Haixian Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Haixian Zhang has a total of 93 co-authors. Here are the top co-authors by number of shared publications:
- Zhang Yi
Sichuan University, Chengdu, China
9 shared publications - Han Wang
State Key Laboratory of Natural and Biomimetic Drugs, School of Pharmaceutical Sciences, Peking University
4 shared publications - Xinyang Li
School of Computer Science and Technology, Sichuan University, Chengdu, China
3 shared publications - Yi Zhang
School of Computer Science and Technology, Sichuan University, Chengdu, China
3 shared publications - Chengrong Yu
Yunnan University, Kunming, China
2 shared publications Dawen Wu
Department of Ophthalmology, West China Hospital, Sichuan University, Chengdu, China
2 shared publications
Latest Publications
- Open Access
- 29 September 2026
- Published by Springer Nature in npj Digital Medicine
Strabismus-Net mobile application for nine-gaze photograph-based strabismus diagnosis using dual-stream cross-attention and adaptive feature aggregation
- 29 September 2026
- Published by Springer Nature in npj Digital Medicine
Open AccessStrabismus-Net mobile application for nine-gaze photograph-based strabismus diagnosis using dual-stream cross-attention and adaptive feature aggregation
- 25 September 2026
- Published by Springer Nature
- p. 226-235
IgG4-QPC: Quantitative Pathology-Consistent Virtual Staining from H&E to IgG4
- 25 September 2026
- Published by Springer Nature
- p. 226-235
IgG4-QPC: Quantitative Pathology-Consistent Virtual Staining from H&E to IgG4
- 21 September 2025
- Published by Springer Nature
- p. 35-44
Asymmetric Matching in Abdominal Lymph Nodes of Follow-up CT Scans
- 21 September 2025
- Published by Springer Nature
- p. 35-44
Asymmetric Matching in Abdominal Lymph Nodes of Follow-up CT Scans
- Open Access
- 17 July 2025
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 27, e74402
Deep Learning–Based Precision Cropping of Eye Regions in Strabismus Photographs: Algorithm Development and Validation Study for Workflow Optimization
Background: Traditional ocular gaze photograph preprocessing, relying on manual cropping and head tilt correction, is time-consuming and inconsistent, limiting artificial intelligence (AI) model development and clinical application. Objective: This study aimed to address these challenges using an advanced preprocessing algorithm to enhance the accuracy, efficiency, and standardization of eye region cropping for clinical workflows and AI data preprocessing. Methods: This retrospective and prospective cross-sectional study utilized 5832 images from 648 inpatients and outpatients, capturing 3 gaze positions under diverse conditions, including obstructions and varying distances. The preprocessing algorithm, based on a rotating bounding box detection framework, was trained and evaluated using precision, recall, and mean average precision (mAP) at various intersections over union thresholds. A 5-fold cross-validation was performed on an inpatient dataset, with additional testing on an independent outpatient dataset and an external cross-population dataset of 500 images from the IMDB-WIKI collection, representing diverse ethnicities and ages. Expert validation confirmed alignment with clinical standards across 96 images (48 images from a Chinese dataset of patients with strabismus and 48 images from IMDB-WIKI). Gradient-weighted class activation mapping heatmaps were used to assess model interpretability. A control experiment with 5 optometry specialists compared manual and automated cropping efficiency. Downstream task validation involved preprocessing 1000 primary gaze photographs using the Dlib toolkit, faster region-based convolutional neural network (R-CNN; both without head tilt correction), and our model (with correction), evaluating the impact of head tilt correction via the vision transformer strabismus screening network through 5-fold cross-validation. Results: The model achieved exceptional performance across datasets: on the 5-fold cross-validation set, it recorded a mean precision of 1.000 (95% CI 1.000‐1.000), recall of 1.000 (95% CI 1.000‐1.000), mAP50 of 0.995 (95% CI 0.995‐0.995), and mAP95 of 0.893 (95% CI 0.870‐0.918); on the internal independent test set, precision and recall were 1.000, with mAP50 of 0.995 and mAP95 of 0.801; and on the external cross-population test set, precision and recall were 1.000, with mAP50 of 0.937 and mAP95 of 0.792. The control experiment reduced image preparation time from 10 hours for manual cropping of 900 photos to 30 seconds with the automated model. Downstream strabismus screening task validation showed our model (with head tilt correction) achieving an area under the curve of 0.917 (95% CI 0.901‐0.933), surpassing Dlib-toolkit and faster R-CNN (both without head tilt correction) with an area under the curve of 0.856 (P=.02) and 0.884 (P=.05), respectively. Heatmaps highlighted core ocular focus, aligning with head tilt directions. Conclusions: This study delivers an AI-driven platform featuring a preprocessing algorithm that automates eye region cropping, correcting head tilt variations to improve image quality for AI development and clinical use. Integrated with electronic archives and patient-physician interaction, it enhances workflow efficiency, ensures telemedicine privacy, and supports ophthalmological research and strabismus care.
