한국인에 있어서의 동맥경화증의 특징 - 한국인의 안저 및 대동맥의 동맥경화성 병변 소견과 혈압혈청, 지질 및 년령과의 상관관계에 관한 관찰

손의석 ( E Suk Sohn ),강종명 ( Chong Myung Kang ),한철수 ( Cheol Su Han ),허봉렬 ( Hong Yul Huh ) 대한내과학회 1979 대한내과학회지 Vol.22 No.6

연결주의와 계산신경과학

한철(Cheol E. Han) 대한전자공학회 2017 대한전자공학회 학술대회 Vol.2017 No.6

The connectionism is a philosophical approach to understand the human’s intelligence can be explained by the connections in our brain. It inspired the development of early artificial neural network studies and still inspires deep learning researchers. However, due to lack ot understanding of our brains, it is less emphasized recently. In this paper, I will introduce a few recent use of the connectionism in the computational neuroscience field, which is laid between the neuroscience and the artificial intelligence. First, the neuroimaging studies related to the connectionism was briefly explained. There are two different levels: micro-scale and macro-scale. In the micro-scale analysis, a neuron and its connection is the main theme of the study, while in the macro-scale analysis, brain regions and their connections are. Second, the use of the connectionism in the neural modeling studies were briefly covered. To promote the artificial intelligence research in the next level, instead of exploiting current models, exploring to the other fields will be beneficial and crucial.

뉴럴모델링이란 무엇인가?

한철(Cheol E. Han) 한국정보과학회 2020 정보과학회지 Vol.38 No.10

그래프 합성곱 신경망을 이용한 알츠하이머 치매환자 분류

박재희(Jaehee Park),김대겸(Daegyeom Kim),정현강(Hyun-Ghang Jeong),한철(Cheol E. Han) 대한전자공학회 2021 대한전자공학회 학술대회 Vol.2021 No.6

Alzheimer’s disease is a neuro-degenerative disease with severe memory deficits and cognitive declines. These symptoms are generally along with various structural changes including the weakened connectivity between brain regions, that may lead lowered information processing and thus consequent cognitive declines. In this paper, we developed a deep learning model to classify Alzheimer’s disease. We extracted brain networks from diffusion-weighted MR image (DWI) of each individual, and used a recently developed deep learning algorithm, graph convolutional neural network (GCN). Our model achieved 90.7% accuracy on average. We also investigate which brain regions were used to make decisions. We extracted the brain regions with statistically significant differences between groups through Grad-CAM on the results of GCN. This showed that GCN also considers the brain regions similar to those found in the traditional statistical analyses.

특권 정보를 이용하는 딥러닝 모델을 통한 폐렴 검출

고명섭(Myeongseob Ko),정병창(ByeongChang Jeong),김대겸(Daegyeom Kim),한철(Cheol E. Han) 대한전자공학회 2020 대한전자공학회 학술대회 Vol.2020 No.8

In a recent computer vision society, there has been an incredible improvement in the performance of image classification tasks along with the development of deep learning technology. In the medical field, these classification techniques have been widely exploited to detect and diagnose several kinds of diseases. In this paper, we propose a method for detecting Pneumonia, leveraging privileged information which is predetermined by humans in order to provide a beneficial area of a given image for detection similarly to attention. The overall method is based on Learning Under Privileged Information (LUPI) framework, including the information bottleneck, gaussian dropout, and the reparameterization trick. We build our dataset to incorporate privileged information and provide the F1 score and test accuracy about our experiment to demonstrate that our experimental results show reliable improvement in efficiency as well as accuracy when we additionally use privileged information in the training stage.

3D U-Net을 이용한 비등방성(non-isotropic) 의료 영상의 등방성(isotropic) 의료 영상으로의 변환기법 연구

김대겸(Daegyeom Kim),최명원(Myeoungwon Choi),김지현(Ji Hyun Kim),한철(Cheol E. Han) 대한전자공학회 2018 대한전자공학회 학술대회 Vol.2018 No.6

Fluid Attenuated Inversion Recovery (FLAIR) images are widely used for diagnostic and medical imaging studies related to brain diseases. However, FLAIR images are often inappropriate for research aims, since their spatial resolution in the z-axis is lower than the other two axes to reduce the acquisition time in most hospitals. Thus, conversion from non-isotropic FLAIR images to isotropic FLAIR images is quite useful. On the other hands, since the medical image is often required to be high-resolution, and thus the size of image is quite large. It causes memory space deficiency for analysis medical images using deep learning, which must be solved for medical image research. In this study, we employed and modified U-net that is widely used for super-resolution applications. We not only restored images that were not obtained, but also suggested a solution for memory deficiency for handling large-sized medical images.

인공지능 마스크 착용 판독 시스템

노현정(Hyeonjeong Noh),전주예(Juye Jeon),최유정(Yujeong Choi),하민영(Minyoung Ha),한철(Cheol E. Han) 대한전자공학회 2020 대한전자공학회 학술대회 Vol.2020 No.8

It is now well-proven that wearing the mask is highly effective to prevent disease spread. In this study, we proposed an artificial intelligence-based system to classify whether a person wears a mask or not. We used pretrained convolutional neural network (CNN) to classify whether a person in given images wears a mask or not. The final model achieved 92.28% accuracy. We believe that this system can be used for the entrance of buildings to achieve successful national quarantine.

