Chu-Song Chen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Chu-Song Chen has a total of 44 co-authors. Here are the top co-authors by number of shared publications:
- Jun-Cheng Chen
Academia Sinica, Taipei, Taiwan
3 shared publications - Chen-Hao Chiang
Department of Orthopaedics, Ditmanson Medical Foundation Chia-Yi Christian Hospital, Chiayi, Taiwan
2 shared publications - Chen-Kun Liaw
Department of Orthopedics, School of Medicine, College of Medicine, Taipei Medical University, Taipei City 11031, Taiwan
2 shared publications - Chien-Hua Huang
National Taiwan University
2 shared publications - Chiou-Shann Fuh
National Taiwan University
2 shared publications - Kuei-Lin Yeh
Department of Orthopaedics, Ditmanson Medical Foundation Chia-Yi Christian Hospital, Chiayi City 600, Taiwan
2 shared publications
Latest Publications
- 10 October 2025
- Published by Springer Nature in Journal of Medical Systems
- Vol. 49 (1) , 1-12
Real-Time Identification of Cricothyrotomy Landmarks in Emergency Care and Obstetric Patients Using Wireless Handheld Ultrasound and Edge-Computing Artificial Intelligence: A Prospective Observational Study
This study aimed to develop machine learning-based algorithms to assist physicians in ultrasound-guided localization of the cricoid cartilage (CC), thyroid cartilage (TC), and cricothyroid membrane (CTM) for cricothyroidotomy. Adult female participants presenting to the emergency department with dyspnea or to the obstetrics and gynecology department for a scheduled cesarean section between August 2022 and July 2024 were prospectively recruited. Ultrasonographic images were collected using a wireless handheld ultrasound device connected to an edge computing tablet. Three You Only Look Once (YOLO) model variants—v5n6, v8n, and v10n—were selected for development and evaluation. A total of 608 participants (median age: 58.0 years, interquartile range [IQR]: 40.0–73.0; median body mass index: 23.2 kg/m², IQR: 20.2–26.5) contributed 117,094 ultrasonographic frames. All three YOLO-based models demonstrated high accuracy in detecting CC, TC, and CTM, with area under the receiver operating characteristic curve values exceeding 0.88. In correctly identified frames, the models effectively localized CC (IOU values: YOLOv5n6, 0.713 [95% confidence interval (CI): 0.698–0.726]; YOLOv8n, 0.718 [95% CI: 0.702–0.733]; YOLOv10n, 0.718 [95% CI: 0.701–0.734]; p value: 0.03) and TC (YOLOv5n6, 0.700 [95% CI: 0.683–0.717]; YOLOv8n, 0.706 [95% CI: 0.687–0.725]; YOLOv10n, 0.703 [95% CI: 0.783–0.721] ; p value: 0.037), though localization accuracy was lower for CTM (YOLOv5n6, 0.364 [95% CI: 0.333–0.394]; YOLOv8n, 0.363 [95% CI: 0.331–0.394]; YOLOv10n, 0.354 [95% CI: 0.325–0.381] ; p value: 0.053). The mean frames per second for YOLOv5n6, YOLOv8n, and YOLOv10n were 3.67, 13.83, and 14.13, respectively, when deployed on the handheld ultrasound platform. YOLO-based models demonstrated high accuracy in detecting and localizing CC, TC, and CTM. YOLOv8n and YOLOv10n achieved clinically acceptable real-time imaging performance when deployed on a wireless handheld ultrasound device with an edge computing tablet. Further studies are needed to assess whether this favorable performance translates into actual clinical benefits.
