• Clinical Endoscopy
• /
• v.55 no.3
• /
• pp.390-400
• /
• 2022

Background/Aims: Previous artificial intelligence (AI) models attempting to segment gastric intestinal metaplasia (GIM) areas have failed to be deployed in real-time endoscopy due to their slow inference speeds. Here, we propose a new GIM segmentation AI model with inference speeds faster than 25 frames per second that maintains a high level of accuracy. Methods: Investigators from Chulalongkorn University obtained 802 histological-proven GIM images for AI model training. Four strategies were proposed to improve the model accuracy. First, transfer learning was employed to the public colon datasets. Second, an image preprocessing technique contrast-limited adaptive histogram equalization was employed to produce clearer GIM areas. Third, data augmentation was applied for a more robust model. Lastly, the bilateral segmentation network model was applied to segment GIM areas in real time. The results were analyzed using different validity values. Results: From the internal test, our AI model achieved an inference speed of 31.53 frames per second. GIM detection showed sensitivity, specificity, positive predictive, negative predictive, accuracy, and mean intersection over union in GIM segmentation values of 93%, 80%, 82%, 92%, 87%, and 57%, respectively. Conclusions: The bilateral segmentation network combined with transfer learning, contrast-limited adaptive histogram equalization, and data augmentation can provide high sensitivity and good accuracy for GIM detection and segmentation.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.13 no.2
• /
• pp.827-833
• /
• 2012

In this paper, we propose an improved optical flow algorithm which reduces computational complexity as well as noise level. This algorithm reduces computational time by applying level simplification technique and removes noise by using eigenvectors of objects. Optical flow is one of the accurate algorithms used to generate depth information from two image frames using the vectors which track the motions of pixels. This technique, however, has disadvantage of taking very long computational time because of the pixel-based calculation and can cause some noise problems. The level simplifying technique is applied to reduce the computational time, and the noise is removed by applying optical flow only to the area of having eigenvector, then using the edge image to generate the depth information of background area. Three-dimensional images were created from two-dimensional images using the proposed method which generates the depth information first and then converts into three-dimensional image using the depth information and DIBR(Depth Image Based Rendering) technique. The error rate was obtained using the SSIM(Structural SIMilarity index).

• Journal of Mechanical Science and Technology
• /
• v.14 no.10
• /
• pp.1104-1113
• /
• 2000

This paper proposes an efficient reverse engineering technique for compound surfaces using a boundary detection method. This approach consists in extracting geometric edge information using a vision system, which can be used in order to drastically reduce geometric errors in the vicinity of compound surface boundaries. Through the image-processing technique and the interpolation process, boundaries are reconstructed by either analytic curves (e. g. circle, ellipse, line) or parametric curves (B-spline curve). In other regions, except boundaries, geometric data are acquired on CMM as points inspected using a touch type probe, and then they are interpolated on several surfaces using a B-spline skinning method. Finally, the boundary edge and the skinned surfaces are combined to reconstruct the final compound surface. Through simulations and experimental works, the effectiveness of the proposed method is confirmed.

• Journal of Institute of Control, Robotics and Systems
• /
• v.9 no.7
• /
• pp.536-546
• /
• 2003

We propose an object extraction technique adequate for the radial shape's radar signal structure for the purpose of implementing ARPA(Automatic Radar Plotting Aid) installed in the vessel. The radar signal data are processed by interpolation and accumulation to acquire a qualified image. The objects of the radar image have characteristics of having different shape and size as it gets far from the center, and it is not adequate for clustering generally. Therefore, this study designs a new vigilance distance model of elliptical shape and adopts this model in the ART2 neural network. We prove that the proposed clustering method makes it possible to extract objects adaptively and to separate the connected objects effectively.

• Journal of the Korea Society of Computer and Information
• /
• v.21 no.12
• /
• pp.11-18
• /
• 2016

