Myung Jin Chung

Sungkyunkwan University, Sŏul, Seoul, South Korea

Are you Myung Jin Chung?

Claim your profile

Publications (340)343.15 Total impact

  • [Show abstract] [Hide abstract]
    ABSTRACT: Subcentimeter nodules without change in size during long-term follow-up period (for minimum 2 years) are assumed as benign lesions. However, the 2-year stability rule has not been fully verified so far and is still questionable. Thus, we aimed to retrospectively investigate long-term follow-up results for 2-year stable subcentimeter nodules at screening low-dose computed tomography (LDCT).
    Respirology 06/2014; · 2.78 Impact Factor
  • Yungeun Choe, Seunguk Ahn, Myung Jin Chung
    [Show abstract] [Hide abstract]
    ABSTRACT: Urban object recognition is the ability to categorize ambient objects into several classes and it plays an important role in various urban robotic missions, such as surveillance, rescue, and SLAM. However, there were several difficulties when previous studies on urban object recognition in point clouds were adopted for robotic missions: offline-batch processing, deterministic results in classification, and necessity of many training examples. The aim of this paper is to propose an urban object recognition algorithm for urban robotic missions with useful properties: online processing, classification results with probabilistic outputs, and training with a few samples based on a generative model. To achieve this, the proposed algorithm utilizes the consecutive point information (CPI) of a 2D LIDAR sensor. This additional information was useful for designing an online algorithm consisting of segmentation and classification. Experimental results show that the proposed algorithm using CPI enhances the applicability of urban object recognition for various urban robotic missions.
    Robotics and Autonomous Systems 01/2014; · 1.16 Impact Factor
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: In recent years, the research of 3D mapping technique in urban environments obtained by mobile robots equipped with multiple sensors for recognizing the robot's surroundings is being studied actively. However, the map generated by simple integration of multiple sensors data only gives spatial information to robots. To get a semantic knowledge to help an autonomous mobile robot from the map, the robot has to convert low-level map representations to higher-level ones containing semantic knowledge of a scene. Given a 3D point cloud of an urban scene, this research proposes a method to recognize the objects effectively using 3D graph model for autonomous mobile robots. The proposed method is decomposed into three steps: sequential range data acquisition, normal vector estimation and incremental graph-based segmentation. This method guarantees the both real-time performance and accuracy of recognizing the objects in real urban environments. Also, it can provide plentiful data for classifying the objects. To evaluate a performance of proposed method, computation time and recognition rate of objects are analyzed. Experimental results show that the proposed method has efficiently in understanding the semantic knowledge of an urban environment.
    The Journal of Korea Robotics Society. 01/2014; 9(1).
  • [Show abstract] [Hide abstract]
    ABSTRACT: PURPOSE To compare the radiologists’ performance for the evaluation of diffuse interstitial lung disease (DILD) on CT images reconstructed with Model Based Iterative Reconstruction (MBIR), Adaptive Statistical Iterative Reconstruction (ASIR), and conventional Filtered Back Projection (FBP) METHOD AND MATERIALS Included were 23 subjects (5 normal, 11 usual intersitial pneumonia(UIP), and 7 nonspecific interstitial pneumonia(NSIP); 13 men, 20 women; mean age, 64 ± 6 years) who underwent high dose thin section non-helical CT (HRCT, 1.25mm thickness, 100 – 300mA with dose modulation) and low dose helical CT (LDCT, 0.625mm x 64 rows with 0.984 pitch, fixed 40mA with 0.5 sec rotation). CTDI was 1.87 ± 0.01 mGy and 21.5 ± 5.5 mGy for LDCT and HRCT, respectively. LDCT was reconstructed using FBP, 50% ASIR, and MBIR. Matching images from four subsets were blinded and randomly sorted. Four board certified radiologists independently reviewed shuffled images. Semi-quantitative results for the extent of abnormalities (GGO, ground glass opacity; RET, reticular or honeycombing opacity; CON, consolidation) and confidence scores for disease status (normal vs. abnormal; UIP vs. NSIP) were compared. RESULTS The difference of overall diagnosis in presence of disease did not reach statistical significance (The AUCs for detection of DILD by HRCT, FBP, ASIR and MBIR were as follows: 0.978, 0.979, 0.972 and 0.963.). There was a trend of underestimation than HRCT of RET (-2.8%, -4.1%, and -5.3% in FBP, ASIR, and MBIR, respectively) and overestimation of GGO (+4.6%, +8.9%, and +8.5% in FBP, ASIR, and MBIR, respectively). However, no significant difference was found among above mentioned reconstruction methods for decision making on type of DILD (The AUCs for differentiation of UIP vs. NSIP by HRCT, FBP, ASIR and MBIR were as follows: 0.780, 0.805, 0.785 and 0.778). CONCLUSION LDCT with MBIR showed similar diagnostic performance to HRCT for typical DILD cases with reduced image noise in spite of much lower dose (less than 1/10). However, our study confirmed more caution is needed for comparing the disease extent especially on follow up studies reconstructed with iterative recontruction (IR) due to its possible influence on characterization of the interstitial lung disease pattern. CLINICAL RELEVANCE/APPLICATION IR has potential to distort the perception of extent and character of DILD. Further caution is needed for radiologists in evaluating DILD using IR.
