Zhouqianwei

2024.08–2026.08. Postdoctoral Associate, Department of Anesthesiology, Weill Cornell Medicine, New York, USA. Mentor Jyun-you Liou. Synchronized recording and analysis of multimodal neural signals.

2022.12–present. Associate Professor, Institute of Computer Vision, College of Computer Science and Technology, Zhejiang University of Technology. Doctoral supervisor (academic degree programs) since 2023.

2018.08–2019.08. Visiting Scholar, Imaging Research Division, Department of Radiology, University of Pittsburgh, USA. Mentor Shandong Wu. Intelligent understanding of medical images.

2014.07–2022.12. Lecturer, Institute of Computer Vision, College of Computer Science and Technology, Zhejiang University of Technology.

2013.08–2013.10. IBM Research China, Shanghai. Internet of Vehicles project group, big-data algorithms for connected vehicles.

2009.09–2014.07. Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences. Combined master's and doctoral program in Communication and Information Systems, Ph.D. in Engineering conferred by the University of Chinese Academy of Sciences. Advisor Xiaobing Yuan.

2005.09–2009.07. Hangzhou Dianzi University. B.Eng. in Communication Engineering.

Qianwei Zhou, Ph.D., is an Associate Professor and doctoral supervisor in the College of Computer Science and Technology at Zhejiang University of Technology (ZJUT). Zhou studies robust representations of multimodal neural and medical signals. The current focus is robust feature extraction from the electroencephalogram (EEG) under multimodal constraints.

The research is organized around one question. Why does a model fail when the subject, the recording device, or the recording session changes, and how can that failure be prevented? Observations of complex systems commonly carry three defects. Acquisition cannot be fully controlled, so sensor damage, occlusion, and disconnection produce missing data. The structure of the system is highly dynamic, so the regions observed by different modalities do not line up in space, which produces channel misalignment. The state of the system fluctuates strongly over time, so data from different subjects and sessions differ in their statistics, which produces distribution shift. Existing methods compensate for these defects with manual preprocessing, namely manual rejection of bad samples, manual registration, and manual normalization. However, the preprocessing parameters fit only the data they were tuned on, so a system stops being reliable as soon as it leaves its original setting. Zhou's position is that manual preprocessing limits robustness. The handling of all three defects should instead be built into the model itself, so that the space in which the model represents its input, referred to below as the latent space, does not change with acquisition conditions. Zhou's early work used sensing signals such as acoustic arrays, seismic vibration, and geomagnetic fields. The work then moved to medical imaging, and in recent years it has concentrated on multiscale neural signals, including EEG, electrocorticography (ECoG), and neural calcium imaging.

Four representative results answer four successive questions along one chain. The first asks why the latent space is not robust. The second asks how a network should be designed to extract a robust latent space within one dataset. The third asks how the latent space can stay stable when the data source and the acquisition time change. The fourth asks how it can stay stable when the subject changes. The cause of a failure must be understood before a design can target it, and features must be stable within one data domain before stability across domains becomes meaningful. These two dependencies fix the order of the four results.

Result 1. Why the latent space is not robust. A reader study that Zhou published as first author in Nature Communications (2021) gave the answer, which is that deep representations lack structural constraints. The study used a generative adversarial network (GAN) to modify only the lesion region of breast images, keeping all other tissue identical pixel by pixel, and then compared how artificial intelligence (AI) models and radiologists read these images. If a model truly judged a lesion by its structural relation to the surrounding tissue, its output should change when the lesion appears at an anatomically impossible location. The experiment showed the opposite. The participating radiologists pointed out that some synthetic images placed lesions in regions where no breast tissue should exist, yet the diagnostic models scored these images just as they scored real ones. It follows that the models' latent space does not encode anatomical structure. The work also established a reusable evaluation principle. A generative model constructs semantically controlled counterfactual samples, and the model's responses to those samples measure which information its latent space actually relies on. The work was presented orally at the 2019 Annual Meeting of the Radiological Society of North America (RSNA), and its follow-up direction is funded by a General Program grant from the National Natural Science Foundation of China (NSFC).

