Xinfeng Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xinfeng Zhang has a total of 38 co-authors. Here are the top co-authors by number of shared publications:
- Xiaomin Liu
Beijing University of Technology
6 shared publications - Xiangsheng Li
Medical Imaging Department, Air Force Medical Center, PLA, China
5 shared publications - Fei Wang
The Department of TCM Manipulative Orthopedic, Air Force Medical Center, Air Force Medical University, Beijing, China.
4 shared publications - Wei Guo
Air Force Medical Center
4 shared publications - Maoshen Jia
School of Information Science and Technology, Beijing University of Technology, Beijing, China
2 shared publications - Tianyu Zhu
School of Information and Communication Engineering, Beijing University of Technology, China
2 shared publications
Latest Publications
- 10 May 2025
- Published by Springer Nature
- p. 208-215
Spine-Ridge Extraction Based on the New Method of Curvature
- 10 May 2025
- Published by Springer Nature
- p. 208-215
Spine-Ridge Extraction Based on the New Method of Curvature
- Open Access
- 19 January 2024
- Published by Association for Computing Machinery (ACM)
- p. 111-115
UNet-based MRI spinal vertebral body center extraction
The spine is a vital part of the body's skeletal structure, connecting the entire skeletal framework and providing stability to the torso, head, and neck. When the spine is diseased, the spatial morphology of the spine will change accordingly. Therefore, spinal spatial pattern analysis is a crucial method to assess the health status of the spine. In the spatial morphology description of the spine, the line connecting each vertebra's center is generally used as the spatial morphology of the spine. Hence, extracting the center of the spine's vertebra becomes crucial. Currently, the vertebral centers in three-dimensional (3D) space are mainly extracted using the extraction of vertebral centers in a single image on a two-dimensional (2D) MRI or constraints on vertebral shapes based on prior knowledge. Therefore, this paper proposes a more stable and objective vertebral center extraction method with strong subjective factors and poor extraction accuracy for the current whole spine vertebral center extraction; it also analyzes and demonstrates its accuracy and stability. The specific method follows: firstly, the spine MR images in the sagittal plane are trained with the neural network by binary classification; after that, segmentation is carried out by using a neural network, and then each spine MRI obtained by segmentation is used to perform the pixel level's output of probability maps based on each vertebral body; finally, the 3D coordinates of all the pixel points of the whole spine MRI sequences are extracted and superimposed to form a body, which is combined with the probability maps outputted before to obtain the center of the entire spine. The experimental results showed that the average measurement error of this method was 1.48 mm for 23 vertebrae, and the measurement results remained unchanged after repeating the measurement three times, which was better than other measurement methods in terms of accuracy and stability.
- 19 January 2024
- Published by Association for Computing Machinery (ACM)
- p. 111-115
Open AccessUNet-based MRI spinal vertebral body center extraction
The spine is a vital part of the body's skeletal structure, connecting the entire skeletal framework and providing stability to the torso, head, and neck. When the spine is diseased, the spatial morphology of the spine will change accordingly. Therefore, spinal spatial pattern analysis is a crucial method to assess the health status of the spine. In the spatial morphology description of the spine, the line connecting each vertebra's center is generally used as the spatial morphology of the spine. Hence, extracting the center of the spine's vertebra becomes crucial. Currently, the vertebral centers in three-dimensional (3D) space are mainly extracted using the extraction of vertebral centers in a single image on a two-dimensional (2D) MRI or constraints on vertebral shapes based on prior knowledge. Therefore, this paper proposes a more stable and objective vertebral center extraction method with strong subjective factors and poor extraction accuracy for the current whole spine vertebral center extraction; it also analyzes and demonstrates its accuracy and stability. The specific method follows: firstly, the spine MR images in the sagittal plane are trained with the neural network by binary classification; after that, segmentation is carried out by using a neural network, and then each spine MRI obtained by segmentation is used to perform the pixel level's output of probability maps based on each vertebral body; finally, the 3D coordinates of all the pixel points of the whole spine MRI sequences are extracted and superimposed to form a body, which is combined with the probability maps outputted before to obtain the center of the entire spine. The experimental results showed that the average measurement error of this method was 1.48 mm for 23 vertebrae, and the measurement results remained unchanged after repeating the measurement three times, which was better than other measurement methods in terms of accuracy and stability.
