• Journal of Intelligence and Information Systems
• /
• v.22 no.2
• /
• pp.127-142
• /
• 2016

Deep learning model is a kind of neural networks that allows multiple hidden layers. There are various deep learning architectures such as convolutional neural networks, deep belief networks and recurrent neural networks. Those have been applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition and bioinformatics where they have been shown to produce state-of-the-art results on various tasks. Among those architectures, convolutional neural networks and recurrent neural networks are classified as the supervised learning model. And in recent years, those supervised learning models have gained more popularity than unsupervised learning models such as deep belief networks, because supervised learning models have shown fashionable applications in such fields mentioned above. Deep learning models can be trained with backpropagation algorithm. Backpropagation is an abbreviation for "backward propagation of errors" and a common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent. The method calculates the gradient of an error function with respect to all the weights in the network. The gradient is fed to the optimization method which in turn uses it to update the weights, in an attempt to minimize the error function. Convolutional neural networks use a special architecture which is particularly well-adapted to classify images. Using this architecture makes convolutional networks fast to train. This, in turn, helps us train deep, muti-layer networks, which are very good at classifying images. These days, deep convolutional networks are used in most neural networks for image recognition. Convolutional neural networks use three basic ideas: local receptive fields, shared weights, and pooling. By local receptive fields, we mean that each neuron in the first(or any) hidden layer will be connected to a small region of the input(or previous layer's) neurons. Shared weights mean that we're going to use the same weights and bias for each of the local receptive field. This means that all the neurons in the hidden layer detect exactly the same feature, just at different locations in the input image. In addition to the convolutional layers just described, convolutional neural networks also contain pooling layers. Pooling layers are usually used immediately after convolutional layers. What the pooling layers do is to simplify the information in the output from the convolutional layer. Recent convolutional network architectures have 10 to 20 hidden layers and billions of connections between units. Training deep learning networks has taken weeks several years ago, but thanks to progress in GPU and algorithm enhancement, training time has reduced to several hours. Neural networks with time-varying behavior are known as recurrent neural networks or RNNs. A recurrent neural network is a class of artificial neural network where connections between units form a directed cycle. This creates an internal state of the network which allows it to exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can use their internal memory to process arbitrary sequences of inputs. Early RNN models turned out to be very difficult to train, harder even than deep feedforward networks. The reason is the unstable gradient problem such as vanishing gradient and exploding gradient. The gradient can get smaller and smaller as it is propagated back through layers. This makes learning in early layers extremely slow. The problem actually gets worse in RNNs, since gradients aren't just propagated backward through layers, they're propagated backward through time. If the network runs for a long time, that can make the gradient extremely unstable and hard to learn from. It has been possible to incorporate an idea known as long short-term memory units (LSTMs) into RNNs. LSTMs make it much easier to get good results when training RNNs, and many recent papers make use of LSTMs or related ideas.

• Korean Journal of Remote Sensing
• /
• v.38 no.6_1
• /
• pp.1125-1139
• /
• 2022

In an unmanned aerial vehicles (UAVs) system, a physical offset can be existed between the global positioning system/inertial measurement unit (GPS/IMU) sensor and the observation sensor such as a hyperspectral sensor, and a lidar sensor. As a result of the physical offset, a misalignment between each image can be occurred along with a flight direction. In particular, in a case of multi-sensor system, an observation sensor has to be replaced regularly to equip another observation sensor, and then, a high cost should be paid to acquire a calibration parameter. In this study, we establish a precise sensor model equation to apply for a multiple sensor in common and propose an independent physical offset estimation method. The proposed method consists of 3 steps. Firstly, we define an appropriate rotation matrix for our system, and an initial sensor model equation for direct-georeferencing. Next, an observation equation for the physical offset estimation is established by extracting a corresponding point between a ground control point and the observed data from a sensor. Finally, the physical offset is estimated based on the observed data, and the precise sensor model equation is established by applying the estimated parameters to the initial sensor model equation. 4 region's datasets(Jeon-ju, Incheon, Alaska, Norway) with a different latitude, longitude were compared to analyze the effects of the calibration parameter. We confirmed that a misalignment between images were adjusted after applying for the physical offset in the sensor model equation. An absolute position accuracy was analyzed in the Incheon dataset, compared to a ground control point. For the hyperspectral image, root mean square error (RMSE) for X, Y direction was calculated for 0.12 m, and for the point cloud, RMSE was calculated for 0.03 m. Furthermore, a relative position accuracy for a specific point between the adjusted point cloud and the hyperspectral images were also analyzed for 0.07 m, so we confirmed that a precise data mapping is available for an observation without a ground control point through the proposed estimation method, and we also confirmed a possibility of multi-sensor fusion. From this study, we expect that a flexible multi-sensor platform system can be operated through the independent parameter estimation method with an economic cost saving.

