Publications

More publications are listed in CV (click here for pdf file)

LOCALIZATION OF RETINAL NERVE FIBER LAYER DEFECT IN FUNDUS IMAGE BY VISUAL FIELD GUIDED LEARNING NETWORK

Published in The 32st IPPR Conference on Computer Vision, Graphics and Image Processing, 2019

Retinal Nerve Fiber Layer Defect (RNFLD) can be an earliest sign to detect the ongoing glaucomatous damage. However, existing measurements, including visual field test and optic cup/disc ratio, fail to reflect RNFLD. Although optical coherence tomography (OCT) may provide information about RNFLD, the field of view (FOV) of OCT is smaller than that of fundus camera. This means early RNFLD may be undetected by OCT. In order to screen out patients with early-stage glaucoma, we propose to build a deep neural network to both predict glaucoma and locate RNFLD in fundus image by constraining its latent space with visual field map (VFM), which has wider FOV than fundus image and indicates visual field loss led by RNFLD. Since VFM does not match fundus image at pixel level, the challenge of this net-work would be to learn the spatial relationship between fundus image and VFM in addition to the prediction of glaucoma. To tackle this challenge, we compared three encoder-decoder convolutional neural network (CNN) with distinctive architectures in this study: (i) encoder-decoder CNN, (ii) encoder-decoder CNN with spatial transformer network (STN) and (iii) generative adversarial network (GAN), whose generator is the same as (i). The main evaluation metrics in this study was the correlation coefficient between predicted VFM and real VFM. Be-sides, accuracy and AUC of each network for the prediction of glaucoma were measured to make sure the predicted VFMs were closely related to glaucoma. The study was conducted on the dataset we collected from a medical center. Our results demonstrated that the correlation coefficient produced from model (iii) was the highest and it also did well in the prediction of glaucoma. This proposed network would be the first one to predict glaucoma and locate RNFLD simultaneously to provide explainable results for ophthalmologists and address the pixel-level mismatch between fundus images and VFM.

Recommended citation:
Chun-Fu Yeh, Guan-An Chen, Kwou-Yeung Wu, Min-Yu Huang (2019) Localization of Retinal Nerve Fiber Layer Defect in Fundus Image by Visual Field Guided Learning Network. The 32st IPPR Conference on Computer Vision, Graphics and Image Processing, Taitung, Taiwan, August 25-27.</br>

Feasibility of Early Detection for Retinal Nerve Fiber Layer Defect with Digital Fundus Image in Glaucoma Patients

Published in The 31st IPPR Conference on Computer Vision, Graphics and Image Processing, 2018

Due to the fact that retinal nerve fiber layer defect in Glaucoma is not perceived by patients at early stage, a technique to monitor the progress of retinal nerve fiber layer defect is imperative. Currently, the tools or parameters, such as standard automated Perimetry and the cup-to-disc ratio derived from fundus images, can only screen out the glaucomatous patients whose ganglion cells in optic nerve have already lost about 50%. It would be meaningful to clinicians if the loss of optic nerve could be quantified in the early stage efficiently. For this purpose, the fundus images, which has a 45¡ field of view and encircle most of the optic nerve, should be leveraged and a deep learning-based algorithm should be developed to measure the RNFL defect. In this paper, the feasibility of early detection for retinal nerve fiber layer defect with fundus images in glaucoma patients is demonstrated. We built up a deep learning model to detect glaucomatous patients from healthy subjects with fundus images. The accuracy of the deep learning model for classification of Glaucoma patients was 90% and it could be seen that the model treated the features relevant to the retinal nerve fiber layer defect as the essential ones for discriminating glaucoma patients. From this experiment, it is promising to measure the severity of RNFL defects and locate them efficiently with fundus images and deep learning algorithms. Future studies would focus on development of an accurate deep learning-based algorithm, which may take visual field data as reference or labels, for retinal nerve fiber layer defect quantification.

Recommended citation:
Chun-Fu Yeh, Guan-An Chen, Kwou-Yeung Wu, Min-Yu Huang (2018) Feasibility of Early Detection for Retinal Nerve Fiber Layer Defect with Digital Fundus Image in Glaucoma Patients. The 31st IPPR Conference on Computer Vision, Graphics and Image Processing, Tainan, Taiwan, August 19-21.</br>

Body Motion Analysis System, Portable Device and Body Motion Analysis Method.

Published in U.S. Patent, 2018

A body motion analysis system with pressure-sensing mat to quantify the performance of movements of Tai-Chi Practitioner

Recommended citation:
Chun-Fu Yeh, Szu-Han Tzao and Ming-Chieh Tsai. (2016) Body Motion Analysis System, Portable Device and Body Motion Analysis Method. U.S. Patent 2018/0178060A1, filed December 23, 2016. Patent Pending.</br>

Postural control during standing reach in children with Down syndrome

Published in Research in Developmental Disabilities , 2014

The purpose of the present study was to investigate the dynamic postural control of children with Down syndrome (DS). Specifically, we compared postural control and goal- directed reaching performance between children with DS and typically developing children during standing reach. Standing reach performance was analyzed in three main phases using the kinematic and kinetic data collected from a force plate and a motion capture system. Fourteen children with DS, age and gender matched with fourteen typically developing children, were recruited for this study. The results showed that the demand of the standing reach task affected both dynamic postural control and reaching performance in children with DS, especially in the condition of beyond armÍs length reaching. More postural adjustment strategies were recruited when reaching distance was beyond armÍs length. Children with DS tended to use inefficient and conservative strategies for postural stability and reaching. That is, children with DS perform standing reach with increased reaction and execution time and decreased amplitudes of center of pressure displacements. Standing reach resembled functional balance that is required in daily activities. It is suggested to be considered as a part of strength and balance training program with graded task difficulty.

Recommended citation:
Chen, H.-L., Yeh, C.-F., & Howe, T.-H. (2015). Postural control during standing reach in children with Down syndrome. Res Dev Disabil.. 38(0), 345-351. doi: http://dx.doi.org/10.1016/j.ridd.2014.12.024</br>