- 10 October 2025
- Published by Springer Nature in Journal of Medical Systems
- Vol. 49 (1) , 1-12
Real-Time Identification of Cricothyrotomy Landmarks in Emergency Care and Obstetric Patients Using Wireless Handheld Ultrasound and Edge-Computing Artificial Intelligence: A Prospective Observational Study
This study aimed to develop machine learning-based algorithms to assist physicians in ultrasound-guided localization of the cricoid cartilage (CC), thyroid cartilage (TC), and cricothyroid membrane (CTM) for cricothyroidotomy. Adult female participants presenting to the emergency department with dyspnea or to the obstetrics and gynecology department for a scheduled cesarean section between August 2022 and July 2024 were prospectively recruited. Ultrasonographic images were collected using a wireless handheld ultrasound device connected to an edge computing tablet. Three You Only Look Once (YOLO) model variants—v5n6, v8n, and v10n—were selected for development and evaluation. A total of 608 participants (median age: 58.0 years, interquartile range [IQR]: 40.0–73.0; median body mass index: 23.2 kg/m², IQR: 20.2–26.5) contributed 117,094 ultrasonographic frames. All three YOLO-based models demonstrated high accuracy in detecting CC, TC, and CTM, with area under the receiver operating characteristic curve values exceeding 0.88. In correctly identified frames, the models effectively localized CC (IOU values: YOLOv5n6, 0.713 [95% confidence interval (CI): 0.698–0.726]; YOLOv8n, 0.718 [95% CI: 0.702–0.733]; YOLOv10n, 0.718 [95% CI: 0.701–0.734]; p value: 0.03) and TC (YOLOv5n6, 0.700 [95% CI: 0.683–0.717]; YOLOv8n, 0.706 [95% CI: 0.687–0.725]; YOLOv10n, 0.703 [95% CI: 0.783–0.721] ; p value: 0.037), though localization accuracy was lower for CTM (YOLOv5n6, 0.364 [95% CI: 0.333–0.394]; YOLOv8n, 0.363 [95% CI: 0.331–0.394]; YOLOv10n, 0.354 [95% CI: 0.325–0.381] ; p value: 0.053). The mean frames per second for YOLOv5n6, YOLOv8n, and YOLOv10n were 3.67, 13.83, and 14.13, respectively, when deployed on the handheld ultrasound platform. YOLO-based models demonstrated high accuracy in detecting and localizing CC, TC, and CTM. YOLOv8n and YOLOv10n achieved clinically acceptable real-time imaging performance when deployed on a wireless handheld ultrasound device with an edge computing tablet. Further studies are needed to assess whether this favorable performance translates into actual clinical benefits.
- 1 November 2024
- Published by Springer Nature
- p. 323-340
SAM4MLLM: Enhance Multi-Modal Large Language Model for Referring Expression Segmentation
- 1 November 2024
- Published by Springer Nature
- p. 323-340
SAM4MLLM: Enhance Multi-Modal Large Language Model for Referring Expression Segmentation
- Open Access
RDPN6D: Residual-based Dense Point-wise Network for 6Dof Object Pose Estimation Based on RGB-D Images
In this work, we introduce a novel method for calculating the 6DoF pose of an object using a single RGB-D image. Unlike existing methods that either directly predict objects' poses or rely on sparse keypoints for pose recovery, our approach addresses this challenging task using dense correspondence, i.e., we regress the object coordinates for each visible pixel. Our method leverages existing object detection methods. We incorporate a re-projection mechanism to adjust the camera's intrinsic matrix to accommodate cropping in RGB-D images. Moreover, we transform the 3D object coordinates into a residual representation, which can effectively reduce the output space and yield superior performance. We conducted extensive experiments to validate the efficacy of our approach for 6D pose estimation. Our approach outperforms most previous methods, especially in occlusion scenarios, and demonstrates notable improvements over the state-of-the-art methods. Our code is available on https://github.com/AI-Application-and-Integration-Lab/RDPN6D.
