서브메뉴
검색
Weak Supervision in Deep Learning for Medical Imaging and Astrophysics
Weak Supervision in Deep Learning for Medical Imaging and Astrophysics
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211151414
- ISBN
- 9798382761831
- DDC
- 621.3
- 저자명
- Wu, Yunan.
- 서명/저자
- Weak Supervision in Deep Learning for Medical Imaging and Astrophysics
- 발행사항
- [Sl] : Northwestern University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 227 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
- 주기사항
- Advisor: Katsaggelos, Aggelos K.
- 학위논문주기
- Thesis (Ph.D.)--Northwestern University, 2024.
- 초록/해제
- 요약Traditional supervised learning pipelines heavily rely on high-quality, extensively annotated datasets. These datasets, however, are often costly to produce and prone to errors, leading to inaccurate or insufficient labels. As a remedy, weakly supervised learning (WSL) emerges as a robust alternative by leveraging weaker forms of supervision-such as incomplete, inexact, and inaccurate supervision-to effectively train deep learning models. Chapter 1 provides a foundational overview of WSL, highlighting its necessity and exploring its various methods, including multiple instance learning, transfer learning, crowdsourcing, and multi-modal fusion.This thesis explores the domain of WSL with a particular focus on medical imaging and astrophysics. The first part addresses WSL in medical imaging, a critical domain for disease diagnosis, treatment planning, and monitoring. The major challenge in medical imaging arises from the difficulty in obtaining large, accurately labeled datasets, which are time-consuming and expensive to produce and require expert annotation. The variability inherent in medical subjects further complicates the annotation process. Therefore, the first goal of this work is to develop and apply WSL methods across various medical imaging applications, including classification, detection, and segmentation. Chapter 2 introduces an attention-based multiple instance learning (Att-MIL) approach by combining an attention-based convolutional neural network (Att-CNN) with a variational Gaussian process (VGPMIL). This method efficiently uses hemorrhage labels at scan-level to enhance predictive accuracy at both slice and scan levels, significantly outperforming other methods trained similarly and achieving comparable results to those requiring more detailed slice-level annotations. Chapter 3 builds on this by implementing a Smooth Attention Deep MIL (SA-DMIL) model that incorporates first and second order constraints on the attention mechanism to capture spatial dependencies between slices within a scan. Chapter 4 proposes a novel integration of CNN and Long Short-Term Memory (LSTM) networks for image-level CT hemorrhage localization. This method achieves performance comparable to senior neuroradiologists and provides clinically useful attention weights and heatmaps to facilitate rapid diagnostic decisions. Chapter 5 progresses to pixel-level segmentation of CT hemorrhage using conditioned diffusion models, which trained on image-level labels to generate anomaly maps and reconstruct images, thus minimizing the reliance on extensive pixel-level annotations. Chapter 6 explores the use of multi-modal data to assess the health status and predict outcomes of critically ill COVID-19 patients. It demonstrates that a fusion of chest X-rays, respiratory sounds, and ICU clinical variables significantly enhances the prediction accuracy for indirect outcomes. Finally, Chapter 7 introduces DeepCOVID-Fuse, a deep learning fusion model that combines chest x-rays and clinical variables to predict risk levels in COVID-19 patients, effectively managing outcomes even with missing modalities during testing.The second part of the thesis focuses on WSL in astrophysics, particularly analyzing large-scale astronomical datasets. In this field, expert analysis is crucial for understanding the phenomena of space. However, acquiring labeled datasets is particularly challenging due to the complex and dynamic nature of space, which means that labels must be continuously updated as new discoveries are made and understandings evolve. To address this, crowdsourcing becomes an essential tool, leveraging the efforts of volunteers to annotate data and machine learning to guide these efforts. Therefore, the second goal of this work is to apply WSL methods with crowdsourcing. Chapter 8 introduces the crowdsourcing project Gravity Spy, detailing its workflow and its development from version 1.0 to 2.0. Gravity Spy 1.0 primarily focuses on classifying glitches in the main channel, while Gravity Spy 2.0 extends to include glitches in the auxiliary channels, exploring the underlying correlations between glitches across these channels. Chapter 9 presents a multi-view fusion network with attention mechanisms in Gravity Spy 1.0 for classifying glitches during the new O4 observing run. Finally, Chapter 10 describes the Cross-Temporal Spectrogram Autoencoder (CTSAE), an innovative unsupervised model that identifies correlations between main and auxiliary channel glitches through a novel integration of CNN and Vision Transformers.
- 일반주제명
- Electrical engineering
- 일반주제명
- Computer science
- 일반주제명
- Medical imaging
- 키워드
- Deep learning
- 키워드
- Gravity Spy
- 키워드
- CT hemorrhage
- 기타저자
- Northwestern University Electrical and Computer Engineering
- 기본자료저록
- Dissertations Abstracts International. 85-11B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017161571
■00520250211151414
■006m o d
■007cr#unu||||||||
■020 ▼a9798382761831
■035 ▼a(MiAaPQ)AAI31293247
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a621.3
■1001 ▼aWu, Yunan.▼0(orcid)0000-0001-6980-9746
■24510▼aWeak Supervision in Deep Learning for Medical Imaging and Astrophysics
■260 ▼a[Sl]▼bNorthwestern University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a227 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 85-11, Section: B.
