서브메뉴
검색
Artificial Intelligence in Lymphoma Imaging: Incorporating Longitudinal Awareness
Artificial Intelligence in Lymphoma Imaging: Incorporating Longitudinal Awareness
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202104654
- ISBN
- 9798280770447
- DDC
- 616
- 저자명
- Tie, Xin.
- 서명/저자
- Artificial Intelligence in Lymphoma Imaging: Incorporating Longitudinal Awareness
- 발행사항
- [Sl] : The University of Wisconsin - Madison, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 143 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-12, Section: B.
- 주기사항
- Advisor: Bradshaw, Tyler J.
- 학위논문주기
- Thesis (Ph.D.)--The University of Wisconsin - Madison, 2025.
- 초록/해제
- 요약Hodgkin lymphoma is the most commonly diagnosed cancer in adolescents. Over the past decades, significant advances in combined modality therapy, risk stratification, and response-adapted therapy have dramatically improved patient outcomes. In pediatric populations, five-year overall survival exceeds 90%, with even higher rates observed in patients at the early stage of the disease. As a result, the therapeutic focus has shifted from maximizing cure to minimizing late effects from treatment-related toxicities.Effective lymphoma management relies on integrating clinical and imaging data across multiple time points to guide treatment strategies. In practice, this involves complex and time-intensive analysis, such as comparing baseline and interim 18F-fluorodeoxyglucose (18F-FDG) PET/CT scans to assess therapeutic response. Despite the availability of rich longitudinal data, most existing artificial intelligence (AI) algorithms are designed to analyze single-time-point images, overlooking the temporal evolution of the disease. To address this gap, this dissertation systematically investigates the integration of longitudinal awareness into AI models for image-guided lymphoma management. Specifically, it explores how leveraging imaging data across different phases of treatment can improve disease quantification, prognostic modeling, and radiation treatment planning.In the first study, we developed a longitudinally aware segmentation network (LAS-Net) for automatic quantification of serial PET/CT images in high-risk pediatric Hodgkin lymphoma. The algorithm utilized longitudinal cross-attention to jointly analyze baseline and interim PET scans. The model was trained and evaluated on a retrospective dataset of 200 patients from the Children's Oncology Group AHOD1331 clinical trial. Results demonstrated that LAS-Net achieved significant improvements in residual lesion detection on interim PET compared to other methods, without sacrificing segmentation accuracy on baseline PET. In PET quantification, LAS-Net's measurements of several key baseline and interim PET metrics (e.g., total metabolic tumor volume or TMTV, and percentage difference between baseline and interim maximum standardized uptake value or ∆SUVmax) were strongly correlated with physician measurements. These results highlight the potential of LAS-Net to deliver rapid and consistent assessment of PET tumor burden and response.In the second study, we evaluated outcome prediction for high-risk pediatric Hodgkin lymphoma using imaging features derived from both baseline and interim PET scans. LAS-Net was applied to the remaining unlabeled cases in the AHOD1331 trial to generate tumor segmentations, which were then reviewed and corrected by nuclear medicine physicians to ensure accuracy. This process yielded a large, multi-institutional cohort of 559 pediatric patients. We assessed the prognostic value of common clinical risk factors, conventional PET metrics, and radiomics features. Among these, the model incorporating conventional PET metrics, particularly baseline TMTV and interim qPET (peak lesion SUV divided by mean liver SUV), demonstrated the highest prognostic performance, significantly outperforming models based solely on clinical variables. Furthermore, we evaluated the prognostic utility of PET metrics extracted from deep learning (DL)-based segmentations and found that models using DL-derived features could achieve performance comparable to those built on physician-derived features. This study represents one of the most comprehensive survival analyses conducted in pediatric Hodgkin lymphoma, offering valuable insights for enhancing risk stratification.The third study focused on automatic delineation of clinical target volume (CTV) in involved-site radiation therapy (ISRT) for Hodgkin lymphoma. ISRT targets initially involved lymph nodes and tissues while accounting for anatomical changes following chemotherapy. In clinical practice, radiation oncologists rely on FDG PET