Peng-Hsiang Hung
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
Peng-Hsiang Hung has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
Chung-Ming Lo
Graduate Institute of Library, Information and Archival Studies, National Chengchi University, Taipei, Taiwan
5 shared publications- Chung-Ming Lo
Institute of Artificial Intelligence Innovation, Industry Academia Innovation School, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
2 shared publications Chi-Cheng Wang
Department of Radiology, Mackay Memorial Hospital, Taipei, Taiwan
1 shared publicationChuan-Hsien Chen
Department of Computer Science and Information Engineering, National Chung Cheng University, Minhsiung, Taiwan
1 shared publication- Daw-Tung Lin
National Taipei University
1 shared publication Kai-Chun Huang
Department of Computer Science and Information Engineering, National Chung Cheng University, Minhsiung, Taiwan
1 shared publication
Latest Publications
- 5 November 2026
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- p. 1-13
Cross-Center Stroke Evaluation using Carotid Ultrasound Patterns via Scale-Adaptive Frequency-Enhanced Neural Network
Medical imaging datasets vary across institutions due to differences in scanner types, posing challenges for model generalization. To address the issue, this study proposed a scale-adaptive frequency-enhanced neural network (SAFE-Net), which leveraged a DWT-ViT encoder to incorporate frequency-domain features into a transformer-based architecture for cross-center stroke evaluation. The discrete wavelet transform (DWT) decomposed carotid color Doppler (CCD) images into multiresolution frequency sub-bands. The extracted high-resolution (Xh) and low-resolution (Xl) features are concatenated to enhance both local and global feature extraction. Experimental results demonstrated that SAFE-Net achieves the highest accuracy, with 92.5% accuracy on Institution A and 93.75% on Institution B, outperforming other models in stroke evaluation. SAFE-Net maintains cross-center generalization, with an accuracy of 90.73% when trained on Institution A and tested on Institution B, and 89.16% in the reverse setting. Also, the model required 25% fewer parameters and 2.0× fewer FLOPs than conventional models. These findings underscored the effectiveness of SAFE-Net across diverse medical imaging datasets.
- 5 November 2026
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- p. 1-13
Cross-Center Stroke Evaluation using Carotid Ultrasound Patterns via Scale-Adaptive Frequency-Enhanced Neural Network
Medical imaging datasets vary across institutions due to differences in scanner types, posing challenges for model generalization. To address the issue, this study proposed a scale-adaptive frequency-enhanced neural network (SAFE-Net), which leveraged a DWT-ViT encoder to incorporate frequency-domain features into a transformer-based architecture for cross-center stroke evaluation. The discrete wavelet transform (DWT) decomposed carotid color Doppler (CCD) images into multiresolution frequency sub-bands. The extracted high-resolution (Xh) and low-resolution (Xl) features are concatenated to enhance both local and global feature extraction. Experimental results demonstrated that SAFE-Net achieves the highest accuracy, with 92.5% accuracy on Institution A and 93.75% on Institution B, outperforming other models in stroke evaluation. SAFE-Net maintains cross-center generalization, with an accuracy of 90.73% when trained on Institution A and tested on Institution B, and 89.16% in the reverse setting. Also, the model required 25% fewer parameters and 2.0× fewer FLOPs than conventional models. These findings underscored the effectiveness of SAFE-Net across diverse medical imaging datasets.
