Haichuan Zhang
Haichuan Zhang
Computer Vision & Imaging Scientist
Computer Vision & Imaging Scientist
Ph.D., Electrical Engineering · San Jose, California
Resume · Email · Google Scholar · LinkedIn · GitHub
Embedded Files
Biography
Biography
I am a computer vision and imaging scientist with a Ph.D. in Electrical Engineering from Penn State. I combine imaging physics, optimization, and deep learning to build interpretable, efficient methods for image restoration and sensor-data recovery. My doctoral advisor was Prof. Vishal Monga at the Information Processing and Algorithms Laboratory (iPAL).
My industry experience spans event-based tracking and on-chip vision at OMNIVISION, event-guided motion deblurring at Dolby Laboratories, and diffusion-based fiber-sensor restoration at NEC Laboratories America. At Penn State, my research includes convergent unrolled networks, phase-informed deblurring, ultrasound phase correction, and NTIRE dehazing and bokeh challenge solutions. I work across mathematical modeling, model implementation, and experimental evaluation under practical sensing and compute constraints.
Earlier at Xi’an Jiaotong University, I developed computational pathology software and renal cancer analysis models, including a whole-slide image annotation framework used in the AIPath Dataset. This work led to a U.S. patent on collaborative pathology annotation, granted in 2022.
Expertise
Image restoration · Computational imaging · Event-based vision · Optimization & algorithm unrolling · Generative models · Medical image analysis
Core tools: Python · PyTorch · C++ · OpenCV · MATLAB
Education
Education
Training in electrical engineering, computer science, and research software development.
Penn State University
Penn State University
Ph.D. in Electrical Engineering
Aug 2021 – May 2026 · GPA 3.90/4.00
Advisor: Prof. Vishal Monga · Information Processing and Algorithms Laboratory (iPAL).
Dissertation: Model-Based Deep Learning for Imaging Inverse Problems.
Doctoral research: computational imaging and image restoration using physical models, optimization, and interpretable unrolled networks with convergence analysis.
Recipient of the Milton and Albertha Langdon Memorial Graduate Fellowship (June 2024).
Xi’an Jiaotong University
Xi’an Jiaotong University
B.Eng. in Computer Science and Technology
Honor Science Program
Aug 2016 – Jun 2020 · GPA 3.60/4.00
Contributed to OpenHI, an open-source platform for collaborative annotation of whole-slide pathology images.
Honors include the Siyuan Scholarship for Academic Excellence, the 2018 VEX U Robot Skills world championship, and the VEX U Excellence Award.
Duke University
Duke University
Summer Session · Jul–Aug 2018 · GPA 4.00/4.00
Work Experience
Work Experience
Image restoration, sensor algorithms, and research software across industry and academia.
Senior Image Scientist · OMNIVISION
Senior Image Scientist · OMNIVISION
Jun–Aug 2026 · Santa Clara, California, United States
Designed event-based tracking within a region of interest (ROI), improving target consistency and robustness to nearby objects and distractors.
Developed on-chip camera algorithms under strict power, memory, and compute constraints.
Research Intern · Dolby Laboratories
Research Intern · Dolby Laboratories
May–Aug 2025 · Sunnyvale, California, United States
Developed EUMD, an interpretable event-guided deblurring framework. Improved PSNR by 0.31 dB over ClearSight on synthetic GoPro and 1.04 dB over CMTA on real EVRB.
Combined long-exposure images with asynchronous event streams through a physical blur model and unrolled optimization, supporting sharp-image recovery and short high-speed frame-sequence reconstruction.
Published the work at WACV Workshops 2026 and co-invented technology described in a U.S. patent application.
Research Intern · NEC Laboratories America, Inc.
Research Intern · NEC Laboratories America, Inc.
Jun–Aug 2024 · Princeton, New Jersey, United States
Built a conditional diffusion model to restore low-quality fiber-sensor measurements while preserving task-relevant signal structure.
Enabled classification models trained on high-quality measurements to use restored inputs without retraining or redesign.
Co-invented technology described in a U.S. patent application.
Documented data preparation, model design, evaluation, and deployment considerations in an internal technical report.
