心臟超音波影像左心室之自動圈選與低品質影像之輪廓修補

Automatic Left Ventricle Segmentation in the Presence of Poor Echocardiographic Image Quality Using Machine Learning and Contour Fitting

蔡旻潔(Min-Chieh Tsai)

指導教授 ： 李百祺

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202200463

摘要

心臟超音波能透過不同切面角度的影像評估整體心臟功能及結構，其中心尖四腔室視圖(Apical 4-chamber views, A4C)是定量左心室收縮功能的標準切面，可以透過圈選心內膜輪廓，進而計算左心室射出分率(Left Ventricle Ejection Fraction, LVEF)、整體縱向型變(Global Longitudinal Strain, GLS)臨床指標來定量左心室功能。然而影像雜訊及人為操作不當，導致不良品質的出現，進而影響診斷的正確率。目前臨床常見的商用軟體可自動圈選具有清楚輪廓的左心室影像，但在不良品質的影像中，會因無法判斷左心室輪廓而無法進行自動圈選或圈選錯誤，導致診斷錯誤率提高。為了解決上述問題，本論文提出的演算法分為兩步驟，首先透過臨床收集的良好品質影像與調整時間增益補償(temporal gain compensation, TGC)及側向增益補償(lateral gain compensation, LGC)的不良品質影像，訓練深度學習模型計算區域影像品質，再依據良好區域的輪廓及預先計算的平均輪廓模型，重建不良區域的左心室輪廓，以輔助商用軟體因左心室輪廓不清楚無法進行自動圈選或圈選錯誤之問題。研究結果顯示，對於良好品質的心尖四腔室視圖影像，本論文提出的方法可以即時的自動化圈選左心室和計算臨床指標。利用 Dice 指標做分析，發現在相同收案機器中獲得的影像，自動化圈選可達到 93.7%的準確度，而從不同機器中獲得的影像能達到 82.7%準確度，利用臨床指標分析，演算法計算整體縱向型變和左心室射出分率的誤差分別是 13.91% 和 8.31%，圈選結果達到現有文獻標準。對於不良品質的心尖四腔室視圖影像，在調整時間增益補償的影像，利用商用軟體僅能正確自動化圈選 30%受試者之左心室影像，而利用本論文提出之演算法可重建 60%受試者左心室輪廓，在臨床指標上，整體縱向型變和左心室射出分率的誤差分別是 23.08%、7.04%，重建結果達到臨床標準，可以輔助商用軟體。而調整 LGC 影像後，則由於不良品質的範圍過大，不適合用此方法修補。目前演算法會受到不良品質範圍之限制，且僅適用於區域性不良品質的假影，因此未來期望能擴展方法於其他類型假影以及三維超音波影像，並且應用於可攜式裝置上，更廣泛輔助臨床診斷。

關鍵字

心臟超音波影像 ； 左心室圈選 ； 低品質影像 ； 深度學習網路 ； 輪廓修補

並列摘要

In echocardiography, the apical 4-chamber view (A4C) is used to segment the left ventricle (LV) and calculate parameters such as the Left Ventricle Ejection Fraction (LVEF) and the Global Longitudinal Strain (GLS) as quantitative clinical indicators for the systolic functions. However, image artifacts and improper scanning lead to questionable clinical diagnosis. Currently, even the commonly used commercial software cannot correctly segment poor-quality images, leading to diagnostic errors. A two-step algorithm is proposed in this research to solve the above problems. First, the algorithm trains a deep learning model to quantify the regional image quality. Next, a set of good-quality images was manipulated by artificially adjusting the temporal gain compensation (TGC) and lateral gain compensation (LGC) of individual images to obtain a corresponding poor-quality dataset while the original images were kept for subsequent comparisons. Then, based on the contour of the good-quality region and the pre-calculated average contour model, the LV contour of the poor-quality area is reconstructed. Results show that for good-quality A4C images, the proposed method can automatically segment the LV and calculate clinical parameters in real-time. The images collected by the same devices achieved 93.7% accuracy in the dice coefficient, and the images collected by different devices achieved 82.7% accuracy. The GLS and LVEF errors were 13.91% and 8.31%, respectively, and the segmentation results achieved the existing clinical standards. For poor-quality A4C images, in the TGC testing data, only 30% were successfully segmented by the commercial software. The proposed algorithm achieved a 60% successful reconstruction. The GLS error and the LVEF error were 23.08% and 7.04%, respectively. The reconstruction results achieved the clinical standards. However, for the LGC testing data, the angular extent of the missing contour is too large to be successfully reconstructed by this method. Finally, we expect to apply the method to test other types of artifacts and extend to portable devices and 3D imaging in the future.

並列關鍵字

Cardiac ultrasound ； left ventricle segmentation ； deep learning ； poor image quality ； contour reconstruction

參考文獻

[1] I. Aly et al., "Cardiac ultrasound: an anatomical and clinical review," Translational Research in Anatomy, vol. 22, p. 100083, 2021, doi: /10.1016/j.tria.2020.100083.

[2] N. J. Soni, R. Arntfield, and P. Kory, Point of care ultrasound e-book. Elsevier Health Sciences, 2019.

[3] J. A. Noble and D. Boukerroui, "Ultrasound image segmentation: a survey," IEEE Transactions on medical imaging, vol. 25, no. 8, pp. 987-1010, 2006, doi: 10.1109/tmi.2006.877092.

[4] R. Abou, P. van der Bijl, J. J. Bax, and V. Delgado, "Global Longitudinal Strain: clinical use and prognostic Implications in contemporary practice," Heart, vol. 106, no. 18, pp. 1438-1444, 2020, doi: 10.1136/heartjnl-2019-316215.

[5] A. Kosaraju, A. Goyal, Y. Grigorova, and A. N. Makaryus, Left ventricular ejection fraction. 2017.