Yuwen Zeng
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Yuwen Zeng has a total of 42 co-authors. Here are the top co-authors by number of shared publications:
- Noriyasu Homma
Department of Radiological Imaging and Informatics, Tohoku University Graduate School of Medicine, Sendai, Japan
3 shared publications Xiaoyong Zhang
National Institute of Technology, Sendai College
3 shared publicationsKei Ichiji
Department of Radiological Imaging and Informatics, Tohoku University Graduate School of Medicine, 2-1 Seiryo-machi, Aoba-ku, Sendai, Miyagi, 980-8575, Japan
2 shared publications- Kei Ichiji
Graduate School of Medicine, Tohoku University, Sendai, Japan
2 shared publications Masato Funayama
Department of Radiological Imaging and Informatics, Tohoku University Graduate School of Medicine, Sendai, Japan
2 shared publicationsNoriyasu Homma
Department of Radiological Imaging and Informatics, Tohoku University Graduate School of Medicine, 2-1 Seiryo-machi, Aoba-ku, Sendai, Miyagi, 980-8575, Japan
2 shared publications
Latest Publications
- Open Access
- 17 November 2026
- Published by MDPI AG in Preprints.org
A Survey on Recursive Self-Improvement in Healthcare AI Agents
Advances in persistent memory, reusable skills and adaptive training are making recursive self-improvement (RSI) increasingly relevant to healthcare agents. RSI concerns changes to the procedures that generate future improvements. This survey connects these developments with healthcare evidence by distinguishing component updates, training-generator changes and revisions of improvement procedures. We examine applications in diagnosis, reporting and question answering, together with challenges in evaluation, reliability, privacy and update governance. Current evidence supports persistent component updates and selected generator gains, while the effects of revised improvement procedures on subsequent learning remain untested in the mapped studies. An accompanying evidence map and evaluation templates support study comparison and future research.
- 17 November 2026
- Published by MDPI AG in Preprints.org
Open AccessA Survey on Recursive Self-Improvement in Healthcare AI Agents
Advances in persistent memory, reusable skills and adaptive training are making recursive self-improvement (RSI) increasingly relevant to healthcare agents. RSI concerns changes to the procedures that generate future improvements. This survey connects these developments with healthcare evidence by distinguishing component updates, training-generator changes and revisions of improvement procedures. We examine applications in diagnosis, reporting and question answering, together with challenges in evaluation, reliability, privacy and update governance. Current evidence supports persistent component updates and selected generator gains, while the effects of revised improvement procedures on subsequent learning remain untested in the mapped studies. An accompanying evidence map and evaluation templates support study comparison and future research.
- Open Access
- 19 December 2025
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 39 (5) , 4038-4051
Transformer-based Deep Learning Models with Shape Guidance for Predicting Breast Cancer in Mammography Images
Recent breast cancer research has investigated shape-based attention guidance in Vision Transformer (ViT) models, focusing on anatomical structures and the heterogeneity surrounding tumors. However, few studies have clarified the optimal transformer encoder layer stage for applying attention guidance. Our study aimed to evaluate the effectiveness of shape-guidance strategies by varying the combinations of encoder layers that guide attention to breast structures and by comparing the proposed models with conventional models. For the shape-guidance strategy, we applied breast masks to the attention mechanism to emphasize spatial dependencies and enhance the learning of positional relationships within breast anatomy. We then compared the representative models—Masked Transformer models that demonstrated the best performance across layer combinations—with the conventional ResNet50, ViT, and SwinT V2. In our study, a total of 2,436 publicly available mammography images from the Chinese Mammography Database via The Cancer Imaging Archive were analyzed. Three-fold cross-validation was employed, with a patient-wise split of 70% for training and 30% for validation. Model performance on differentiating breast cancer from non-cancer images was assessed by the area under the receiver-operating characteristic curve (AUROC). The results showed that applying masks at the Shallow and Deep stages gave the highest AUROC for Masked ViT. The Masked ViT achieved an AUROC of 0.885 [95% confidence interval: 0.849–0.918], a sensitivity of 0.876, and a specificity of 0.802, outperforming all other conventional models. These results indicate that incorporating mask guidance into particular Transformer encoders promotes representation learning, highlighting their potential as decision-support tools in breast cancer diagnosis.
- 19 December 2025
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 39 (5) , 4038-4051
Open AccessTransformer-based Deep Learning Models with Shape Guidance for Predicting Breast Cancer in Mammography Images
Recent breast cancer research has investigated shape-based attention guidance in Vision Transformer (ViT) models, focusing on anatomical structures and the heterogeneity surrounding tumors. However, few studies have clarified the optimal transformer encoder layer stage for applying attention guidance. Our study aimed to evaluate the effectiveness of shape-guidance strategies by varying the combinations of encoder layers that guide attention to breast structures and by comparing the proposed models with conventional models. For the shape-guidance strategy, we applied breast masks to the attention mechanism to emphasize spatial dependencies and enhance the learning of positional relationships within breast anatomy. We then compared the representative models—Masked Transformer models that demonstrated the best performance across layer combinations—with the conventional ResNet50, ViT, and SwinT V2. In our study, a total of 2,436 publicly available mammography images from the Chinese Mammography Database via The Cancer Imaging Archive were analyzed. Three-fold cross-validation was employed, with a patient-wise split of 70% for training and 30% for validation. Model performance on differentiating breast cancer from non-cancer images was assessed by the area under the receiver-operating characteristic curve (AUROC). The results showed that applying masks at the Shallow and Deep stages gave the highest AUROC for Masked ViT. The Masked ViT achieved an AUROC of 0.885 [95% confidence interval: 0.849–0.918], a sensitivity of 0.876, and a specificity of 0.802, outperforming all other conventional models. These results indicate that incorporating mask guidance into particular Transformer encoders promotes representation learning, highlighting their potential as decision-support tools in breast cancer diagnosis.