Graduate Research Assistant · Penn State University
Graduate Research Assistant · Penn State University
Aug 2021 – May 2026 · State College, Pennsylvania, United States
Developed interpretable networks for non-blind image restoration using half-quadratic splitting and structured parameter learning, with theoretical and empirical convergence analysis (ICIP 2023; IEEE TCI 2024).
Co-authored phase-informed blind deblurring research that jointly recovers image and blur information in an unrolled network (ECCV 2026).
Contributed architecture design, training, ablations, and report preparation to NTIRE challenge solutions. Team IPAL-Bokeh achieved the lowest real-image LPIPS (0.2161) in the 2023 lens-to-lens bokeh challenge.
Team iPAL-LightDehaze ranked 4th in LPIPS among 17 NTIRE 2023 dehazing finalists using TransER: 2.6M parameters and no extra training data.
Team iPAL-GridFFA ranked 2nd in SSIM (0.839) among 23 NTIRE 2021 finalists and tied for the lowest reported VGG-based LPIPS (0.194).
Developed methods, software, data preparation, and validation for CT-informed ultrasound phase correction (Neuroelectronics 2025); contributed to MRI-based brain-growth analysis (Journal of Neurosurgery: Pediatrics 2026).
Scientific Software Engineer · Xi’an Jiaotong University
Scientific Software Engineer · Xi’an Jiaotong University
Jul 2020 – Jun 2021 · Xi’an, Shaanxi, China
Built cloud-based pathology annotation software with AI-assisted labeling for consistent, large-scale collaboration on whole-slide images.
Developed machine-learning models for renal cell carcinoma analysis, including nuclei analysis and whole-slide image assessment.
Contributed to OpenHI/OpenHI2 and related pathology research; co-invented collaborative annotation technology granted as U.S. patent US11392634B2.
Software Engineer Intern · Xi’an Suoer Software Technology Co., Ltd.
Software Engineer Intern · Xi’an Suoer Software Technology Co., Ltd.
Jun–Jul 2019 · Xi’an, Shaanxi, China
Led a five-person team to deliver an e-commerce web application from requirements through deployment.
Coordinated implementation and testing across the team.
Received the Outstanding Production Internship Award in July 2019.
Research
Research
Interpretable imaging algorithms, medical image analysis, and applied machine learning.
Selected Research
Selected Research
Peer-reviewed work in image restoration, computational imaging, medical image analysis, and robust learning.
Phase-informed Unrolled Deblurring — UPADNet
Phase-informed Unrolled Deblurring — UPADNet
ECCV 2026 · Penn State
Recovers sharp images and unknown blur kernels by learning Fourier phase and amplitude updates; improves deblurring under noise and limited training data.
Problem and method. Recovering a sharp image from unknown blur requires estimating both the image and its blur kernel. In this co-authored work, UPADNet separates Fourier phase and amplitude, derives linear minimum mean-square error (LMMSE) estimators, and learns the updates of an alternating optimization algorithm. Each stage refines image and kernel estimates, with generated weights and multiple scales adapting the reconstruction.
Results. The 27M-parameter model reaches PSNR/SSIM of 34.63 dB/0.975 on GoPro, 41.29 dB/0.979 on RealBlur-R, and 34.35 dB/0.946 on RealBlur-J. Controlled COCO experiments show stronger restoration under added noise and reduced training data; using 60% of the training set, it reaches 32.12 dB/0.926.
Architecture, restoration examples, and robustness experiments.
Event-guided Motion Deblurring — EUMD
Event-guided Motion Deblurring — EUMD
WACV Workshops 2026 · Dolby Laboratories / Penn State
Recovers sharp images from motion blur by combining event-camera measurements with physical blur modeling and learned image priors.
Problem and method. Long exposures merge fast motion into a blurred frame; event cameras retain the brightness changes within the exposure. I developed EUMD to connect these measurements to a physical blur operator, then unroll half-quadratic splitting into eight reconstruction stages with a Transformer-enhanced U-Net prior. Training-time timestamp correction aligns sharp labels with the event-derived reconstruction; evaluation uses the exposure midpoint.