Result 2. How to extract a robust latent space within one dataset. Because the failure comes from missing structural constraints, robustness should be built into the network design, so that the network itself learns features that withstand perturbation. In a feed-forward feature extractor, the designer controls only three components, namely the architecture, the activation function, and the loss function. The representative result is a layer-wise adaptive activation function method, published by Zhou as corresponding author in IEEE Transactions on Neural Networks and Learning Systems (2023). It determines each layer's activation function automatically from the network structure and the target task, instead of applying one hand-picked function throughout the network. For the other two components, Zhou proposed, as first author, the redundant and missing feature decoupling network RMFDNet (Engineering Applications of Artificial Intelligence, 2025) and a non-binary intersection-over-union (IoU) loss (Expert Systems with Applications, 2023). Channel misalignment mixes information from different locations into one feature channel and creates redundant components, whereas missing channels leave expected information empty and create missing components. The two have opposite causes, so RMFDNet separates them explicitly and treats each one on its own. The non-binary IoU loss gives correct gradients in regions with continuous values, such as object edges. This line of work goes back to Zhou's early robust feature extraction for seismic signals (IEEE Signal Processing Letters, 2012). At that time, features that resist noise and deformation were designed by hand. Now the rules that let a network learn such features are designed by hand.

Result 3. How the latent space stays stable across data sources and acquisition times. The first two results stabilize the latent space within one dataset. The real test is whether that stability survives a change of data source and acquisition time. Zhou is a co-corresponding author, and the only corresponding author from computer science, of a multicenter study in eClinicalMedicine (published by The Lancet, 2025). The study predicts the response of breast cancer to neoadjuvant chemotherapy early and noninvasively. A Siamese network extracts dynamic contrast-enhanced magnetic resonance imaging (MRI) features at two time points, before chemotherapy and early in chemotherapy, and a Transformer-based multi-head attention module models the spatiotemporal interaction between the two time points. The study enrolled 1044 patients from 5 medical centers. On 3 external validation cohorts, the area under the receiver operating characteristic curve (AUC) was 0.892–0.923. In two public cohorts independent of the training data (I-SPY1 and I-SPY2), the cases that the model predicted as good responders showed upregulated immune-related genes and stronger immune cell infiltration. This indicates that the latent space of macroscopic images captured microscopic biological processes related to treatment response, rather than fitting one particular dataset.

Result 4. How the latent space stays stable across subjects. EEG signals differ markedly between individuals, and the failure to generalize across subjects is a bottleneck that keeps brain-computer interfaces (BCIs) from practical use. As corresponding author, Zhou proposed a masked self-supervised contrastive learning framework for EEG motor imagery at the 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Its three components correspond one to one with the three data defects. Manifold-aware masking simulates missing data, domain alignment handles channel misalignment, and multi-view spatiotemporal attention handles distribution shift. With scarce labels, the framework still generalized across subjects better than the supervised baseline.

From the four results to the current direction. Result 3 checked the latent space of macroscopic images against microscopic molecular and cellular evidence after training, and the two agreed. From August 2024 to August 2026, Zhou was a postdoctoral researcher at Weill Cornell Medicine and took part in synchronized recordings of mouse cortex with widefield calcium imaging, two-photon calcium imaging, and transparent ECoG electrode arrays. As co-first author, Zhou completed a cross-scale cortical electrophysiology study (bioRxiv preprint) in which the synchronized recordings directly showed a correspondence between microscopic population neural activity and macroscopic field potentials. Because this correspondence exists, there is no need to wait until training ends to check it. Instead, microscopic neural activity can constrain the EEG latent space directly during training. This is Zhou's current research direction, robust EEG feature extraction under multimodal constraints. In practical BCI use only EEG is available, so the cross-scale constraint acts as a regularizer during training only, and inference needs no microscopic recordings.

Zhou has published or had accepted more than 80 papers, with 2076 Google Scholar citations, an h-index of 20, and an i10-index of 36. Zhou has led one NSFC General Program grant and one NSFC Young Scientists Fund grant, and holds 20 granted Chinese invention patents as first inventor, 4 of which have been transferred to industry. During doctoral study, Zhou received the National Scholarship for Doctoral Students, the President's Excellence Award of the Chinese Academy of Sciences, and the institute director's first-class scholarship. In 2023, Zhou was selected as a Category D talent in the ZJUT 14th Five-Year High-Level Talent Development Program.

