Medical Image Processing
Advanced CT Reconstruction
In the past, we reconstructed CT images using several hundred projection images with Cone beam CT reconstruction algorithm (FDK). The low dose algorithm can show the same image quality by using statistical-based compressed sensing algorithm and using only several tens of projection images.The time required for reconstruction of the CBCT data using the GPU was reduced to 1/6.
Intelligent Medical Imaging
We have been studying image processing using machine learning based algorithm called ApellesNET by modifying existing UNET.
We are studying Auto-Contouring of tumor and normal tissues essential for radiotherapy, reconstruction of CT images form MR images, and machine learning based low does CBCT reconstruction methods for image-guided radiation therapy (IGRT).
External Beam Therapy
Radiation Therapy
Although the cancer treatment method is very important because there are 4 million companion animals diagnosed as cancer, which is 2.5 times as many as the number of people in the United States, there is no radiation cancer treatment system optimized for companion animals.
We are studying customized low energy radiotherapy using 3D printing compensator.
Particle Therapy
Yonsei Medical Center is conducting research for the first time in Korea to start baryon therapy in 2022.
We are studying Monte Carlo simulation to calculate accurate doses, optimization of baryon therapy plan, and integrated study of hardware and software to treat moving tumors and organs.
X-ray Image Systems
Cone-Beam CT
Produce volumetric three-dimensional images. Provide distinctive volumetric and morphological information of the patient than conventional two-dimensional images.
Tomosynthesis
Provide cross sectional images from a finite number of projection views within a limited data acquisition range.
Spectral CT
Utilize the multiple energy spectrum of X-ray and Material Differentiation is possible. Increase the contrast of the medical images and enable the noise reduction.
Image Quality Evaluation
MTF
Ratio of the output modulation. Quantitative metric describing the resolution performance of X-ray based medical imaging system.
NPS
Quantitative metric describing the noise including correlation introduced by filtering and other processing steps.
Observer
Evaluate lesion detection performance of medical imaging system using human or mathematical models mimicking the human visual system.
Computer Graphics
Inverse Lighting
This paper introduces a novel approach for robustly extracting lighting conditions from an RGB-D (RGB + depth) image. The proposed method takes non-homogeneous surface objects into account in the inverse lighting framework via segment-based scene representation. Moreover, we employ outlier removal and appropriate region selection to achieve robust lighting estimation in the presence of inter-reflections and noise.
Generative Adversarial Networks
Face Recognition for Low-shot Learning
Recently, low-shot learning has been proposed for handling the lack of training data in machine learning. In this paper, we aim to increase the size of training dataset in various ways to improve the accuracy and robustness of face recognition. In detail, we adapt a generator from the Generative Adversarial Network (GAN) to increase the size of training dataset, which includes a base set, a widely available dataset, and a novel set, a given limited dataset, while adopting transfer learning as a backend.
Object localization using GAN
This paper proposes an improved technique for weakly-supervised object localization. Conventional methods have a limitation that they focus only on most discriminative parts of the target objects. The proposed method employes an effective data augmentation for improving the accuracy of the object localization. In addition, we introduce improved learning methods by optimizing CNN based on the state-of-the-art model.
Improved Training of Generative Adversarial Networks Using Representative Features
We aim to enhance both visual quality and image diversity simultaneously by improving the stability of training GANs. A key idea of the proposed approach is to implicitly regularizing the discriminator using a representative feature.
Identity preserving face relighting using GAN
Identity preserving is a challenging problem of face image generation using GAN. Realistic synthesis of relighting face image is one way to augment low-shot data. Physical-based intrinsic image disentangling approach induce semantic separation of latent variables.
Environment Map From A Single Image
We propose a new approach to predicting the surrounding environment map from a single image of the scene using deep adversarial networks. Using the state-of-the-art model as a baseline network, we employ the adversarial loss and confidence masks for improving the quality of estimated environment maps. Based on the comparisons on the benchmark database, we show the effectiveness of the proposed approach for increasing the visual quality and stability of training.
Multimodal Image Reconstruction
Depth reconstruction of Translucent object
This study presents a robust approach to reconstructing a three dimensional (3-D) translucent object using a single time-of-flight depth camera with simple user marks. In this study, we introduce a ground plane and piece-wise linear surface model as priors and construct a robust 3-D reconstruction framework for translucent objects. These two depth priors are combined with the depth error model built on the time-of-flight principle. Extensive evaluation of various real data reveals that the proposed method substantially improves the accuracy and reliability of 3-D reconstruction for translucent objects.