    Radiological Society of North America 2013 Scientific Assembly and Annual Meeting; 12/2013
  • [Show abstract] [Hide abstract]
    ABSTRACT: PURPOSE To compare the capability of lung nodule detection and characterization between Dual-energy radiography with color-representation (DCR) and conventional gray scale chest radiography (GSR) METHOD AND MATERIALS A total of 130 paired chest radiographs (DCR and GSR) taken from 14 normal subjects and 51 patients with pulmonary nodules were evaluated. The 51 patients had 45 non-calcified and 21 calcified nodules. DCR was obtained by adding color space within material-decomposed data (applying blue color for high attenuating component and red color for low attenuating component) and by compounding the manipulated data to one color image. Three radiologists marked suggested nodules on radiographic images and assessed the confidence of lesion presence and probability of nodule calcification by using a nine-point rating scale. The jackknife active free-response receiver operating characteristics (JAFROC) analysis was used to evaluate lesion detectability, and multi-reader multi-case receiver operating characteristics (MRMC ROC) analysis was used for the evaluation of the accuracy of nodule calcification prediction. RESULTS Figures of merit (FOM) from JAFROC were 0.806 for DCR and 0.811 for GSR, respectively; thus, nodule detectability was not significantly different between DCR and GSR (P = .93). Areas under curve (AUC) from MRMC ROC were 0.927 for DCR and 0.781 for GSR, respectively; thus, performance of DCR in predicting lung nodule calcification was significantly better than that of GSR (P = .04). CONCLUSION DCR shows similar performance in lung nodule detection compared with GSR. However, DCR does provide a significant benefit in predicting the presence of nodule calcification. CLINICAL RELEVANCE/APPLICATION DCR has additional value for lung nodule characterization to conventional gray scale radiography.
    Radiological Society of North America 2013 Scientific Assembly and Annual Meeting; 12/2013
  • [Show abstract] [Hide abstract]
    ABSTRACT: PURPOSE To create virtual non-contrast CT, two-material decomposition is allowed from Spectral CT based on sinogram space, instead of three-material decomposition method allowed in image space. However, various virtual non-contrast images can be made from various material decomposition (MD) methods in spectral CT. We evaluated the feasibilities of three different virtual non-contrast (VNC) images derived from single source dual energy spectral CT compared to true non-contrast (TNC) image. METHOD AND MATERIALS This HIPAA-compliant study was approved by institutional review board and informed consent was provided from all patients. Twenty-four patients prospectively underwent non-contrast CT followed by contrast enhanced chest CT using single source fast kVp switching dual energy scan. Iodine eliminated images so called as VNC were reconstructed using two kinds of 2-material decomposition algorithms (MDW, material density-iodine/water; MDC, material density-iodine/calcium) and material suppression algorithm (MSI, iodine suppression image). Using third party workstation, semiautomatic calcium measurements were performed. RESULTS Quantified calcium scores (AJ score) from all three VNCs correlated well with that of TNC (R2 = 0.95, 0.88, and 0.88 for MDW, MDC, and MSI, respectively). However correlation coefficients were less than 0.9 (C = 0.83, 0.62, and 0.63 for MDW, MDC, and MSI, respectively). Measured calcium volumes on VNCs also correlated well with that of TNC (R2 = 0.94, 0.87, and 0.90 for MDW, MDC, and MSI, respectively), with correlation coefficients of 0.78, 0.59, and 0.63 for MDW, MDC, and MSI, respectively. Among the three VNCs, MDW correlated best with TNC. CONCLUSION VNC image from contrast enhanced CT using dual energy material decomposition/suppression is feasible for coronary calcium scoring. However, among various methods to make virtual noncontrast image from spectral CT, material quantifications are different depending on the decomposition methods. Furthermore, the absolute value on VNC tends to be smaller than that on TNC and should be considered with calibration. CLINICAL RELEVANCE/APPLICATION Absolute values of calcium scoring on VNC tend to be smaller than that on TNC and should be considered with calibration.