Results. EUMD reaches 37.35 dB/0.9801 on synthetic GoPro and 32.42 dB/0.937 on real EVRB. PSNR improves by 0.31 dB over ClearSight on GoPro and 1.04 dB over CMTA on EVRB. The Transformer and training timestamp correction contribute 0.41 dB and 1.02 dB, respectively, in the paper’s ablations.
Event-based reconstruction architecture and visual comparisons.
Brain MRI Segmentation & Growth Assessment
Brain MRI Segmentation & Growth Assessment
Journal of Neurosurgery: Pediatrics 2026 · Penn State
Uses the unaffected brain hemisphere to assess growth in pediatric hydrocephalus when shunt artifacts prevent reliable whole-brain MRI measurements.
Problem and method. Metallic shunt hardware can obscure part of the brain in MRI. This co-authored study measures the artifact-free hemisphere as a proxy for whole-brain growth. A Dense U-Net segments brain tissue and cerebrospinal fluid (CSF), a second model identifies the hemibrain, and voxel dimensions convert the masks into volumes. The training loss combines Dice overlap with CSF-focused total-variation regularization.
Results and scope. MRI from 75 ESTHI trial patients supports the hemisphere-based approach: postoperative hemisphere ratios stabilize, and hemibrain and whole-brain growth show similar patterns where both are measurable. Brain and CSF segmentation achieve Dice scores of 94.7% and 94.2%. My contributions include data interpretation and statistical analysis. Some scans require manual refinement; validation in larger, more diverse cohorts remains future work.
Volumetry, segmentation, and brain-growth analysis.
Skull-induced Phase Aberration Correction — DB-SIPAC
Skull-induced Phase Aberration Correction — DB-SIPAC
Neuroelectronics 2025 · Penn State
Predicts CT-informed phase corrections in 0.45 seconds to improve transcranial ultrasound focusing in acoustic simulations.
Problem and method. The skull distorts focused ultrasound, while simulation-based correction is slow. DB-SIPAC combines a direct-propagation pathway branch with a full-skull branch that captures broader reflections and refractions. Iterative Time Delay Search (ITDS) refines the training targets. My contributions included methodology, software, data preparation, and validation.
Results and scope. The dataset contains 1,260 skull/focal-point pairs from 36 CT-derived slices of 12 adults. In 2D k-Wave simulations, DB-SIPAC predicts delays in 0.45 seconds, versus 32 seconds for time reversal and 249 seconds for hybrid angular spectrum, and reaches mean normalized focal pressure of 0.9923. Reduced-data tests and saliency maps support the anatomical design. These are simulation results; clinical validation remains future work.
Network design, acoustic setup, and simulation results.
Convergent Unrolled Networks for Image Deconvolution
Convergent Unrolled Networks for Image Deconvolution
IEEE TCI 2024; ICIP 2023 · Penn State
Restores images with known blur kernels using a compact unrolled network that retains provable convergence and performs well with limited training data.
Problem and method. Learning independent parameters at each layer can remove an iterative algorithm’s convergence guarantees. In this co-authored work, DCUNBD (ICIP 2023) and its journal extension DECUN (IEEE TCI 2024) unroll half-quadratic splitting and constrain learned filters and thresholds to approach limiting values. The analysis proves convergence as network depth increases and characterizes its rate.
Results. A 30-layer DECUN configuration uses only 1,117 trainable parameters. Tests with linear and nonlinear blur kernels show competitive restoration quality; experiments using 10% of the training images demonstrate the model’s data efficiency. Convergence curves and image comparisons connect the theoretical guarantees to observed results.
Network design, convergence curves, and restoration results.
Lens-to-lens Bokeh Transformation — SGLMS
Lens-to-lens Bokeh Transformation — SGLMS
NTIRE 2023 · IPAL-Bokeh / Penn State
Transforms out-of-focus lens bokeh while preserving a sharp foreground; team IPAL-Bokeh achieved the lowest real-image LPIPS in NTIRE 2023.
My participation and method. As a member of IPAL-Bokeh, I contributed to the Segmentation-Guided Lens-Mapping Scheme (SGLMS). Detail and semantic branches estimate the foreground mask; lens-specific encoders and decoders transform the background bokeh. Mask-guided fusion preserves the sharp foreground, while gated convolutions and selective-kernel fusion improve the mapping.