- Open Access
- 20 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 63-68
Automatic Breast Tumor Segmentation System with Breast Localization Module Based on Dynamic Contrast Enhanced Magnetic Resonance Imaging
Segmentation of breast tumors from dynamic contrast-enhanced magnetic resonance (DCE-MR) images is a critical step in the diagnosis of breast cancer and subsequent efficacy assessment. However, the irregular shape and size of breast tumors, as well as the inhomogeneity of the background, pose challenges to accurately segmenting tumors in DCE-MR images. To address this, our study proposes a breast tumor segmentation model based on U-net++ and a breast region localization (BRL) module for more precise segmentation in DCE-MRI. This model helps doctors accurately locate the tumor position and perform lesion region segmentation. The BRL module effectively localizes the breast region, inhibits confusion of tissues in the chest cavity, and accelerates the convergence of the subsequent segmentation network. U-Net++ applies a residual network structure to segment the lesion region. Experimental results show that our proposed segmentation model achieves accurate segmentation of breast tumors, with a Dice coefficient of 0.78 and a sensitivity of 0.89.
- 20 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 63-68
Open AccessAutomatic Breast Tumor Segmentation System with Breast Localization Module Based on Dynamic Contrast Enhanced Magnetic Resonance Imaging
Segmentation of breast tumors from dynamic contrast-enhanced magnetic resonance (DCE-MR) images is a critical step in the diagnosis of breast cancer and subsequent efficacy assessment. However, the irregular shape and size of breast tumors, as well as the inhomogeneity of the background, pose challenges to accurately segmenting tumors in DCE-MR images. To address this, our study proposes a breast tumor segmentation model based on U-net++ and a breast region localization (BRL) module for more precise segmentation in DCE-MRI. This model helps doctors accurately locate the tumor position and perform lesion region segmentation. The BRL module effectively localizes the breast region, inhibits confusion of tissues in the chest cavity, and accelerates the convergence of the subsequent segmentation network. U-Net++ applies a residual network structure to segment the lesion region. Experimental results show that our proposed segmentation model achieves accurate segmentation of breast tumors, with a Dice coefficient of 0.78 and a sensitivity of 0.89.
- Open Access
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 559-564
An Aneurysm Localization Algorithm Based on Faster R-CNN Network for Cerebral Small Vessels
The use of artificial intelligence algorithm to determine whether the lesion has cerebral aneurysm, especially small aneurysms, is still not completely solved. In this paper, the Faster R-CNN network was used as the localization network, and the model was trained by adjusting the network parameters, and the appropriate feature extraction network and classification network were selected to finally solve the localization problem of small aneurysms. Compared with most 3D methods, this method had the characteristics of shorter training cycle and faster image recognition. The experimental results show that the algorithm has a high accuracy in discriminating whether the lesion has cerebral aneurysm, but the false positive phenomenon may occur in the identification of single image localization. Finally, the paper discusses the experimental results and puts forward some conjecture ideas to solve the problem.
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 559-564
Open AccessAn Aneurysm Localization Algorithm Based on Faster R-CNN Network for Cerebral Small Vessels
The use of artificial intelligence algorithm to determine whether the lesion has cerebral aneurysm, especially small aneurysms, is still not completely solved. In this paper, the Faster R-CNN network was used as the localization network, and the model was trained by adjusting the network parameters, and the appropriate feature extraction network and classification network were selected to finally solve the localization problem of small aneurysms. Compared with most 3D methods, this method had the characteristics of shorter training cycle and faster image recognition. The experimental results show that the algorithm has a high accuracy in discriminating whether the lesion has cerebral aneurysm, but the false positive phenomenon may occur in the identification of single image localization. Finally, the paper discusses the experimental results and puts forward some conjecture ideas to solve the problem.
- Open Access
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 656-662
A Symmetric Dual-Attention Generative Adversarial Network with Channel and Spatial Features Fusion
Many existing generative adversarial networks (GANs) lack effective semantic modeling, leading to unnatural local details and blurring in generated images. In this work, based on DivCo, we propose a Symmetric Dual-Attention Generative Adversarial Network (DivCo-SDAGAN) with channel and spatial feature fusion in which the Dual-Attention Module (DAM) is introduced to strengthen the feature representation ability of the network to synthesize photo-realistic images with more natural local details. The Channel Weighted Aggregation Module (CWAM) and the Spatial Attention Module (SAM) of the DAM are designed to capture the semantic information of channel dimension and spatial dimension, respectively, and they can be easily integrated into other GANs-based models. Extensive experiments show that the proposed DivCo-SDAGAN can produce more diverse images under the same input, achieving more satisfactory results than other existing methods.