Magnetic resonance image (MRI) is widely used in brain research field and medical image. Especially, non-invasive brain activation acquired image technique, which is functional magnetic resonance image (fMRI) is used in brain study. In this study, we investigate brain activation occurred by LED light stimulation. For investigate of brain activation in experimental small animal, we used high magnetic field 9.4T MRI. Experimental small animal is Balb/c mouse, method of fMRI is using echo planar image (EPI). EPI method spend more less time than any other MRI method. For this reason, however, EPI data has low contrast. Due to the low contrast, image pre-processing is very hard and inaccuracy. In this study, we planned the study protocol, which is called block design in fMRI research field. The block designed has 8 LED light stimulation session and 8 rest session. All block is consist of 6 EPI images and acquired 1 slice of EPI image is 16 second. During the light session, we occurred LED light stimulation for 1 minutes 36 seconds. During the rest session, we do not occurred light stimulation and remain the light off state for 1 minutes 36 seconds. This session repeat the all over the EPI scan time, so the total spend time of EPI scan has almost 26 minutes. After acquired EPI data, we performed the analysis of this image data. In this study, we analysis of EPI data using statistical parametric map (SPM) software and performed image pre-processing such as realignment, co-registration, normalization, smoothing of EPI data. The pre-processing of fMRI data have to segmented using this software. However this method has 3 different method which is Gaussian nonparametric, warped modulate, and tissue probability map. In this study we performed the this 3 different method and compared how they can change the result of fMRI analysis results. The result of this study show that LED light stimulation was activate superior colliculus region in mouse brain. And the most higher activated value of segmentation method was using tissue probability map. this study may help to improve brain activation study using EPI and SPM analysis.

• Journal of Digital Convergence
• /
• v.16 no.2
• /
• pp.19-26
• /
• 2018

We proposed a numerical calculation of the proportion of necrotic cells in pulmonary segmentation, pulmonary vessel segmentation lung disease site for diagnosis of lung disease from chest CT images. The first step is to separate the lungs and bronchi by applying a three-dimensional labeling technique from a chest CT image and a three-dimensional region growing method. The second step is to divide the pulmonary vessels by applying the rate of change using the first order polynomial regression, perform noise reduction, and divide the final pulmonary vessels. The third step is to find a disease prediction factor in a two-step image and calculate the proportion of necrotic cells.

• Journal of KIISE:Computing Practices and Letters
• /
• v.15 no.9
• /
• pp.695-699
• /
• 2009

In this paper, we propose an automatic prostate segmentation technique using image intensity and gradient information. Our method is composed of four steps. First, rays at regular intervals are generated. To minimize the effect of noise, the start and end positions of the ray are calculated. Second, the profiles on each ray are sorted based on the gradient. And priorities are applied to the sorted gradient in the profile. Third, boundary points are extracted by using gradient priority and intensity distribution. Finally, to reduce the error, the extracted boundary points are corrected by using B-spline interpolation. For accuracy evaluation, the average distance differences and overlapping region ratio between results of manual and automatic segmentations are calculated. As the experimental results, the average distance difference error and standard deviation were 1.09mm $\pm0.20mm$. And the overlapping region ratio was 92%.

• The Journal of the Korea Contents Association
• /
• v.10 no.1
• /
• pp.1-9
• /
• 2010

In this paper we present a shading technique for realistic rendering of the surfaces with irregular and complex textures using a single photograph. So far, most works have been using many photographs or special photographing equipment to render the surfaces with irregular and complex textures as well as dividing texture regions manually. We present an automatic selection method of the region segmentation techniques according to properties of materials. As our technique produces a reflectance model and the approximated Bidirectional Reflection Distribution Function(BRDF) parameters, it allows the recovery of the photometric properties of diffuse, specular, isotropic or anisotropic textured objects. Also it make it possible to present several synthetic images with novel lighting conditions and views.

• The Journal of the Acoustical Society of Korea
• /
• v.15 no.2E
• /
• pp.85-91
• /
• 1996

New very low bit rate segmentation video image coding technique is proposed by segmenting image into textually homogeneous regions. Regions are classified into one of three perceptually distinct texture classes(perceived constant intensity, smooth texture, and rough texture) using the Human Visual System(HVS) and the fractals. To design very low bit rate video image coder, it is very important to determine the best block size for estimation the fractal dimension and the thresholding of the fractal dimension for each texture class. Good quality reconstructed images are obtained with about 0.10 to 0.21 bit per pixel(bpp) for many different types of imagery.

• Journal of IKEEE
• /
• v.11 no.4
• /
• pp.307-314
• /
• 2007

In this paper, we propose an image contrast enhancement algorithm for an LCD TV. The proposed algorithm consists of two processes: the image segmentation process and the tone curve control process. The first process uses an automatic threshold technique to decompose an input image into two regions and then utilizes a hierarchical structure for real-time processing. The second process generates a gray level tone curve for contrast enhancement using a weighted sum of average tone curves for two segmented regions. Experimental result shows that the proposed algorithm outperforms the conventional contrast enhancement methods for an LCD TV.