- 17 July 2025
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 27, e74402
Open AccessDeep Learning–Based Precision Cropping of Eye Regions in Strabismus Photographs: Algorithm Development and Validation Study for Workflow Optimization
Background: Traditional ocular gaze photograph preprocessing, relying on manual cropping and head tilt correction, is time-consuming and inconsistent, limiting artificial intelligence (AI) model development and clinical application. Objective: This study aimed to address these challenges using an advanced preprocessing algorithm to enhance the accuracy, efficiency, and standardization of eye region cropping for clinical workflows and AI data preprocessing. Methods: This retrospective and prospective cross-sectional study utilized 5832 images from 648 inpatients and outpatients, capturing 3 gaze positions under diverse conditions, including obstructions and varying distances. The preprocessing algorithm, based on a rotating bounding box detection framework, was trained and evaluated using precision, recall, and mean average precision (mAP) at various intersections over union thresholds. A 5-fold cross-validation was performed on an inpatient dataset, with additional testing on an independent outpatient dataset and an external cross-population dataset of 500 images from the IMDB-WIKI collection, representing diverse ethnicities and ages. Expert validation confirmed alignment with clinical standards across 96 images (48 images from a Chinese dataset of patients with strabismus and 48 images from IMDB-WIKI). Gradient-weighted class activation mapping heatmaps were used to assess model interpretability. A control experiment with 5 optometry specialists compared manual and automated cropping efficiency. Downstream task validation involved preprocessing 1000 primary gaze photographs using the Dlib toolkit, faster region-based convolutional neural network (R-CNN; both without head tilt correction), and our model (with correction), evaluating the impact of head tilt correction via the vision transformer strabismus screening network through 5-fold cross-validation. Results: The model achieved exceptional performance across datasets: on the 5-fold cross-validation set, it recorded a mean precision of 1.000 (95% CI 1.000‐1.000), recall of 1.000 (95% CI 1.000‐1.000), mAP50 of 0.995 (95% CI 0.995‐0.995), and mAP95 of 0.893 (95% CI 0.870‐0.918); on the internal independent test set, precision and recall were 1.000, with mAP50 of 0.995 and mAP95 of 0.801; and on the external cross-population test set, precision and recall were 1.000, with mAP50 of 0.937 and mAP95 of 0.792. The control experiment reduced image preparation time from 10 hours for manual cropping of 900 photos to 30 seconds with the automated model. Downstream strabismus screening task validation showed our model (with head tilt correction) achieving an area under the curve of 0.917 (95% CI 0.901‐0.933), surpassing Dlib-toolkit and faster R-CNN (both without head tilt correction) with an area under the curve of 0.856 (P=.02) and 0.884 (P=.05), respectively. Heatmaps highlighted core ocular focus, aligning with head tilt directions. Conclusions: This study delivers an AI-driven platform featuring a preprocessing algorithm that automates eye region cropping, correcting head tilt variations to improve image quality for AI development and clinical use. Integrated with electronic archives and patient-physician interaction, it enhances workflow efficiency, ensures telemedicine privacy, and supports ophthalmological research and strabismus care.