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 656-662
Open AccessA Symmetric Dual-Attention Generative Adversarial Network with Channel and Spatial Features Fusion
Many existing generative adversarial networks (GANs) lack effective semantic modeling, leading to unnatural local details and blurring in generated images. In this work, based on DivCo, we propose a Symmetric Dual-Attention Generative Adversarial Network (DivCo-SDAGAN) with channel and spatial feature fusion in which the Dual-Attention Module (DAM) is introduced to strengthen the feature representation ability of the network to synthesize photo-realistic images with more natural local details. The Channel Weighted Aggregation Module (CWAM) and the Spatial Attention Module (SAM) of the DAM are designed to capture the semantic information of channel dimension and spatial dimension, respectively, and they can be easily integrated into other GANs-based models. Extensive experiments show that the proposed DivCo-SDAGAN can produce more diverse images under the same input, achieving more satisfactory results than other existing methods.
- Open Access
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 58-64
A 3D U-Net-Based Approach for Intracranial Aneurysm Detection
Intracranial aneurysm refers to a neoplastic protrusion of the arterial wall caused by a localized abnormal enlargement of the cerebral artery lumen. In clinical practice, patients in the early stage of onset generally have no obvious symptoms, which is very easy to miss diagnosis. In medicine, methods such as MRA, CTA and DSA can be used to display the images of blood vessels. Among them, magnetic resonance angiography (MRA) has the advantages of low cost and small damage to the human body. Which can display the images of blood vessels in the brain. The data set used herein was based on images provided by a three-dimensional time-of-flight magnetic resonance angiography system. The main contributions of this paper are as follows: (1) We improved a classic 3D U-Net model with the combination of attention gate, residual connection, and the changes of size. Which achieved automatic segmentation of aneurysms in MRA. In the detection of aneurysms with mean diameters of 6.10mm and 7.69mm, the sensitivity was 83.4% and 86.4% respectively. (2) On the basis of this sensitivity, we achieved a low false positive rate which was 0.36 FPs/case and 0.34 FPs/case respectively. CCS CONCEPTS • Computing methodologies∼Computer graphics∼Image manipulation∼Image processing
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 58-64
Open AccessA 3D U-Net-Based Approach for Intracranial Aneurysm Detection
Intracranial aneurysm refers to a neoplastic protrusion of the arterial wall caused by a localized abnormal enlargement of the cerebral artery lumen. In clinical practice, patients in the early stage of onset generally have no obvious symptoms, which is very easy to miss diagnosis. In medicine, methods such as MRA, CTA and DSA can be used to display the images of blood vessels. Among them, magnetic resonance angiography (MRA) has the advantages of low cost and small damage to the human body. Which can display the images of blood vessels in the brain. The data set used herein was based on images provided by a three-dimensional time-of-flight magnetic resonance angiography system. The main contributions of this paper are as follows: (1) We improved a classic 3D U-Net model with the combination of attention gate, residual connection, and the changes of size. Which achieved automatic segmentation of aneurysms in MRA. In the detection of aneurysms with mean diameters of 6.10mm and 7.69mm, the sensitivity was 83.4% and 86.4% respectively. (2) On the basis of this sensitivity, we achieved a low false positive rate which was 0.36 FPs/case and 0.34 FPs/case respectively. CCS CONCEPTS • Computing methodologies∼Computer graphics∼Image manipulation∼Image processing
- Open Access
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 264-272
Key Points Positioning: A Two-Stage Algorithm For Single-view Point Cloud of Human Back Based on Point-wise Network
Point cloud data is a collection of massive points containing the spatial position of each point on the target surface, which contains abundant spatial information. At present, it is also applied to the digital modeling of human surface in medical imaging, as the data basis for subsequent human body measurement, morphology estimation and data analysis. Among them, the key points is defined as the landmark position of the surface morphology analysis, those key points provides a reference position for the analysis work, and also reflects the symmetry of the body to a certain extent and morphology information. Aiming at the back shape analysis in clinical diagnosis, this paper proposes a two-stage key points positioning scheme of coarse segmentation and fine positioning. We design and build an pointwise artificial neural network to roughly locate the body part, in this part, we propose a maximum pooling module based on spatial location coding to express local features more strongly. Farther, we propose a gray distance and curvature based operator to match the position of key points. The experiment, shows that our method can effectively enhance the distinctiveness of features and meanwhile, reduce the influence from background.