    Radiological Society of North America 2013 Scientific Assembly and Annual Meeting; 12/2013
  • [Show abstract] [Hide abstract]
    ABSTRACT: PURPOSE To evaluate the clinical characteristics and prognostic implications of patients with lung adenocarcinoma presenting with concurrent lymphangitic metastasis at CT METHOD AND MATERIALS We retrospectively reviewed the clinical data base of patients who were newly diagnosed to have non-small cell lung cancer (NSCLC) from 2007 through 2010. We searched for the cases in which radiologic report text harbored the term “lymphangitic metastasis” and clinically ensured that the findings are suggestive of lymphangitic metastasis by excluding those who were proven to have pulmonary edema on follow-up study. The extent of lymphangitic metastasis was classified into having involved the same lobe, the ipsilateral lung but the different lobe, and the different lung. Two chest radiologists reviewed CT scans and decisions were reached by consensus. We excluded the patients with non-adenocarcinoma pathology or with surgery or irradiation history. After restaging with CT and PET/CT results by the new 7th TNM classification, CT and clinico-demographic data were investigated for those in subject group (with lymphangitic metastasis) and in control group (without lymphangitic metastasis). New clinical stage group, sex, age and history of chemotherapy were matched to assess the impact of lymphangitic metastasis on patient overall survival. RESULTS The subject group (n = 54) consisted of four stage IIIB (7.4%) and 50 stage IV (93.6%) lung adenocarcinoma patients (34 men, 20 women; mean age 59 ± 10 years). Of these patients, lymphangitic metastasis was depicted in the same lobe in 10, in the ipsilateral lung but different lobe in 32 and in the different lung in 12 patients. Forty two patients received chemotherapy and 12 patients did not. There was no significant difference in overall survival between lung adenocarcinoma patients with and without lymphangitic metastasis (p = .758). The extent of lymphangitic metastasis also showed no significant correlation with patient prognosis (p = .121). CONCLUSION Stage IIIB or IV lung adenocarcinoma patients with lymphangitic metastasis at the initial CT does not show significantly different prognosis compared with those without the condition, and the extent of lymphangitic metastasis also does not affect on patient overall survival. CLINICAL RELEVANCE/APPLICATION In advanced stage (IIIB and IV) lung adenocarcinoma, the presence of lymphangitic metastasis or its extent does not have any impact on patient prognosis.
    Radiological Society of North America 2013 Scientific Assembly and Annual Meeting; 12/2013
  • [Show abstract] [Hide abstract]
    ABSTRACT: PURPOSE To retrospectively investigate long-term follow-up results for 2-year-stable subcentimeter nodules seen at screening low-dose CT (LDCT). METHOD AND MATERIALS A total of 635 subjects, who had follow-up low-dose CT (LDCT) for the initial two-year screening period and for additional three years thereafter and who had noncalcified subcentimeter nodules, were included. By using nodule volumetry software, we measured interval change of nodule volume. Positive growth was defined as an increase in volume of at least 25% between two volume measurements. RESULTS A total of 1107 nodules (1037 solid, 70 ground-glass opacity nodules [GGNs]) were detected at baseline CT. Of 1037 solid nodules, 1032 (99.5%) showed no growth at initial two-year follow-up CT, while of 70 GGNs, 59 (84.3%) showed no growth. Of 1032 solid subcentimeter nodules showing no growth for initial two-year follow-up period, none showed further growth during additional three-year follow-up CT, whereas of 59 GGNs stable for initial two years, two (3.4%) showed growth to be proved as adenocarcinomas. Of five solid nodules depicting growth at initial two-year follow-up CT, one (20%) proved to be adenocarcinoma, whereas of 11 GGNs demonstrating growth for the initial two-year follow-up CT, four (36.4%) showed growth and proved to be lung cancers. CONCLUSION All solid subcentimeter nodules having initial two-year stability at screening LDCT can be considered benign, because none shows growth at further follow-up CT. On the other hand, subcentimeter GGNs have a more chance of growth than solid nodules and need further follow-up CT for more than two years. CLINICAL RELEVANCE/APPLICATION This study demonstrates the two-year stability rule for subcentimeter solid nodules in LDCT using volumetry and CT follow up for more than two years seems to be mandatory for subcentimeter GGNs.