- 4 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 33 (10577149) , 4116-4130
Spectral Embedding Fusion for Incomplete Multiview Clustering
Incomplete multiview clustering (IMVC) aims to reveal the underlying structure of incomplete multiview data by partitioning data samples into clusters. Several graph-based methods exhibit a strong ability to explore high-order information among multiple views using low-rank tensor learning. However, spectral embedding fusion of multiple views is ignored in low-rank tensor learning. In addition, addressing missing instances or features is still an intractable problem for most existing IMVC methods. In this paper, we present a unified spectral embedding tensor learning (USETL) framework that integrates the spectral embedding fusion of multiple similarity graphs and spectral embedding tensor learning for IMVC. To remove redundant information from the original incomplete multiview data, spectral embedding fusion is performed by introducing spectral rotations at two different data levels, i.e., the spectral embedding feature level and the clustering indicator level. The aim of introducing spectral embedding tensor learning is to capture consistent and complementary information by seeking high-order correlations among multiple views. The strategy of removing missing instances is adopted to construct multiple similarity graphs for incomplete multiple views. Consequently, this strategy provides an intuitive and feasible way to construct multiple similarity graphs. Extensive experimental results on multiview datasets demonstrate the effectiveness of the two spectral embedding fusion methods within the USETL framework.
- 4 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 33 (10577149) , 4116-4130
Spectral Embedding Fusion for Incomplete Multiview Clustering
Incomplete multiview clustering (IMVC) aims to reveal the underlying structure of incomplete multiview data by partitioning data samples into clusters. Several graph-based methods exhibit a strong ability to explore high-order information among multiple views using low-rank tensor learning. However, spectral embedding fusion of multiple views is ignored in low-rank tensor learning. In addition, addressing missing instances or features is still an intractable problem for most existing IMVC methods. In this paper, we present a unified spectral embedding tensor learning (USETL) framework that integrates the spectral embedding fusion of multiple similarity graphs and spectral embedding tensor learning for IMVC. To remove redundant information from the original incomplete multiview data, spectral embedding fusion is performed by introducing spectral rotations at two different data levels, i.e., the spectral embedding feature level and the clustering indicator level. The aim of introducing spectral embedding tensor learning is to capture consistent and complementary information by seeking high-order correlations among multiple views. The strategy of removing missing instances is adopted to construct multiple similarity graphs for incomplete multiple views. Consequently, this strategy provides an intuitive and feasible way to construct multiple similarity graphs. Extensive experimental results on multiview datasets demonstrate the effectiveness of the two spectral embedding fusion methods within the USETL framework.
- Open Access
- 4 July 2024
- Published by Springer Nature in Artificial Intelligence Review
- Vol. 57 (8) , 1-42
From single to universal: tiny lesion detection in medical imaging
Accurate and automatic detection of tiny lesions in medical imaging plays a critical role in comprehensive cancer diagnosis, staging, treatment, follow-up, and prognosis. Numerous deep learning (DL) based studies have focused on single lesions, providing highly sensitive identification and precise localization. On the other hand, some recent studies have started to concentrate on whole-body lesions, as they could provide systemic clinical support. This paper presents a single-to-universal review of DL studies on tiny lesion detection in medical imaging, with a particular emphasis on detection models and techniques, as well as the data-related aspects such as modality, dimension, and dataset. A wide range of tasks are covered, including traditional single lesion detection tasks such as lung nodules, breast masses, thyroid nodules, and diseased lymph nodes, as well as the emerging task of universal lesion detection. Moreover, the paper provides in-depth analysis and discussion, with the hope of inspiring future research endeavors.
