Kaixu Chen (陈 凯旭)

Kaixu Chen (陈 凯旭)

he/him

Computer vision researcher at the University of Tsukuba. Clinical gait analysis, sports analytics, and multi-view 3D human pose estimation.

Publications

You can also find my articles on my Google Scholar profile.
Journal and conference papers are listed below. For forthcoming conference papers, the accepted status is indicated explicitly; publication links point to the publisher or official manuscript record.

Journal Articles

Clinical knowledge-guided attention for gait-based adult spinal deformity assessment from monocular video

Published in Journal of Medical Imaging, 2026

An anatomy-informed attention framework for interpretable adult spinal deformity assessment from monocular gait video.

Recommended citation: Chen K, Asada T, Miura K, Kadone H, Kikuchi N, Homma Y, Yamazaki M, Kitahara I, Kuroda Y. Clinical knowledge-guided attention for gait-based adult spinal deformity assessment from monocular video. Journal of Medical Imaging, 13(5), 054002 (2026). doi:10.1117/1.JMI.13.5.054002.
View Paper

Incision trajectory tracing for electrosurgical navigation by CNN-based knife contacting frames extraction method

Published in Computer Methods and Programs in Biomedicine, 2026

CNN-based detection of tool–tissue contact in thermal video to improve incision trajectory tracing for surgical navigation.

Recommended citation: Wang YC, Chen K, Ienaga N, Kuroda Y. Incision trajectory tracing for electrosurgical navigation by CNN-based knife contacting frames extraction method. Computer Methods and Programs in Biomedicine, 286, 109612 (2026). doi:10.1016/j.cmpb.2026.109612.
View Paper

PhaseMix: A Periodic Motion Fusion Method for Adult Spinal Deformity Classification

Published in IEEE Access, 2024

PhaseMix introduces a periodic motion fusion approach for video-based classification of adult spinal deformity. It models gait cycles to better capture clinically relevant temporal patterns, supporting quantitative analysis of walking impairment.

Recommended citation: Chen K, Xu J, Asada T, et al. PhaseMix: A Periodic Motion Fusion Method for Adult Spinal Deformity Classification[J]. IEEE Access, 2024.
View Paper

Two-stage video-based convolutional neural networks for adult spinal deformity classification

Published in Frontiers in Neuroscience, 2023

This study develops a two-stage video-based CNN framework for classifying adult spinal deformity using spatiotemporal information from gait recordings. The work investigates video-based assessment as a complementary, non-invasive approach to clinical gait analysis.

Recommended citation: Chen K, Asada T, Ienaga N, et al. Two-stage video-based convolutional neural networks for adult spinal deformity classification[J]. Frontiers in Neuroscience, 2023, 17: 1278584.
View Paper

Enhanced Full Attention Generative Adversarial Networks

Published in IEICE TRANSACTIONS on Information and Systems, 2023

In this paper, we propose improved Generative Adversarial Networks with attention module in Generator, which can enhance the effectiveness of Generator. Furthermore, recent work has shown that Generator conditioning affects GAN performance. Leveraging this insight, we explored the effect of different normalization (spectral normalization, instance normalization) on Generator and Discriminator. Moreover, an enhanced loss function called Wasserstein Divergence distance, can alleviate the problem of difficult to train module in practice.

Recommended citation: Chen K X, Yamane S. Enhanced Full Attention Generative Adversarial Networks[J]. IEICE TRANSACTIONS on Information and Systems, 2023, 106(5): 813-817.
View Paper

Conference Papers

CanonFuse3D: Human-Centric Canonical Fusion for Calibration-Free Dual-View 3D Pose Estimation in Skiing Videos

Accepted at MMSports 2026 (ACM Multimedia Workshop), 2026 (forthcoming)

Accepted for oral presentation: calibration-free fusion of complementary 3D pose streams estimated from mobile skiing videos.

Recommended citation: Chen K, Luo N, Kimura T, Naito E, Xie C, Kitahara I. CanonFuse3D: Human-Centric Canonical Fusion for Calibration-Free Dual-View 3D Pose Estimation in Skiing Videos. MMSports 2026, ACM Multimedia Workshop (accepted, 2026).
View Paper

A clinical knowledge-guided attention framework for gait-based adult spinal deformity diagnosis

Published in SPIE Medical Imaging 2026: Image Processing, 2026

Clinician-informed joint attention maps for gait-video classification and more interpretable adult spinal deformity analysis.