Results. Our 7M-parameter submission uses no ensembling and achieves the lowest real-image LPIPS (0.2161) among submitted methods. This distinction measures perceptual similarity on real captures; the overall challenge ranking uses PSNR/SSIM. The challenge report documents our team’s participating method.
Segmentation, lens mapping, and challenge results.
Efficient Non-homogeneous Dehazing — TransER
Efficient Non-homogeneous Dehazing — TransER
CVPR Workshops / NTIRE 2023 · iPAL-LightDehaze / Penn State
Removes spatially varying haze with a lightweight physical-model and Transformer approach; team iPAL-LightDehaze ranked 4th in LPIPS at NTIRE 2023.
My participation and method. As a member of iPAL-LightDehaze, I contributed to TransER for spatially varying haze. A shared Transformer–convolution encoder and three decoders estimate atmospheric light, transmission, and a haze-free image. A lightweight ensemble network then fuses physical-model and direct reconstructions, using feature distillation from a clean-image reconstruction teacher.
Results. The 2.6M-parameter model reaches PSNR/SSIM of 17.03 dB/0.597 on Dense-Haze and 21.64 dB/0.743 on NH-Haze. Our NTIRE 2023 submission ranks 4th in LPIPS among 17 finalists (0.384), without extra training data. Reported inference takes 0.72 seconds on a Titan XP GPU.
Method paper · Challenge report · Code
Architecture, feature-fusion modules, and benchmark results.
Attention-based Dehazing — GridFFA
Attention-based Dehazing — GridFFA
NTIRE 2021 · iPAL-GridFFA / Penn State
Removes non-uniform haze with grid-based feature fusion and attention; team iPAL-GridFFA placed 2nd in SSIM and tied for the lowest reported VGG-based LPIPS at NTIRE 2021.
My participation and method. As a member of iPAL-GridFFA, I contributed to a generative adversarial network (GAN) for non-uniform haze. Its 3 × 6 GridDehazeNet-based generator combines feature-fusion groups, channel and pixel attention, and skip connections. Each group contains 15 basic blocks, while a PatchGAN-style discriminator guides restoration quality.
Results. Our submission ranks 2nd in SSIM (0.839) among 23 finalists and ties for the lowest VGG-based LPIPS (0.194 at the precision reported). These metric-specific results appear in Table 1; the report’s final perceptual ranking is based on mean opinion score.
GridFFA, Grid Net, feature attention, and challenge results.
Computational Pathology — OpenHI & AIPath
Computational Pathology — OpenHI & AIPath
2018–2021 · Xi’an Jiaotong University
Developed collaborative whole-slide annotation software and renal cancer analysis models, with peer-reviewed studies, AIPath datasets, and a granted annotation patent.
Collaborative annotation. Gigapixel pathology slides are difficult to inspect and annotate consistently. I contributed to OpenHI/OpenHI2 and developed cloud-based collaborative annotation software. Multi-scale superpixels support pixel-level semantic labeling and responsive region retrieval; OpenHI2 adds calibrated magnification, shared diagnostic regions, agreement tests, and modular image analysis. OpenHI reports roughly 300 ms typical response time under its test conditions.
AI-assisted workflows. My software work included AI-assisted labeling and annotation workflows. Related co-authored PIMIP research integrates slide management, multi-device collaboration, automatic nucleus labeling, manual correction, extensible analysis, and linked patient/report information.
Nuclei grading and datasets. I developed renal cell carcinoma (RCC) machine-learning models and co-authored related nuclei-grading studies. CHR-Net uses W-Net to separate crowded nuclei, then combines high-resolution features and two cross-category classification heads. The MICCAI study contains 1,000 patches and 70,945 annotated nuclei, including clear-cell and supplementary papillary RCC samples. CHR-Net reaches Dice 0.8790 and average class-wise panoptic quality 0.5458 in the reported evaluation.
