以機器學習方法應用於小兒中央靜脈導管 插管深度評估之研究

Abstract

中央靜脈導管（Central Venous Catheter, CVC）廣泛應用於兒童重症醫療中，用於給藥、營養補充、血流動力監測與抽血等多種臨床需求。然而，導管插入後若未正確定位，可能導致心律不整、血管穿孔、栓塞、藥物無效或其他併發症。因此，術後影像確認導管尖端位置是臨床照護中不可或缺的步驟。儘管目前最常用的影像工具為胸腔X光（Chest Radiograph, CXR），但其影像對比度較低，使得人工解讀耗時、易受主觀經驗影響，特別在兒科影像中更為困難。兒童胸腔解剖與成人存在差異，例如氣管與心房距離較短、肋骨與椎體比例不同、膈肌位置較高等，使得成人常用之定位基準（如氣管隆突或 T4–T5 椎體）不再適用。因此，需發展專為兒童設計的導管定位與深度評估方法。本研究提出一套以肋骨為參考的深度學習系統，用於兒童 中央靜脈導管位置的自動偵測與深度分類。我們建立一個多任務模型，結合導管尖端偵測（使用 YOLOv9）與肋骨分割（使用 U-Net）模組，再透過自行開發的 Rib Score Algorithm產生肋骨代表點，進而將尖端分類為過淺、適當或過深。資料來自台大兒童醫院胸腔X光影像，我們共建構三種測試組合：（a）混合組（包含未遮蔽、遮蔽與無導管影像）、（b）未遮蔽組、（c）遮蔽組，以評估模型在不同可視條件下的泛化能力。評估指標包含平均精確率（mAP50）、IoU、中心距離誤差、準確率、召回率與F1-score。實驗結果顯示，本系統在混合組達到 mAP50 = 0.987、mAP50–95 = 0.515，平均 IoU 為 0.71，中心誤差為 5.8 ± 2.1 像素，並在遮蔽條件下仍維持高準確性，顯示本系統具備良好空間準確性。進一步地，導管深度分類的平均準確率達 96%，大幅減少傳統人工判讀所需的時間與主觀誤差。系統亦具備可視化功能，可顯示導管與肋骨結構之對應關係，提供放射科醫師直觀且可解釋的判讀依據。本研究除提供一套即時、自動且具臨床適用性的 中央靜脈導管定位工具外，亦展現以肋骨為參考架構的潛力，未來可應用於其他胸腔結構之定位，例如肺炎區塊對應肺葉位置、膈肌抬升程度監測等。此外，肋骨標記機制亦可推進序列性影像追蹤分析，使臨床醫師可監控導管在病患住院期間的移動情形，進一步發展預測導管移位風險的模型。總結而言，本研究建立了一個具準確性、可擴展性與臨床價值的影像分析系統，為兒童中央靜脈導管定位問題提供創新解決方案。

Central venous catheters (CVCs) are widely used in pediatric intensive care for various clinical purposes, including drug administration, nutritional support, hemodynamic monitoring, and blood sampling. However, improper placement of the catheter tip after insertion can lead to serious complications such as arrhythmia, vascular perforation, embolism, treatment failure, or other adverse events. Therefore, post-operative imaging to confirm the tip position is an essential step in clinical care. Although chest radiographs (CXR) are the most commonly used imaging modality, their low contrast makes manual interpretation time-consuming and highly subjective—especially in pediatric cases. Anatomical differences between children and adults, such as shorter distances between the trachea and atrium, differing rib-to-spine proportions, and higher diaphragm position, render traditional adult-based landmarks (e.g., the carina or T4–T5 vertebrae) unsuitable for pediatric use. This necessitates the development of dedicated catheter localization and depth assessment methods for children.

In this study, we propose a rib-based deep learning system for automatic detection and depth classification of pediatric CVC tip positions. We developed a multi-task model combining a YOLOv9-based tip detection module and a U-Net-based rib segmentation module. A custom-designed Rib Score Algorithm is used to generate anatomical reference points along the ribs, allowing the tip position to be classified as shallow, appropriate, or deep. The dataset comprises pediatric chest X-ray images from National Taiwan University Children’s Hospital. We constructed three test sets to evaluate generalizability under different visibility conditions: (a) a mixed group (including unoccluded, occluded, and no-catheter images), (b) an unoccluded group, and (c) an occluded group.

Evaluation metrics include mean average precision (mAP50), mAP50–95, intersection over union (IoU), center point error, accuracy, recall, and F1-score. Our system achieved strong spatial performance in the mixed group (mAP50 = 0.987, mAP50–95 = 0.515), with a mean IoU of 0.71 and a center point error of 5.8 ± 2.1 pixels. Even under occluded conditions, the system maintained high accuracy. For depth classification, the average accuracy reached 96%, significantly reducing the time and subjectivity associated with manual interpretation. The system also provides visual outputs linking the catheter and rib structures, offering intuitive and interpretable references for radiologists. Beyond being a real-time, automated, and clinically applicable CVC localization tool, our rib-based framework shows promise for broader thoracic applications, such as identifying pneumonia regions by lung lobe, monitoring diaphragmatic elevation, and enabling longitudinal tracking. The rib labeling mechanism may further support sequential imaging analysis to monitor catheter migration during hospitalization and facilitate the development of predictive models for dislocation risk. In summary, this study presents an accurate, scalable, and clinically valuable image analysis system, offering an innovative solution for pediatric CVC localization.