imaging, especially baseline PET/CT, for CTV delineation. In this work, we investigated various DL architectures and evaluated the impact of incorporating baseline and interim PET images alongside the planning CT. Comparative analysis showed that models incorporating PET/CT images significantly improved CTV segmentation accuracy compared to models based on planning CT alone. Evaluation against inter-observer variability and independent clinical review supported that DL-generated contours were of comparable quality to those manually created by radiation oncologists. Importantly, over half of the DL-generated CTVs were considered clinically acceptable with minor or no changes needed, highlighting the model's real-world potential to facilitate ISRT planning by providing high-quality initial contours for radiation oncologists to review and refine.The final study broadened the scope of this dissertation by applying the longitudinally aware segmentation framework to MR-guided adaptive radiation therapy in head and neck cancer. Specifically, we focused on improving the segmentation accuracy of mid-treatment gross tumor volume (GTV) on MR images by incorporating pre-treatment GTV information. Consistent with earlier findings, the integration of prior imaging significantly enhanced model performance, achieving segmentation accuracy comparable to the inter-observer variability among expert radiation oncologists. These results demonstrate the strong potential of our model to streamline clinical workflows in radiation oncology and, more importantly, validate the broader applicability of the proposed framework beyond lymphoma.Collectively, the studies presented in this dissertation highlight the importance of integrating longitudinal imaging data into AI model development. These findings lay the foundation for future research and provide a compelling case for the prospective clinical integration and evaluation of longitudinally aware AI tools to advance personalized cancer care.
- 일반주제명
- Medical imaging
- 일반주제명
- Physics
- 키워드
- Deep learning
- 키워드
- Lymphoma
- 키워드
- Nuclear medicine
- 키워드
- Outcome analysis
- 키워드
- Segmentation
- 기타저자
- The University of Wisconsin - Madison Medical Physics
- 기본자료저록
- Dissertations Abstracts International. 86-12B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017358390
■00520260202104654
■006m o d
■007cr#unu||||||||
■020 ▼a9798280770447
■035 ▼a(MiAaPQ)AAI32115687
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a616
■1001 ▼aTie, Xin.
■24510▼aArtificial Intelligence in Lymphoma Imaging: Incorporating Longitudinal Awareness
■260 ▼a[Sl]▼bThe University of Wisconsin - Madison▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a143 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-12, Section: B.
■500 ▼aAdvisor: Bradshaw, Tyler J.
■5021 ▼aThesis (Ph.D.)--The University of Wisconsin - Madison, 2025.
■520 ▼aHodgkin lymphoma is the most commonly diagnosed cancer in adolescents. Over the past decades, significant advances in combined modality therapy, risk stratification, and response-adapted therapy have dramatically improved patient outcomes. In pediatric populations, five-year overall survival exceeds 90%, with even higher rates observed in patients at the early stage of the disease. As a result, the therapeutic focus has shifted from maximizing cure to minimizing late effects from treatment-related toxicities.Effective lymphoma management relies on integrating clinical and imaging data across multiple time points to guide treatment strategies. In practice, this involves complex and time-intensive analysis, such as comparing baseline and interim 18F-fluorodeoxyglucose (18F-FDG) PET/CT scans to assess therapeutic response. Despite the availability of rich longitudinal data, most existing artificial intelligence (AI) algorithms are designed to analyze single-time-point images, overlooking the temporal evolution of the disease. To address this gap, this dissertation systematically investigates the integration of longitudinal awareness into AI models for image-guided lymphoma management. Specifically, it explores how leveraging imaging data across different phases of treatment can improve disease quantification, prognostic modeling, and radiation treatment planning.In the first study, we developed a longitudinally aware segmentation network (LAS-Net) for automatic quantification of serial PET/CT images in high-risk pediatric Hodgkin lymphoma. The algorithm utilized longitudinal cross-attention to jointly analyze baseline and interim PET scans. The model was trained and evaluated on a retrospective dataset of 200 patients from the Children's Oncology Group AHOD1331 clinical