Open AccessRDPN6D: Residual-based Dense Point-wise Network for 6Dof Object Pose Estimation Based on RGB-D Images
In this work, we introduce a novel method for calculating the 6DoF pose of an object using a single RGB-D image. Unlike existing methods that either directly predict objects' poses or rely on sparse keypoints for pose recovery, our approach addresses this challenging task using dense correspondence, i.e., we regress the object coordinates for each visible pixel. Our method leverages existing object detection methods. We incorporate a re-projection mechanism to adjust the camera's intrinsic matrix to accommodate cropping in RGB-D images. Moreover, we transform the 3D object coordinates into a residual representation, which can effectively reduce the output space and yield superior performance. We conducted extensive experiments to validate the efficacy of our approach for 6D pose estimation. Our approach outperforms most previous methods, especially in occlusion scenarios, and demonstrates notable improvements over the state-of-the-art methods. Our code is available on https://github.com/AI-Application-and-Integration-Lab/RDPN6D.
- 28 January 2024
- Published by Springer Nature
- p. 503-516
Adversarially Robust Deepfake Detection via Adversarial Feature Similarity Learning
- 28 January 2024
- Published by Springer Nature
- p. 503-516
Adversarially Robust Deepfake Detection via Adversarial Feature Similarity Learning
- Open Access
- 10 January 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 37 (1) , 363-373
Application of Machine Learning to Ultrasonography in Identifying Anatomical Landmarks for Cricothyroidotomy Among Female Adults: A Multi-center Prospective Observational Study
We aimed to develop machine learning (ML)-based algorithms to assist physicians in ultrasound-guided localization of cricoid cartilage (CC) and thyroid cartilage (TC) in cricothyroidotomy. Adult female volunteers were prospectively recruited from two hospitals between September and December, 2020. Ultrasonographic images were collected via a modified longitudinal technique. You Only Look Once (YOLOv5s), Faster Regions with Convolutional Neural Network features (Faster R-CNN), and Single Shot Detector (SSD) were selected as the model architectures. A total of 488 women (mean age: 36.0 years) participated in the study, contributing to a total of 292,053 frames of ultrasonographic images. The derived ML-based algorithms demonstrated excellent discriminative performance for the presence of CC (area under the receiver operating characteristic curve [AUC]: YOLOv5s, 0.989, 95% confidence interval [CI]: 0.982–0.994; Faster R-CNN, 0.986, 95% CI: 0.980–0.991; SSD, 0.968, 95% CI: 0.956–0.977) and TC (AUC: YOLOv5s, 0.989, 95% CI: 0.977–0.997; Faster R-CNN, 0.981, 95% CI: 0.965–0.991; SSD, 0.982, 95% CI: 0.973–0.990). Furthermore, in the frames where the model could correctly indicate the presence of CC or TC, it also accurately localized CC (intersection-over-union: YOLOv5s, 0.753, 95% CI: 0.739–0.765; Faster R-CNN, 0.720, 95% CI: 0.709–0.732; SSD, 0.739, 95% CI: 0.726–0.751) or TC (intersection-over-union: YOLOv5s, 0.739, 95% CI: 0.722–0.755; Faster R-CNN, 0.709, 95% CI: 0.687–0.730; SSD, 0.713, 95% CI: 0.695–0.730). The ML-based algorithms could identify anatomical landmarks for cricothyroidotomy in adult females with favorable discriminative and localization performance. Further studies are warranted to transfer this algorithm to hand-held portable ultrasound devices for clinical use.
