Xuyao Yu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xuyao Yu has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Hui Yu
Department of Biomedical Engineering, Tianjin University, Tianjin, China
2 shared publications - Yongzheng Ding
Tianjin Key Laboratory of Biomedical Detecting Techniques and Instruments, Tianjin University, China
2 shared publications - Chong Chen
Institute of Medical Engineering and Translational Medicine, Tianjin University, China
1 shared publication - Dagong Jia
Tianjin University
1 shared publication - Fengling Zhu
Tianjin Hospital, China
1 shared publication - Huanming Li
Tianjin Joint Laboratory of intelligent medicine, Tianjin 4TH Centre Hospital, China
1 shared publication
Latest Publications
- Open Access
- 10 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 55-60
Deep-learning algorithm for rib fracture detection in low-dose chest CT images
The detection of rib fractures is of great significance for the diagnosis and treatment of patients with blunt chest trauma. The while application of low-dose CT has become a trend. But there are few studies focusing on automatic fracture detection under low-dose CT. In this paper, a deep-learning model was proposed based on Mask R-CNN and its performance was evaluated by public dataset RibFrac containing 480 chest CT images with 3D rib fracture annotations. In addition, another dataset including 100 cases were collected from Tianjin Hospital for extensibility verification.FROC curve was used to compare the effects of different doses of CT images while the precision, recall, F1 score, IoU index and Dice coefficient were calculated. On the rib detection task,the detection rates achieved 93.5%, 90.8%, and 90.5% at normal dose, 10% mAs, and 5% mAs, respectively. On the external dataset,the recall rates of detection achieved of 95.7%, 93.9% and 92.7%.Our proposed network can effectively realize the task of detecting rib fractures in low-dose CT, and provide help for the application of deep learning for rib fracture detection in low-dose CT .
- 10 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 55-60
Open AccessDeep-learning algorithm for rib fracture detection in low-dose chest CT images
The detection of rib fractures is of great significance for the diagnosis and treatment of patients with blunt chest trauma. The while application of low-dose CT has become a trend. But there are few studies focusing on automatic fracture detection under low-dose CT. In this paper, a deep-learning model was proposed based on Mask R-CNN and its performance was evaluated by public dataset RibFrac containing 480 chest CT images with 3D rib fracture annotations. In addition, another dataset including 100 cases were collected from Tianjin Hospital for extensibility verification.FROC curve was used to compare the effects of different doses of CT images while the precision, recall, F1 score, IoU index and Dice coefficient were calculated. On the rib detection task,the detection rates achieved 93.5%, 90.8%, and 90.5% at normal dose, 10% mAs, and 5% mAs, respectively. On the external dataset,the recall rates of detection achieved of 95.7%, 93.9% and 92.7%.Our proposed network can effectively realize the task of detecting rib fractures in low-dose CT, and provide help for the application of deep learning for rib fracture detection in low-dose CT .
- Open Access
- 10 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 17-21
Unsupervised Cross-Domain White Blood Cells Classification Using DANN
The classification of white blood cells (WBCs) from microscopic blood image provides invaluable information for diagnosis of various diseases. Deep Convolutional Neural Networks are often used to classify WBCs automatically and have obtained certain achievements. However, when the training (source) dataset and test (target) dataset fall from different data distributions (i.e. domain shift), deep convolution neural networks adapt poorly. To solve the problem, we proposed a DANN-based method aiming to help our classifier learn domain-invariant information by using adversarial training. Two datasets were tested and our method achieved 97.1% accuracy, 97.2% recall, 97.2% precision and 97.4%f1-score, respectively. Domain adaptation verification shows that the proposed method has higher performance than other adaptive methods, and has broad application prospects in WBC classification.
- 10 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 17-21
Open AccessUnsupervised Cross-Domain White Blood Cells Classification Using DANN
The classification of white blood cells (WBCs) from microscopic blood image provides invaluable information for diagnosis of various diseases. Deep Convolutional Neural Networks are often used to classify WBCs automatically and have obtained certain achievements. However, when the training (source) dataset and test (target) dataset fall from different data distributions (i.e. domain shift), deep convolution neural networks adapt poorly. To solve the problem, we proposed a DANN-based method aiming to help our classifier learn domain-invariant information by using adversarial training. Two datasets were tested and our method achieved 97.1% accuracy, 97.2% recall, 97.2% precision and 97.4%f1-score, respectively. Domain adaptation verification shows that the proposed method has higher performance than other adaptive methods, and has broad application prospects in WBC classification.