Zhongjie Shi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Zhongjie Shi has a total of 25 co-authors. Here are the top co-authors by number of shared publications:
Xin Gao
Department of Neurosurgery, The First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China
3 shared publications- Zhanxiang Wang
Xiamen University
3 shared publications Deyong Xiao
Department of Neurosurgery, The First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China
2 shared publicationsXiaojun Li
Department of Neurosurgery, The First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China
2 shared publications- Zhanxiang Wang
Department of Neurosurgery, The First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, China
2 shared publications Zhengbo Yuan
Department of Neurosurgery, The First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University
2 shared publications
Latest Publications
- Open Access
- 17 April 2026
- Published by Frontiers Media SA in Frontiers in Neuroscience
- Vol. 20, 1783020
Brain-computer interfaces: an engineering black-box swindle or a lone advance guided by deep learning
Brain-computer interfaces (BCIs) have evolved from speculative science fiction to tangible neurotechnology, with clinical systems now enabling paralyzed individuals to control assistive devices. Yet a persistent question remains: do BCIs truly "understand" neural intent, or do they merely fit statistical models to observable signals? This debate often polarizes into two camps: one dismissing BCIs as "engineering black-box swindles," and the other heralding them as "lone advances guided by deep learning." Here, we argue that both extremes overlook the essence of modern Brain-computer interface (BCI) systems-bidirectional co-adaptation between brain and machine. This opinion article examines the interpretive and methodological challenges in BCI research while also translating them into practical priorities for clinically meaningful, stable, and scalable BCI development. Specifically, we focus on three linked themes: (i) why "black-box" decoding can be a principled approximation problem rather than a philosophical failure; (ii) how recording constraints and low-dimensional structure jointly shape what BCIs can infer and control; and (iii) why non-stationarity and co-adaptation should be treated as central design variables rather than afterthoughts. Building on these themes, we further outline concrete future directions that may help move the field from conceptual debate toward constructive clinical and technological advancement.A common critique of BCIs is their reliance on "black-box" models-such as Kalman filters or deep neural networks-that perform dimensionality reduction on high-dimensional neural data. Detractors label this as overfitting or "hard-coding." However, we reframe this process as necessary function approximation. Consider the following formulation: Teacher Model: the brain (unknown architecture, unknown weights, highly nonlinear, and state-dependent). Student Model: the BCI decoder (for example, a Kalman filter or a transformer-based model). Input: recorded neural activity (for example, spikes or local field potentials). Target: task-relevant variables such as cursor velocity, prosthetic-arm kinematics, or speech-related features.According to the Universal Approximation Theorem, as long as the student model (decoder) has sufficient capacity and enough observed features (neurons), one can fit the teacher model's (brain's) projection under a specific task. This holds mathematically.In practice, dimensionality reduction is not a mechanistic shortcut but a computational strategy to extract task-relevant information from noisy, high-dimensional inputs. For instance, the Unscented Kalman Filter improves upon linear models by capturing nonlinear neural dynamics [1,2], while deep learning architectures enable robust feature learning from large-scale neural recordings. These approaches do not claim to reveal the brain's internal logic; rather, they establish usable mappings that facilitate real-time control under hardware constraints.We argue that mapping complex neural activity into a low-dimensional space currently enables real-time decoding of electrophysiological signals for external device control under constrained computational resources [3,4]. In this context, dimensionality reduction should not be viewed as a violent simplification of the brain, but as a principled way to extract task-relevant structure from high-dimensional and noisy activity. Describing the closed-loop BCI paradigm as "engineering hard-coding" and "forced dimensionality-reduction fitting of chaotic signals" is an oversimplification; rather, it it reflects a pragmatic scientific strategy for extracting task-relevant structure from complex neural activity. The key scientific question is therefore not whether a decoder fully recovers the brain's entire internal mechanism, but whether it captures sufficient structure to support reliable and clinically meaningful control. Future work in this area should move beyond the binary opposition between "interpretability" and "performance," and instead ask which levels of representation are necessary for robust decoding, generalization across sessions, and practical bedside usability.A given neuron may contribute to different task variables across adjacent moments and behavioral contexts. Modern neuroscience emphasizes