DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | 이문용 | - |
dc.contributor.author | Kim, Mujin | - |
dc.contributor.author | 김무진 | - |
dc.date.accessioned | 2024-08-08T19:30:59Z | - |
dc.date.available | 2024-08-08T19:30:59Z | - |
dc.date.issued | 2024 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=1098139&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/321984 | - |
dc.description | 학위논문(박사) - 한국과학기술원 : 데이터사이언스대학원, 2024.2,[v, 105 p. :] | - |
dc.description.abstract | Cancer is one of the world's leading causes of death. Pathologists manually review biopsy slides to diagnose cancer. In order to reduce the workload of pathologists, Deep learning based diagnostic assistance system is used. However, convolutional neural network-based models require a large number of training data. Deep active learning is a method to effectively train deep learning classification models through active learning, with the goal of allowing deep learning models train with only a small amount of data to achieve high performance. However, there is a problem when active learning is performed in the field, the effect may decrease. Therefore, this study conducted two studies to perform active learning more effectively in the field. In the first study, we identify the difference between the dataset used in prior studies and the dataset in the field, and propose a method to mitigate the effect reduction caused by the difference. The pathology field dataset can include noise data during the manufacturing process. Therefore, in order to effectively perform deep active learning on noisy dataset, we propose a method to compute thresholds using predictive loss values and select informative data in intervals where less noise data is distributed. As a result, proposed method reduced the probability of noise data being selected, and achieved better classification accuracy of trained model. A second study proposes a deep active learning framework that can infuse expert knowledge about informative data. This study actually identifies that there is a difference between expert selection and model selection for informative data to model training. And we train a selection model that learns expert selection records to reflects expertise of pathologist. As a result, we can develop a framework that includes the corresponding selection model to perform effective deep active learning in the field. Overall these studies contribute to supporting field pathologists and improving the effectiveness of their work, by providing practical and effective deep active learning in the field. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | 조직 병리학 이미지 분류▼a딥 러닝▼a심층 능동 학습▼a의료 진단 지원 | - |
dc.subject | Pathology image classification▼aDeep learning▼aDeep active learning▼aClinical decision supporting system | - |
dc.title | Framework studies on the application of deep active learning to pathology for field optimization | - |
dc.title.alternative | 현업 최적화를 위한 심층 능동 학습의 병리학 적용 프레임워크 연구 | - |
dc.type | Thesis(Ph.D) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :데이터사이언스대학원, | - |
dc.contributor.alternativeauthor | Yi, Mun Yong | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.