Selected Publications

Listed in reverse chronological order. * marks a corresponding author and # marks a co-first author.

1. Wu S#, Zhou Q#, Ryu J, Li G, Iyer A, Gill B, Ma H, Fang H, Schevon CA*, Schwartz TH*, Liou JY*. Suppressing cortical glutamatergic neurons produces paradoxical interictal discharges and seizures. bioRxiv, 2025. DOI: 10.1101/2025.10.02.676593.

2. Tang W, Jin C, Kong Q, Liu C, Chen S, Ding S, Liu B, Feng Z, Li Y, Dai Y, Zhang L, Chen Y, Han X, Liu S, Chen D, Weng Z, Liu W, Wei X, Jiang X, Zhou Q*, Mao N*, Guo Y*. Development and validation of an MRI spatiotemporal interaction model for early noninvasive prediction of neoadjuvant chemotherapy response in breast cancer: a multicentre study. eClinicalMedicine, 2025, 85: 103298.

3. Zhou Q, Wang J, Li J, Zhou C, Hu H, Hu K*. RMFDNet: Redundant and missing feature decoupling network for salient object detection. Engineering Applications of Artificial Intelligence, 2025, 139: 109459.

4. Zhang K, Zhou Q*, Hu H. A novel self-supervised contrastive learning framework for masked EEG motor imagery modeling. 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Hyderabad, India, 2025: 1-5. DOI: 10.1109/ICASSP49660.2025.10888531.

5. Zhou Q, Zhou C, Yang Z, Xu Y, Guan Q*. Non-binary IoU and progressive coupling and refining network for salient object detection. Expert Systems with Applications, 2023, 230: 120370.

6. Hu H, Liu A, Guan Q, Qian H, Li X, Chen S, Zhou Q*. Adaptively customizing activation functions for various layers. IEEE Transactions on Neural Networks and Learning Systems, 2023, 34(9): 6096-6107.

7. Zhou Q, Li B, Tao P, Xu Z, Zhou C, Wu Y, Hu H*. Residual-recursive autoencoder for accelerated evolution in Savonius wind turbines optimization. Neurocomputing, 2022, 500: 909-920.

8. Hu H, Shen L, Guan Q, Li X, Zhou Q*, Ruan S*. Deep co-supervision and attention fusion strategy for automatic COVID-19 lung infection segmentation on CT images. Pattern Recognition, 2022, 124: 108452.

9. Hu K*, Zhao L, Feng S, Zhang S, Zhou Q*, Gao X, Guo Y. Colorectal polyp region extraction using saliency detection network with neutrosophic enhancement. Computers in Biology and Medicine, 2022, 147: 105760. (Highly Cited Paper, Essential Science Indicators)

10. Guan Q, Chen Y, Wei Z, Heidari AA, Hu H, Yang XH, Zheng J, Zhou Q*, Chen H*, Chen F*. Medical image augmentation for lesion detection using a texture-constrained multichannel progressive GAN. Computers in Biology and Medicine, 2022, 145: 105444.

11. Zhou Q, Zuley M, Guo Y, Yang L, Nair B, Vargo A, Ghannam S, Arefan D, Wu S*. A machine and human reader study on AI diagnosis model safety under attacks of adversarial images. Nature Communications, 2021, 12(1): 7281.

12. Zhou Q, Liu Y, Hu H, Guan Q, Guo Y, Zhang F*. Unsupervised multimodal MR images synthesizer using knowledge from higher dimension. 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2021: 1633-1636.

13. Zhou Q, Chen Y, Li B, Li X, Zhou C, Huang J, Hu H*. Training deep neural networks for wireless sensor networks using loosely and weakly labeled images. Neurocomputing, 2021, 427: 64-73.

14. Zhou Q, Yuan G, Yang L, Nebbia G, Wu S*. Are your AI diagnosis models safe under attack of manipulated images? Radiological Society of North America 2019 Scientific Assembly and Annual Meeting (RSNA 2019), Chicago, IL, 2019. Oral presentation.