- 4 July 2024
- Published by Springer Nature in Artificial Intelligence Review
- Vol. 57 (8) , 1-42
Open AccessFrom single to universal: tiny lesion detection in medical imaging
Accurate and automatic detection of tiny lesions in medical imaging plays a critical role in comprehensive cancer diagnosis, staging, treatment, follow-up, and prognosis. Numerous deep learning (DL) based studies have focused on single lesions, providing highly sensitive identification and precise localization. On the other hand, some recent studies have started to concentrate on whole-body lesions, as they could provide systemic clinical support. This paper presents a single-to-universal review of DL studies on tiny lesion detection in medical imaging, with a particular emphasis on detection models and techniques, as well as the data-related aspects such as modality, dimension, and dataset. A wide range of tasks are covered, including traditional single lesion detection tasks such as lung nodules, breast masses, thyroid nodules, and diseased lymph nodes, as well as the emerging task of universal lesion detection. Moreover, the paper provides in-depth analysis and discussion, with the hope of inspiring future research endeavors.
- 21 June 2024
- Published by World Scientific Pub Co Pte Ltd in International Journal of Neural Systems
- Vol. 34 (09) , 2450048
A Forward Learning Algorithm for Neural Memory Ordinary Differential Equations
The deep neural network, based on the backpropagation learning algorithm, has achieved tremendous success. However, the backpropagation algorithm is consistently considered biologically implausible. Many efforts have recently been made to address these biological implausibility issues, nevertheless, these methods are tailored to discrete neural network structures. Continuous neural networks are crucial for investigating novel neural network models with more biologically dynamic characteristics and for interpretability of large language models. The neural memory ordinary differential equation (nmODE) is a recently proposed continuous neural network model that exhibits several intriguing properties. In this study, we present a forward-learning algorithm, called nmForwardLA, for nmODE. This algorithm boasts lower computational dimensions and greater efficiency. Compared with the other learning algorithms, experimental results on MNIST, CIFAR10, and CIFAR100 demonstrate its potency.
- 21 June 2024
- Published by World Scientific Pub Co Pte Ltd in International Journal of Neural Systems
- Vol. 34 (09) , 2450048
A Forward Learning Algorithm for Neural Memory Ordinary Differential Equations
The deep neural network, based on the backpropagation learning algorithm, has achieved tremendous success. However, the backpropagation algorithm is consistently considered biologically implausible. Many efforts have recently been made to address these biological implausibility issues, nevertheless, these methods are tailored to discrete neural network structures. Continuous neural networks are crucial for investigating novel neural network models with more biologically dynamic characteristics and for interpretability of large language models. The neural memory ordinary differential equation (nmODE) is a recently proposed continuous neural network model that exhibits several intriguing properties. In this study, we present a forward-learning algorithm, called nmForwardLA, for nmODE. This algorithm boasts lower computational dimensions and greater efficiency. Compared with the other learning algorithms, experimental results on MNIST, CIFAR10, and CIFAR100 demonstrate its potency.
- Open Access
- 26 April 2024
- Published by Wiley in International Journal of Intelligent Systems
- Vol. 2024, 1-11
Incorporating Adaptive Sparse Graph Convolutional Neural Networks for Segmentation of Organs at Risk in Radiotherapy
Precisely segmenting the organs at risk (OARs) in computed tomography (CT) plays an important role in radiotherapy’s treatment planning, aiding in the protection of critical tissues during irradiation. Renowned deep convolutional neural networks (DCNNs) and prevailing transformer-based architectures are widely utilized to accomplish the segmentation task, showcasing advantages in capturing local and contextual characteristics. Graph convolutional networks (GCNs) are another specialized model designed for processing the nongrid dataset, e.g., citation relationship. The DCNNs and GCNs are considered as two distinct models applicable to the grid and nongrid datasets, respectively. Motivated by the recently developed dynamic-channel GCN (DCGCN) that attempts to leverage the graph structure to enhance the feature extracted by the DCNNs, this paper proposes a novel architecture termed adaptive sparse GCN (ASGCN) to mitigate the inherent limitations in DCGCN from the aspect of node’s representation and adjacency matrix’s construction. For the node’s representation, the global average pooling used in the DCGCN is replaced by the learning mechanism to accommodate the segmentation task. For the adjacency matrix, an adaptive regularization strategy is leveraged to penalize the coefficient in the adjacency matrix, resulting in a sparse one that can better exploit the relationships between nodes. Rigorous experiments on multiple OARs’ segmentation tasks of the head and neck demonstrate that the proposed ASGCN can effectively improve the segmentation accuracy. Comparison between the proposed method and other prevalent architectures further confirms the superiority of the ASGCN.