뇌의 회백질 두께를 이용한 다발성경화증과 시신경척수염범주질환 분류 모델 개발

오지석(JiSeok Oh),조은빈(Eunbin Cho),민주홍(JuHong Min),한철(Cheol E. Han) 대한전자공학회 2021 대한전자공학회 학술대회 Vol.2021 No.6

Both multiple sclerosis (MS) and neuromyelitis optica spectrum disease (NMOSD) deteriorates neural fibers in the white matter. However, they are fundamentally different disease, and thus require distinct medical treatments. In this paper, we investigate the efficacy of the machine learning model to classify MS and NMOSD only based on gray-matter characteristics, cortical thickness, extracted from the widely used T1-weighted MR images. We employed a simple multi-layer perceptron (MLP) and support vector machine (SVM). We also compared a 2-step binary classification model and 1-step multi-class classification model. We found that the former out-performed the latter. We achieved the accuracy of 83% on average.

특권 정보를 이용하는 딥러닝 모델을 통한 폐렴 검출

고명섭(Myeongseob Ko),정병창(ByeongChang Jeong),김대겸(Daegyeom Kim),한철(Cheol E. Han) 대한전자공학회 2021 전자공학회논문지 Vol.58 No.3

최근 컴퓨터 비전분야에선 딥러닝의 발달과 함께 이미지 분류 임무에 대한 성능이 급격한 발전을 이루고 있다. 의학 분야에서는 이러한 분류 임무가 여러 종류의 질병을 검출하고 진단하는 데 널리 이용되어 왔다. 본 논문에서는 기존의 이미지 분류를 위해 많이 사용되고 있는 딥러닝 네트워크에 특권 정보를 추가적으로 이용하여 폐렴을 검출하는 방법을 제안한다. 특권 정보는 이미지내에서 분류 임무와 직접적으로 관련된 영역으로, 본 연구에서는 이미지 내 폐 영역으로 설정하였다. 이와 같은 특권 정보는 근래에 많이 활용되는 내재적 주의집중(implicit attention)의 역할을 함으로써 모델로 하여금 분류 임무와 직접적인 관련이 있는 영역에 집중하도록 도와준다. 본 연구에서는 파라미터가 공유된 VGG-16 모델을 두 개 사용하였는데, 이 중 한 네트워크에는 주 정보인 이미지 자체를 제공하고, 또 하나의 네트워크는 정보 병목, 가우시안 드롭아웃, 리파라미터라이제이션 기법을 이용하여 특권정보를 제공하였다. 원본 데이터 셋보다 작은 다양한 크기의 데이터 셋을 특권 정보를 제공하였을 때와 제공하지 않았을 때를 비교하였다. 특권정보를 제공하였을 때, 테스트 정확도와 F1점수가 모두 향상되었는데, 데이터셋이 작을수록 특권정보로 인한 성능향상의 폭이 커졌다 (1,000장의 이미지를 사용했을 때는 테스트 정확도 3.5%, F1 점수 0.0285만큼 향상, 100장의 이미지를 사용했을 때는 테스트 정확도 3%, F1 점수 0.0173만큼 향상, 75장의 이미지를 사용했을 때는 테스트 정확도 16.72%, F1 점수 0.0629만큼 향상). 또한 특권 정보를 활용할 경우 모델의 판단에 기준이 된 영역을 활성화 맵을 통해 제시함으로써 해석 가능성을 보여주었다. In a recent computer vision society, there has been an rapid improvement in the performance of image classification tasks along with the development of deep learning technology. In the medical field, these classification techniques have been widely exploited to detect and diagnose several types of diseases. In this paper, we propose a method to detect pneumonia by additionally providing predetermined privileged information with off-the-shelf deep learning networks based on Learning Under Privileged Information (LUPI) framework. The privileged information is a designated area within an image, and can serve as implicit attention, encouraging the model to focus on the area directly related to the task, and thus may improve the classification performance. As an example, in this paper, we designated lung areas as our privileged information. Our proposed model consists of two shared VGG-16 models; one is for processing main information, image itself, and the other is for processing privileged information through information bottleneck, Gaussian dropout, and reparameterization trick. We provided various sized datasets but smaller than the original dataset by resampling it and compared model performances with and without privileged information. Our experiment showed that privileged information improves the test accuracy and F1 score, and the performance gain by the privileged information remarkably increases as the size of dataset gets smaller: increasing test accuracy and F1 score respectively by 3.5% and 0.0285 with 1000 training images, by 3% and 0.0173 with 100 training images, and by 16.72% and 0.0629 with 75 training images. We also demonstrated that our model can be interpretable through the activation maps of our model with the privileged information.

교육이 뇌의 정보전달능력에 미치는 효과 연구

김대겸(Daegyeom Kim),박재희(Jaehee Park),정병창(ByeongChang Jeong),한철(Cheol E. Han) 대한전자공학회 2020 대한전자공학회 학술대회 Vol.2020 No.8

An essential part of our brain is a complex structure of neural fibers connecting parts of the brain. This complex structure forms a brain network that has an important role in information communication between brain regions. Since the efficient information communication in the brain is associated with the higher intelligence and longer education duration, it is still in debate what aspect of brain networks is crucial in cognitive development. In this study, we analyzed the effects of the education duration on the information communication patterns using communicability that captures the total information flows through all possible communication pathways between brain regions, not only through the shortest path. We observed that the more-educated individuals have increased communicability in the left hemisphere, and decreased interhemispheric communicability, implying that the more specialization on the left hemisphere as education duration increases.