Whole-slide and multi-source analysis. Other co-authored studies combine tumor detection, RCC subtyping, grading, and whole-case summaries using TCGA and hospital slides. A personalized framework links similar histology with clinical reports and survival analysis. An attention-based graph convolutional network (GCN) extracts structured information from 3,632 TCGA reports across four cancer types, improving macro-F1 on most evaluated tasks. Annotation-granularity experiments show how precise labels affect model performance.
Outcomes. This work connects reusable research software, annotated datasets, and image/text analysis. I am a co-inventor on the collaborative annotation patent US11392634B2, granted in 2022.
OpenHI paper · OpenHI2 paper · CHR-Net paper · PIMIP paper · RCC framework · Personalized analysis · Report extraction · Annotation study · OpenHI code · AIPath datasets · ccRCC grading dataset
Platform workflows, annotation tools, and research evaluations.
Robust Sentiment Analysis — BERT Meta-learning
Robust Sentiment Analysis — BERT Meta-learning
NLPCC 2021 · Xi’an Jiaotong University
Uses meta-learning to handle noisy labels and class imbalance in literary-review sentiment analysis, improving macro-F1 by 13.51 percentage points.
Problem and method. Star ratings produce noisy sentiment labels, while positive reviews dominate the training data. This co-authored study introduces BERT-MLB: a BERT classifier with a Looking Back meta-model. An LSTM learns each sample’s training weight from historical features and its label embedding, guided by a small, balanced set of manually verified reviews.
Results. The dataset contains 109,286 reviews of 187 Chinese literary books, with 600 verified meta samples and a balanced 6,000-review test set. Macro-F1 rises from 61.12% for BERT to 74.63% for BERT-MLB, a gain of 13.51 percentage points. Sample-weight analysis shows reduced influence from noisy labels and greater attention to minority classes.
Meta-learning architecture and noisy-label analysis.
Exploratory Research
Exploratory Research
Poisson Diffusion for Generative Modeling
Poisson Diffusion for Generative Modeling
Preliminary research on Poisson-noise formulations for diffusion-based generation, motivated by signal-dependent imaging noise. Explored connections between Poisson processes and diffusion models, with low-light and night-scene generation as target applications.
Personal Projects
Personal Projects
AI Video Creation Agent
AI Video Creation Agent
TypeScript · React Native · Agentic AI
Built an Android application that converts multi-turn text and image conversations into structured prompts and AI video-generation tasks. Implemented persistent task state, asynchronous job tracking, retries, recovery, and configurable workflows with validation and runtime tests.
Multi-Agent Novel Generation
Multi-Agent Novel Generation
LLMs · Multi-Agent Systems
Built a novel-generation workflow with planning, writing, review, and revision agents. Persistent character profiles, world state, and event timelines support narrative continuity; review loops identify and address inconsistencies in long-form output.
Publications & Patents
Publications & Patents
Peer-reviewed research in computational imaging, medical image analysis, and applied machine learning. Complete list on Google Scholar.
Image Restoration & Computational Imaging
Image Restoration & Computational Imaging
Leveraging Phase Information to Boost Unrolled Network Learning for Image Deblurring
Leveraging Phase Information to Boost Unrolled Network Learning for Image Deblurring
Malek S, Zhang H, Lee C, Monga V.
European Conference on Computer Vision (ECCV), 2026
To recover images from unknown blur, UPADNet separates Fourier phase and amplitude and learns the updates of an alternating optimization algorithm. It reaches 34.63 dB PSNR on GoPro and shows stronger restoration under noise and limited training data in controlled COCO experiments.
EUMD: Event-based Unrolled Motion Deblurring via Learnable Prior Optimization
EUMD: Event-based Unrolled Motion Deblurring via Learnable Prior Optimization
Zhang H, Fu D, Miller J, Choudhury A, Lee C, Monga V.
IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), 2026
EUMD recovers motion-blurred images by combining event-conditioned blur modeling, unrolled half-quadratic splitting, and a Transformer-enhanced U-Net prior. It reaches 37.35 dB PSNR on synthetic GoPro and 32.42 dB on real EVRB, improving on ClearSight and CMTA by 0.31 dB and 1.04 dB, respectively.