- 19 March 2026
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- p. 1-17
Wavelet-Transformed Frequency Linked Attention with Selective Hierarchy for Abdominal Organ Segmentation
The clinical use of computed tomography (CT) scans has been steadily rising due to their ability to provide detailed three-dimensional (3D) representations of organs, along with advantages such as speed and high resolution. With CT, physicians can identify potential diseased organs and plan appropriate treatments. However, accurate quantitative evaluation is time-consuming, and precise automatic segmentation can significantly enhance the effectiveness of CT scans. Frequency linked attention with selective hierarchy (FLASH) was introduced as a novel deep learning architecture for multi-organ segmentation in abdominal CT scans. FLASH incorporated the 3D discrete wavelet transform into Transformer blocks as the backbone of the encoder. By distinguishing 3D frequency bands, the network addressed the challenge of unclear organ boundaries in abdominal CT scans caused by low contrast. Two-directional skip connections were introduced to link the multi-scale frequency features into the bottleneck and to assign adaptive weights to frequency components using attention mechanisms. By training from scratch on the organ dataset with five-fold validation, FLASH achieved the highest dice similarity coefficient = 0.826, normalized surface distance = 0.698 with the lowest standard deviation of dice similarity coefficient = 0.034, standard deviation of normalized surface distance = 0.046 compared to other networks in the experiments. With better segmentation results, FLASH had fewer parameters (96.86 M vs. 139.45 M) and computation time (36.3 h vs. 50.4 h) compared to Swin UNETR, which is more suitable in clinical use with limited resources and waiting time. Our code can be found at https://github.com/jorden0721/FLASH-UNETR.git
- 19 March 2026
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- p. 1-17
Wavelet-Transformed Frequency Linked Attention with Selective Hierarchy for Abdominal Organ Segmentation
The clinical use of computed tomography (CT) scans has been steadily rising due to their ability to provide detailed three-dimensional (3D) representations of organs, along with advantages such as speed and high resolution. With CT, physicians can identify potential diseased organs and plan appropriate treatments. However, accurate quantitative evaluation is time-consuming, and precise automatic segmentation can significantly enhance the effectiveness of CT scans. Frequency linked attention with selective hierarchy (FLASH) was introduced as a novel deep learning architecture for multi-organ segmentation in abdominal CT scans. FLASH incorporated the 3D discrete wavelet transform into Transformer blocks as the backbone of the encoder. By distinguishing 3D frequency bands, the network addressed the challenge of unclear organ boundaries in abdominal CT scans caused by low contrast. Two-directional skip connections were introduced to link the multi-scale frequency features into the bottleneck and to assign adaptive weights to frequency components using attention mechanisms. By training from scratch on the organ dataset with five-fold validation, FLASH achieved the highest dice similarity coefficient = 0.826, normalized surface distance = 0.698 with the lowest standard deviation of dice similarity coefficient = 0.034, standard deviation of normalized surface distance = 0.046 compared to other networks in the experiments. With better segmentation results, FLASH had fewer parameters (96.86 M vs. 139.45 M) and computation time (36.3 h vs. 50.4 h) compared to Swin UNETR, which is more suitable in clinical use with limited resources and waiting time. Our code can be found at https://github.com/jorden0721/FLASH-UNETR.git
- 9 December 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 38 (4) , 2170-2182
Fusion Learning from Non-contrast CT Scans for the Detection of Hemorrhagic Transformation in Stroke Patients
Hemorrhagic transformation (HT) is a potentially catastrophic complication after acute ischemic stroke. Prevention of HT risk is crucial because it worsens prognosis and increases mortality. This study aimed at developing and validating a computer-aided diagnosis system using pretreatment non-contrast computed tomography (CT) scans for HT prediction in stroke patients undergoing revascularization. This retrospective study included all acute ischemic stroke patients with non-contrast CT before reperfusion therapy who also underwent follow-up MRI from January 2018 to December 2022. Among the 188 evaluated patients, any degree of HT at follow-up imaging was observed in 103 patients. HT diagnosis via MRI was defined as the reference standard for neuroradiologists. Using a database of 2076 serial non-contrast CT images of the brain, pretrained deep learning architectures such as convolutional neural networks and vision transformers (ViTs) were used for feature extraction. The performance of the predictive HT risk model was evaluated via tenfold cross-validation in machine learning classifiers. The accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were evaluated. Using an individual deep learning architecture, DenseNet201 features achieved the highest accuracy of 87% and an AUC of 0.8863 in the classifier of the subspace ensemble k-nearest neighbor. By combining the DenseNet201 and ViT features, the accuracy and AUC can be improved to 88% and 0.8987, respectively, which are significantly better than those of using ViT alone. Detecting HT in stroke patients is a meaningful but challenging issue. On the basis of the model approach, HT diagnosis would be more automatic, efficient, and consistent, which would be helpful in clinic use.
