under-sampled k-space data. Subsequently, a cascaded complex convolutional neural network with dense connections is utilized for further reconstruction of the calibrated k-space data. Besides, a data consistency layer ensures the fidelity of the reconstructed image in both k-space and image domains. The final magnitude image is fused from the two parts of the reconstruction magnitude images by the sequentialized model-based Bayesian optimization fusion module at a certain ratio. Experimental results on different knee datasets show that DSFNet can bring about a substantial improvement in the visualization quality, peak signal-to-noise ratio and structural similarity of the reconstructed images compared to the SPIRiT, Deepcomplex and DONet models. References | Related Articles | Metrics

Select

Fabrication of Functionally Fused Porous Tantalum/Lithium Magnesium Silicate Polymer Bone Repair Scaffolds JING Lei, HAI Jizhe, SHAN Chunlong, XU Qingyu, LEI Yueheng, LI Haijie 2026, 31 (5):  1290-1301.  doi: 10.1007/s12204-025-2834-4 Abstract ( 35 )   PDF (2947KB) ( 13 )   The repair of severe bone defects remains a significant challenge in the field of orthopedics. 3D-printed metal-based scaffolds have been widely applied in bone repair due to their excellent structural and mechanical adaptability. However, conventional porous scaffolds lack the microstructure that mimics the extracellular matrix (ECM), limiting their osteointegration potential. Therefore, the development of composite bone repair materials with both biomimetic structures and biological activity has become a research focus. This study is based on the design of triply periodic minimal surfaces (TPMS) and constructs a melt-processed porous tantalum scaffold using gyroid and diamond units. Additionally, a gelatin-alginate-magnesium-lithium silicate (GE-SA-LAP) composite polymer is infiltrated into the scaffold, followed by freeze-drying to form a new composite bone scaffold material with both structural biomimicry and functionality. The TPMS structure mimics the microstructure and mechanical properties of natural trabecular bone, while the polymer filling further mimics the ECM microenvironment, optimizing conditions for cell attachment and proliferation. The characteristics and biocompatibility of the porous tantalum-polymer composite scaffold were evaluated through finite element simulation, mechanical property testing, scanning electron microscopy, and in vitro cell experiments. Finite element analysis revealed that the elastic modulus of the scaffolds in each group ranged from 1.21 to 1.55 GPa, yield strength ranged from 29.81 to 39.68 MPa, and permeability ranged from 2.72×10⁻⁸ to 3.45×10⁻⁸ m², falling within the range of human cortical and trabecular bone. The GE-SA-LAP composite polymer, which fills the pores, exhibits high porosity and provides an ideal biological microenvironment for cell attachment. In vitro cell experiments further validated that the composite scaffold outperforms the pure tantalum scaffold in terms of cell attachment and proliferation, demonstrating exceptional potential for bone tissue repair. References | Related Articles | Metrics

Select

Boundary-Aware Voxel-Wise Contrastive Learning for Infant Brain Tissue Segmentation HONG Guofei, PENG Bo, LIU Surui, LIU Yan, ZHOU Zhiyong, DAI Yakang 2026, 31 (5):  1302-1314.  doi: 10.1007/s12204-024-2782-4 Abstract ( 31 )   PDF (2323KB) ( 14 )   The accurate segmentation of infant brain magnetic resonance (MR) images is crucial for understanding the structure and early detection of neurodevelopmental disorders. However, it remains challenging mainly because of the inverted contrast between white matter and gray matter, and high-intensity inhomogeneity. Existing deep-learning segmentation methods do not fully learn the boundary features of brain tissues, resulting in difficulties in improving the segmentation accuracy of brain tissue boundaries. In this study, a boundary-aware voxel-wise contrastive learning method is proposed for infant brain tissue segmentation by designing a voxel-wise contrastive learning loss to represent the distance among positive samples, negative samples, and anchors in the embedding space. The proposed method was tested on a 0.35 T low-field infant brain MR dataset and the MICCAI iSeg-2019 challenge dataset. Compared with the baseline methods, the proposed method achieved an average improvement of 1.92% and 0.81% in the Dice index for different brain tissues on the respective datasets, thereby significantly enhancing the accuracy of brain tissue segmentation. References | Related Articles | Metrics