    Radiological Society of North America 2013 Scientific Assembly and Annual Meeting; 12/2013
  • Sijong Kim, Jungwon Kang, Myung Jin Chung
    [Show abstract] [Hide abstract]
    ABSTRACT: In this paper, we describe a probabilistic voxel mapping algorithm using an adaptive confidence measure of stereo matching. Most of the 3D mapping algorithms based on stereo matching usually generate a map formed by point cloud. There are many reconstruction errors. The reconstruction errors are due to stereo reconstruction error factors such as calibration errors, stereo matching errors, and triangulation errors. A point cloud map with reconstruction errors cannot accurately represent structures of environments and needs large memory capacity. To solve these problems, we focused on the confidence of stereo matching and probabilistic representation. For evaluation of stereo matching, we propose an adaptive confidence measure that is suitable for outdoor environments. The confidence of stereo matching can be reflected in the probability of restoring structures. For probabilistic representation, we propose a probabilistic voxel mapping algorithm. The proposed probabilistic voxel map is a more reliable representation of environments than the commonly used voxel map that just contains the occupancy information. We test the proposed confidence measure and probabilistic voxel mapping algorithm in outdoor environments.
    Intelligent Service Robotics 04/2013; 6(2).
  • [Show abstract] [Hide abstract]
    ABSTRACT: Background Digital tomosynthesis considerably reduces problems created by overlapping anatomy compared with chest X-ray (CXR). However, digital tomosynthesis requires a longer scan time compared with CXR, and thus may be vulnerable to motion artifacts.PurposeTo compare the diagnostic performance of digital tomosynthesis in subjects with and without respiratory motion artifacts.Material and Methods The institutional review board approved this retrospective study, and the requirement for written informed consent was waived. A total of 46 subjects with imaging containing respiratory motion artifacts were enrolled in this study, 18 of whom were positive and 28 of whom were negative for lung nodules on computed tomography (CT). The control group was comprised of 92 age-matched subjects with imaging devoid of motion artifacts. Of these, 36 were positive and 56 were negative for lung nodules on subsequent CT scan. The size criteria of nodules were 4-10 mm. Three chest radiologists independently evaluated the radiographs and digital tomosynthesis images for the presence of pulmonary nodules. Multireader multicase receiver-operating characteristic (ROC) analyses was used for statistical comparisons.ResultsWithin the control group, the areas under curve (AUC) for observer performances in detecting lung nodules on digital tomosynthesis was higher than that on CXR (P = 0.017). Within the study group, there were no significant differences in AUCs for observer performances (P = 0.576).Conclusion When no motion artifacts are present, the detection performance of nodules (4-10 mm) on digital tomosynthesis is significantly better than that on CXR, whereas there is not a significant difference in cases with motion artifacts.
    Acta Radiologica 03/2013; · 1.33 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: This paper suggests an interactive facial robot system on a smart device which has a touch screen and a built-in microphone. The recognition process for touch inputs is enhanced by analyzing input patterns and a built-in microphone. Recognized results from the input procedure are converted to emotional states of the system, and then the emotional states are reactively expressed at a facial simulator displayed on the device's touch screen. Therefore, the proposed facial system can be implemented in one smart device at which input sensors and a visual output are on a same display component.