• KIPS Transactions on Computer and Communication Systems
• /
• v.11 no.9
• /
• pp.317-322
• /
• 2022

The rapid development of information technology is also bringing about many changes in the medical environment. In particular, it is leading the rapid change of medical image information systems using big data and artificial intelligence (AI). The prescription delivery system (OCS), which consists of an electronic medical record (EMR) and a medical image storage and transmission system (PACS), has rapidly changed the medical environment from analog to digital. When combined with multiple solutions, PACS represents a new direction for advancement in security, interoperability, efficiency and automation. Among them, the combination with artificial intelligence (AI) using big data that can improve the quality of images is actively progressing. In particular, AI PACS, a system that can assist in reading medical images using deep learning technology, was developed in cooperation with universities and industries and is being used in hospitals. As such, in line with the rapid changes in the medical image information system in the medical environment, structural changes in the medical market and changes in medical policies to cope with them are also necessary. On the other hand, medical image information is based on a digital medical image transmission device (DICOM) format method, and is divided into a tomographic volume image, a volume image, and a cross-sectional image, a two-dimensional image, according to a generation method. In addition, recently, many medical institutions are rushing to introduce the next-generation integrated medical information system by promoting smart hospital services. The next-generation integrated medical information system is built as a solution that integrates EMR, electronic consent, big data, AI, precision medicine, and interworking with external institutions. It aims to realize research. Korea's medical image information system is at a world-class level thanks to advanced IT technology and government policies. In particular, the PACS solution is the only field exporting medical information technology to the world. In this study, along with the analysis of the medical image information system using big data, the current trend was grasped based on the historical background of the introduction of the medical image information system in Korea, and the future development direction was predicted. In the future, based on DICOM big data accumulated over 20 years, we plan to conduct research that can increase the image read rate by using AI and deep learning algorithms.

• Radiation Oncology Journal
• /
• v.15 no.1
• /
• pp.19-26
• /
• 1997

Purpose : This study is to report experience with Fractionated Stereotactic Radiation Therapy (FSRT) for locally recurrent nasopharynx cancer after curative conventional radiation therapy. Materials and Methods : Three Patients with locally recurrent and symptomatic nasopharynx cancer were given FSRT as reirradiation method between the Period of September of 1995 and August of 1996 For two Patients, application of FSRT is their third radiation therapy directed to the nasopharynx. Two patients were given low dose chemotherapy as radiation sensitizer concurrently with FSRT Authors used 3-dimensional coordinate system by individually made, relocatable Gill-Thomas-Cosman (GTC) stereotactic frame and multiple non-coplanar arc therapy dose Planning was done using Xknife-3. Total of 45 Gy/18 fractions or 50 Gy/20 fractions were given. Results : Authors observed satisfactory symptomatic improvement and remarkable objective tumor size decrease by follow-up MR images taken 1 month Post-FSRT in ali three patients, while no neurologic side effect attributable to reirradiation was noticed. Two died at 7 and 9 months with loco-regional and distant seeding outside FSRT field, while one patient is living for 4 month. Conclusion Authors experienced satisfactory therapeutic effectiveness and safety of FSRT as reirradiatlon method for locally recurrent nasopharynx cancer Development of more effective systemic chemotherapeutic regimen is desired for distant metastasis