mixed selectivity, namely that single neurons can encode multiple task variables in a context-dependent manner, thereby expanding representational capacity [5]. For BCI systems, this implies that the mapping between recorded activity and behavioral variables is not strictly fixed. Across minutes, days, and sessions, changes in neural recruitment, recording conditions, or latent population structure can produce neural drift, causing a previously effective decoder to lose accuracy and forcing repeated recalibration. The key challenge, therefore, is not to assume a static brain, but to design dynamic models that can track non-stationarity while preserving stable control.Critically, drift adaptation does not require full observation of internal brain states. Instead, it leverages the geometric stability of neural manifolds-low-dimensional subspaces that capture task-relevant neural variability.By monitoring manifold structure, decoders can infer drift direction and adjust without explicit supervision. This approach aligns with the principle that effective BCI control depends not on reading "all neurons," [6] but on extracting consistent computational motifs from population activity. Importantly, a low-dimensional manifold does not mean that we can simply reduce the number of recorded signals, because it only indicates that activity does not fill the entire ambient space, while any given activation unit may still be critically engaged. For example, the raw image space is 256×256×3, which contains vast amounts of meaningless content (e.g., pure noise images); thus, natural images occupy a low-dimensional manifold...
- 17 April 2026
- Published by Frontiers Media SA in Frontiers in Neuroscience
- Vol. 20, 1783020
Open AccessBrain-computer interfaces: an engineering black-box swindle or a lone advance guided by deep learning
Brain-computer interfaces (BCIs) have evolved from speculative science fiction to tangible neurotechnology, with clinical systems now enabling paralyzed individuals to control assistive devices. Yet a persistent question remains: do BCIs truly "understand" neural intent, or do they merely fit statistical models to observable signals? This debate often polarizes into two camps: one dismissing BCIs as "engineering black-box swindles," and the other heralding them as "lone advances guided by deep learning." Here, we argue that both extremes overlook the essence of modern Brain-computer interface (BCI) systems-bidirectional co-adaptation between brain and machine. This opinion article examines the interpretive and methodological challenges in BCI research while also translating them into practical priorities for clinically meaningful, stable, and scalable BCI development. Specifically, we focus on three linked themes: (i) why "black-box" decoding can be a principled approximation problem rather than a philosophical failure; (ii) how recording constraints and low-dimensional structure jointly shape what BCIs can infer and control; and (iii) why non-stationarity and co-adaptation should be treated as central design variables rather than afterthoughts. Building on these themes, we further outline concrete future directions that may help move the field from conceptual debate toward constructive clinical and technological advancement.A common critique of BCIs is their reliance on "black-box" models-such as Kalman filters or deep neural networks-that perform dimensionality reduction on high-dimensional neural data. Detractors label this as overfitting or "hard-coding." However, we reframe this process as necessary function approximation. Consider the following formulation: Teacher Model: the brain (unknown architecture, unknown weights, highly nonlinear, and state-dependent). Student Model: the BCI decoder (for example, a Kalman filter or a transformer-based model). Input: recorded neural activity (for example, spikes or local field potentials). Target: task-relevant variables such as cursor velocity, prosthetic-arm kinematics, or speech-related features.According to the Universal Approximation Theorem, as long as the student model (decoder) has sufficient capacity and enough observed features (neurons), one can fit the teacher model's (brain's) projection under a specific task. This holds mathematically.In practice, dimensionality reduction is not a mechanistic shortcut but a computational strategy to extract task-relevant information from noisy, high-dimensional inputs. For instance, the Unscented Kalman Filter improves upon linear models by capturing nonlinear neural dynamics [1,2], while deep learning architectures enable robust feature learning from large-scale neural recordings. These approaches do not claim to reveal the brain's internal logic; rather, they establish usable mappings that facilitate real-time control under hardware constraints.We argue that mapping complex neural activity into a low-dimensional space currently enables real-time decoding of electrophysiological signals for external device control under constrained computational resources [3,4]. In this context, dimensionality reduction should not be viewed as a violent simplification of the brain, but as a principled way to extract task-relevant structure from high-dimensional and noisy activity. Describing the closed-loop BCI paradigm as "engineering hard-coding" and "forced dimensionality-reduction fitting of chaotic signals" is an oversimplification; rather, it