    Human-Robot Interaction (HRI), 2013 8th ACM/IEEE International Conference on; 01/2013
  • Hyun Chul Roh, Chang Hun Sung, Myung Jin Chung
    [Show abstract] [Hide abstract]
    ABSTRACT: Simultaneous Localization and Mapping is the one of essential techniques for mobile robot navigation. In this paper, we propose rapid SLAM using simple map representation in indoor environment for mobile robot. This approach offers a new way to look at the problem focusing on the issues that have caused the use of 3D laser scanner which provide lot of 3D point data. We have tried to create the simple segments of line and range table for scan-matching in a way that allows a robust solution to the problem. In this article, two important issues of this work on 2D registrations are made. First, it is shown that the algorithm performs very well on the transformations of segmented line map from lot of 3D point cloud data. Second the mapping logic and sequence are explained in this paper to a line component accumulation for a map building. Experimental results from 3-D sensor Velodyne and wheel-odometry data logger Racelogic Vbox are given. Experimental results show that this approach is not only robust for line mapping but it is also fast, requires significantly less memory.
    Frontiers of Computer Vision, (FCV), 2013 19th Korea-Japan Joint Workshop on; 01/2013
  • Jungwon Kang, Si Jong Kim, Myung Jin Chung
    [Show abstract] [Hide abstract]
    ABSTRACT: This paper deals with the improvement of clustering results by enhancing performance of sampling. The proposed clustering framework is established on a J-linkage framework [4] with guided sampling technique for multi-structures [7]. We tested the proposed method on publicly available dataset, verifying the validness of the proposed method.
    Ubiquitous Robots and Ambient Intelligence (URAI), 2013 10th International Conference on; 01/2013
  • [Show abstract] [Hide abstract]
    ABSTRACT: This paper suggests a system to detect multiple planes fast and roughly based on a depth map from depth camera ‘Kinect’. In order to improve the speed of the plane detection process, we compute local normal vectors of points of the depth map and we classify 3D point cloud data whether it includes in the same plane using these local normal vectors. It is possible to detect multiple planes simultaneously. Experimental results show that the proposed method is faster than previous plane detection methods such as 3D Hough Transform and RANSAC, and also works in real-time.
    Robotics (ISR), 2013 44th International Symposium on; 01/2013
  • Source
    Hye Sun Hwang, Myung Jin Chung, Kyung Soo Lee
    [Show abstract] [Hide abstract]
    ABSTRACT: To determine the optimum low dose (LD) digital tomosynthesis (DT) setting, and to compared the image quality of the LD DT with that of the standard default (SD) DT. Nine DT settings, by changing tube voltage, copper filter, and dose ratio, were performed for determining the LD setting. Among combinations of DT setting, a condition providing the lowest radiation dose was determined. Eighty artificial nodules less than 1 cm in diameter (subcentimeter nodules: 40, micronodules less than 4 mm: 40) were attached to a Styrofoam and a diaphragm of the phantom. Among these, 38 nodules were located at the periphery of the lung (thin area) and 42 nodules were located at the paravertebral or sub-diaphragmatic area (thick area). Four observers counted the number of nodules detected in the thick and thin areas. The detection sensitivity in SD and LD settings were calculated separately. Data were analyzed statistically. The lowest LD setting was a combination of 100 kVp, 0.3 mm additional copper filter, and a 1 : 5 dose ratio. The effective dose for the LD and SD settings were 62 µSv and 140 µSv, separately. A 56.7% dose reduction was achieved in the LD setting compared with the SD setting. Detection sensitivities were not different between the SD and the LD settings except between observers 1 and 2 for the detection of micronodules in the thick area. LD DT can be effective in nodule detection bigger than 4 mm without a significant decrease in image quality compared with SD DT.
    Korean journal of radiology: official journal of the Korean Radiological Society 01/2013; 14(3):525-531. · 1.32 Impact Factor
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: To compare the diagnostic performance of light emitting diode (LED) backlight monitors and cold cathode fluorescent lamp (CCFL) monitors for the interpretation of digital chest radiographs. We selected 130 chest radiographs from health screening patients. The soft copy image data were randomly sorted and displayed on a 3.5 M LED (2560 × 1440 pixels) monitor and a 3 M CCFL (2048 × 1536 pixels) monitor. Eight radiologists rated their confidence in detecting nodules and abnormal interstitial lung markings (ILD). Low dose chest CT images were used as a reference standard. The performance of the monitor systems was assessed by analyzing 2080 observations and comparing them by multi-reader, multi-case receiver operating characteristic analysis. The observers reported visual fatigue and a sense of heat. Radiant heat and brightness of the monitors were measured. Measured brightness was 291 cd/m(2) for the LED and 354 cd/m(2) for the CCFL monitor. Area under curves for nodule detection were 0.721 ± 0.072 and 0.764 ± 0.098 for LED and CCFL (p = 0.173), whereas those for ILD were 0.871 ± 0.073 and 0.844 ± 0.068 (p = 0.145), respectively. There were no significant differences in interpretation time (p = 0.446) or fatigue score (p = 0.102) between the two monitors. Sense of heat was lower for the LED monitor (p = 0.024). The temperature elevation was 6.7℃ for LED and 12.4℃ for the CCFL monitor. Although the LED monitor had lower maximum brightness compared with the CCFL monitor, soft copy reading of the digital chest radiographs on LED and CCFL showed no difference in terms of diagnostic performance. In addition, LED emitted less heat.
