서브메뉴
검색
Enhancing Accuracy and Plausibility of Unsupervised Deep-Learning-Based Deformable Image Registration With Registration-Specific Designs and Anatomical Priors
Enhancing Accuracy and Plausibility of Unsupervised Deep-Learning-Based Deformable Image Registration With Registration-Specific Designs and Anatomical Priors
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103618
- ISBN
- 9798315754503
- DDC
- 616
- 저자명
- Liu, Hengjie.
- 서명/저자
- Enhancing Accuracy and Plausibility of Unsupervised Deep-Learning-Based Deformable Image Registration With Registration-Specific Designs and Anatomical Priors
- 발행사항
- [Sl] : University of California, Los Angeles, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 206 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-11, Section: B.
- 주기사항
- Advisor: Sheng, Ke;Ruan, Dan.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Los Angeles, 2025.
- 초록/해제
- 요약Deformable image registration (DIR) is the computational process of aligning different images to a unified coordinate system through locally varying, non-linear (hence deformable) displacement fields. It is fundamental to many medical image analysis workflows where images from different time points, modalities, or patients need to be compared after spatial alignment. Conventional optimization-based DIR methods have seen limited clinical adoption due to persistent challenges in accuracy and efficiency. Recently, deep-learning-based deformable image registration (DL-DIR) has attracted interest for its speed and ability to leverage rich, data-driven feature representation. Unsupervised and weakly supervised approaches have become the main paradigms of DL-DIR.DIR faces three core challenges: the absence of ground truth, the problem's inherent ill-posedness, and its non-convexity. These intrinsic limitations not only complicate the evaluation of registration accuracy but also hinder the assessment of deformation plausibility. The most common accuracy surrogate for DIR is segmentation label matching (e.g., Dice score) due to its wide availability. As such, weak supervision via contour matching losses becomes popular to boost the Dice metric. Despite improving the apparent accuracy, this strategy often leads to unrealistic deformations and compromises generalizability, especially when applied to datasets with different labeling protocols. Consequently, unsupervised DL-DIR has regained prominence. However, current literature lacks fair, standardized comparisons among unsupervised DL-DIR methods, and offers little consensus on best practices for unsupervised DL-DIR. It is even unclear whether unsupervised DL-DIR truly outperforms conventional methods, and if so, what drives that advantage. Moreover, most DL-DIR methods remain intensity-driven with only simple smoothness regularization, neglecting rich physiological priors in human anatomy. This oversight not only limits accuracy but can yield anatomically implausible deformation fields, significantly hindering clinical translation.To address these challenges, we propose three specific aims. We first conduct a comprehensive ablation-type study to identify the true drivers for accurate unsupervised DL-DIR, which highlights the importance of registration-specific designs. Building upon that, we then tailor DL-DIR to two radiation-therapy applications at two distinct anatomical sites, enhancing accuracy and anatomical plausibility by integrating relevant anatomical information into the registration framework. The three aims are:Specific Aim 1: Enhance accuracy of unsupervised DL-DIR with registration-specific designs. We hypothesize that registration-specific designs such as the multi-resolution pyramid, correlation calculation, and inverse-consistency constraints are more important than complex network architectures. We propose a comprehensive ablation-type study to identify the key modules for unsupervised mono-modal DL-DIR and demonstrate that simple models, when properly equipped with these design choices, can achieve state-of-the-art performance.Specific Aim 2: Improve bladder trigone MRI registration by integrating anatomical context through multi-task learning. We hypothesize that the features learned in landmark prediction and segmentation tasks will improve registration accuracy. We propose a multi-task learning framework for joint landmark regression, segmentation, and deformation registration for the bladder trigone and demonstrate improved registration accuracy.Specific Aim 3: Improve accuracy and plausibility in head-and-neck CT registration using a MUsculo-Skeleton-Aware (MUSA) framework. We hypothesize registration of the complex and heterogeneous head-and-neck deformations can benefit from distinguishing the tissue types and motion types. We propose to decompose such complex deformations into bulk posture changes and residual fine deformations and distinguish between the rigidity of bone structures and the flexibility of soft tissues. With such anatomical knowledge incorporated into the DIR optimization process, we demonstrate both enhanced accuracy and markedly better plausibility of deformation fields.
- 일반주제명
- Medical imaging
- 일반주제명
- Biophysics
- 일반주제명
- Biomedical engineering
- 일반주제명
- Computer science
- 키워드
- Deep learning
- 기타저자
- University of California, Los Angeles Physics and Biology in Medicine 009Y
- 기본자료저록
- Dissertations Abstracts International. 86-11B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357923
■00520260202103618
■006m o d
■007cr#unu||||||||
■020 ▼a9798315754503
■035 ▼a(MiAaPQ)AAI32045025
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a616
■1001 ▼aLiu, Hengjie.
