• Korean Journal of Radiology
• /
• v.25 no.4
• /
• pp.363-373
• /
• 2024

Objective: To develop and evaluate a deep learning model for automated segmentation and detection of bone metastasis on spinal MRI. Materials and Methods: We included whole spine MRI scans of adult patients with bone metastasis: 662 MRI series from 302 patients (63.5 ± 11.5 years; male:female, 151:151) from three study centers obtained between January 2015 and August 2021 for training and internal testing (random split into 536 and 126 series, respectively) and 49 MRI series from 20 patients (65.9 ± 11.5 years; male:female, 11:9) from another center obtained between January 2018 and August 2020 for external testing. Three sagittal MRI sequences, including non-contrast T1-weighted image (T1), contrast-enhanced T1-weighted Dixon fat-only image (FO), and contrast-enhanced fat-suppressed T1-weighted image (CE), were used. Seven models trained using the 2D and 3D U-Nets were developed with different combinations (T1, FO, CE, T1 + FO, T1 + CE, FO + CE, and T1 + FO + CE). The segmentation performance was evaluated using Dice coefficient, pixel-wise recall, and pixel-wise precision. The detection performance was analyzed using per-lesion sensitivity and a free-response receiver operating characteristic curve. The performance of the model was compared with that of five radiologists using the external test set. Results: The 2D U-Net T1 + CE model exhibited superior segmentation performance in the external test compared to the other models, with a Dice coefficient of 0.699 and pixel-wise recall of 0.653. The T1 + CE model achieved per-lesion sensitivities of 0.828 (497/600) and 0.857 (150/175) for metastases in the internal and external tests, respectively. The radiologists demonstrated a mean per-lesion sensitivity of 0.746 and a mean per-lesion positive predictive value of 0.701 in the external test. Conclusion: The deep learning models proposed for automated segmentation and detection of bone metastases on spinal MRI demonstrated high diagnostic performance.

• Korean Journal of Remote Sensing
• /
• v.36 no.2_1
• /
• pp.199-215
• /
• 2020

A seagrass that is marine vascular plants plays an important role in the marine ecosystem, so periodic monitoring ofseagrass habitatsis being performed. Recently, the use of dronesthat can easily acquire very high-resolution imagery is increasing to efficiently monitor seagrass habitats. And deep learning based on a convolutional neural network has shown excellent performance in semantic segmentation. So, studies applied to deep learning models have been actively conducted in remote sensing. However, the segmentation accuracy was different due to the hyperparameter, various deep learning models and imagery. And the normalization of the image and the tile and batch size are also not standardized. So,seagrass habitats were segmented from drone-borne imagery using a deep learning that shows excellent performance in this study. And it compared and analyzed the results focused on normalization and tile size. For comparison of the results according to the normalization, tile and batch size, a grayscale image and grayscale imagery converted to Z-score and Min-Max normalization methods were used. And the tile size isincreased at a specific interval while the batch size is allowed the memory size to be used as much as possible. As a result, IoU was 0.26 ~ 0.4 higher than that of Z-score normalized imagery than other imagery. Also, it wasfound that the difference to 0.09 depending on the tile and batch size. The results were different according to the normalization, tile and batch. Therefore, this experiment found that these factors should have a suitable decision process.

• Geophysics and Geophysical Exploration
• /
• v.24 no.3
• /
• pp.89-97
• /
• 2021

Fault detection in seismic data is well suited to the application of machine learning algorithms. Accordingly, various machine learning techniques are being developed. In recent studies, machine learning models, which utilize synthetic data, are the particular focus when training with deep learning. The use of synthetic training data has many advantages; Securing massive data for training becomes easy and generating exact fault labels is possible with the help of synthetic training data. To interpret real data with the model trained by synthetic data, the synthetic data used for training should be geologically realistic. In this study, we introduce a method to generate realistic synthetic seismic data. Initially, reflectivity models are generated to include realistic fault structures, and then, a one-way wave equation is applied to efficiently generate seismic stack sections. Next, a migration algorithm is used to remove diffraction artifacts and random noise is added to mimic actual field data. A convolutional neural network model based on the U-Net structure is used to verify the generated synthetic data set. From the results of the experiment, we confirm that realistic synthetic data effectively creates a deep learning model that can be applied to field data.