it reflects a pragmatic scientific strategy for extracting task-relevant structure from complex neural activity. The key scientific question is therefore not whether a decoder fully recovers the brain's entire internal mechanism, but whether it captures sufficient structure to support reliable and clinically meaningful control. Future work in this area should move beyond the binary opposition between "interpretability" and "performance," and instead ask which levels of representation are necessary for robust decoding, generalization across sessions, and practical bedside usability.A given neuron may contribute to different task variables across adjacent moments and behavioral contexts. Modern neuroscience emphasizes mixed selectivity, namely that single neurons can encode multiple task variables in a context-dependent manner, thereby expanding representational capacity [5]. For BCI systems, this implies that the mapping between recorded activity and behavioral variables is not strictly fixed. Across minutes, days, and sessions, changes in neural recruitment, recording conditions, or latent population structure can produce neural drift, causing a previously effective decoder to lose accuracy and forcing repeated recalibration. The key challenge, therefore, is not to assume a static brain, but to design dynamic models that can track non-stationarity while preserving stable control.Critically, drift adaptation does not require full observation of internal brain states. Instead, it leverages the geometric stability of neural manifolds-low-dimensional subspaces that capture task-relevant neural variability.By monitoring manifold structure, decoders can infer drift direction and adjust without explicit supervision. This approach aligns with the principle that effective BCI control depends not on reading "all neurons," [6] but on extracting consistent computational motifs from population activity. Importantly, a low-dimensional manifold does not mean that we can simply reduce the number of recorded signals, because it only indicates that activity does not fill the entire ambient space, while any given activation unit may still be critically engaged. For example, the raw image space is 256×256×3, which contains vast amounts of meaningless content (e.g., pure noise images); thus, natural images occupy a low-dimensional manifold...
- Open Access
- 16 December 2025
- Published by Frontiers Media SA in Frontiers in Medicine
- Vol. 12, 1700166
Digital neurosurgery in the era of intelligent medicine: a scoping review
Background: Digital neurosurgery represents a transformative shift in modern neurosurgical practice, integrating advanced technologies, such as three-dimensional (3D) imaging, robotics, artificial intelligence (AI), and digital twin technology (DTT) models. These technologies offer innovative solutions for preoperative planning, intraoperative navigation, and postoperative management, with an emphasis on precision, personalization, and efficiency. Methods: We conducted a scoping review in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses-Scoping Review (PRISMA-ScR) checklist and guidance from the Joanna Briggs Institute (JBI) Manual for Evidence Synthesis. PubMed, Web of Science (WOS), and China National Knowledge Infrastructure (CNKI) were searched without language or date limits through September 2025. Eligibility was structured using the Population–Concept–Context (PCC) framework. Two reviewers independently screened records in Rayyan with consensus resolution, and data were charted using a prepiloted form. A total of 133 sources were included and mapped. Results: Key technologies reviewed include: (1) 3D reconstruction: Facilitates precise anatomical modeling, improving spatial understanding and surgical planning. (2) 3D printing (3DP): Enables creation of patient-specific models and surgical guides, enhancing preoperative simulation and intraoperative accuracy. (3) Digital twins (DT): Offers dynamic virtual models for real-time surgical simulation, training, and personalized patient management. (4) Intraoperative navigation: Utilizes advanced electromagnetic and AI-enhanced systems to improve tracking accuracy and reduce surgical errors. (5) Robotic-assisted surgery: Includes telesurgical, supervisory, and handheld systems that enhance precision and enable minimally invasive procedures. (6) AI: Supports image registration, subtask automation, and clinical decision-making, improving diagnostic and prognostic accuracy. These technologies demonstrate significant benefits in operative precision, patient outcomes, training efficacy, and interdisciplinary communication, though challenges remain in data integration, regulatory standards, and computational demands. Conclusion: Positioning this study as a scoping review clarifies its objective to map technologies and applications across digital neurosurgery rather than to synthesize effect estimates, thereby providing an evidence-informed overview to guide future systematic evaluations. Digital neurosurgery is rapidly evolving toward greater integration of multimodal data, real-time adaptive systems, and AI-driven automation. Future developments should focus on standardizing regulatory frameworks, enhancing data fusion capabilities, and promoting interdisciplinary collaboration to fully realize the potential of digital technologies in advancing neurosurgical care.