- 26 April 2024
- Published by Wiley in International Journal of Intelligent Systems
- Vol. 2024, 1-11
Open AccessIncorporating Adaptive Sparse Graph Convolutional Neural Networks for Segmentation of Organs at Risk in Radiotherapy
Precisely segmenting the organs at risk (OARs) in computed tomography (CT) plays an important role in radiotherapy’s treatment planning, aiding in the protection of critical tissues during irradiation. Renowned deep convolutional neural networks (DCNNs) and prevailing transformer-based architectures are widely utilized to accomplish the segmentation task, showcasing advantages in capturing local and contextual characteristics. Graph convolutional networks (GCNs) are another specialized model designed for processing the nongrid dataset, e.g., citation relationship. The DCNNs and GCNs are considered as two distinct models applicable to the grid and nongrid datasets, respectively. Motivated by the recently developed dynamic-channel GCN (DCGCN) that attempts to leverage the graph structure to enhance the feature extracted by the DCNNs, this paper proposes a novel architecture termed adaptive sparse GCN (ASGCN) to mitigate the inherent limitations in DCGCN from the aspect of node’s representation and adjacency matrix’s construction. For the node’s representation, the global average pooling used in the DCGCN is replaced by the learning mechanism to accommodate the segmentation task. For the adjacency matrix, an adaptive regularization strategy is leveraged to penalize the coefficient in the adjacency matrix, resulting in a sparse one that can better exploit the relationships between nodes. Rigorous experiments on multiple OARs’ segmentation tasks of the head and neck demonstrate that the proposed ASGCN can effectively improve the segmentation accuracy. Comparison between the proposed method and other prevalent architectures further confirms the superiority of the ASGCN.
- Open Access
- 1 January 2024
- Published by Wiley in International Journal of Intelligent Systems
- Vol. 2024 (1)
An Intelligent System of Predicting Lymph Node Metastasis in Colorectal Cancer Using 3D CT Scans
In colorectal cancer (CRC), accurately predicting lymph node metastasis (LNM) contributes to developing appropriate treatment plans and serves as the key to long‐term survival of patients. In the clinical settings, preoperative LNM diagnosis in CRC predominantly depends on computed tomography (CT). Nevertheless, lymph nodes are small in size and difficult to identify on 3D CT scans, and CT‐based diagnosis of metastatic lymph nodes is prone to a significant misdiagnosis rate and lacks consistency across clinicians. Currently, there is no automatic system available for LNM prediction in CRC via 3D CT scans. In addition, existing deep learning‐ (DL‐) based lymph node detection models present low detection accuracy and high false‐positive rates, and most existing DL‐based lymph node metastasis prediction models mainly use tumor area characteristics but fail to adequately utilize lymph node information, thus not yielding satisfactory results. To tackle these issues, we propose an intelligent diagnosis system for this challenging task, mainly including a lymph node detection (LND) model and a lymph node metastasis prediction (LNMP) model. In detail, the LND model utilizes an encoder‐decoder network to detect lymph nodes, and the LNMP model employs an innovative attention‐based multiple instance learning (MIL) network. An instance‐level self‐attention feature enhancement module is designed to extract and augment lymph node features as a bag of instances. Furthermore, a bag‐level MIL prediction module is employed to extract instance features and create a bag representation for the ultimate LNM prediction. As far as we know, the proposed intelligent system represents the pioneering method for addressing this complex clinical challenge. In experiments, our proposed intelligent system achieves the AUC of 75.4% and the accuracy of 73.9%, showcasing a significant enhancement compared to physicians specialising in CRC and highlighting its strong clinical applicability. The accessible code can be found at https://github.com/SCU-MI/IS-LNM.