Recommended citation: Chen K, Asada T, Miura K, Kadone H, Sakashita K, Sunami T, Yamazaki M, Kuroda Y, Kitahara I. A clinical knowledge-guided attention framework for gait-based adult spinal deformity diagnosis. Proceedings of SPIE, vol. 13925, pp. 256–261 (2026). doi:10.1117/12.3083822.
View Paper

FilterNet: A Filtered Gait Motion Fusion Network for Classifying Adult Spinal Deformity

Published in 2026 International Conference on Electronics, Information, and Communication (ICEIC), 2026

Analyzing periodic human motion from videos is vital for applications such as action recognition and healthcare. In gait analysis, models must capture subtle phase-specific motion patterns. However, processing entire sequences often introduces noise and reduces accuracy. We propose FilterNet, a frameselection framework that identifies phase-relevant frames to improve temporal modeling and enhance discriminative representation. On a clinical gait video dataset, FilterNet achieves 74.5% accuracy, 75.2% precision, and 74.3% F1-score, outperforming baseline methods. Though demonstrated in a medical context, the framework is broadly applicable to other periodic motion analysis tasks. Code is available at: FilterNet_ASD_PyTorch.

Recommended citation: K. Chen et al., "FilterNet: A Filtered Gait Motion Fusion Network for Classifying Adult Spinal Deformity," 2026 International Conference on Electronics, Information, and Communication (ICEIC), Macau, China, 2026, pp. 1-6, doi: 10.1109/ICEIC69189.2026.11386438.
View Paper

Feel What You See: A Novel Sensory Interface Linking Visual Heat Cues and Instant Thermal Feedback

Published in SIGGRAPH Asia 2025 Emerging Technologies, 2025

We present a novel sensory interface that allows users to truly feel what they see by transforming visually implied heat cues into physical temperature sensations. Using a vision-language model, the system semantically analyzes video content to detect heat-related elements—such as fire, ice, or sunlight—and determines the appropriate thermal response based on both type and screen prominence. The core of this experience is a custom-designed, non-contact thermal device combining two physical mechanisms: an ethanol-based aerosol module for cooling, and a visible-spectrum radiation module for heating. The system dynamically adjusts the type and intensity of stimulation to reflect the semantic properties of each visual scene. This prototype demonstrates a new form of semantic-driven multisensory interaction, extending the expressive power of visual media. It opens new possibilities in immersive storytelling, sensory accessibility, and emotion-enhanced interfaces—where the ambient temperature of a scene can be directly felt, not just seen.

Recommended citation: Xu J, Chen K, Kuroda Y, et al. Feel What You See: A Novel Sensory Interface Linking Visual Heat Cues and Instant Thermal Feedback[M]//Proceedings of the SIGGRAPH Asia 2025 Emerging Technologies. 2025: 1-2.
View Paper

Spinal Disease Classification Using Deep Learning on Dual-View Videos

Published in Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2025

Adult spinal deformity (ASD) is characterized by spinal deformities that cause symptoms such as balance disorders. Although gait observation is used for the diagnosis of ASD, no quantitative evaluation protocol has been established. In this study, we propose a novel approach for distinguishing ASD from non-ASD using a neural network with gait videos captured from two perspectives, enabling the detection of unique gait fluctuations specific to each patient. By integrating spatiotemporal gait dynamics, the proposed method enhances quantitative diagnostics and provides a more comprehensive gait assessment. The experimental results indicate that, relative to a model employing a singular perspective, there was a modest enhancement in classification accuracy accompanied by a notable improvement in the F1 score. Specifically, the F1 score increased by 10% over the single-view model, achieving an F1 score of 71.86%. These results confirm the efficacy of integrating two-perspective videos for gait analysis in ASD and suggest that further investigation into fusion techniques may be beneficial. Codes and models are available at https://github.com/TsuguTsukumo/2stream_3D_CNN_Walk_Pytorch.git.

Recommended citation: Tsukumo T, Chen K, Asada T, et al. Spinal Disease Classification Using Deep Learning on Dual-View Videos[C]//Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference. 2025, 2025: 1-5.
View Paper