• Korean Journal of Remote Sensing
• /
• v.36 no.2_2
• /
• pp.249-261
• /
• 2020

Coastal monitoring using multiple platforms/sensors is a very important tools for accurately understanding the changes in offshore marine environment and disaster with high temporal and spatial resolutions. However, integrated observation studies using multiple platforms and sensors are insufficient, and none of them have been evaluated for efficiency and limitation of convergence. In this study, we aimed to suggest an integrated observation method with multi-remote sensing platform and sensors, and to diagnose the utility and limitation. Integrated in situ surveys were conducted using Rhodamine WT fluorescent dye to simulate various marine disasters. In September 2019, the distribution and movement of RWT dye patches were detected using satellite (Kompsat-2/3/3A, Landsat-8 OLI, Sentinel-3 OLCI and GOCI), unmanned aircraft (Mavic 2 pro and Inspire 2), and manned aircraft platforms after injecting fluorescent dye into the waters of the South Sea-Yeosu Sea. The initial patch size of the RWT dye was 2,600 ㎡ and spread to 62,000 ㎡ about 138 minutes later. The RWT patches gradually moved southwestward from the point where they were first released,similar to the pattern of tidal current flowing southwest as the tides gradually decreased. Unmanned Aerial Vehicles (UAVs) image showed highest resolution in terms of spatial and time resolution, but the coverage area was the narrowest. In the case of satellite images, the coverage area was wide, but there were some limitations compared to other platforms in terms of operability due to the long cycle of revisiting. For Sentinel-3 OLCI and GOCI, the spectral resolution and signal-to-noise ratio (SNR) were the highest, but small fluorescent dye detection was limited in terms of spatial resolution. In the case of hyperspectral sensor mounted on manned aircraft, the spectral resolution was the highest, but this was also somewhat limited in terms of operability. From this simulation approach, multi-platform integrated observation was able to confirm that time,space and spectral resolution could be significantly improved. In the future, if this study results are linked to coastal numerical models, it will be possible to predict the transport and diffusion of contaminants, and it is expected that it can contribute to improving model accuracy by using them as input and verification data of the numerical models.

• Journal of the korean academy of Pediatric Dentistry
• /
• v.33 no.3
• /
• pp.498-503
• /
• 2006

This in vitro study compared the remineralization of incipient interproximal caries in the presence of three glass ionomer cements (highly-filled glass ionomer cement, compomer, resin-modified glass ionomer cement) and a resin composite(control). The long-term changes in remineralization caused by each material were evaluated by microtomography. Proximal restoration was simulated by placing tooth specimens and the various glass ionomer cements in closed containers with artificial saliva at $37^{\circ}C$ and pH 7.0 for 30 days with constant circulation Tomographic images were obtained with a micro CT scanner at 90, 180, and 270 days, and density-measuring software was used to calculate the micro-density of artificial caries lesions in the specimens. The mean density changes were compared between groups in order to evaluate the effects of remineralization. All data were analyzed using one-way ANOVA and the post-HOC Tukey multiple comparison test at p<0.05. While the density of artificial caries lesions increased for all treatments, the increases for the three glass ionomer groups were significantly higher than that for the resin group in each three month period. As time went on, the amount of density increase of the glass ionomer groups decreased, and significant differences were found between the remineralization effects of the glass ionomer groups.

