基於 CCTA 的冠狀動脈分割與 AI 代理式狹窄診斷之拓 樸感知多階段框架

Abstract

冠狀動脈電腦斷層血管攝影（coronary computed tomography angiography, CCTA）可提供冠狀動脈的三維解剖資訊，支援中心線擷取、管腔量測與狹窄分析。然而，遠端細小分支及局部狹窄管腔容易受到低對比、部分容積效應、鈣化與運動偽影影響，導致自動分割結果出現遺漏或斷裂，進而降低下游狹窄定位的可靠性。 本研究提出一個拓樸感知的三階段冠狀動脈分割框架，並將其整合至代理式冠狀動脈狹窄診斷流程。第一階段使用 SwinUNETR 建立全域粗分割；第二階段將粗分割結果骨架化，並利用中心線資訊選取局部三維區塊；第三階段使用 3D U-Net 細化所選區域，再透過機率層級融合整合全域與局部預測。此外，本研究提出 Thin-Branch-Aware Topology Dice（TBA-Dice），利用真實標註遮罩的距離轉換產生逆半徑權重，以加強細小分支與局部狹窄區域的拓樸監督。 在 ImageCAS 資料集上，完整框架達到 84.23% Dice、86.69% clDice 與 87.38% TBA-Dice。相較於僅使用第一階段的分割結果，完整框架的 Dice 提升 0.54 個百分點，TBA-Dice 提升 2.59 個百分點。在 NTUH 臨床測試資料中，平均 Dice 為 83.16%。當分割結果整合至以 CAD-RADS 2.0 為外部醫療知識來源的代理式系統後，狹窄節段定位率達到 48.48%，高於僅使用第一階段分割結果的 42.42%。研究結果顯示，較佳的血管連續性與結構完整性可為中心線式管腔分析及代理式冠狀動脈狹窄診斷提供更可靠的解剖基礎。

Coronary computed tomography angiography (CCTA) provides three-dimensional anatomical information of the coronary arteries and supports centerline extraction, lumen measurement, and stenosis analysis. However, distal small branches and locally narrowed lumen regions are susceptible to low contrast, partial-volume effects, calcification, and motion artifacts, resulting in omissions or discontinuities in automatic segmentation and reducing the reliability of downstream stenosis localization. This study proposes a topology-aware three-stage coronary artery segmentation framework and integrates it into an agentic coronary stenosis diagnosis workflow. In the first stage, SwinUNETR generates a global coarse segmentation. In the second stage, the coarse segmentation is skeletonized, and centerline information is used to select local three-dimensional patches. In the third stage, a 3D U-Net refines the selected regions, after which the global and local predictions are integrated through probability-level fusion. This study further introduces Thin-Branch-Aware Topology Dice (TBA-Dice), which derives inverse-radius weights from the distance transform of the ground-truth vessel mask to strengthen topological supervision for small branches and locally narrowed regions. On ImageCAS, the framework achieved 84.23% Dice, 86.69% clDice, and 87.38% TBA-Dice, improving Dice and TBA-Dice over the first stage by 0.54 and 2.59 percentage points. It achieved 83.16% Dice on NTUH. When the segmentation output was integrated into an agentic system using CAD-RADS 2.0 as an external medical knowledge source, the stenotic-segment localization rate reached 48.48%, compared with 42.42% using the first-stage segmentation alone. These results suggest that more complete vessel representations provide a more reliable anatomical basis for downstream lumen analysis and agentic coronary stenosis diagnosis.