Select

Breast Cancer Classification Using Graph Convolutional Networks nd DenseNet121 with Pruning KAVITHA S. J., SRIDEVI S. 2026, 31 (5):  1315-1332.  doi: 10.1007/s12204-025-2826-4 Abstract ( 26 )   PDF (1539KB) ( 14 )   Breast cancer remains a significant global health burden, particularly among women, necessitating accurate and accessible diagnostic systems for early detection and intervention. Despite advancements in medical technology, challenges persist such as high dimensionality, variability, and complexity. Hence, this study introduces a novel approach that amalgamates DenseNet121 architecture within graph convolutional networks (GCNs) for breast cancer classification. The proposed method combines the strengths of a pre-trained DenseNet121 model, pruning techniques, and GCN with a cross-space filter (CSF). DenseNet121 is employed for feature extraction, while pruning enhances efficiency and prevents overfitting. GCNs capture structural and semantic information from the data and the CSF aids in regularization. Furthermore, the incorporation of the CSF enhances model performance by effectively integrating information from both graph topology and node attributes. The proposed approach was trained and evaluated using the BreakHis dataset, a widely recognized resource for breast cancer classification through histopathological images. By leveraging several performance measures consisting of precision, recall, F1-score, AUC, MCC, and accuracy, the proposed approach was compared against existing models to assess its effectiveness. The results indicate that the proposed method, particularly when combined with the CSF, outperforms existing approaches, demonstrating its potential as an advanced tool for accurate and interpretable breast cancer diagnosis. References | Related Articles | Metrics

Select

Prototype Device for Neonatal Jaundice Phototherapy with Dynamic Physiological Monitoring Function GUO Kunming, HE Ziliang, ZOU Zuofeng, ZHANG Benjin, ZHANG Jing, XIAO Zhenghua, YANG Gang 2026, 31 (5):  1333-1345.  doi: 10.1007/s12204-024-2774-4 Abstract ( 37 )   PDF (968KB) ( 14 )   Treating neonatal jaundice through phototherapy requires a personalized approach that considers both effectiveness and potential side effects and is adaptable based on the infant's physiological condition. However, there is a current lack of standardized measurement protocols for evaluating phototherapy's effectiveness and side effects in clinical practice. This study developed a neonatal jaundice phototherapy system with integrated dynamic physiological monitoring capabilities to address this gap. A neonatal jaundice phototherapy system was designed and constructed based on previous research. The dynamic physiological monitoring component was based on established methodologies, and the phototherapy system utilized light-emitting diode arrays, undergoing thorough performance and functional testing. Furthermore, specialized software was developed to process and present physiological monitoring data and regulate phototherapy parameters. Performance testing confirmed the stability of the phototherapy system, its ability to achieve therapeutic light levels, and the precise control over light intensity through software adjustments. Functional testing demonstrated the system's efficacy in promoting bilirubin degradation in an in vitro model and regulating the degradation rate by adjusting light intensity levels. The physiological monitoring component successfully tracked various physiological indicators (jaundice, body positioning, and skin condition). The innovative system developed in this study enables real-time monitoring of essential physiological parameters in neonates, offering a fresh perspective on standardized and personalized phototherapy for neonatal jaundice. This system holds promise for broad application in clinical settings. References | Related Articles | Metrics

Select

Real-Time Online Intelligent Segmentation of Carotid Artery Ultrasound Video GUO Jin, LIN Yanping, LI Zhaojun 2026, 31 (5):  1346-1355.  doi: 10.1007/s12204-024-2758-4 Abstract ( 37 )   PDF (1682KB) ( 16 )   The morphology and structure of carotid arteries can provide valuable information about the degree of arterial hardening and plaque formation. This information holds significant clinical importance in the diagnosis and treatment of cardiovascular diseases such as atherosclerosis. Ultrasound image segmentation technology precisely extracts carotid artery contours from background images, providing physicians with anatomical and pathological insights. Traditional carotid artery ultrasound image segmentation methods rely on offline volume data, posing issues related to low efficiency, poor accuracy, and dependence on physician expertise during clinical use. This study introduces a real-time online intelligent segmentation approach for carotid artery ultrasound, aimed at enhancing clinical physicians' operational efficiency and diagnostic accuracy during ultrasound examinations. Utilizing space-time memory, we propose an end-to-end video segmentation network that simplifies panoramic segmentation into localized segmentation through coordinate regression, providing relative spatial position information. After processing and training on 60 expert-annotated ultrasound volume data samples, the model achieved an average Dice of 90.51% on the carotid artery ultrasound image test set. The network processed images at a speed of 40 frames per second, achieving the goal of real-time online intelligent and precise segmentation of carotid artery ultrasound images. References | Related Articles | Metrics