• Journal of KIISE:Software and Applications
• /
• v.31 no.10
• /
• pp.1332-1344
• /
• 2004

Upon development of information super-highway and multimedia-related technoiogies in recent years, more efficient technologies to transmit, store and retrieve the multimedia data are required. Among such technologies, firstly, it is common that the semantic-based image retrieval is annotated separately in order to give certain meanings to the image data and the low-level property information that include information about color, texture, and shape Despite the fact that the semantic-based information retrieval has been made by utilizing such vocabulary dictionary as the key words that given, however it brings about a problem that has not yet freed from the limit of the existing keyword-based text information retrieval. The second problem is that it reveals a decreased retrieval performance in the content-based image retrieval system, and is difficult to separate the object from the image that has complex background, and also is difficult to extract an area due to excessive division of those regions. Further, it is difficult to separate the objects from the image that possesses multiple objects in complex scene. To solve the problems, in this paper, I established a content-based retrieval system that can be processed in 5 different steps. The most critical process of those 5 steps is that among RGB images, the one that has the largest and the smallest background are to be extracted. Particularly. I propose the method that extracts the subject as well as the background by using an Image, which has the largest background. Also, to solve the second problem, I propose the method in which multiple objects are separated using RGB channel selection techniques having optimized the excessive division of area by utilizing Watermerge's threshold value with the object separation using the method of RGB channels separation. The tests proved that the methods proposed by me were superior to the existing methods in terms of retrieval performances insomuch as to replace those methods that developed for the purpose of retrieving those complex objects that used to be difficult to retrieve up until now.

• Journal of Korean Society of Forest Science
• /
• v.112 no.4
• /
• pp.426-441
• /
• 2023

This study investigated the predictive accuracy of a model of landslide displacement in Jecheon-si, where a great number of landslides were triggered by heavy rain on both natural (non-clear-cut) and clear-cut slopes during August 2020. This was accomplished by applying three flow direction methods (single flow direction, SFD; multiple flow direction, MFD; infinite flow direction, IFD) and the degree of root cohesion to an infinite slope stability equation. The application assumed that the soil saturation and any changes in root cohesion occurred following the timber harvest (clear-cutting). In the study area, 830 landslide locations were identified via landslide inventory mapping from satellite images and 25 cm resolution aerial photographs. The results of the landslide modeling comparison showed the accuracy of the models that considered changes in the root cohesion following clear-cutting to be improved by 1.3% to 2.6% when compared with those not considered in the area under the receiver operating characteristics (AUROC) analysis. Furthermore, the accuracy of the models that used the MFD algorithm improved by up to 1.3% when compared with the models that used the other algorithms in the AUROC analysis. These results suggest that the discriminatory application of the root cohesion, which considers changes in the vegetation condition, and the selection of the flow direction method may influence the accuracy of landslide predictive modeling. In the future, the results of this study should be verified by examining the root cohesion and its dynamic changes according to the tree species using the field hydrological monitoring technique.

• Nuclear Medicine and Molecular Imaging
• /
• v.43 no.1
• /
• pp.10-18
• /
• 2009

Purpose: We investigated quantification of dopaminergic transporter (DAT) and serotonergic transporter (SERT) on $^{123}I$-FP-CIT SPECT for differentiating between multiple systemic atrophy (MSA) and idiopathic Parkinson's disease (IPD). Materials and Methods: N-fluoropropyl-$2{\beta}$-carbomethoxy-$3{\beta}$-4-[$^{123}I$]-iodophenylnortropane SPECT ($^{123}I$-FP-CIT SPECT) was performed in 8 patients with MSA (mean age: $64.0{\pm}4.5yrs$, m:f=6:2), 13 with early IPD (mean age: $65.5{\pm}5.3yrs$, m:f=9:4), and 12 healthy controls (mean age: $63.3{\pm}5.7yrs$, m:f=8:4). Standard regions of interests (ROls) of striatum to evaluate DAT, and hypothalamus and midbrain for SERT were drawn on standard template images and applied to each image taken 4 hours after radiotracer injection. Striatal specific binding for DAT and hypothalamic and midbrain specific binding for SERT were calculated using region/reference ratio based on the transient equilibrium method. Group differences were tested using ANOVA with the postHoc analysis. Results: DAT in the whole striatum and striatal subregions were significantly decreased in both patient groups with MSA and early IPD, compared with healthy control (p<0.05 in all). In early IPD, a significant increase in the uptake ratio in anterior and posterior putamen and a trend of increase in caudate to putamen ratio was observed. In MSA, the decrease of DAT was accompanied with no difference in the striatal uptake pattern compared with healthy controls. Regarding the brain regions where $^{123}I$-FP-CIT binding was predominant by SERT, MSA patients showed a decrease in the binding of $^{123}I$-FP-CIT in the pons compared with controls as well as early IPD patients (MSA: $0.22{\pm}0.1$ healthy controls: $0.33{\pm}0.19$, IPD: $0.29{\pm}0.19$), however, it did not reach the statistical significance. Conclusion: In this study, the differential patterns in the reduction of DAT in the striatum and the reduction of pontine $^{123}I$-FP-CIT binding predominant by SERT could be observed in MSA patients on $^{123}I$-FP-CIT SPECT. We suggest that the quantification of SERT as well as DAT using $^{123}I$-FP-CIT SPECT is helpful to differentiate parkinsonian disorders in early stage.