- 9 December 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 38 (4) , 2170-2182
Fusion Learning from Non-contrast CT Scans for the Detection of Hemorrhagic Transformation in Stroke Patients
Hemorrhagic transformation (HT) is a potentially catastrophic complication after acute ischemic stroke. Prevention of HT risk is crucial because it worsens prognosis and increases mortality. This study aimed at developing and validating a computer-aided diagnosis system using pretreatment non-contrast computed tomography (CT) scans for HT prediction in stroke patients undergoing revascularization. This retrospective study included all acute ischemic stroke patients with non-contrast CT before reperfusion therapy who also underwent follow-up MRI from January 2018 to December 2022. Among the 188 evaluated patients, any degree of HT at follow-up imaging was observed in 103 patients. HT diagnosis via MRI was defined as the reference standard for neuroradiologists. Using a database of 2076 serial non-contrast CT images of the brain, pretrained deep learning architectures such as convolutional neural networks and vision transformers (ViTs) were used for feature extraction. The performance of the predictive HT risk model was evaluated via tenfold cross-validation in machine learning classifiers. The accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were evaluated. Using an individual deep learning architecture, DenseNet201 features achieved the highest accuracy of 87% and an AUC of 0.8863 in the classifier of the subspace ensemble k-nearest neighbor. By combining the DenseNet201 and ViT features, the accuracy and AUC can be improved to 88% and 0.8987, respectively, which are significantly better than those of using ViT alone. Detecting HT in stroke patients is a meaningful but challenging issue. On the basis of the model approach, HT diagnosis would be more automatic, efficient, and consistent, which would be helpful in clinic use.
- Open Access
- 5 February 2024
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 69 (4) , 045004
Interactive content-based image retrieval with deep learning for CT abdominal organ recognition
Objective. Recognizing the most relevant seven organs in an abdominal computed tomography (CT) slice requires sophisticated knowledge. This study proposed automatically extracting relevant features and applying them in a content-based image retrieval (CBIR) system to provide similar evidence for clinical use. Approach. A total of 2827 abdominal CT slices, including 638 liver, 450 stomach, 229 pancreas, 442 spleen, 362 right kidney, 424 left kidney and 282 gallbladder tissues, were collected to evaluate the proposed CBIR in the present study. Upon fine-tuning, high-level features used to automatically interpret the differences among the seven organs were extracted via deep learning architectures, including DenseNet, Vision Transformer (ViT), and Swin Transformer v2 (SwinViT). Three images with different annotations were employed in the classification and query. Main results. The resulting performances included the classification accuracy (94%–99%) and retrieval result (0.98–0.99). Considering global features and multiple resolutions, SwinViT performed better than ViT. ViT also benefited from a better receptive field to outperform DenseNet. Additionally, the use of hole images can obtain almost perfect results regardless of which deep learning architectures are used. Significance. The experiment showed that using pretrained deep learning architectures and fine-tuning with enough data can achieve successful recognition of seven abdominal organs. The CBIR system can provide more convincing evidence for recognizing abdominal organs via similarity measurements, which could lead to additional possibilities in clinical practice.
- 5 February 2024
- Published by IOP Publishing in Physics in Medicine & Biology
- Vol. 69 (4) , 045004
Open AccessInteractive content-based image retrieval with deep learning for CT abdominal organ recognition
Objective. Recognizing the most relevant seven organs in an abdominal computed tomography (CT) slice requires sophisticated knowledge. This study proposed automatically extracting relevant features and applying them in a content-based image retrieval (CBIR) system to provide similar evidence for clinical use. Approach. A total of 2827 abdominal CT slices, including 638 liver, 450 stomach, 229 pancreas, 442 spleen, 362 right kidney, 424 left kidney and 282 gallbladder tissues, were collected to evaluate the proposed CBIR in the present study. Upon fine-tuning, high-level features used to automatically interpret the differences among the seven organs were extracted via deep learning architectures, including DenseNet, Vision Transformer (ViT), and Swin Transformer v2 (SwinViT). Three images with different annotations were employed in the classification and query. Main results. The resulting performances included the classification accuracy (94%–99%) and retrieval result (0.98–0.99). Considering global features and multiple resolutions, SwinViT performed better than ViT. ViT also benefited from a better receptive field to outperform DenseNet. Additionally, the use of hole images can obtain almost perfect results regardless of which deep learning architectures are used. Significance. The experiment showed that using pretrained deep learning architectures and fine-tuning with enough data can achieve successful recognition of seven abdominal organs. The CBIR system can provide more convincing evidence for recognizing abdominal organs via similarity measurements, which could lead to additional possibilities in clinical practice.