Select

Schizophrenia Identification Using Machine Learning Methods with Graph-Theoretic Features ZHU Geng, XU Qi, ZHANG Fengzhu, YANG Yaqin, LI Xiaoou, LI Bin, LI Shengtian 2026, 31 (5):  1356-1365.  doi: 10.1007/s12204-025-2811-y Abstract ( 29 )   PDF (1193KB) ( 16 )   To realize the automatic identification of electroencephalogram (EEG) signals from patients with schizophrenia, this study calculated graph-theoretic features derived from event-related potentials (ERP) recorded during a Go/NoGo task. The ERP data were collected from 42 individuals diagnosed with schizophrenia and 29 healthy controls. Different functional networks of single-frequency band and all band of these EEG data, were constructed by phase locking value method. Their graph-theoretic features were extracted. Three machine learning classifiers were used to automatically distinguish the patients and health controls. The classification accuracies of different models with graph-theoretic features of different networks were compared. The results revealed that F-score combined with support vector machine (SVM) model achieved the highest accuracy. This model with graph-theoretic features of all-band network and theta band networks resulted in the highest classification accuracy, which was 95.45%. It suggested that the SVM with graph-theoretic features can be considered as a potential method for schizophrenia identification. References | Related Articles | Metrics

Select

Interference Detection Method for Eye Tracking System in External Radiation Therapy LUO Jiachen, SHI Dequan, WANG Xu, WANG Kundong 2026, 31 (5):  1366-1373.  doi: 10.1007/s12204-024-2753-9 Abstract ( 33 )   PDF (1607KB) ( 17 )   In the process of external radiation therapy, eye movements need to be detected in real time using an eye tracking system. The complex eye tracking system may interfere with the treatment radiation beam, so it is necessary to build a system model to detect possible interference and ensure that the input dose to the diseased site is as expected. Firstly, the scene needs to be modeled, and the control robotic arm and its end structure are modeled using the Denavit-Hartenberg (D-H) parameters. The interference detection is implemented using the Gilbert-Johnson-Keerthi algorithm and the detection results are displayed using the Godot Engine, while the simulation scene built in Simscape using the same model is compared and verified. The results of the two simulations are compared under the condition that the robot arm is in the initial position, and the incident rays are at different angles, respectively, proving the effectiveness of the D-H parametric modeling and the interference detection algorithm. References | Related Articles | Metrics

Select

Information Analysis System of Human Respiratory Motion for PET/CT Scanning MAO Yuxuan, LIU Weiping, CHEN Xiaojun 2026, 31 (5):  1374-1385.  doi: 10.1007/s12204-025-2833-5 Abstract ( 37 )   PDF (1096KB) ( 13 )   Positron emission tomography - computed tomography (PET/CT) is an advanced molecular imaging technique that is mainly used for the early detection and diagnosis of major heart diseases and tumors. To reduce imaging artifacts caused by the patient's respiratory motion, the mainstream respiratory control method such as gating technique has been widely used. Based on the optical fiber pressure sensor, a real-time analysis software system of respiratory motion information has been developed in this study. A gate alignment algorithm for correlating respiratory motion in PET/CT scanning was investigated. In this study, a real-time extraction algorithm of smooth respiratory waveform and a fast temporal phase segmentation algorithm were proposed, reducing the computational complexity and guaranteeing low latency and high recognition accuracy, and then they were implemented in the lower computer by programming. The software and hardware integration of the respiratory acquisition system with the fiber-optic sensor-based mattress was realized, and the whole prototype product was developed for experimental studies. Compared with the existing traditional methods, the average detection latency of the method in this study is less than 150 ms (about 300 ms for existing methods), and the relative error of temporal phase segmentation is about 1% (about 5% for existing methods). This study offers significant potential for enhancing PET/CT imaging quality and efficiency, with future steps aimed at integrating the system into commercial platforms for clinical validation. References | Related Articles | Metrics