• Investigative Magnetic Resonance Imaging
• /
• v.8 no.2
• /
• pp.100-108
• /
• 2004

Purpose : Early degeneration of articular cartilage is accompanied by a loss of glycosaminoglycan (GAG) and the consequent change of the integrity. The purpose of this study was to biochemically quantify the loss of GAG, and to evaluate the $Gd(DTPA)^{2-}$-enhanced, and T1, T2, rho relaxation map for detection of the early degeneration of cartilage. Materials and Methods : A cartilage-bone block in size of $8mm\;\times\;10mm$ was acquired from the patella in each of three pigs. Quantitative analysis of GAG of cartilage was performed at spectrophotometry by use of dimethylmethylene blue. Each of cartilage blocks was cultured in one of three different media: two different culture media (0.2 mg/ml trypsin solution, 1mM Gd $(DTPA)^{2-}$ mixed trypsin solution) and the control media (phosphate buffered saline (PBS)). The cartilage blocks were cultured for 5 hrs, during which MR images of the blocks were obtained at one hour interval (0 hr, 1 hr, 2 hr, 3 hr, 4 hr, 5 hr). And then, additional culture was done for 24 hrs and 48 hrs. Both T1-weighted image (TR/TE, 450/22 ms), and mixed-echo sequence (TR/TE, 760/21-168ms; 8 echoes) were obtained at all times using field of view 50 mm, slice thickness 2 mm, and matrix $256\times512$. The MRI data were analyzed with pixel-by-pixel comparisons. The cultured cartilage-bone blocks were microscopically observed using hematoxylin & eosin, toluidine blue, alcian blue, and trichrome stains. Results : At quantitation analysis, GAG concentration in the culture solutions was proportional to the culture durations. The T1-signal of the cartilage-bone block cultured in the $Gd(DTPA)^{2-}$ mixed solution was significantly higher ($42\%$ in average, p<0.05) than that of the cartilage-bone block cultured in the trypsin solution alone. The T1, T2, rho relaxation times of cultured tissue were not significantly correlated with culture duration (p>0.05). However the focal increase in T1 relaxation time at superficial and transitional layers of cartilage was seen in $Gd(DTPA)^{2-}$ mixed culture. Toluidine blue and alcian blue stains revealed multiple defects in whole thickness of the cartilage cultured in trypsin media. Conclusion : The quantitative analysis showed gradual loss of GAG proportional to the culture duration. Microimagings of cartilage with $Gd(DTPA)^{2-}$-enhancement, relaxation maps were available by pixel size of $97.9\times195\;{\mu}m$. Loss of GAG over time better demonstrated with $Gd(DTPA)^{2-}$-enhanced images than with T1, T2, rho relaxation maps. Therefore $Gd(DTPA)^{2-}$-enhanced T1-weighted image is superior for detection of early degeneration of cartilage.