- 21 June 2022
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 147, 105779
Computer-aided diagnosis of ischemic stroke using multi-dimensional image features in carotid color Doppler
- 21 June 2022
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 147, 105779
Computer-aided diagnosis of ischemic stroke using multi-dimensional image features in carotid color Doppler
- 1 August 2021
- Published by Elsevier in Ultrasound in Medicine & Biology
- Vol. 47 (8) , 2266-2276
Assessing Ischemic Stroke with Convolutional Image Features in Carotid Color Doppler
- 1 August 2021
- Published by Elsevier in Ultrasound in Medicine & Biology
- Vol. 47 (8) , 2266-2276
Assessing Ischemic Stroke with Convolutional Image Features in Carotid Color Doppler
- 7 May 2021
- Published by Springer Nature in Journal of Digital Imaging
- Vol. 34 (3) , 637-646
Rapid Assessment of Acute Ischemic Stroke by Computed Tomography Using Deep Convolutional Neural Networks
Acute stroke is one of the leading causes of disability and death worldwide. Regarding clinical diagnoses, a rapid and accurate procedure is necessary for patients suffering from acute stroke. This study proposes an automatic identification scheme for acute ischemic stroke using deep convolutional neural networks (DCNNs) based on non-contrast computed tomographic (NCCT) images. Our image database for the classification model was composed of 1254 grayscale NCCT images from 96 patients (573 images) with acute ischemic stroke and 121 normal controls (681 images). According to the consensus of critical stroke findings by two neuroradiologists, a gold standard was established and used to train the proposed DCNN using machine-generated image features. Including the earliest DCNN, AlexNet, the popular Inception-v3, and ResNet-101 were proposed. To train the limited data size, transfer learning with ImageNet parameters was also used. The established models were evaluated by tenfold cross-validation and tested on an independent dataset containing 50 patients with acute ischemic stroke (108 images) and 58 normal controls (117 images) from another institution. AlexNet without pretrained parameters achieved an accuracy of 97.12%, a sensitivity of 98.11%, a specificity of 96.08%, and an area under the receiver operating characteristic curve (AUC) of 0.9927. Using transfer learning, transferred AlexNet, transferred Inception-v3, and transferred ResNet-101 achieved accuracies between 90.49 and 95.49%. Tested with a dataset from another institution, AlexNet showed an accuracy of 60.89%, a sensitivity of 18.52%, and a specificity of 100%. Transferred AlexNet, Inception-v3, and ResNet-101 achieved accuracies of 81.77%, 85.78%, and 80.89%, respectively. The proposed DCNN architecture as a computer-aided diagnosis system showed that training from scratch can generate a customized model for a specific scanner, and transfer learning can generate a more generalized model to provide diagnostic suggestions of acute ischemic stroke to radiologists.
- 7 May 2021
- Published by Springer Nature in Journal of Digital Imaging
- Vol. 34 (3) , 637-646
Rapid Assessment of Acute Ischemic Stroke by Computed Tomography Using Deep Convolutional Neural Networks
Acute stroke is one of the leading causes of disability and death worldwide. Regarding clinical diagnoses, a rapid and accurate procedure is necessary for patients suffering from acute stroke. This study proposes an automatic identification scheme for acute ischemic stroke using deep convolutional neural networks (DCNNs) based on non-contrast computed tomographic (NCCT) images. Our image database for the classification model was composed of 1254 grayscale NCCT images from 96 patients (573 images) with acute ischemic stroke and 121 normal controls (681 images). According to the consensus of critical stroke findings by two neuroradiologists, a gold standard was established and used to train the proposed DCNN using machine-generated image features. Including the earliest DCNN, AlexNet, the popular Inception-v3, and ResNet-101 were proposed. To train the limited data size, transfer learning with ImageNet parameters was also used. The established models were evaluated by tenfold cross-validation and tested on an independent dataset containing 50 patients with acute ischemic stroke (108 images) and 58 normal controls (117 images) from another institution. AlexNet without pretrained parameters achieved an accuracy of 97.12%, a sensitivity of 98.11%, a specificity of 96.08%, and an area under the receiver operating characteristic curve (AUC) of 0.9927. Using transfer learning, transferred AlexNet, transferred Inception-v3, and transferred ResNet-101 achieved accuracies between 90.49 and 95.49%. Tested with a dataset from another institution, AlexNet showed an accuracy of 60.89%, a sensitivity of 18.52%, and a specificity of 100%. Transferred AlexNet, Inception-v3, and ResNet-101 achieved accuracies of 81.77%, 85.78%, and 80.89%, respectively. The proposed DCNN architecture as a computer-aided diagnosis system showed that training from scratch can generate a customized model for a specific scanner, and transfer learning can generate a more generalized model to provide diagnostic suggestions of acute ischemic stroke to radiologists.