- 16 December 2025
- Published by Frontiers Media SA in Frontiers in Medicine
- Vol. 12, 1700166
Open AccessDigital neurosurgery in the era of intelligent medicine: a scoping review
Background: Digital neurosurgery represents a transformative shift in modern neurosurgical practice, integrating advanced technologies, such as three-dimensional (3D) imaging, robotics, artificial intelligence (AI), and digital twin technology (DTT) models. These technologies offer innovative solutions for preoperative planning, intraoperative navigation, and postoperative management, with an emphasis on precision, personalization, and efficiency. Methods: We conducted a scoping review in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses-Scoping Review (PRISMA-ScR) checklist and guidance from the Joanna Briggs Institute (JBI) Manual for Evidence Synthesis. PubMed, Web of Science (WOS), and China National Knowledge Infrastructure (CNKI) were searched without language or date limits through September 2025. Eligibility was structured using the Population–Concept–Context (PCC) framework. Two reviewers independently screened records in Rayyan with consensus resolution, and data were charted using a prepiloted form. A total of 133 sources were included and mapped. Results: Key technologies reviewed include: (1) 3D reconstruction: Facilitates precise anatomical modeling, improving spatial understanding and surgical planning. (2) 3D printing (3DP): Enables creation of patient-specific models and surgical guides, enhancing preoperative simulation and intraoperative accuracy. (3) Digital twins (DT): Offers dynamic virtual models for real-time surgical simulation, training, and personalized patient management. (4) Intraoperative navigation: Utilizes advanced electromagnetic and AI-enhanced systems to improve tracking accuracy and reduce surgical errors. (5) Robotic-assisted surgery: Includes telesurgical, supervisory, and handheld systems that enhance precision and enable minimally invasive procedures. (6) AI: Supports image registration, subtask automation, and clinical decision-making, improving diagnostic and prognostic accuracy. These technologies demonstrate significant benefits in operative precision, patient outcomes, training efficacy, and interdisciplinary communication, though challenges remain in data integration, regulatory standards, and computational demands. Conclusion: Positioning this study as a scoping review clarifies its objective to map technologies and applications across digital neurosurgery rather than to synthesize effect estimates, thereby providing an evidence-informed overview to guide future systematic evaluations. Digital neurosurgery is rapidly evolving toward greater integration of multimodal data, real-time adaptive systems, and AI-driven automation. Future developments should focus on standardizing regulatory frameworks, enhancing data fusion capabilities, and promoting interdisciplinary collaboration to fully realize the potential of digital technologies in advancing neurosurgical care.
- Open Access
- 3 November 2025
- Published by Frontiers Media SA in Frontiers in Neurology
- Vol. 16, 1691434
NP-guide: a portable projection-based navigation system for neurosurgery and beyond
Background: Stereotactic systems and various robot-assisted navigation platforms in neurosurgery have enabled high-precision localization. However, these systems, while highly accurate, are expensive, technically demanding, and procedurally complex, making them less practical for routine use. This study introduced and evaluated the Navigation and Projection Guide (NP-Guide), a projection-based augmented reality (AR) system designed to provide a portable and accessible solution for surgical navigation. Methods: NP-Guide, a mobile application, projects patient imaging data and three-dimensional (3D) reconstructions onto the patient’s head surface to assist with localization. This proof-of-concept study prospectively enrolled 52 neurosurgical patients, randomized to the NP-Guide group ( n = 27) or the freehand localization group ( n = 25). Two physicians with different training backgrounds performed the procedures. Localization error and operating time were measured using a commercial optical navigation system (ONS). Bland–Altman analysis was applied to assess inter-operator agreement, and learning curves were generated to evaluate proficiency. Results: Baseline characteristics were comparable (all p > 0.05). In the NP-Guide group, mean localization error was 4.1 ± 2.1 mm for Physician A and 3.4 ± 1.8 mm for Physician B, with mean times of 1.2 ± 0.5 min and 1.1 ± 0.4 min, respectively. Compared with freehand localization, NP-Guide significantly improved the accuracy and efficiency (all p < 0.001). Bland–Altman analysis demonstrated good inter-operator agreement; no significant difference was observed ( p = 0.25). Learning curves showed that operating times plateaued at approximately 1 min after about 15 cases. Conclusion: The NP-Guide demonstrated accurate, efficient, and reproducible projection-based localization in this proof-of-concept study. Its portability, low cost, and ease of use suggest potential value, particularly in resource-limited settings. However, these findings should be interpreted as preliminary, and further phantom experiments and multicenter clinical studies are required before widespread adoption in routine practice.