- 1 January 2024
- Published by Wiley in International Journal of Intelligent Systems
- Vol. 2024 (1)
Open AccessAn Intelligent System of Predicting Lymph Node Metastasis in Colorectal Cancer Using 3D CT Scans
In colorectal cancer (CRC), accurately predicting lymph node metastasis (LNM) contributes to developing appropriate treatment plans and serves as the key to long‐term survival of patients. In the clinical settings, preoperative LNM diagnosis in CRC predominantly depends on computed tomography (CT). Nevertheless, lymph nodes are small in size and difficult to identify on 3D CT scans, and CT‐based diagnosis of metastatic lymph nodes is prone to a significant misdiagnosis rate and lacks consistency across clinicians. Currently, there is no automatic system available for LNM prediction in CRC via 3D CT scans. In addition, existing deep learning‐ (DL‐) based lymph node detection models present low detection accuracy and high false‐positive rates, and most existing DL‐based lymph node metastasis prediction models mainly use tumor area characteristics but fail to adequately utilize lymph node information, thus not yielding satisfactory results. To tackle these issues, we propose an intelligent diagnosis system for this challenging task, mainly including a lymph node detection (LND) model and a lymph node metastasis prediction (LNMP) model. In detail, the LND model utilizes an encoder‐decoder network to detect lymph nodes, and the LNMP model employs an innovative attention‐based multiple instance learning (MIL) network. An instance‐level self‐attention feature enhancement module is designed to extract and augment lymph node features as a bag of instances. Furthermore, a bag‐level MIL prediction module is employed to extract instance features and create a bag representation for the ultimate LNM prediction. As far as we know, the proposed intelligent system represents the pioneering method for addressing this complex clinical challenge. In experiments, our proposed intelligent system achieves the AUC of 75.4% and the accuracy of 73.9%, showcasing a significant enhancement compared to physicians specialising in CRC and highlighting its strong clinical applicability. The accessible code can be found at https://github.com/SCU-MI/IS-LNM.
- 2 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 43 (3) , 1180-1190
FAOT-Net: A 1.5-Stage Framework for 3D Pelvic Lymph Node Detection With Online Candidate Tuning
Accurate and automatic detection of pelvic lymph nodes in computed tomography (CT) scans is critical for diagnosing lymph node metastasis in colorectal cancer, which in turn plays a crucial role in its staging, treatment planning, surgical guidance, and postoperative follow-up of colorectal cancer. However, achieving high detection sensitivity and specificity poses a challenge due to the small and variable sizes of these nodes, as well as the presence of numerous similar signals within the complex pelvic CT image. To tackle these issues, we propose a 3D feature-aware online-tuning network (FAOT-Net) that introduces a novel 1.5-stage structure to seamlessly integrate detection and refinement via our online candidate tuning process and takes advantage of multi-level information through the tailored feature flow. Furthermore, we redesign the anchor fitting and anchor matching strategies to further improve detection performance in a nearly hyperparameter-free manner. Our framework achieves the FROC score of 52.8 and the sensitivity of 91.7% with 16 false positives per scan on the PLNDataset. Code will be available at: github.com/SCUsomebody/FAOT-Net/.
- 2 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 43 (3) , 1180-1190
FAOT-Net: A 1.5-Stage Framework for 3D Pelvic Lymph Node Detection With Online Candidate Tuning
Accurate and automatic detection of pelvic lymph nodes in computed tomography (CT) scans is critical for diagnosing lymph node metastasis in colorectal cancer, which in turn plays a crucial role in its staging, treatment planning, surgical guidance, and postoperative follow-up of colorectal cancer. However, achieving high detection sensitivity and specificity poses a challenge due to the small and variable sizes of these nodes, as well as the presence of numerous similar signals within the complex pelvic CT image. To tackle these issues, we propose a 3D feature-aware online-tuning network (FAOT-Net) that introduces a novel 1.5-stage structure to seamlessly integrate detection and refinement via our online candidate tuning process and takes advantage of multi-level information through the tailored feature flow. Furthermore, we redesign the anchor fitting and anchor matching strategies to further improve detection performance in a nearly hyperparameter-free manner. Our framework achieves the FROC score of 52.8 and the sensitivity of 91.7% with 16 false positives per scan on the PLNDataset. Code will be available at: github.com/SCUsomebody/FAOT-Net/.