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 264-272
Open AccessKey Points Positioning: A Two-Stage Algorithm For Single-view Point Cloud of Human Back Based on Point-wise Network
Point cloud data is a collection of massive points containing the spatial position of each point on the target surface, which contains abundant spatial information. At present, it is also applied to the digital modeling of human surface in medical imaging, as the data basis for subsequent human body measurement, morphology estimation and data analysis. Among them, the key points is defined as the landmark position of the surface morphology analysis, those key points provides a reference position for the analysis work, and also reflects the symmetry of the body to a certain extent and morphology information. Aiming at the back shape analysis in clinical diagnosis, this paper proposes a two-stage key points positioning scheme of coarse segmentation and fine positioning. We design and build an pointwise artificial neural network to roughly locate the body part, in this part, we propose a maximum pooling module based on spatial location coding to express local features more strongly. Farther, we propose a gray distance and curvature based operator to match the position of key points. The experiment, shows that our method can effectively enhance the distinctiveness of features and meanwhile, reduce the influence from background.
- Open Access
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 282-289
Tongue Image Retrieval Based On Reinforcement Learning
In Chinese medicine, the patient's body constitution plays a crucial role in determining the course of treatment because it is so intrinsically linked to the patient's physiological and pathological processes. Traditional Chinese medicine practitioners use tongue diagnosis to determine a person's constitutional type during an examination. An effective solution is needed to overcome the complexity of this setting before the tongue image constitution recognition system can be deployed on a non-invasive mobile device for fast, efficient, and accurate constitution recognition. We will use deep deterministic policy gradients to implement tongue retrieval techniques. We suggested a new method for image retrieval systems based on Deep Deterministic Policy Gradients (DDPG) in an effort to boost the precision of database searches for query images. We present a strategy for enhancing image retrieval accuracy that uses the complexity of individual instances to split the dataset into two subsets for independent classification using Deep reinforcement learning. Experiments on tongue datasets are performed to gauge the efficacy of our suggested approach; in these experiments, deep reinforcement learning techniques are applied to develop a retrieval system for pictures of tongues affected by various disorders. Using our proposed strategy, it may be possible to enhance image retrieval accuracy through enhanced recognition of tongue diseases. Databases containing pictures of tongues affected by a wide range of disorders will be used as examples. The experimental results suggest that the new approach to computing the main colour histogram outperforms the prior one. Though the difference is tiny statistically, the enhanced retrieval impact is clear to the human eye. The tongue is similarly brought to the fore to emphasise the importance of the required verbal statement. Both investigations used tongue images classified into five distinct categories.
- 17 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 282-289
Open AccessTongue Image Retrieval Based On Reinforcement Learning
In Chinese medicine, the patient's body constitution plays a crucial role in determining the course of treatment because it is so intrinsically linked to the patient's physiological and pathological processes. Traditional Chinese medicine practitioners use tongue diagnosis to determine a person's constitutional type during an examination. An effective solution is needed to overcome the complexity of this setting before the tongue image constitution recognition system can be deployed on a non-invasive mobile device for fast, efficient, and accurate constitution recognition. We will use deep deterministic policy gradients to implement tongue retrieval techniques. We suggested a new method for image retrieval systems based on Deep Deterministic Policy Gradients (DDPG) in an effort to boost the precision of database searches for query images. We present a strategy for enhancing image retrieval accuracy that uses the complexity of individual instances to split the dataset into two subsets for independent classification using Deep reinforcement learning. Experiments on tongue datasets are performed to gauge the efficacy of our suggested approach; in these experiments, deep reinforcement learning techniques are applied to develop a retrieval system for pictures of tongues affected by various disorders. Using our proposed strategy, it may be possible to enhance image retrieval accuracy through enhanced recognition of tongue diseases. Databases containing pictures of tongues affected by a wide range of disorders will be used as examples. The experimental results suggest that the new approach to computing the main colour histogram outperforms the prior one. Though the difference is tiny statistically, the enhanced retrieval impact is clear to the human eye. The tongue is similarly brought to the fore to emphasise the importance of the required verbal statement. Both investigations used tongue images classified into five distinct categories.