- 10 January 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 37 (1) , 363-373
Open AccessApplication of Machine Learning to Ultrasonography in Identifying Anatomical Landmarks for Cricothyroidotomy Among Female Adults: A Multi-center Prospective Observational Study
We aimed to develop machine learning (ML)-based algorithms to assist physicians in ultrasound-guided localization of cricoid cartilage (CC) and thyroid cartilage (TC) in cricothyroidotomy. Adult female volunteers were prospectively recruited from two hospitals between September and December, 2020. Ultrasonographic images were collected via a modified longitudinal technique. You Only Look Once (YOLOv5s), Faster Regions with Convolutional Neural Network features (Faster R-CNN), and Single Shot Detector (SSD) were selected as the model architectures. A total of 488 women (mean age: 36.0 years) participated in the study, contributing to a total of 292,053 frames of ultrasonographic images. The derived ML-based algorithms demonstrated excellent discriminative performance for the presence of CC (area under the receiver operating characteristic curve [AUC]: YOLOv5s, 0.989, 95% confidence interval [CI]: 0.982–0.994; Faster R-CNN, 0.986, 95% CI: 0.980–0.991; SSD, 0.968, 95% CI: 0.956–0.977) and TC (AUC: YOLOv5s, 0.989, 95% CI: 0.977–0.997; Faster R-CNN, 0.981, 95% CI: 0.965–0.991; SSD, 0.982, 95% CI: 0.973–0.990). Furthermore, in the frames where the model could correctly indicate the presence of CC or TC, it also accurately localized CC (intersection-over-union: YOLOv5s, 0.753, 95% CI: 0.739–0.765; Faster R-CNN, 0.720, 95% CI: 0.709–0.732; SSD, 0.739, 95% CI: 0.726–0.751) or TC (intersection-over-union: YOLOv5s, 0.739, 95% CI: 0.722–0.755; Faster R-CNN, 0.709, 95% CI: 0.687–0.730; SSD, 0.713, 95% CI: 0.695–0.730). The ML-based algorithms could identify anatomical landmarks for cricothyroidotomy in adult females with favorable discriminative and localization performance. Further studies are warranted to transfer this algorithm to hand-held portable ultrasound devices for clinical use.
- Open Access
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 15345-15358
ACCEPT: Adaptive Codebook for Composite and Efficient Prompt Tuning
Yu-Chen Lin, Wei-Hua Li, Jun-cheng Chen, Chu-Song Chen. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024.
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 15345-15358
Open AccessACCEPT: Adaptive Codebook for Composite and Efficient Prompt Tuning
Yu-Chen Lin, Wei-Hua Li, Jun-cheng Chen, Chu-Song Chen. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024.
- Open Access
Towards Adversarial Robustness for Multi-Mode Data through Metric Learning
Adversarial attacks have become one of the most serious security issues in widely used deep neural networks. Even though real-world datasets usually have large intra-variations or multiple modes, most adversarial defense methods, such as adversarial training, which is currently one of the most effective defense methods, mainly focus on the single-mode setting and thus fail to capture the full data representation to defend against adversarial attacks. To confront this challenge, we propose a novel multi-prototype metric learning regularization for adversarial training which can effectively enhance the defense capability of adversarial training by preventing the latent representation of the adversarial example changing a lot from its clean one. With extensive experiments on CIFAR10, CIFAR100, MNIST, and Tiny ImageNet, the evaluation results show the proposed method improves the performance of different state-of-the-art adversarial training methods without additional computational cost. Furthermore, besides Tiny ImageNet, in the multi-prototype CIFAR10 and CIFAR100 where we reorganize the whole datasets of CIFAR10 and CIFAR100 into two and ten classes, respectively, the proposed method outperforms the state-of-the-art approach by 2.22% and 1.65%, respectively. Furthermore, the proposed multi-prototype method also outperforms its single-prototype version and other commonly used deep metric learning approaches as regularization for adversarial training and thus further demonstrates its effectiveness.
Open AccessTowards Adversarial Robustness for Multi-Mode Data through Metric Learning
Adversarial attacks have become one of the most serious security issues in widely used deep neural networks. Even though real-world datasets usually have large intra-variations or multiple modes, most adversarial defense methods, such as adversarial training, which is currently one of the most effective defense methods, mainly focus on the single-mode setting and thus fail to capture the full data representation to defend against adversarial attacks. To confront this challenge, we propose a novel multi-prototype metric learning regularization for adversarial training which can effectively enhance the defense capability of adversarial training by preventing the latent representation of the adversarial example changing a lot from its clean one. With extensive experiments on CIFAR10, CIFAR100, MNIST, and Tiny ImageNet, the evaluation results show the proposed method improves the performance of different state-of-the-art adversarial training methods without additional computational cost. Furthermore, besides Tiny ImageNet, in the multi-prototype CIFAR10 and CIFAR100 where we reorganize the whole datasets of CIFAR10 and CIFAR100 into two and ten classes, respectively, the proposed method outperforms the state-of-the-art approach by 2.22% and 1.65%, respectively. Furthermore, the proposed multi-prototype method also outperforms its single-prototype version and other commonly used deep metric learning approaches as regularization for adversarial training and thus further demonstrates its effectiveness.