■24510▼aEnhancing Accuracy and Plausibility of Unsupervised Deep-Learning-Based Deformable Image Registration With Registration-Specific Designs and Anatomical Priors
■260 ▼a[Sl]▼bUniversity of California, Los Angeles▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a206 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-11, Section: B.
■500 ▼aAdvisor: Sheng, Ke;Ruan, Dan.
■5021 ▼aThesis (Ph.D.)--University of California, Los Angeles, 2025.
■520 ▼aDeformable image registration (DIR) is the computational process of aligning different images to a unified coordinate system through locally varying, non-linear (hence deformable) displacement fields. It is fundamental to many medical image analysis workflows where images from different time points, modalities, or patients need to be compared after spatial alignment. Conventional optimization-based DIR methods have seen limited clinical adoption due to persistent challenges in accuracy and efficiency. Recently, deep-learning-based deformable image registration (DL-DIR) has attracted interest for its speed and ability to leverage rich, data-driven feature representation. Unsupervised and weakly supervised approaches have become the main paradigms of DL-DIR.DIR faces three core challenges: the absence of ground truth, the problem's inherent ill-posedness, and its non-convexity. These intrinsic limitations not only complicate the evaluation of registration accuracy but also hinder the assessment of deformation plausibility. The most common accuracy surrogate for DIR is segmentation label matching (e.g., Dice score) due to its wide availability. As such, weak supervision via contour matching losses becomes popular to boost the Dice metric. Despite improving the apparent accuracy, this strategy often leads to unrealistic deformations and compromises generalizability, especially when applied to datasets with different labeling protocols. Consequently, unsupervised DL-DIR has regained prominence. However, current literature lacks fair, standardized comparisons among unsupervised DL-DIR methods, and offers little consensus on best practices for unsupervised DL-DIR. It is even unclear whether unsupervised DL-DIR truly outperforms conventional methods, and if so, what drives that advantage. Moreover, most DL-DIR methods remain intensity-driven with only simple smoothness regularization, neglecting rich physiological priors in human anatomy. This oversight not only limits accuracy but can yield anatomically implausible deformation fields, significantly hindering clinical translation.To address these challenges, we propose three specific aims. We first conduct a comprehensive ablation-type study to identify the true drivers for accurate unsupervised DL-DIR, which highlights the importance of registration-specific designs. Building upon that, we then tailor DL-DIR to two radiation-therapy applications at two distinct anatomical sites, enhancing accuracy and anatomical plausibility by integrating relevant anatomical information into the registration framework. The three aims are:Specific Aim 1: Enhance accuracy of unsupervised DL-DIR with registration-specific designs. We hypothesize that registration-specific designs such as the multi-resolution pyramid, correlation calculation, and inverse-consistency constraints are more important than complex network architectures. We propose a comprehensive ablation-type study to identify the key modules for unsupervised mono-modal DL-DIR and demonstrate that simple models, when properly equipped with these design choices, can achieve state-of-the-art performance.Specific Aim 2: Improve bladder trigone MRI registration by integrating anatomical context through multi-task learning. We hypothesize that the features learned in landmark prediction and segmentation tasks will improve registration accuracy. We propose a multi-task learning framework for joint landmark regression, segmentation, and deformation registration for the bladder trigone and demonstrate improved registration accuracy.Specific Aim 3: Improve accuracy and plausibility in head-and-neck CT registration using a MUsculo-Skeleton-Aware (MUSA) framework. We hypothesize registration of the complex and heterogeneous head-and-neck deformations can benefit from distinguishing the tissue types and motion types. We propose to decompose such complex deformations into bulk posture changes and residual fine deformations and distinguish between the rigidity of bone structures and the flexibility of soft tissues. With such anatomical knowledge incorporated into the DIR optimization process, we demonstrate both enhanced accuracy and markedly better plausibility of deformation fields.
■590 ▼aSchool code: 0031.
■650 4▼aMedical imaging
■650 4▼aBiophysics
■650 4▼aBiomedical engineering
■650 4▼aComputer science
■653 ▼aDeep learning
■653 ▼aDeformable image registration
■653 ▼aMedical image analysis
■653 ▼aRadiation therapy
■653 ▼aUnsupervised learning
■690 ▼a0574
■690 ▼a0786
■690 ▼a0541
■690 ▼a0984
■71020▼aUniversity of California, Los Angeles▼bPhysics and Biology in Medicine 009Y.
■7730 ▼tDissertations Abstracts International▼g86-11B.
■790 ▼a0031
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357923▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