■500 ▼aAdvisor: Katsaggelos, Aggelos K.
■5021 ▼aThesis (Ph.D.)--Northwestern University, 2024.
■520 ▼aTraditional supervised learning pipelines heavily rely on high-quality, extensively annotated datasets. These datasets, however, are often costly to produce and prone to errors, leading to inaccurate or insufficient labels. As a remedy, weakly supervised learning (WSL) emerges as a robust alternative by leveraging weaker forms of supervision-such as incomplete, inexact, and inaccurate supervision-to effectively train deep learning models. Chapter 1 provides a foundational overview of WSL, highlighting its necessity and exploring its various methods, including multiple instance learning, transfer learning, crowdsourcing, and multi-modal fusion.This thesis explores the domain of WSL with a particular focus on medical imaging and astrophysics. The first part addresses WSL in medical imaging, a critical domain for disease diagnosis, treatment planning, and monitoring. The major challenge in medical imaging arises from the difficulty in obtaining large, accurately labeled datasets, which are time-consuming and expensive to produce and require expert annotation. The variability inherent in medical subjects further complicates the annotation process. Therefore, the first goal of this work is to develop and apply WSL methods across various medical imaging applications, including classification, detection, and segmentation. Chapter 2 introduces an attention-based multiple instance learning (Att-MIL) approach by combining an attention-based convolutional neural network (Att-CNN) with a variational Gaussian process (VGPMIL). This method efficiently uses hemorrhage labels at scan-level to enhance predictive accuracy at both slice and scan levels, significantly outperforming other methods trained similarly and achieving comparable results to those requiring more detailed slice-level annotations. Chapter 3 builds on this by implementing a Smooth Attention Deep MIL (SA-DMIL) model that incorporates first and second order constraints on the attention mechanism to capture spatial dependencies between slices within a scan. Chapter 4 proposes a novel integration of CNN and Long Short-Term Memory (LSTM) networks for image-level CT hemorrhage localization. This method achieves performance comparable to senior neuroradiologists and provides clinically useful attention weights and heatmaps to facilitate rapid diagnostic decisions. Chapter 5 progresses to pixel-level segmentation of CT hemorrhage using conditioned diffusion models, which trained on image-level labels to generate anomaly maps and reconstruct images, thus minimizing the reliance on extensive pixel-level annotations. Chapter 6 explores the use of multi-modal data to assess the health status and predict outcomes of critically ill COVID-19 patients. It demonstrates that a fusion of chest X-rays, respiratory sounds, and ICU clinical variables significantly enhances the prediction accuracy for indirect outcomes. Finally, Chapter 7 introduces DeepCOVID-Fuse, a deep learning fusion model that combines chest x-rays and clinical variables to predict risk levels in COVID-19 patients, effectively managing outcomes even with missing modalities during testing.The second part of the thesis focuses on WSL in astrophysics, particularly analyzing large-scale astronomical datasets. In this field, expert analysis is crucial for understanding the phenomena of space. However, acquiring labeled datasets is particularly challenging due to the complex and dynamic nature of space, which means that labels must be continuously updated as new discoveries are made and understandings evolve. To address this, crowdsourcing becomes an essential tool, leveraging the efforts of volunteers to annotate data and machine learning to guide these efforts. Therefore, the second goal of this work is to apply WSL methods with crowdsourcing. Chapter 8 introduces the crowdsourcing project Gravity Spy, detailing its workflow and its development from version 1.0 to 2.0. Gravity Spy 1.0 primarily focuses on classifying glitches in the main channel, while Gravity Spy 2.0 extends to include glitches in the auxiliary channels, exploring the underlying correlations between glitches across these channels. Chapter 9 presents a multi-view fusion network with attention mechanisms in Gravity Spy 1.0 for classifying glitches during the new O4 observing run. Finally, Chapter 10 describes the Cross-Temporal Spectrogram Autoencoder (CTSAE), an innovative unsupervised model that identifies correlations between main and auxiliary channel glitches through a novel integration of CNN and Vision Transformers.
■590 ▼aSchool code: 0163.
■650 4▼aElectrical engineering
■650 4▼aComputer science
■650 4▼aMedical imaging
■653 ▼aDeep learning
■653 ▼aGravity Spy
■653 ▼aCT hemorrhage
■653 ▼aWeakly supervised learning
■690 ▼a0544
■690 ▼a0984
■690 ▼a0574
■690 ▼a0800
■71020▼aNorthwestern University▼bElectrical and Computer Engineering.
■7730 ▼tDissertations Abstracts International▼g85-11B.
■790 ▼a0163
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17161571▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