- Open Access
- 26 June 2023
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Conference on Artificial Intelligence
- Vol. 37 (1) , 369-377
Scalable Spatial Memory for Scene Rendering and Navigation
Neural scene representation and rendering methods have shown promise in learning the implicit form of scene structure without supervision. However, the implicit representation learned in most existing methods is non-expandable and cannot be inferred online for novel scenes, which makes the learned representation difficult to be applied across different reinforcement learning (RL) tasks. In this work, we introduce Scene Memory Network (SMN) to achieve online spatial memory construction and expansion for view rendering in novel scenes. SMN models the camera projection and back-projection as spatially aware memory control processes, where the memory values store the information of the partial 3D area, and the memory keys indicate the position of that area. The memory controller can learn the geometry property from observations without the camera's intrinsic parameters and depth supervision. We further apply the memory constructed by SMN to exploration and navigation tasks. The experimental results reveal the generalization ability of our proposed SMN in large-scale scene synthesis and its potential to improve the performance of spatial RL tasks.
- 26 June 2023
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Conference on Artificial Intelligence
- Vol. 37 (1) , 369-377
Open AccessScalable Spatial Memory for Scene Rendering and Navigation
Neural scene representation and rendering methods have shown promise in learning the implicit form of scene structure without supervision. However, the implicit representation learned in most existing methods is non-expandable and cannot be inferred online for novel scenes, which makes the learned representation difficult to be applied across different reinforcement learning (RL) tasks. In this work, we introduce Scene Memory Network (SMN) to achieve online spatial memory construction and expansion for view rendering in novel scenes. SMN models the camera projection and back-projection as spatially aware memory control processes, where the memory values store the information of the partial 3D area, and the memory keys indicate the position of that area. The memory controller can learn the geometry property from observations without the camera's intrinsic parameters and depth supervision. We further apply the memory constructed by SMN to exploration and navigation tasks. The experimental results reveal the generalization ability of our proposed SMN in large-scale scene synthesis and its potential to improve the performance of spatial RL tasks.
- Open Access
- 27 October 2022
- Published by MDPI AG in Journal of Clinical Medicine
- Vol. 11 (21) , 6353
Degree of Pelvic Rotation in the Coronal Plane on Postoperative Radiographs Obtained after Total Hip Arthroplasty
There are many published cup anteversion measurements for postoperative total hip arthroplasty (THA), including Liaw’s, Lewinnek’s, and Murray’s methods. However, most measurements ignore the potential pelvic rotation on radiographs except in Liaw’s method. Without considering pelvic rotation, clinicians can miscalculate cup anteversion. Therefore, we aimed to quantify the mean degree of pelvic rotation. Herein, we collected 388 radiographs of 98 postoperative THA hips of 77 patients and measured pelvic rotation by measuring h, the horizontal displacement of the sacrococcygeal junction associated with the upper pole of the symphysis pubis, and ssd, the distance between the sacrococcygeal junction and pubic symphysis. The angle θ of pelvic rotation was defined as θ = arc sin (h/ssd) × (180°/π). The mean degree of pelvic rotation was then calculated. The standard deviation of h was 7.84 mm, and the mean ssd was 158 mm. The potential pelvic rotation was 2.50°. The p-values from the paired t-test were all >0.05 when interobserver and intraobserver errors were assessed. This is the first study to quantify the potential pelvic rotation in the coronal plane on postoperative plain radiographs. The potential pelvic rotation was too large to be neglected during the measurement of cup anteversion.