- 3 November 2025
- Published by Frontiers Media SA in Frontiers in Neurology
- Vol. 16, 1691434
Open AccessNP-guide: a portable projection-based navigation system for neurosurgery and beyond
Background: Stereotactic systems and various robot-assisted navigation platforms in neurosurgery have enabled high-precision localization. However, these systems, while highly accurate, are expensive, technically demanding, and procedurally complex, making them less practical for routine use. This study introduced and evaluated the Navigation and Projection Guide (NP-Guide), a projection-based augmented reality (AR) system designed to provide a portable and accessible solution for surgical navigation. Methods: NP-Guide, a mobile application, projects patient imaging data and three-dimensional (3D) reconstructions onto the patient’s head surface to assist with localization. This proof-of-concept study prospectively enrolled 52 neurosurgical patients, randomized to the NP-Guide group ( n = 27) or the freehand localization group ( n = 25). Two physicians with different training backgrounds performed the procedures. Localization error and operating time were measured using a commercial optical navigation system (ONS). Bland–Altman analysis was applied to assess inter-operator agreement, and learning curves were generated to evaluate proficiency. Results: Baseline characteristics were comparable (all p > 0.05). In the NP-Guide group, mean localization error was 4.1 ± 2.1 mm for Physician A and 3.4 ± 1.8 mm for Physician B, with mean times of 1.2 ± 0.5 min and 1.1 ± 0.4 min, respectively. Compared with freehand localization, NP-Guide significantly improved the accuracy and efficiency (all p < 0.001). Bland–Altman analysis demonstrated good inter-operator agreement; no significant difference was observed ( p = 0.25). Learning curves showed that operating times plateaued at approximately 1 min after about 15 cases. Conclusion: The NP-Guide demonstrated accurate, efficient, and reproducible projection-based localization in this proof-of-concept study. Its portability, low cost, and ease of use suggest potential value, particularly in resource-limited settings. However, these findings should be interpreted as preliminary, and further phantom experiments and multicenter clinical studies are required before widespread adoption in routine practice.
- Open Access
- 28 October 2025
- Published by Springer Nature in European Journal of Medical Research
- Vol. 30 (1) , 1-11
Preliminary evaluation of mixed reality navigation with single-vertebra registration in spinal surgery
To develop a mixed reality navigation system based on single-vertebra registration and to preliminarily validate its feasibility in clinical surgery, providing a potential reference for spinal surgical navigation registration methods. We developed a portable, stand-alone, optical-based mixed reality navigation system for spinal surgery using a head-mounted display. The system integrates registration and real-time localization functions, enabling single-vertebra registration via anatomical landmarks and interactive visualization through the HoloLens 2 platform. System performance was evaluated in both laboratory and clinical settings by measuring localization time, fiducial registration error (FRE), and target registration error (TRE). Preoperative assessments were conducted on patient-specific 3D-printed vertebral models, and intraoperative evaluations were performed on exposed vertebrae during surgery. Additional trials with novice surgeons were performed to assess the learning curve. For selected cases requiring pedicle screw placement, surgeons qualitatively evaluated the intraoperative visualization of preplanned screw trajectories in terms of anatomical congruence, feasibility, and display clarity. A total of 37 spinal surgery patients were included, covering cervical, thoracic, and lumbar segments. The average preoperative navigation time based on 3D-printed models was 2.96 ± 0.26 min, compared with 3.11 ± 0.25 min intraoperatively (p = 0.002). The FRE was 1.37 ± 0.54 mm preoperatively and 1.42 ± 0.52 mm intraoperatively (p = 0.178). The TRE was 1.52 ± 0.56 mm preoperatively and 1.75 ± 0.61 mm intraoperatively (p = 0.001). Among four pedicle screw placement cases, three were rated satisfactory intraoperatively, and one showed an approximately 5 mm deviation due to minor patient positional changes. The learning curve indicated that procedure time stabilized at approximately 2.7 min after about 15 practice cases. The single-vertebra registration-based mixed reality navigation system enables rapid and accurate vertebral localization in both preoperative and intraoperative settings, with TREs within clinically acceptable limits. The system is portable, easy to operate, and provides an intuitive, immersive navigation interface without relying on external tracking devices. These features suggest its potential as a cost-effective and versatile tool for spinal surgery navigation.