Domain Enriched Learning for Skull Induced Phase Aberration Correction in Ultrasound Neuromodulation
Domain Enriched Learning for Skull Induced Phase Aberration Correction in Ultrasound Neuromodulation
Zhang H, Biswas R, Kiani M, Monga V.
Neuroelectronics, 2025
To correct skull-induced distortion of focused ultrasound, DB-SIPAC combines CT-derived propagation paths with full-skull features and refined time-delay targets. In 2D acoustic simulations, it predicts delays in 0.45 seconds and achieves mean normalized focal pressure of 0.9923.
Deep, convergent, unrolled half-quadratic splitting for image deconvolution
Deep, convergent, unrolled half-quadratic splitting for image deconvolution
Zhao Y, Li Y, Zhang H, Monga V, Eldar YC.
IEEE Transactions on Computational Imaging, 2024
DECUN addresses the loss of convergence guarantees in learned iterative restoration. Structured half-quadratic-splitting parameters retain provable convergence and its rate as network depth increases. A 30-layer configuration uses 1,117 trainable parameters and achieves competitive non-blind restoration, including under reduced training data.
A convergent neural network for non-blind image deblurring
A convergent neural network for non-blind image deblurring
Zhao Y, Li Y, Zhang H, Monga V, Eldar YC.
IEEE International Conference on Image Processing (ICIP), 2023
DCUNBD restores images with known blur kernels through structured, learned half-quadratic splitting. The analysis establishes convergence as depth increases; experiments support the theory and show improved restoration over the compared conventional, convolutional-network, and unrolled baselines.
TransER: Hybrid Model and Ensemble-Based Sequential Learning for Non-Homogenous Dehazing
TransER: Hybrid Model and Ensemble-Based Sequential Learning for Non-Homogenous Dehazing
Hoang T, Zhang H, Yazdani A, Monga V.
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023
TransER removes spatially varying haze by combining physical scene estimation with a Transformer–convolution model and distilled ensemble reconstruction. Its 2.6M-parameter model reaches PSNR/SSIM of 17.03 dB/0.597 on Dense-Haze and 21.64 dB/0.743 on NH-Haze.
NTIRE 2023 HR NonHomogeneous Dehazing Challenge Report
NTIRE 2023 HR NonHomogeneous Dehazing Challenge Report
Ancuti CO, Ancuti C, Vasluianu FA, et al.
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023
I participated with team iPAL-LightDehaze using TransER, a lightweight combination of physical haze estimation, Transformer–convolution features, and distilled reconstruction. Our submission ranked 4th in LPIPS among 17 finalists (0.384), with 2.6M parameters and no extra training data.
Lens-to-lens bokeh effect transformation. NTIRE 2023 challenge report
Lens-to-lens bokeh effect transformation. NTIRE 2023 challenge report
Conde MV, Kolmet M, Seizinger T, et al.
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023
I participated with team IPAL-Bokeh using SGLMS, which combines foreground segmentation, lens-specific mappings, and mask-guided fusion to transform bokeh while preserving the foreground. Our 7M-parameter submission achieved the lowest real-image LPIPS (0.2161) among submitted challenge methods.
NTIRE 2021 nonhomogeneous dehazing challenge report
NTIRE 2021 nonhomogeneous dehazing challenge report
Ancuti CO, et al.
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2021
I participated with team iPAL-GridFFA using a 3 × 6 grid generator with feature-fusion attention and a patch-based discriminator for non-homogeneous dehazing. Our submission ranked 2nd in SSIM (0.839) among 23 finalists and tied for the lowest reported VGG-based LPIPS (0.194).
Medical Image Analysis
Medical Image Analysis
Hemibrain growth as a biomarker for whole-brain growth
Hemibrain growth as a biomarker for whole-brain growth
Bajaj US, Yu M, Templeton K, Mukherjee S, Zhang H, Nunn N, Kulkarni AV, Kestle JRW, Monga V, Schiff SJ.
Journal of Neurosurgery: Pediatrics, 2026, pp. 1–11. Advance online publication.
