東華大學圖書館 |

語系: 繁體中文

說明(常見問題)

回圖書館首頁

手機版館藏查詢

登入

回首頁

切換: 標籤 | MARC模式 | ISBD

Multimodal Vision-Language Modeling ...

Huemann, Zachary,

FindBook

Google Book

Amazon

博客來

Multimodal Vision-Language Modeling for Advanced Quantitative Analysis of Positron Emission Tomography Imaging /

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Multimodal Vision-Language Modeling for Advanced Quantitative Analysis of Positron Emission Tomography Imaging // Zachary Huemann.
作者:	Huemann, Zachary,
面頁冊數:	1 electronic resource (109 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-09, Section: B.
Contained By:	Dissertations Abstracts International86-09B.
標題:	Physics. -
電子資源:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31845698
ISBN:	9798310120358

Multimodal Vision-Language Modeling for Advanced Quantitative Analysis of Positron Emission Tomography Imaging /
Huemann, Zachary,

Multimodal Vision-Language Modeling for Advanced Quantitative Analysis of Positron Emission Tomography Imaging /Zachary Huemann. - 1 electronic resource (109 pages)

Source: Dissertations Abstracts International, Volume: 86-09, Section: B.

Positron emission tomography (PET) provides powerful functional imaging that can be critical in diagnosing and monitoring many diseases. However, extracting quantitative biomarkers from PET/CT scans, such as total metabolic tumor volume, remains time-consuming. Radiology reports, in turn, offer a rich yet underutilized summary of all the imaging findings. We hypothesized that these reports can serve as a valuable source of guidance to aid in the automated extraction of quantitative biomarkers from PET images. This dissertation explores how to integrate these textual descriptions with PET images using multimodal vision-language models to advance automated quantitative analysis.We first demonstrate that large language models (LLMs) are well-suited for interpreting domain-specific nuclear medicine text, outperforming vision-only methods on the task of Deauville score classification. We further show that adapting LLMs with additional pretraining on nuclear medicine corpora improves performance compared to generic language models.Next, we then explored vision-language fusion methods and presented ConTEXTual Net, a U-Net-based segmentation framework that integrated physician-generated text with 2D images. We worked in the simplified 2D setting of pneumothorax segmentation in chest X-rays where our multimodal approach achieves higher segmentation accuracy than vision-only models and performs similar to inter-radiologist agreement. Finally, we extend this framework to 3D PET/CT using a novel weak-labeling pipeline, which leverages radiology report findings containing standardized uptake values and slice references to generate lesion-level annotations. This large-scale dataset enables the training of a full 3D vision-language segmentation model, ConTEXTual Net 3D, which accurately localizes lesions based on their textual descriptions in a task called visual grounding. By integrating imaging and text data, the proposed multimodal model not only reduces the burden of manual lesion delineation but also provides interactive and interpretable outputs. These methods highlight the untapped potential of radiology reports as a source of domain expertise. They point to future directions where multimodal models can extract biomarkers, improve workflows, and enhance physician-patient communication.

English

ISBN: 9798310120358Subjects--Topical Terms:

516296
Physics.
Subjects--Index Terms:

Positron emission tomography

Multimodal Vision-Language Modeling for Advanced Quantitative Analysis of Positron Emission Tomography Imaging /
LDR:03826nmm a22004213i 4500 001 2432285
005 20250707081335.5
006 m o d
007 cr|nu||||||||
008 261011s2025 miu||||||m |||||||eng d
020 $a 9798310120358
035 $a (MiAaPQD)AAI31845698
035 $a AAI31845698
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Huemann, Zachary, $e author. $3 3821445
245 1 0 $a Multimodal Vision-Language Modeling for Advanced Quantitative Analysis of Positron Emission Tomography Imaging / $c Zachary Huemann.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2025
300 $a 1 electronic resource (109 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-09, Section: B.
500 $a Advisors: Dasu, Sridhara; Bradshaw, Tyler J. Committee members: McMillan, Alan B.; Hu, Junjie.
502 $b Ph.D. $c The University of Wisconsin - Madison $d 2025.
520 $a Positron emission tomography (PET) provides powerful functional imaging that can be critical in diagnosing and monitoring many diseases. However, extracting quantitative biomarkers from PET/CT scans, such as total metabolic tumor volume, remains time-consuming. Radiology reports, in turn, offer a rich yet underutilized summary of all the imaging findings. We hypothesized that these reports can serve as a valuable source of guidance to aid in the automated extraction of quantitative biomarkers from PET images. This dissertation explores how to integrate these textual descriptions with PET images using multimodal vision-language models to advance automated quantitative analysis.We first demonstrate that large language models (LLMs) are well-suited for interpreting domain-specific nuclear medicine text, outperforming vision-only methods on the task of Deauville score classification. We further show that adapting LLMs with additional pretraining on nuclear medicine corpora improves performance compared to generic language models.Next, we then explored vision-language fusion methods and presented ConTEXTual Net, a U-Net-based segmentation framework that integrated physician-generated text with 2D images. We worked in the simplified 2D setting of pneumothorax segmentation in chest X-rays where our multimodal approach achieves higher segmentation accuracy than vision-only models and performs similar to inter-radiologist agreement. Finally, we extend this framework to 3D PET/CT using a novel weak-labeling pipeline, which leverages radiology report findings containing standardized uptake values and slice references to generate lesion-level annotations. This large-scale dataset enables the training of a full 3D vision-language segmentation model, ConTEXTual Net 3D, which accurately localizes lesions based on their textual descriptions in a task called visual grounding. By integrating imaging and text data, the proposed multimodal model not only reduces the burden of manual lesion delineation but also provides interactive and interpretable outputs. These methods highlight the untapped potential of radiology reports as a source of domain expertise. They point to future directions where multimodal models can extract biomarkers, improve workflows, and enhance physician-patient communication.
546 $a English
590 $a School code: 0262
650 4 $a Physics. $3 516296
650 4 $a Medical imaging. $3 3172799
653 $a Positron emission tomography
653 $a Vision-language modeling
653 $a Large language models
690 $a 0605
690 $a 0800
690 $a 0574
710 2 $a The University of Wisconsin - Madison. $b Physics. $e degree granting institution. $3 3816229
720 1 $a Dasu, Sridhara $e degree supervisor.
720 1 $a Bradshaw, Tyler J. $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-09B.
790 $a 0262
791 $a Ph.D.
792 $a 2025
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31845698