電腦斷層血管造影之頸動脈自動分割與管徑評估：TotalSegmentator 與 SAM2 的比較研究;Automatic Segmentation and Diameter Estimation of Carotid Arteries in CT Angiography: A Comparative Study of TotalSegmentator and SAM2

NCU Institutional Repository > 生醫理工學院 > 生物醫學工程研究所 > 博碩士論文 > Item 987654321/97535

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/97535

題名:	電腦斷層血管造影之頸動脈自動分割與管徑評估：TotalSegmentator 與 SAM2 的比較研究;Automatic Segmentation and Diameter Estimation of Carotid Arteries in CT Angiography: A Comparative Study of TotalSegmentator and SAM2
作者:	胡俊生;Hu, Jun-Sheng
貢獻者:	生醫科學與工程學系
關鍵詞:	自動分割;管徑評估;頸動脈;電腦斷層血管造影;Automatic Segmentation;Diameter Estimation;Carotid Arteries;CT Angiography
日期:	2025-07-31
上傳時間:	2025-10-17 11:30:00 (UTC+8)
出版者:	國立中央大學
摘要:	頭頸部電腦斷層血管造影（CTA）是評估頸動脈狹窄與阻塞的重要影像工具，對中風等腦血管疾病的臨床診斷具有關鍵意義。頸動脈的精確分割對於電腦輔助診斷系統的建立與準確性至為重要。本研究提出一套兩階段的頸動脈自動分割流程，結合深度學習模型 TotalSegmentator 與 SAM2 ，應用於無人工標註的 CTA 影像。首先，使用 TotalSegmentator 初步偵測總頸動脈與內頸動脈，取得遮罩與中心點資訊；接著，沿 Z 軸逐切片以多階段提示策略導入 SAM2 進行細緻分割，根據遮罩面積與中心點距離進行多重判斷與提示優化。研究資料包括 16 受試者 32 頸動脈，其中有支架 1 處。從 16 組 CTA 資料中，TotalSegmentator 成功辨識所有的左右頸動脈與左右內頸動脈的位置而產生 9249 張切片。接著，SAM2 從 9249 張切片中正確分割出 8911 張切片，分割正確率達 96.35%。為解決骨質遮蔽造成的內頸動脈遠端穿骨血管之辨識困難，本研究進一步使用 Matched Mask Bone Elimination（MMBE）前處理技術，有效抑制骨骼干擾問題，增加內頸動脈分割的範圍。此外，本研究亦針對血管邊界進行評估管徑之量化比較分析，提出以 Piecewise Linear Model（PLM）擬合灰階變化曲線的方式估計血管管壁橫切面邊界，並提出峰值 (ridge)、中間值(mid)、基底(base)三種管壁位置判定策略。實驗結果顯示，在無支架血管中，SAM2 所估邊界介於 base 和 mid 位置之間，血管直徑評估為7.06 ± 0.41mm 比 mid 血管直徑評估5.55 ± 0.38mm 略高1.51 ± 0.56mm。本研究所提出之全自動整合性分割與血管管壁邊界辨識方法，能兼顧準確性與運算效率，並具備高度的臨床應用潛力。此方法有望進一步發展為頸動脈病變的量化篩檢與治療決策輔助系統。;Head and neck computed tomography angiography (CTA) is a critical imaging modality for evaluating carotid artery stenosis and occlusion. It plays an essential role in the clinical diagnosis of cerebrovascular diseases like ischemic stroke. Accurate segmentation of the carotid arteries is crucial for the development of precise computer-aided diagnosis (CAD) systems. This study proposes a two-stage automatic segmentation pipeline that integrates deep learning models—TotalSegmentator and SAM2—applied to unlabeled CTA images. Initially, TotalSegmentator is employed to detect the common and internal carotid arteries, providing initial segmentation masks and center point information. Subsequently, SAM2 is applied sliceby-slice along the Z-axis with a multi-prompt strategy, using mask area and center point distance as criteria for iterative refinement and validation. This study included 16 subjects, encompassing 32 carotid arteries (one of which contained a stent). TotalSegmentator successfully identified all left and right common and internal carotid arteries, generating 9249 cross-sectional patches of carotid arterial images. SAM2 was able to segment carotid artery, achieving a success rate of 96.35%. To address the challenge of bone interference in vascular recognition, we applied a preprocessing technique, Matched Mask Bone Elimination (MMBE), to suppress bone artifacts and improve segmentation performance in complex anatomical regions. In addition, we propose a fine-grained boundary estimation method using a Piecewise Linear Model (PLM) to fit grayscale intensity curves in polar coordinates. Based on this model, the arterial diameter was estimated using three boundary determination strategies: ridge, mid, iii and base. Experimental results show that SAM2’s diameter estimations fell between base- and mid-value diameter estimates. Specifically, SAM2’s diameter estimates were 7.06 ± 0.41mm, compared to 5.55 ± 0.38mm for the mid-value method, resulting in a difference of 1.51 ± 0.56mm. The proposed integrated framework achieved both high segmentation accuracy and computational efficiency. It has the potential to serve as a foundation for automated carotid disease screening and decision support systems.
顯示於類別:	[生物醫學工程研究所 ] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	2	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....