- 21 September 2025
- Published by Springer Nature
- p. 355-364
MGG-Net: A Multi-modal Feature Extraction and Global-Aware Feature Graph-Based Deep Learning Network for MGMT Status Classification in Glioma
- 21 September 2025
- Published by Springer Nature
- p. 355-364
MGG-Net: A Multi-modal Feature Extraction and Global-Aware Feature Graph-Based Deep Learning Network for MGMT Status Classification in Glioma
- Open Access
- 27 August 2025
- Published by Springer Nature in Medical & Biological Engineering & Computing
- Vol. 63 (12) , 3857-3872
Deep learning-based dual-energy subtraction synthesis from single-energy kV x-ray fluoroscopy for markerless tumor tracking
Markerless tumor tracking in x-ray fluoroscopic images is an important technique for achieving precise dose delivery for moving lung tumors during radiation therapy. However, accurate tumor tracking is challenging due to the poor visibility of the target tumor overlapped by other organs such as rib bones. Dual-energy (DE) x-ray fluoroscopy can enhance tracking accuracy with improved tumor visibility by suppressing bones. However, DE x-ray imaging requires special hardware, limiting its clinical use. This study presents a deep learning-based DE subtraction (DES) synthesis method to avoid hardware limitations and enhance tracking accuracy. The proposed method employs a residual U-Net model trained on a simulated DES dataset from a digital phantom to synthesize DES from single-energy (SE) fluoroscopy. Experimental results using a digital phantom showed quantitative evaluation results of synthesis quality. Also, experimental results using clinical SE fluoroscopic images of ten lung cancer patients showed improved tumor tracking accuracy using synthesized DES images, reducing errors from 1.80 to 1.68 mm on average. The tracking success rate within a 25% movement range increased from 50.2% (SE) to 54.9% (DES). These findings indicate the feasibility of deep learning-based DES synthesis for markerless tumor tracking, offering a potential alternative to hardware-dependent DE imaging. Graphical The U-Net model was first trained on simulated single-energy and dual-energy subtraction image pairs, then applied to synthesize DES from clinical X-ray images for improving tumor tracking
- 27 August 2025
- Published by Springer Nature in Medical & Biological Engineering & Computing
- Vol. 63 (12) , 3857-3872
Open AccessDeep learning-based dual-energy subtraction synthesis from single-energy kV x-ray fluoroscopy for markerless tumor tracking
Markerless tumor tracking in x-ray fluoroscopic images is an important technique for achieving precise dose delivery for moving lung tumors during radiation therapy. However, accurate tumor tracking is challenging due to the poor visibility of the target tumor overlapped by other organs such as rib bones. Dual-energy (DE) x-ray fluoroscopy can enhance tracking accuracy with improved tumor visibility by suppressing bones. However, DE x-ray imaging requires special hardware, limiting its clinical use. This study presents a deep learning-based DE subtraction (DES) synthesis method to avoid hardware limitations and enhance tracking accuracy. The proposed method employs a residual U-Net model trained on a simulated DES dataset from a digital phantom to synthesize DES from single-energy (SE) fluoroscopy. Experimental results using a digital phantom showed quantitative evaluation results of synthesis quality. Also, experimental results using clinical SE fluoroscopic images of ten lung cancer patients showed improved tumor tracking accuracy using synthesized DES images, reducing errors from 1.80 to 1.68 mm on average. The tracking success rate within a 25% movement range increased from 50.2% (SE) to 54.9% (DES). These findings indicate the feasibility of deep learning-based DES synthesis for markerless tumor tracking, offering a potential alternative to hardware-dependent DE imaging. Graphical The U-Net model was first trained on simulated single-energy and dual-energy subtraction image pairs, then applied to synthesize DES from clinical X-ray images for improving tumor tracking
- Open Access
- 9 February 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 37 (3) , 1-10
Inconsistency between Human Observation and Deep Learning Models: Assessing Validity of Postmortem Computed Tomography Diagnosis of Drowning
Drowning diagnosis is a complicated process in the autopsy, even with the assistance of autopsy imaging and the on-site information from where the body was found. Previous studies have developed well-performed deep learning (DL) models for drowning diagnosis. However, the validity of the DL models was not assessed, raising doubts about whether the learned features accurately represented the medical findings observed by human experts. In this paper, we assessed the medical validity of DL models that had achieved high classification performance for drowning diagnosis. This retrospective study included autopsy cases aged 8–91 years who underwent postmortem computed tomography between 2012 and 2021 (153 drowning and 160 non-drowning cases). We first trained three deep learning models from a previous work and generated saliency maps that highlight important features in the input. To assess the validity of models, pixel-level annotations were created by four radiological technologists and further quantitatively compared with the saliency maps. All the three models demonstrated high classification performance with areas under the receiver operating characteristic curves of 0.94, 0.97, and 0.98, respectively. On the other hand, the assessment results revealed unexpected inconsistency between annotations and models’ saliency maps. In fact, each model had, respectively, around 30%, 40%, and 80% of irrelevant areas in the saliency maps, suggesting the predictions of the DL models might be unreliable. The result alerts us in the careful assessment of DL tools, even those with high classification performance.