15. Zhou Q, Xu Z, Cheng S, Huang Y, Xiao J. Innovative Savonius rotors evolved by genetic algorithm based on 2D-DCT encoding. Soft Computing, 2018, 22(23): 8001-8010.

16. Zhou Q, Li B, Liu H, Chen S, Huang J. Microphone-based vibration sensor for UGS applications. IEEE Transactions on Industrial Electronics, 2017, 64(8): 6565-6572.

17. Zhou Q, Li B, Kuang Z, Xie D, Tong G, Hu L, Yuan X. A quarter-car vehicle model based feature for wheeled and tracked vehicles classification. Journal of Sound and Vibration, 2013, 332(26): 7279-7289.

18. Zhou Q, Tong G, Xie D, Li B, Yuan X. A seismic-based feature extraction algorithm for robust ground target classification. IEEE Signal Processing Letters, 2012, 19(10): 639-642.

19. Zhou Q, Tong G, Li B, Yuan X. A practicable method for ferromagnetic object moving direction identification. IEEE Transactions on Magnetics, 2012, 48(8): 2340-2345.

Granted Invention Patents

Zhou holds 20 granted Chinese invention patents as first inventor, listed below in reverse order of grant year. The English titles are translations of the Chinese originals.

1. Anatomical plausibility detection method and system based on self-clustering graph convolution. ZL202211495980.8, 2025.

2. Texture continuity detection method based on a self-optimizing deep image prior. ZL202210919570.5, 2025.

3. Ambiguity elimination method for unsupervised semantic segmentation based on adaptive clustering of hard points. ZL202411794144.9, 2025.

4. Causal plausibility detection method for medical images based on dual-channel conditional fusion. ZL202211175481.0, 2025.

5. Method and device for building a game-theoretic decision model for forced merging at highway on-ramps. ZL202411218461.6, 2025.

6. Vulnerability analysis method and device for anomalous load signals in electric vehicle powertrains. ZL202410306885.1, 2025.

7. Unsupervised multimodal image translation method based on generative adversarial networks. ZL202110333549.2, 2024.

8. Unpaired image style transfer method based on generative adversarial networks. ZL202011391478.3, 2024.

9. Clustering method based on three-dimensional vehicle contours. ZL202010564925.4, 2024.

10. Cross-dimensional method for transferring knowledge from high-dimensional to low-dimensional deep learning models. ZL202011487738.7, 2024.

11. Differential feature fusion method for acoustic signals based on deep learning and random arrays. ZL202011483184.3, 2024.

12. Spatial filtering method and system for acoustic arrays based on virtual subarray interleaving. ZL202011430717.1, 2023.

13. Automatic fine segmentation method for images. ZL201910950415.8, 2022.

14. Similarity enhancement method for small neighborhoods on a manifold. ZL201810446278.X, 2022.

15. Sparse grayscale image encoding and decoding method and system based on reconstruction residuals. ZL202010041098.0, 2022.

16. Automatic image dataset construction method based on deep neural networks. ZL201910655806.7, 2021.

17. Training method for image codecs based on deep neural networks. ZL201810446279.4, 2021.

18. Multiscale adaptive near-lossless encoding and decoding method and system. ZL201810293916.9, 2020.

19. Method for extracting binary shape contours in the clockwise direction. ZL201711120452.3, 2020.

20. Automatic floating debris cleaning device. ZL201610241403.4, 2018.

Zhou has led 7 projects, funded by the NSFC, key laboratory open programs, and industry, and has participated in 12 national or provincial projects.

As Principal Investigator

1. Plausibility Assessment of the Content of Computer-Synthesized Medical Images

NSFC General Program (62271448), 2023.01–2026.12,  ongoing.

Computer-synthesized medical images may contain content that violates human anatomy or the natural development of tissue, and AI diagnostic models cannot detect such content. This project develops plausibility assessment methods that need no manual annotation. The methods judge whether a synthetic image can be trusted at three levels, namely texture continuity, anatomical plausibility, and causal plausibility between two successive scans. Preliminary work was published as the Nature Communications paper and the RSNA 2019 oral presentation. The project has produced 3 granted invention patents, one each for the detection of texture continuity, anatomical plausibility, and causal plausibility.

2. Constraint Methods for the Parameterization of Functional Shape Cross-Sections

NSFC Young Scientists Fund (61802347), 2019.01–2021.12, completed.