When shunt artifacts obscure brain MRI, this study combines Dense U-Net segmentation with measurement of the unaffected hemisphere. Results from 75 pediatric hydrocephalus patients support hemibrain growth as a proxy for whole-brain growth; brain and cerebrospinal-fluid segmentation achieve Dice scores of 94.7% and 94.2%.
Nuclei Grading of Clear Cell Renal Cell Carcinoma in Histopathological Image by Composite High-Resolution Network
Nuclei Grading of Clear Cell Renal Cell Carcinoma in Histopathological Image by Composite High-Resolution Network
Gao Z, Shi J, Zhang X, Li Y, Zhang H, Wu J, et al.
International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), 2021
CHR-Net addresses crowded nuclei and similar grades in clear-cell renal cell carcinoma using W-Net instance segmentation, high-resolution features, and cross-category classification. The study contains 1,000 patches and 70,945 annotated nuclei; reported performance reaches Dice 0.8790 and average class-wise panoptic quality 0.5458.
Pimip: An open source platform for pathology information management and integration
Pimip: An open source platform for pathology information management and integration
Wu J, Mao A, Bao X, Zhang H, Gao Z, Wang C, Gong T, Li C.
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2021
PIMIP addresses fragmented pathology workflows by integrating whole-slide viewing, collaborative annotation, extensible deep-learning analysis, and patient/report management. The open-source platform demonstrates automatic nucleus labeling, manual correction, tumor analysis, and linked structured information in one web environment.
A precision diagnostic framework of renal cell carcinoma on whole-slide images using deep learning
A precision diagnostic framework of renal cell carcinoma on whole-slide images using deep learning
Wu J, Zhang R, Gong T, Bao X, Gao Z, Zhang H, Wang C, Li C.
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2021
This whole-slide renal cancer framework combines tumor detection, subtype classification, ISUP grading, and case-level report generation. Using 667 TCGA slides and 632 hospital slides, its base tumor/non-tumor classifier achieves validation accuracy of 92.2% and 92.7%, respectively; targeted augmentation improves uncertain predictions.
A personalized diagnostic generation framework based on multi-source heterogeneous data
A personalized diagnostic generation framework based on multi-source heterogeneous data
Wu J, Zhang R, Gong T, Zhang H, Wang C, Li C.
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2021
To personalize renal cancer prognosis, this framework retrieves histologically similar cases and links images with pathology and clinical reports. Kaplan–Meier and Lasso-Cox analyses generate patient-specific prognostic weights and risk stratification, evaluated using matched data from 859 TCGA patients.
Structured Information Extraction of Pathology Reports with Attention-based Graph Convolutional Network
Structured Information Extraction of Pathology Reports with Attention-based Graph Convolutional Network
Wu J, Tang K, Zhang H, et al.
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2020
An attention-based graph convolutional network combines semantic, syntactic, and sequential text relationships to extract structured information from pathology reports. On 3,632 annotated TCGA reports across four cancer types, it improves macro-F1 on most evaluated tasks, reaching 0.823 for cancer type and 0.871 for histological grade.
Effects of annotation granularity in deep learning models for histopathological images
Effects of annotation granularity in deep learning models for histopathological images
Shi J, Gao Z, Zhang H, et al.
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2019
This study compares bounding-box, ellipse, and pixel-level annotations for histopathology classification and segmentation. Across the tested models, finer annotations generally improve performance; U-Net segmentation reaches 95.79% accuracy with pixel-level labels, versus 89.58% with bounding-box labels.
OpenHI2—Open source histopathological image platform
OpenHI2—Open source histopathological image platform
Puttapirat P, Zhang H, Deng J, et al.
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2019
OpenHI2 extends whole-slide annotation with calibrated virtual magnification, shared diagnostic-region selection, inter-rater agreement tests, and modular machine-learning analysis. An annotation trial with two pathologists demonstrates the workflow across 20 diagnostic regions and 1,883 selected clusters.
OpenHI: Open Platform for Histopathological Image Annotation
OpenHI: Open Platform for Histopathological Image Annotation
Puttapirat P, Zhang H, Deng J, et al.