• Korean Journal of Remote Sensing
• /
• v.39 no.5_3
• /
• pp.979-995
• /
• 2023

As wildfires are difficult to predict, real-time monitoring is crucial for a timely response. Geostationary satellite images are very useful for active fire detection because they can monitor a vast area with high temporal resolution (e.g., 2 min). Existing satellite-based active fire detection algorithms detect thermal outliers using threshold values based on the statistical analysis of brightness temperature. However, the difficulty in establishing suitable thresholds for such threshold-based methods hinders their ability to detect fires with low intensity and achieve generalized performance. In light of these challenges, machine learning has emerged as a potential-solution. Until now, relatively simple techniques such as random forest, Vanilla convolutional neural network (CNN), and U-net have been applied for active fire detection. Therefore, this study proposed an active fire detection algorithm using state-of-the-art (SOTA) deep learning techniques using data from the Advanced Himawari Imager and evaluated it over East Asia and Australia. The SOTA model was developed by applying EfficientNet and lion optimizer, and the results were compared with the model using the Vanilla CNN structure. EfficientNet outperformed CNN with F1-scores of 0.88 and 0.83 in East Asia and Australia, respectively. The performance was better after using weighted loss, equal sampling, and image augmentation techniques to fix data imbalance issues compared to before the techniques were used, resulting in F1-scores of 0.92 in East Asia and 0.84 in Australia. It is anticipated that timely responses facilitated by the SOTA deep learning-based approach for active fire detection will effectively mitigate the damage caused by wildfires.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2018.10a
• /
• pp.637-640
• /
• 2018

The paper proposes a new deep network-based model that rates periorbital wrinkles in order to alleviate the shortcomings of the evaluation by human experts as well as to facilitate the automation. Periorbital wrinkles still need to be classified by human experts. Furthermore, the classification results from experts are different from each other in many cases due to the inter-interpreter variability and the absence of quantification criteria. Unlike existing classification methods which classify original images, the proposed model consists of a cascade of two deep networks: U-Net for the enhancement of wrinkles on an input image and VGG16 for final classification based on the wrinkle information. Experiments of the proposed model are made with a data set that consists of 433 images rated by experts, showing the promising performance.

• Journal of the Korean Geotechnical Society
• /
• v.40 no.3
• /
• pp.33-39
• /
• 2024

Inspecting cracks to determine a structure's condition is crucial for accurate safety diagnosis. However, visual crack inspection methods can be subjective and are dependent on field conditions, thereby resulting in low reliability. To address this issue, this study automates the detection of concrete cracks in image data using ResNet, FPN, and the Mask R-CNN components as the backbone, neck, and head of a convolutional neural network. The performance of the proposed model is analyzed using the intersection over the union (IoU). The experimental dataset contained 1,203 images divided into training (70%), validation (20%), and testing (10%) sets. The model achieved an IoU value of 95.83% for testing, and there were no cases where the crack was not detected. These findings demonstrate that the proposed model realized highly accurate detection of concrete cracks in image data.

• Korean Journal of Remote Sensing
• /
• v.38 no.6_1
• /
• pp.1273-1283
• /
• 2022

Recently, forest fires have frequently occurred due to climate change, leading to human and property damage every year. The forest fire monitoring technique using remote sensing can obtain quick and large-scale information of fire-damaged areas. In this study, the Gangneung and Donghae forest fires that occurred in March 2022 were analyzed using the spectral band of Sentinel-2, the normalized difference vegetation index (NDVI), and the normalized difference water index (NDWI) to classify the affected areas of forest fires. The U-net based convolutional neural networks (CNNs) model was simulated for the fire-damaged areas. The accuracy of forest fire classification in Donghae and Gangneung classification was high at 97.3% (f₁=0.486, IoU=0.946). The same model used in Donghae and Gangneung was applied to Uljin and Samcheok areas to get rid of the possibility of overfitting often happen in machine learning. As a result, the portion of overlap with the forest fire damage area reported by the National Institute of Forest Science (NIFoS) was 74.4%, confirming a high level of accuracy even considering the uncertainty of the model. This study suggests that it is possible to quantitatively evaluate the classification of forest fire-damaged area using a spectral band and indices similar to that of the Compact Advanced Satellite 500 (CAS500-4) in the Sentinel-2.