An image autoencoder compresses wind turbine blade profiles into a low-dimensional latent space, and an evolutionary algorithm searches that space for high-performance profiles, which removes the need for repeated modeling and testing. Results were published in Neurocomputing (2022) and produced 5 granted invention patents on contour extraction and image encoding and decoding.

3. Simulation and Deep Learning Algorithms for Ultra-Dense Micro Acoustic Arrays

Key laboratory open project, started 2020, completed.

This project built simulation software and deep learning algorithms for ultra-dense micro acoustic arrays, with the goal of high-precision direction finding and localization of low-frequency wideband acoustic targets. Zhou designed the software architecture, the hardware platform, and the algorithmic principles. Results were published in Measurement Science and Technology (2025) and produced 2 granted invention patents.

4. Hybrid Target Recognition Algorithms Based on Compact Deep Networks

Key laboratory open project, started 2016, completed.

This project trained and compressed deep neural networks with loosely and weakly labeled images downloaded free from the internet, so that the networks run on wireless sensor network nodes at lower annotation and computation cost. Results were published in Neurocomputing (2021) and produced 1 granted invention patent.

5. Cloud Service Software for Biometric Recognition Terminals

Industry-funded project, started 2025, ongoing.

6. Design of a Seismic-Acoustic Energy Converter (started 2014) and Improvement of the Seismic-Acoustic Energy Converter (started 2017)

Two industry-funded projects, both completed.

Both projects exploit the conversion between vibration and sound. They produced a new vibration sensor built around a microphone that uses the device battery as its proof mass, so the sensor is lighter and smaller than a conventional moving-coil geophone and allows smaller wireless sensor network nodes for security applications. Zhou verified the physical principle and completed the structural design, simulation, circuit design, algorithm development, and field experiments. Results were published in IEEE Transactions on Industrial Electronics (2017).

As Participant (Provincial Level and Above)

1. Research and Demonstration of One Sensor, Multiple Recognitions Monitoring and Early-Warning Equipment for Grassroots Community Governance. National Key R&D Program of China (sub-project), PI Xiaoqin Zhang, started 2025, ongoing.

2. Domain-Adaptive Recognition of Dermoscopic Signs and Intelligent Diagnosis of Hair Loss Disorders. NSFC General Program, PI Haigen Hu, started 2023, ongoing.

3. Traffic Situation Assessment by Spatiotemporal Feature Fusion with Attention and Multi-Structure Graph Convolution. NSFC General Program, PI Guojiang Shen, started 2020, completed.

4. Locating Sensitive Units of Gate-Level Circuits within a Memetic Framework. NSFC General Program, PI Jie Xiao, started 2019, completed.

5. Three-Dimensional Ultrasound Assessment of Sagittal Characteristics in Adolescent Idiopathic Scoliosis. NSFC Young Scientists Fund, PI Weiwei Jiang, started 2017, completed.

6. High-Precision Fast Reliability Evaluation of Generalized Gate Circuits Based on an Extended Probabilistic Transfer Matrix Model. NSFC Young Scientists Fund, PI Jie Xiao, started 2016, completed.

7. Diagnosis and New Technologies for Musculoskeletal Diseases, sub-project on a Swimmer's Shoulder Early-Warning System Based on Motion Capture and Machine Learning. Pioneer and Leading Goose R&D Program of Zhejiang Province, PI Haigen Hu, started 2025, ongoing.

8. Key Technologies for Next-Generation Intelligent Medical Big Data and Multimodal Knowledge Fusion. Pioneer and Leading Goose R&D Program of Zhejiang Province, PI Qiu Guan, started 2024, ongoing.

9. Modern Traditional Chinese Medicine Equipment, sub-project on an Intelligent Robot Based on Tuina Theory and Simulation of Expert Techniques. Pioneer and Leading Goose R&D Program of Zhejiang Province, PI Wanliang Wang, started 2021, ongoing.

10. Interpretable Depression Prediction Based on Multimodal Edge-Centric Brain Networks. Zhejiang Provincial Natural Science Foundation, Exploration Project, PI Haixia Long, started 2022, completed.