Select

Optimizing Implant-Mandibular Model with Sub-Model Approach Enhances Bone Quality and Implant Stability LIAO Chenjian, LIU Xiaoying, XIE Ye, LU Jiancheng 2026, 31 (5):  1386-1396.  doi: 10.1007/s12204-025-2829-1 Abstract ( 47 )   PDF (1515KB) ( 18 )   In this study, an implant-mandibular coupling integral model was constructed. The biomechanical properties of each structure in the implant-mandibular holistic model were explored under dental implant structuring conditions and bone quality. Based on the analysis results of the overall implant-mandibular model, the local area around the implant was studied using the sub-modeling method. By combining the sub-modeling and topology optimization methods, we obtained an efficient and accurate sub-model of four types of bone mandibles; Further, the occlusal conditions of implants for these bone mandibles were numerically simulated, obtaining the optimized structure of implant multi-conditional topology under different bone qualities according to the numerical simulation results; The biomechanical properties of implant topology were evaluated under different bone qualities through the reconstructive design of implant structure to determine optimal distribution of implant materials. The implant sub-modeling analysis is efficient and accurate and can serve as the basis for optimizing implant topology. For the initial reconstructed topology, the primary trend leading to increased stresses in the bone and implant was not optimized. In contrast, for secondary reconstruction topologies with different material removal ratios, the relationship between the influence of the material removal ratio and biomechanical properties of the bone and implant was obtained. The biomechanical performance of the secondary reconstruction topology with a material removal ratio of approximately 20% was superior for all four types of bone mandibles; the better the bone quality, the higher the suitability for the implant topology. References | Supplementary Material | Related Articles | Metrics

Select

Probabilistic Undersampling Mask Optimization with Weight Adaptation for MRI Reconstruction DUAN Jizhong, ZHU Qinyao 2026, 31 (5):  1397-1408.  doi: 10.1007/s12204-025-2831-7 Abstract ( 31 )   PDF (2212KB) ( 14 )   In order to enhance the quality of reconstructed images, a probabilistic undersampling mask optimization with weight adaptation for magnetic resonance imaging (MRI) reconstruction network (PUOR-Net) is proposed based on the undersampling mask optimization strategy and the iterative shrinkage-thresholding algorithm. PUOR-Net comprises two modules: the undersampling subnet and the reconstruction subnet. Within the undersampling subnet, PUOR-Net dynamically generates an optimized undersampling mask through adaptive weight adjustment mechanisms, overcoming the limitations of traditional methods reliant on pre-designed undersampling mask. The reconstruction subnet integrates complex convolution and channel attention mechanisms to further leverage data potential and accurately recover image detail. Experimental results show that PUOR-Net outperforms better on two MRI datasets than the traditional method using the fixed undersampling mask, which verifies the effectiveness of its adaptive optimization strategy. In addition, PUOR-Net demonstrates performance advantages compared to similar optimized mask joint learning networks. References | Related Articles | Metrics

Select

Optimal Design of Coils for Wireless Energy Transfer Urethral Valve YIN Mao, LI Xiao 2026, 31 (5):  1409-1424.  doi: 10.1007/s12204-024-2741-0 Abstract ( 33 )   PDF (1972KB) ( 16 )   This study optimizes a wireless energy transmission coil to address the low efficiency in percutaneous energy transmission of existing urethral valves. Based on the response surface method and genetic algorithm, the optimal model of wireless energy transmission efficiency is established, and the optimal structure parameters of the coil are determined. The influence of adding ferrite substrate and optimizing the structure on the link transmission efficiency (LTE) for the initial coil, as well as the impact of transmitting coil location on the LTE, was studied theoretically and experimentally. The experimental analysis focuses on the optimized coil's LTE and the urethral valve's driving performance. Additionally, a simulation is conducted to assess the electromagnetic safety of the optimized coil. The results show that the structural optimization approach suggested in this research is around 13% more efficient than merely adding substrate, and combining the two enhances the LTE. The LTE decreases as the transmitting coil deflection angle increases. Adding substrate and optimizing the structure can minimize the impact of transmitting coil position change on the LTE. At an axial distance of 30 mm, the optimized coil exhibits an LTE of 74.4%. Compared with the hollow solenoid coil, the urethral valve indicates a 24.4% increase in LTE, a 9.1% increase in driving force, and a reduction of 9.2 s in urethral opening time. After optimization, the coil's radiation dose to the abdomen's subcutaneous tissue is less than the fundamental limit, meeting the safe use requirements. This study can provide theoretical guidance for optimizing the design of efficient, portable, and reliable urethral valves. References | Related Articles | Metrics