- 9 February 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 37 (3) , 1-10
Open AccessInconsistency between Human Observation and Deep Learning Models: Assessing Validity of Postmortem Computed Tomography Diagnosis of Drowning
Drowning diagnosis is a complicated process in the autopsy, even with the assistance of autopsy imaging and the on-site information from where the body was found. Previous studies have developed well-performed deep learning (DL) models for drowning diagnosis. However, the validity of the DL models was not assessed, raising doubts about whether the learned features accurately represented the medical findings observed by human experts. In this paper, we assessed the medical validity of DL models that had achieved high classification performance for drowning diagnosis. This retrospective study included autopsy cases aged 8–91 years who underwent postmortem computed tomography between 2012 and 2021 (153 drowning and 160 non-drowning cases). We first trained three deep learning models from a previous work and generated saliency maps that highlight important features in the input. To assess the validity of models, pixel-level annotations were created by four radiological technologists and further quantitatively compared with the saliency maps. All the three models demonstrated high classification performance with areas under the receiver operating characteristic curves of 0.94, 0.97, and 0.98, respectively. On the other hand, the assessment results revealed unexpected inconsistency between annotations and models’ saliency maps. In fact, each model had, respectively, around 30%, 40%, and 80% of irrelevant areas in the saliency maps, suggesting the predictions of the DL models might be unreliable. The result alerts us in the careful assessment of DL tools, even those with high classification performance.
- Open Access
- 29 November 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 27 (2) , 1026-1035
A 2.5D Deep Learning-Based Method for Drowning Diagnosis Using Post-Mortem Computed Tomography
It is challenging to diagnose drowning in autopsy even with the help of post-mortem multi-slice computed tomography (MSCT) due to the complex pathophysiology and the shortage of forensic specialists equipped with radiology knowledge. Therefore, a computer-aided diagnosis (CAD) system was developed to help with diagnosis. Most deep learning-based CAD systems only utilize 2D information, which is proper for 2D data such as chest X-ray images. However, 3D information should also be considered for 3D data like CT. Conventional 3D methods require a huge amount of data and computational cost when using 3D methods. In this paper, we proposed a 2.5D method that converts 3D data into 2D images to train 2D deep learning models for drowning diagnosis. The key point of this 2.5D method is that it uses a subset to represent the whole case, covering this case as much as possible while avoiding other repetitive information. To evaluate the effectiveness of the proposed method, conventional 2D, previous 2.5D, and 3D deep learning-based methods were tested using an MSCT dataset obtained from Tohoku university. Then, to provide explainable diagnosis results, a visualization method called Gradient-weighted Class Activation Mapping was employed to visualize features relevant to drowning in CT images. Results on drowning diagnosis showed that our proposed method achieved the best performance compared to other 2D, 2.5D, and 3D methods. The visual assessment also demonstrated that our method could find the saliency regions corresponding to drowning.
- 29 November 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 27 (2) , 1026-1035
Open AccessA 2.5D Deep Learning-Based Method for Drowning Diagnosis Using Post-Mortem Computed Tomography
It is challenging to diagnose drowning in autopsy even with the help of post-mortem multi-slice computed tomography (MSCT) due to the complex pathophysiology and the shortage of forensic specialists equipped with radiology knowledge. Therefore, a computer-aided diagnosis (CAD) system was developed to help with diagnosis. Most deep learning-based CAD systems only utilize 2D information, which is proper for 2D data such as chest X-ray images. However, 3D information should also be considered for 3D data like CT. Conventional 3D methods require a huge amount of data and computational cost when using 3D methods. In this paper, we proposed a 2.5D method that converts 3D data into 2D images to train 2D deep learning models for drowning diagnosis. The key point of this 2.5D method is that it uses a subset to represent the whole case, covering this case as much as possible while avoiding other repetitive information. To evaluate the effectiveness of the proposed method, conventional 2D, previous 2.5D, and 3D deep learning-based methods were tested using an MSCT dataset obtained from Tohoku university. Then, to provide explainable diagnosis results, a visualization method called Gradient-weighted Class Activation Mapping was employed to visualize features relevant to drowning in CT images. Results on drowning diagnosis showed that our proposed method achieved the best performance compared to other 2D, 2.5D, and 3D methods. The visual assessment also demonstrated that our method could find the saliency regions corresponding to drowning.