11. Prediction of Populations at High Risk of Depression from Multimodal Data. Zhejiang Provincial Natural Science Foundation, Exploration Project (Youth), PI Haixia Long, started 2017, completed.

12. Reliability Computation of Transistor-Level Generalized Gate Circuits Based on Probabilistic Transfer Matrices. Zhejiang Provincial Natural Science Foundation, Exploration Project (Youth), PI Jie Xiao, started 2015, completed.

Zhou has taught undergraduate and graduate courses at ZJUT since 2014.

Undergraduate courses

• Survey of Artificial Intelligence. Elective for economics, management, and law majors, 2 credits, 32 class hours, 2 sections in the first semester of 2026–2027.

• Introduction to Artificial Intelligence

• Java Programming, for software engineering majors and the experimental class

• Java Programming Course Project

• Fundamentals of Programming (Python)

Graduate course

• Principles and Applications of Artificial Intelligence. Bilingual core course for professional degree programs, co-taught with Qiu Guan.

Zhou also supervises undergraduate graduation projects and serves as an undergraduate academic advisor and class advisor.

Graduate Supervision

Zhou has been a doctoral supervisor for academic degree programs since 2023. As of August 2026, the group had 2 doctoral students and 5 master's students, and at least 9 master's students had graduated under Zhou's supervision. 

Student Competitions and Innovation Projects (Zhou as Faculty Advisor)

1. Third Prize, World Robot Contest Championships 2024, BCI Brain-Controlled Robot Contest, technical track on motor rehabilitation training based on non-invasive BCIs, August 2024. Students Kunkun Zhang, Kehao Wu, and Zhitao Chen.

2. National Second Prize (Problem D), 21st China Post-Graduate Mathematical Contest in Modeling, 2024. Students Kunkun Zhang, Kehao Wu, and Chen Jin.

3. Third Prize, Enterprise-Proposed Topics (Category A), Eastern Regional Contest of the 12th China College Students' Service Outsourcing Innovation and Entrepreneurship Competition, June 2021. Students Chen Zhou, Jun Tao, Yanzhuang Wu, and Yibo Liu.

4. National College Students' Innovation and Entrepreneurship Training Program project Mechanical Arm Calibration System Based on Deep Learning and Ultrasonic Micro Arrays (202210337034), started 2022. Student Hanmo Chen. The project produced 1 granted invention patent (ZL202310596248.8, 2026).

5. Student projects funded by the Zhejiang Provincial College Students' Science and Technology Innovation Program (Xinmiao Talent Program) and by the ZJUT Yunhe Cup Extracurricular Academic Science and Technology Fund.

Mentoring Abroad

During the postdoctoral appointment in the United States, Zhou supervised an undergraduate from Rutgers University, New Brunswick, on brain surface and vessel segmentation from widefield calcium imaging (since April 2025). Zhou also supervised a New York high school student on two-photon calcium imaging analysis and neuronal population modeling (since May 2025).

Academic Appointments

• Youth Member, Youth Committee of the Third Radiology Professional Committee, Zhejiang Society of Mathematical Medicine (since 2023)

Memberships

• IEEE Member (joined 2017)

• China Computer Federation (CCF) Member (joined 2017)

• Member, Zhejiang Association for Artificial Intelligence

Journal Reviewing

Zhou serves as a reviewer for IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on Medical Imaging, IEEE Journal of Biomedical and Health Informatics, IEEE/ACM Transactions on Computational Biology and Bioinformatics, IEEE Internet of Things Journal, IEEE Signal Processing Letters, Pattern Recognition, Neurocomputing, Expert Systems with Applications, Engineering Applications of Artificial Intelligence, Applied Soft Computing, Soft Computing, Energy, Journal of Magnetic Resonance Imaging, and Journal of Sound and Vibration.

Review Panels

• Expert reviewer, Ministry of Education random inspection of undergraduate theses (2023)

• Expert reviewer, project acceptance for the Zhejiang Provincial Natural Science Foundation (2024)

• Expert reviewer, interim acceptance of the Zhejiang Lab project Research on Brain-Inspired Intelligence Guided by the Brainnetome Atlas

Invited Talks

• Invited talk, VAPORS Seminar, Weill Cornell Medicine (2025)