- 28 October 2025
- Published by Springer Nature in European Journal of Medical Research
- Vol. 30 (1) , 1-11
Open AccessPreliminary evaluation of mixed reality navigation with single-vertebra registration in spinal surgery
To develop a mixed reality navigation system based on single-vertebra registration and to preliminarily validate its feasibility in clinical surgery, providing a potential reference for spinal surgical navigation registration methods. We developed a portable, stand-alone, optical-based mixed reality navigation system for spinal surgery using a head-mounted display. The system integrates registration and real-time localization functions, enabling single-vertebra registration via anatomical landmarks and interactive visualization through the HoloLens 2 platform. System performance was evaluated in both laboratory and clinical settings by measuring localization time, fiducial registration error (FRE), and target registration error (TRE). Preoperative assessments were conducted on patient-specific 3D-printed vertebral models, and intraoperative evaluations were performed on exposed vertebrae during surgery. Additional trials with novice surgeons were performed to assess the learning curve. For selected cases requiring pedicle screw placement, surgeons qualitatively evaluated the intraoperative visualization of preplanned screw trajectories in terms of anatomical congruence, feasibility, and display clarity. A total of 37 spinal surgery patients were included, covering cervical, thoracic, and lumbar segments. The average preoperative navigation time based on 3D-printed models was 2.96 ± 0.26 min, compared with 3.11 ± 0.25 min intraoperatively (p = 0.002). The FRE was 1.37 ± 0.54 mm preoperatively and 1.42 ± 0.52 mm intraoperatively (p = 0.178). The TRE was 1.52 ± 0.56 mm preoperatively and 1.75 ± 0.61 mm intraoperatively (p = 0.001). Among four pedicle screw placement cases, three were rated satisfactory intraoperatively, and one showed an approximately 5 mm deviation due to minor patient positional changes. The learning curve indicated that procedure time stabilized at approximately 2.7 min after about 15 practice cases. The single-vertebra registration-based mixed reality navigation system enables rapid and accurate vertebral localization in both preoperative and intraoperative settings, with TREs within clinically acceptable limits. The system is portable, easy to operate, and provides an intuitive, immersive navigation interface without relying on external tracking devices. These features suggest its potential as a cost-effective and versatile tool for spinal surgery navigation.
- Open Access
- 16 October 2025
- Published by Elsevier in Journal of Cranio-Maxillofacial Surgery
- Vol. 53 (12) , 2230-2237
A portable mixed reality navigation system for oral and maxillofacial surgery: Design and preliminary clinical evaluation
- 16 October 2025
- Published by Elsevier in Journal of Cranio-Maxillofacial Surgery
- Vol. 53 (12) , 2230-2237
Open AccessA portable mixed reality navigation system for oral and maxillofacial surgery: Design and preliminary clinical evaluation
- 6 September 2025
- Published by Springer Nature in Multimedia Systems
- Vol. 31 (5) , 1-14
Cross-teaching with dual uncertainty awareness for semi-supervised medical image segmentation
- 6 September 2025
- Published by Springer Nature in Multimedia Systems
- Vol. 31 (5) , 1-14
Cross-teaching with dual uncertainty awareness for semi-supervised medical image segmentation