    Korean journal of radiology: official journal of the Korean Radiological Society 01/2013; 14(6):968-976. · 1.32 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: Here we present 3D world modeling and its quantitative analysis methods in urban environment. If the expensive RTK GPS cannot be prepared, it is difficult to measure the accuracy of the 3D world model due to the blackout of GPS particularly in urban environment. To cope with this difficulty, we combine to process both satellite image and point cloud to compare each other in order to represents accuracy of 3D world model. We also introduce 3D world modeling method through localization algorithm and global registration method in order to validate our quantitative analysis. In the experimental result, we describe our sensor system and evaluate the proposed quantitative analysis method using 3 different localization algorithm. Our framework is suitable of mobile mapping system in urban environment in terms of cost.
    Control Conference (ASCC), 2013 9th Asian; 01/2013
  • Hyun Chul Roh, Yungeun Choe, Myung Jin Chung
    [Show abstract] [Hide abstract]
    ABSTRACT: Digital map can be a good materials for some robotic problems such as localization, segmentation, classification, SLAM algorithm. In this paper, we propose building wall extraction algorithm from land registration map image using corner, edge detection methods and simple matching approach. Experimental results show that this algorithm is valid for general land registration map.
    Ubiquitous Robots and Ambient Intelligence (URAI), 2013 10th International Conference on; 01/2013
  • [Show abstract] [Hide abstract]
    ABSTRACT: In this paper we describe a method for solving a mobile robot localization problem using prior data. By matching 2D image features to a 3D point cloud, the robot position is estimated in the prior point cloud. We prove our method by testing at specific locations over the whole point clod data.
    Ubiquitous Robots and Ambient Intelligence (URAI), 2013 10th International Conference on; 01/2013
  • [Show abstract] [Hide abstract]
    ABSTRACT: In this paper, emotional motion representation is proposed for Human Robot Interaction: HRI. The proposed representation is based on “Laban Movement Analysis: LMA” and trajectories of 3-dimensional whole body joint positions using an RGB-D camera such as a “Microsoft Kinect”. The experimental results show that the proposed method distinguishes two types of human emotional motion well.
    Human-Robot Interaction (HRI), 2013 8th ACM/IEEE International Conference on; 01/2013

Publication Stats

3k Citations
343.15 Total Impact Points

Institutions

  • 2004–2014
    • Sungkyunkwan University
      • • Samsung Medical Center
      • • Department of Radiology
      Sŏul, Seoul, South Korea
    • Seoul National University Hospital
      • Department of Radiology
      Seoul, Seoul, South Korea
  • 2013
    • Hallym University Medical Center
      • Department of Radiology
      Seoul, Seoul, South Korea
  • 1970–2013
    • Korea Advanced Institute of Science and Technology
      • Department of Electrical Engineering
      Sŏul, Seoul, South Korea
  • 2012
    • Bundang Jesaeng Hospital
      Sŏngnam, Gyeonggi Province, South Korea
  • 2006–2008
    • Samsung Medical Center
      • Department of Radiology
      Seoul, Seoul, South Korea
  • 2005
    • Soonchunhyang University
      Onyang, South Chungcheong, South Korea
  • 2002–2004
    • Hanbat National University
      Daiden, Daejeon, South Korea
  • 2001
    • Agency for Defense Development
      Sŏul, Seoul, South Korea
  • 1996–1998
    • Korea Institute of Science and Technology
      Sŏul, Seoul, South Korea