- Open Access
- 1 December 2021
- Published by IOP Publishing in Journal of Physics: Conference Series
- Vol. 2143 (1) , 012033
Application of Computer Algorithm in Fault Diagnosis System of RM Equipment
At present, modern mechanical equipment is gradually developing towards large-scale and intelligent, which leads to more and more complex equipment structure. Therefore, people have higher and higher requirements for intelligent fault diagnosis of mechanical equipment, which leads to the application of various algorithms to mechanical equipment. Among them, rotating machinery (hereinafter referred to as RM) mainly relies on rotating action to complete specific functions, such as gearbox, gas turbine, generator and engine, which are often the core components of mechanical equipment. Therefore, the FSGS (hereinafter referred to as FSGS) of RM equipment has become a very key link in system design and maintenance, which requires designers to constantly overcome the original intelligent diagnosis system. Through a variety of deep learning algorithms, we can improve the diagnosis efficiency of automatic monitoring and diagnosis equipment, which can also reduce the loss caused by untimely diagnosis. Firstly, this paper analyzes the types of application of computer algorithms in the fault body segment system of RM equipment. Then, this paper analyzes an algorithm, which can better improve the diagnosis efficiency of the equipment.
- 1 December 2021
- Published by IOP Publishing in Journal of Physics: Conference Series
- Vol. 2143 (1) , 012033
Open AccessApplication of Computer Algorithm in Fault Diagnosis System of RM Equipment
At present, modern mechanical equipment is gradually developing towards large-scale and intelligent, which leads to more and more complex equipment structure. Therefore, people have higher and higher requirements for intelligent fault diagnosis of mechanical equipment, which leads to the application of various algorithms to mechanical equipment. Among them, rotating machinery (hereinafter referred to as RM) mainly relies on rotating action to complete specific functions, such as gearbox, gas turbine, generator and engine, which are often the core components of mechanical equipment. Therefore, the FSGS (hereinafter referred to as FSGS) of RM equipment has become a very key link in system design and maintenance, which requires designers to constantly overcome the original intelligent diagnosis system. Through a variety of deep learning algorithms, we can improve the diagnosis efficiency of automatic monitoring and diagnosis equipment, which can also reduce the loss caused by untimely diagnosis. Firstly, this paper analyzes the types of application of computer algorithms in the fault body segment system of RM equipment. Then, this paper analyzes an algorithm, which can better improve the diagnosis efficiency of the equipment.
- Open Access
- 30 October 2020
- Published by Association for Computing Machinery (ACM)
- p. 36-42
A Model for Luggage Re-identification Based on the Network Ensemble
Luggage re-identification based on the image recognition is a highly demanding task for security check. The final result of testing model of using single network and loss function is under our expectation. Therefore, network ensemble which integrate multiple networks and loss ensemble which adopt various loss functions are explored and adopted in this paper. Meantime, most of technology of re-identification are adopted in this paper. In order to improve generalization ability, random erasing, batch feature erasing, random flip are adopted in this paper. Multibranch structure is adopted to extract more features and re-ranking is adopted to improve final accuracy. These methods are integrated into the model of luggage re-identification. Experimental results show that the better accuracy is obtained by our method. The model has strong value of practical application.
- 30 October 2020
- Published by Association for Computing Machinery (ACM)
- p. 36-42
Open AccessA Model for Luggage Re-identification Based on the Network Ensemble
Luggage re-identification based on the image recognition is a highly demanding task for security check. The final result of testing model of using single network and loss function is under our expectation. Therefore, network ensemble which integrate multiple networks and loss ensemble which adopt various loss functions are explored and adopted in this paper. Meantime, most of technology of re-identification are adopted in this paper. In order to improve generalization ability, random erasing, batch feature erasing, random flip are adopted in this paper. Multibranch structure is adopted to extract more features and re-ranking is adopted to improve final accuracy. These methods are integrated into the model of luggage re-identification. Experimental results show that the better accuracy is obtained by our method. The model has strong value of practical application.