Projects
In Progress
Out-of-distribution detection in medical image segmentation
Propose joint training of segmentation and energy-based models for out-of-distribution (OOD) detection in medical imaging. The proposed method can reject incorrectly prepared (or preprocessed) medical images. It can also distinguish rare test samples unseen from the training set. Additionally, we show correlations between OOD scores and anatomical regions of medical imaging
* Currently working on manuscript for conference proceedings
* Joint research project at CCIDS, Yonsei University Severance hospital, with professor Jaegul Choo at KAIST AI Graduate school, South Korea.
Developing National Standard of Performance Evaluation for Diagnosis and Prediction of Pandemic (COVID-19)
There exist many studies of developing severity prediction models for pulmonary infectious diseases, such as COVID-19 or viral/bacterial pneumonia. However, reliable comparison between each methods and objective evaluation of predictive performance is difficult, since each of these studies rely on arbitrarily defined evaluation metrics. This project aims at proposing reliable, and unified performance metric that can be adopted to existing algorithms. The other major goal of this project is to develop an integrated model that uses imaging data (CT, X-ray) together with clinical data for virus diagnosis and finding prognostic biomarkers of the disease.
* Research Project at CCIDS, Yonsei University Severance Hospital
* This project is funded by the Ministry of Science and ICT & National IT Industry Promotion Agency
Segmentation and volume estimation of aneurysm
Propose methods for segmentation of aneurysm based on deep-learning. Two possible scenarios on track. 1) Classify image patches with aneurysm then perform lesion segmentation, 2) Detect aneurysm VOI using object detection model then perform lesion segmentation.
* Research Project at Neurophet, South Korea, Seoul
Ischemic stroke segmentation and quantification of severity
Develop automatic system for detecting acute infarct in DWI MR brain imaging based on deep learning. Used U-Net based model for lesion segmentation. Adopted Apparent Diffusion Coefficient (ADC) values for quantification of stroke severity. External validation of the developed method in progress.
* Research Project at CCIDS, Yonsei University Severance Hospital
* Hold South Korean patent for the developed algorithm
*This work was accepted at RSNA 2020.
*Currently working on manuscript for journal publication
Finding integrated imaging biomarker for acute infarct
Find imaging biomarker for acute infarct using radiomics features along with image features extracted from the encoder of segmentation model. Project aim is to calculate risk factor using integrated features.
* Research Project at Neurophet, South Korea, Seoul
LVRR Prediction via Radiomic Analysis of Myocardium region
Predicting left ventricular reverse remodeling (LVRR) through machine learning analysis on radiomics features extracted from myocardium region of left ventricle (LV). LV region can be divided into apical, basal, and mid parts. Three parts can once again divided into six anatomical regions, respectively. Radiomic analysis on each part of LV and ensemble of each result are expected to be a good biomarker for LVRR in myocardium region.
* Research Project at CCIDS, Yonsei University Severance Hospital
Cardiac MRI Standardization
Cardiovascular MR (CMR) imaging protocol varies between institutions and hospitals. Study on how much modern techniques for CMR analysis are sensitive to standardization. Through comparison test on standardized imaging and un-standardized imaging, address the importance of standardization.
* Research Project at CCIDS, Yonsei University Severance Hospital
Past Projects
Automatic diagnosis of cardiomyopathy and prognosis prediction
Track myocardial motion during erection and dilation through image segmentation based on deep-learning. Estimate volume of myocardium region at each phase (erection and dilation), in order to detect cardiovascular disability.
* Research Project at CCIDS, Yonsei University Severance Hospital
MENet: Mixture of Experts of Classification
Lightweight deep learning models, such as MobileNet and MixNet, usually lack in accuracy compared to large-size models. In order to boost the performance of light weight models, use mixture of experts method using Gumbel Softmax trick.
* Graduation Project at Yonsei University, Department of Computer Science
Autonomous Parking System
Deep learning-based autonomous driving system that allows vehicles to park autonomously in an indoor parking lot. Real-time CCTV videos are transmitted to the main frame. Deep learning model analyzes the video then detects the possible parking spot, providing guidance to users in real-time.
* Project at Cube Intelligence, South Korea, Seoul
DEraser - Detect and Erase
Detect the selected object using Mask R-CNN and erase the object using deep-learning based inpainting model. We selected EdgeConnect model as our baseling inpainting model.
* Graduation Project at Yonsei University, Department of Computer Science
Bone Cyst Detection
Multi instance semantic segmentation for detecting cyst out of bone using modified U-Net.
* Joint Research Project at Deep Noid, South Korea, Seoul, with Kyungpook National University Hospital