International Journal of Data Mining and Bioinformatics, 2019
OpenHI combines superpixel-guided region selection, standardized grading labels, responsive region retrieval, and multi-user web annotation for gigapixel slides. Tests demonstrate pixel-level annotation with roughly 300 ms typical response time under the reported viewing conditions, supporting efficient creation of research datasets.
OpenHI—An open source framework for annotating histopathological image
OpenHI—An open source framework for annotating histopathological image
Puttapirat P, Zhang H, Lian Y, et al.
IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2018
The original OpenHI framework enables collaborative, semantically enriched whole-slide annotation through multi-scale superpixels and a virtual-magnification indicator. Testing on TCGA slides demonstrates responsive pixel-level annotation, providing an open-source foundation for large pathology datasets.
Applied Machine Learning
Applied Machine Learning
BERT-Based Meta-Learning Approach with Looking Back for Sentiment Analysis of Literary Book Reviews
BERT-Based Meta-Learning Approach with Looking Back for Sentiment Analysis of Literary Book Reviews
Bao H, He K, Yin X, Li X, Bao X, Zhang H, Wu J, Gao Z.
Natural Language Processing and Chinese Computing (NLPCC), 2021, pp. 235–247
To handle noisy star-derived labels and class imbalance, BERT-MLB learns sample weights from historical features using an LSTM meta-model. On the Chinese literary-review benchmark, macro-F1 rises from 61.12% to 74.63%, a gain of 13.51 percentage points.
Doctoral Dissertation
Doctoral Dissertation
Model-Based Deep Learning for Imaging Inverse Problems
Model-Based Deep Learning for Imaging Inverse Problems
Haichuan Zhang · Ph.D. dissertation, Penn State, 2026.
Doctoral research combining physical imaging models, optimization, and deep learning for interpretable image restoration.
Manuscript Under Review
Manuscript Under Review
Cascaded, convergent unrolled networks for blind image deconvolution
Cascaded, convergent unrolled networks for blind image deconvolution
Malek S, Zhang H, Eldar YC, Monga V.
Under review.
Co-authored research on convergent unrolled networks for blind deconvolution, which jointly estimates a sharp image and its unknown blur kernel.
Patents
Patents
Granted Patents
Granted Patents
Cloud-based large-scale pathological image collaborative annotation method and system
Cloud-based large-scale pathological image collaborative annotation method and system
Li C, Puttapirat P, Zhang H · Xi’an Jiaotong University.
US11392634B2 · Granted Jul 19, 2022.
CN109949907B · Granted Jul 13, 2021.
Patent Applications
Patent Applications
Event-based Image and Video Restoration
Event-based Image and Video Restoration
Co-inventor on U.S. and European patent applications filed by Dolby Laboratories, Inc. Details remain confidential pending publication.
Diffusion-model-based Signal / Data Restoration
Diffusion-model-based Signal / Data Restoration
Co-inventor on a U.S. patent application filed with NEC Laboratories America, Inc. Details remain confidential pending publication.
Pathological Image Prognostic Feature Weight Calculation Method and System
Pathological Image Prognostic Feature Weight Calculation Method and System
Li C, Wu J, Gao Z, Zhang H, Puttapirat P · Xi’an Jiaotong University.
CN112434172A · Filed Oct 29, 2020; published Mar 2, 2021.
Honors & Awards
Honors & Awards
Recognition for research, engineering, and robotics.
Milton and Albertha Langdon Memorial Graduate Fellowship
Department of Electrical Engineering, Penn State · Jun 2024
Outstanding Production Internship Award
Xi’an Suoer Software Technology Co., Ltd. · Jul 2019
VEX U Robot Skills Challenge World Champion
VEX Robotics World Championship · May 2018
VEX U Excellence Award
VEX Robotics World Championship · May 2018
VEX U Robot Skills Challenge Asian Open Champion
VEX Robotics Asian Open · Dec 2017
VEX U Robotics Asian Open League Finalist
VEX Robotics Asian Open · Dec 2017
First Prize, Mathematical Modeling
National College Students Mathematical Modeling Contest
Undergraduate Group, Shaanxi Division · Nov 2017
Siyuan Scholarship for Academic Excellence
Xi’an Jiaotong University · Sep 2017
Page updated
Google Sites
Report abuse