- 27 October 2022
- Published by MDPI AG in Journal of Clinical Medicine
- Vol. 11 (21) , 6353
Open AccessDegree of Pelvic Rotation in the Coronal Plane on Postoperative Radiographs Obtained after Total Hip Arthroplasty
There are many published cup anteversion measurements for postoperative total hip arthroplasty (THA), including Liaw’s, Lewinnek’s, and Murray’s methods. However, most measurements ignore the potential pelvic rotation on radiographs except in Liaw’s method. Without considering pelvic rotation, clinicians can miscalculate cup anteversion. Therefore, we aimed to quantify the mean degree of pelvic rotation. Herein, we collected 388 radiographs of 98 postoperative THA hips of 77 patients and measured pelvic rotation by measuring h, the horizontal displacement of the sacrococcygeal junction associated with the upper pole of the symphysis pubis, and ssd, the distance between the sacrococcygeal junction and pubic symphysis. The angle θ of pelvic rotation was defined as θ = arc sin (h/ssd) × (180°/π). The mean degree of pelvic rotation was then calculated. The standard deviation of h was 7.84 mm, and the mean ssd was 158 mm. The potential pelvic rotation was 2.50°. The p-values from the paired t-test were all >0.05 when interobserver and intraobserver errors were assessed. This is the first study to quantify the potential pelvic rotation in the coronal plane on postoperative plain radiographs. The potential pelvic rotation was too large to be neglected during the measurement of cup anteversion.
- Open Access
Successful Plate Fixation with Long Intramedullary Fibula Bone Graft for Periprosthetic Femur Fracture: A Case Report
Background and objectives: Treatment of a displaced or comminuted periprosthetic distal femur fracture is challenging, especially in patients with osteoporosis. In this case report, we shared our successful surgical experience of using a long intramedullary fibula bone graft in a plate fixation surgery for a periprosthetic distal femur fracture in an extremely elderly patient with osteoporosis. Case report: A 95-year-old woman with severe osteoporosis (bone mineral density level: −3.0) presented with right knee pain and deformity after a fall, and a right periprosthetic distal femur fracture was identified. The patient underwent an open reduction and an internal plate fixation surgery with the application of a long intramedullary fibular bone graft. Due to a solid fixation, immediate weight-bearing was allowed after the surgery. She could walk independently without any valgus or varus malalignment or shortening 3 months after the surgery. A solid union was achieved 4 months postoperatively. Conclusions: We present a case wherein a long intramedullary allogenous fibula strut bone graft was used successfully to treat a right periprosthetic femur fracture in an extremely elderly patient. A long allogenous fibula bone graft can act not only as a firm structure for bridging the bone defect but also as a guide for precise component alignment. We believe this treatment option for periprosthetic fractures is beneficial for achieving biological and mechanical stability and facilitates early mobilization and weight-bearing for the patient.
Open AccessSuccessful Plate Fixation with Long Intramedullary Fibula Bone Graft for Periprosthetic Femur Fracture: A Case Report
Background and objectives: Treatment of a displaced or comminuted periprosthetic distal femur fracture is challenging, especially in patients with osteoporosis. In this case report, we shared our successful surgical experience of using a long intramedullary fibula bone graft in a plate fixation surgery for a periprosthetic distal femur fracture in an extremely elderly patient with osteoporosis. Case report: A 95-year-old woman with severe osteoporosis (bone mineral density level: −3.0) presented with right knee pain and deformity after a fall, and a right periprosthetic distal femur fracture was identified. The patient underwent an open reduction and an internal plate fixation surgery with the application of a long intramedullary fibular bone graft. Due to a solid fixation, immediate weight-bearing was allowed after the surgery. She could walk independently without any valgus or varus malalignment or shortening 3 months after the surgery. A solid union was achieved 4 months postoperatively. Conclusions: We present a case wherein a long intramedullary allogenous fibula strut bone graft was used successfully to treat a right periprosthetic femur fracture in an extremely elderly patient. A long allogenous fibula bone graft can act not only as a firm structure for bridging the bone defect but also as a guide for precise component alignment. We believe this treatment option for periprosthetic fractures is beneficial for achieving biological and mechanical stability and facilitates early mobilization and weight-bearing for the patient.