trial. Results demonstrated that LAS-Net achieved significant improvements in residual lesion detection on interim PET compared to other methods, without sacrificing segmentation accuracy on baseline PET. In PET quantification, LAS-Net's measurements of several key baseline and interim PET metrics (e.g., total metabolic tumor volume or TMTV, and percentage difference between baseline and interim maximum standardized uptake value or ∆SUVmax) were strongly correlated with physician measurements. These results highlight the potential of LAS-Net to deliver rapid and consistent assessment of PET tumor burden and response.In the second study, we evaluated outcome prediction for high-risk pediatric Hodgkin lymphoma using imaging features derived from both baseline and interim PET scans. LAS-Net was applied to the remaining unlabeled cases in the AHOD1331 trial to generate tumor segmentations, which were then reviewed and corrected by nuclear medicine physicians to ensure accuracy. This process yielded a large, multi-institutional cohort of 559 pediatric patients. We assessed the prognostic value of common clinical risk factors, conventional PET metrics, and radiomics features. Among these, the model incorporating conventional PET metrics, particularly baseline TMTV and interim qPET (peak lesion SUV divided by mean liver SUV), demonstrated the highest prognostic performance, significantly outperforming models based solely on clinical variables. Furthermore, we evaluated the prognostic utility of PET metrics extracted from deep learning (DL)-based segmentations and found that models using DL-derived features could achieve performance comparable to those built on physician-derived features. This study represents one of the most comprehensive survival analyses conducted in pediatric Hodgkin lymphoma, offering valuable insights for enhancing risk stratification.The third study focused on automatic delineation of clinical target volume (CTV) in involved-site radiation therapy (ISRT) for Hodgkin lymphoma. ISRT targets initially involved lymph nodes and tissues while accounting for anatomical changes following chemotherapy. In clinical practice, radiation oncologists rely on FDG PET imaging, especially baseline PET/CT, for CTV delineation. In this work, we investigated various DL architectures and evaluated the impact of incorporating baseline and interim PET images alongside the planning CT. Comparative analysis showed that models incorporating PET/CT images significantly improved CTV segmentation accuracy compared to models based on planning CT alone. Evaluation against inter-observer variability and independent clinical review supported that DL-generated contours were of comparable quality to those manually created by radiation oncologists. Importantly, over half of the DL-generated CTVs were considered clinically acceptable with minor or no changes needed, highlighting the model's real-world potential to facilitate ISRT planning by providing high-quality initial contours for radiation oncologists to review and refine.The final study broadened the scope of this dissertation by applying the longitudinally aware segmentation framework to MR-guided adaptive radiation therapy in head and neck cancer. Specifically, we focused on improving the segmentation accuracy of mid-treatment gross tumor volume (GTV) on MR images by incorporating pre-treatment GTV information. Consistent with earlier findings, the integration of prior imaging significantly enhanced model performance, achieving segmentation accuracy comparable to the inter-observer variability among expert radiation oncologists. These results demonstrate the strong potential of our model to streamline clinical workflows in radiation oncology and, more importantly, validate the broader applicability of the proposed framework beyond lymphoma.Collectively, the studies presented in this dissertation highlight the importance of integrating longitudinal imaging data into AI model development. These findings lay the foundation for future research and provide a compelling case for the prospective clinical integration and evaluation of longitudinally aware AI tools to advance personalized cancer care.
■590 ▼aSchool code: 0262.
■650 4▼aMedical imaging
■650 4▼aPhysics
■653 ▼aDeep learning
■653 ▼aLymphoma
■653 ▼aNuclear medicine
■653 ▼aOutcome analysis
■653 ▼aRadiation therapy
■653 ▼aSegmentation
■690 ▼a0574
■690 ▼a0800
■690 ▼a0605
■71020▼aThe University of Wisconsin - Madison▼bMedical Physics.
■7730 ▼tDissertations Abstracts International▼g86-12B.
■790 ▼a0262
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17358390▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