• Smart Structures and Systems
• /
• v.29 no.1
• /
• pp.153-166
• /
• 2022

Identifying fine cracks in steel bridge facilities is a challenging task of structural health monitoring (SHM). This study proposed an end-to-end crack image segmentation framework based on a one-step Convolutional Neural Network (CNN) for pixel-level object recognition with high accuracy. To particularly address the challenges arising from small object detection in complex background, efforts were made in loss function selection aiming at sample imbalance and module modification in order to improve the generalization ability on complicated images. Specifically, loss functions were compared among alternatives including the Binary Cross Entropy (BCE), Focal, Tversky and Dice loss, with the last three specialized for biased sample distribution. Structural modifications with dilated convolution, Spatial Pyramid Pooling (SPP) and Feature Pyramid Network (FPN) were also performed to form a new backbone termed CrackDet. Models of various loss functions and feature extraction modules were trained on crack images and tested on full-scale images collected on steel box girders. The CNN model incorporated the classic U-Net as its backbone, and Dice loss as its loss function achieved the highest mean Intersection-over-Union (mIoU) of 0.7571 on full-scale pictures. In contrast, the best performance on cropped crack images was achieved by integrating CrackDet with Dice loss at a mIoU of 0.7670.

• Journal of radiological science and technology
• /
• v.45 no.2
• /
• pp.151-158
• /
• 2022

The Cardiac Gated Blood Pool (GBP) scintigram, a nuclear medicine imaging, calculates the left ventricular Ejection Fraction (EF) by segmenting the left ventricle from the heart. However, in order to accurately segment the substructure of the heart, specialized knowledge of cardiac anatomy is required, and depending on the expert's processing, there may be a problem in which the left ventricular EF is calculated differently. In this study, using the DeepLabV3 architecture, GBP images were trained on 93 training data with a ResNet-50 backbone. Afterwards, the trained model was applied to 23 separate test sets of GBP to evaluate the reproducibility of the region of interest and left ventricular EF. Pixel accuracy, dice coefficient, and IoU for the region of interest were 99.32±0.20, 94.65±1.45, 89.89±2.62(%) at the diastolic phase, and 99.26±0.34, 90.16±4.19, and 82.33±6.69(%) at the systolic phase, respectively. Left ventricular EF was calculated to be an average of 60.37±7.32% in the ROI set by humans and 58.68±7.22% in the ROI set by the deep learning segmentation model. (p<0.05) The automated segmentation method using deep learning presented in this study similarly predicts the average human-set ROI and left ventricular EF when a random GBP image is an input. If the automatic segmentation method is developed and applied to the functional examination method that needs to set ROI in the field of cardiac scintigram in nuclear medicine in the future, it is expected to greatly contribute to improving the efficiency and accuracy of processing and analysis by nuclear medicine specialists.

• Korean Journal of Remote Sensing
• /
• v.37 no.3
• /
• pp.367-378
• /
• 2021

Field compost is a representative non-point pollution source for livestock. If the field compost flows into the water system due to rainfall, nutrients such as phosphorus and nitrogen contained in the field compost can adversely affect the water quality of the river. In this paper, we propose a method for detecting field compost using unmanned aerial vehicle images and deep learning-based semantic segmentation. Based on 39 ortho images acquired in the study area, about 30,000 data were obtained through data augmentation. Then, the accuracy was evaluated by applying the semantic segmentation algorithm developed based on U-net and the filtering technique of Open CV. As a result of the accuracy evaluation, the pixel accuracy was 99.97%, the precision was 83.80%, the recall rate was 60.95%, and the F1-Score was 70.57%. The low recall compared to precision is due to the underestimation of compost pixels when there is a small proportion of compost pixels at the edges of the image. After, It seems that accuracy can be improved by combining additional data sets with additional bands other than the RGB band.