Depth image enhancement using perceptual texture priors
Due to the limitation of power consumption or hardware, the depth camera presents severe noises, incapable of providing the high quality 3D data. Our concept is replacing bad normals data taken from the depth camera with fine normals data that is best matched one with input from the database.
Transform Photograph
The algorithm which transform a ordinary photograph into an attractive one by the use of database and image completion
We develop an algorithm, which transforms an ordinary photograph into what is likely to draw multiple attentions in social networks using image completion method. We collect a huge database from web, which reflects those favorable features. Unlike general image completion techniques, the algorithm maximizes aesthetic value of a photograph cutting fragments from the picture, patching up the missed parts with matching scenes from our database, but still maintaining the original theme. Any user, even though the one is not good at taking the pictures, can get popular photographs with our algorithm whose performance will get enhanced as more users are involved and give feedback.
Face Detection
Face Classification and Detection in Driving Environment
Face Detection in driving environment is hard problem. In this work, we aim to train face detection and classification network which target to operate well on driving scenario. We first collect data of car environment, processing the data to generate database. We change conventionally used YOLO v2 network structure to fit well on our purpose. Classification IOU is average 83.59% and it can classify and detect each people's face in about 25 fps.
Weight-bearing C-arm CT
Weight-Bearing Imaging Of The Knee Using C-Arm CT
The overarching, long-term goal of this study is to develop novel in-vivo weight-bearing computed tomography (CT) imaging to expand our understanding of the mechanical stresses that affect knee cartilage and meniscus health. In particular, we are developing and optimizing new protocols to obtain quantitatively accurate measurements of in vivo articular cartilage thickness using C-arm CT at the resolution limit of the system of 210 µm isotropic. Specifically, we are working on new algorithms to correct for involuntary motion and for dynamic range-related artifacts, and to automatically segment cartilage. In addition, we are developing strategies to quantify time-dependent deformation of cartilage as a sensitive and early indicator of cartilage health and an important component in the evaluation of osteoarthritis status.
(Left) Acquisition set up in weight-bearing configuration using a C-arm CT. (Right) Sagittal slice of the reconstructed image of the knee. Thanks to contrast injection, the profile of femural and tibial cartilage is clearly visible.
X-ray Phase Contrast Imaging
X-RAY PHASE-CONTRAST (PC) IMAGING AND TOMOGRAPHY
The field of X-ray imaging is currently undergoing a revolution due to the development of X-ray phase-contrast (PC) imaging methods that have dramatic advantages over conventional radiographic X-ray imaging systems. Conventional X-ray imaging methods are based on absorption effects and therefore image contrast is due to differences only in the X-ray absorption properties of the tissues. This can be problematic when imaging soft tissues that do not possess large variations in X-ray absorption properties. There remains an important need for optimization, improvement, characterization, and validation of PC imaging methods, to make them suitable for routine and widespread use. Our research group is actively developing image formation methods for X-ray PC imaging and is collaborating with several groups (including Electronics and Telecommunications Research Institute (ETRI), Yonsei university, and Stanford University) and other institutions to translate the technology to address important needs in clinical and preclinical medicine.
An x-ray far-field interferometer using only phase gratings is based on the phase moiré effect. The mid grating forms Fourier images of the first grating. These images beat with the 3rd grating to produce broad moiré fringes on the detector at the appropriate distance. Phase shifts and de-coherence of the wavefront by the object cause fringe shifts and attenuation of the fringe contrast. (Source: https:
Deep Learning-based Image Processing in Fluoroscopy
This project aims to develop a C-arm-based real-time X-ray fluoroscopy with fast and variable frame-rated digital X-ray sources for continuous patient monitoring. To do this, we will develop: (1) a noise reduction technology for low-dose images using deep learning for real-time image processing, and (2) a detector lag (i.e., residual signal) correction algorithm to improve image quality.
The detailed goals are as follows:
X-ray fluoroscopy will be optimized for a deep learning-based convoluted neural network technology to reduce noise in real time at low-dose images.
A new nonlinear correction algorithm will be developed to compensate for the detector lag on C-arm-based X-ray fluoroscopy images.
Lastly, the developed techniques will be optimized to allow fluoroscopy to be used for real-time monitoring. Particularly, parallel operation using GPU will be realized so that real-time processing can be performed with minimum computation time.
Collaborators: Electronics and Telecommunications Research Institute, Yonsei University
Research Grant: 2017~2019 미래창조과학부 방사선융합기술 사업
Biomedical Image Informatics Technology for Precision Medicine
This project aims to develop core technologies for bio·medical image informatics based on the construction of patient-derived bio·medical heterogeneous big data. By acquiring those original technologies for precision medicine, we finally aim to achieve practical use of personalized medicine which is suitable for the genetic characteristics of Koreans.
