- 學位論文
以電腦視覺強化心臟超音波測量整體縱向形變之信度
Improving the Application Reliability and Validity of Global Longitudinal Strain by Computer Vision
摘要
惡性腫瘤和心血管病都是嚴重危害人類健康之主要的慢性非傳染性疾病，也是我國居民死亡的主要原因。心血管疾病不僅僅是癌症治療最常見的不良反應之一，亦是癌症存活者早期致死致殘的重要原因。在2016年歐洲心臟學會大會上，ESC發布了關於癌症治療與心血管毒性相關問題所持見解的立場文件。專家組將癌症治療相關的心功能障礙定義為：癌症患者首次心臟造影及2-3周後複診確定左心室射出分數降低超過10%，數值降至53%以下。此外，一新興的心臟功能性參數——左心室整體縱向形變相較於左心室射出分率在偵測亞臨床心肌障礙上較為靈敏，亦被列為癌症病患的心臟超音波檢查必備項目之一。左心室整體縱向形變數值比化療前相對值下降15%亦被認為是有意義的左心室功能受損。 左心室整體縱向形變剛發展出來時，不同廠牌機器衍生出來的數值相差甚大，甚至足以導致錯誤的臨床決策。在左心室射出分率無法完全解釋與預測心臟病患預後，而左心室整體縱向形變之信度與效度仍不夠之時，吾人開啟了右心室功能評估的研究。透過三維心臟超音波與右心室對焦擷像，吾人分析了心臟重症病患使用葉克膜循環輔助時之右心室功能，發現右心室射出分率雖然具有臨床預後之價值，但並非僅僅是反映單一右心室的功能，更是代表了左右心室與體液狀態的總合結果。因此，右心室射出分率雖然可以反映預後，但對治療標的的選擇並不能提供更多的資訊。吾人進一步發展了非侵入性的右心室壓力體積分析方法，但因為是原型軟體，暫時無法搭載於超音波主機上進行即時分析與臨床決策。然而就在此時，左心室整體縱向形變的跨廠牌變異度在國際各個心臟學會以及超音波廠家的協同合作下已大幅改善，並有學術期刊發表。 吾人欲確認左心室整體縱向形變是否具備臨床診斷預後之價值，進行了一項回溯性研究。主要是針對台大神經內科診斷之家族性類澱粉多發神經病變病患族群。透過此項研究，吾人發現此一族群因為神經系統侵犯造成的病患活動力下降，因此心臟衰竭症狀往往被忽略，而心肌結構改變卻仍持續進行，此族群病患主要死因也是心臟衰竭。在此一族群之中，左心室整體縱向形變在預後評估的表現明顯優於左心室射出分率，然而，吾人發現有一部分的病患因為心臟超音波檢查時擷像品質不良，無法進行左心室整體縱向形變分析。擷像品質不良對左心室整體縱向形變測量變異的影響開始在國際期刊上被廣泛發表討論。心臟超音波核心實驗室的概念因此被提出，但因為影像分析無法免除人為主觀因素，因此雖然在核心室驗室內的形變分析信度良好，但是跨不同核心實驗室的形變分析仍然存在不可忽視的變異。 在此一同時，人工智慧電腦視覺也迎接了革命性的發展。Geoffrey Everest Hinton教授與其同事在1986年所提出的反向傳播法，奠定了現今深度學習技術的基礎。2006年Hinton教授在Science期刊發表了”Reducing the dimensionality of data with neural network” 實現了非監督學習的可能性。但當時CPU運算效能太差，因此深度學習仍未普及。在2012年，Hinton的學生Alex Krizhevsky利用NVIDIA的GPU加上深度學習卷積神經網路，以懸殊差距獲得ImageNet大規模視覺辨識挑戰賽的優勝，從此開啟了卷積神經網路在圖像辨識的濫觴。Jeffrey Zhang於2018年提出深度學習心臟超音波影像診斷的流程，即1.擷像辨識，2. 影像分割，3.心臟腔室結構與功能量化以及4.疾病偵測。然而，最重要的源頭控管——擷像品質評估仍然未受到重視。 綜合前面所述，吾人假設自動化左心室整體縱向形變分析可以減少整體縱向形變因人為分析技術所導致之測量偏差，如此一來，形變分析的主要偏差因素幾乎只剩下擷像品質本身。吾人認為可解釋人工智慧能提供電腦視覺進行心臟超音波影像擷像辨識時所辨識之特徵。由於特徵辨識之深度學習過程並不直接受到人為標註之影響，因此人工智慧對特徵辨識之過程可視為一種客觀的心臟超音波品質指標。吾人假設電腦視覺所提取之辨識特徵值與臨床心臟超音波擷像品質成正比。因此，若能以可解釋人工智慧提取並檢視分類器使用參數之合理性，將可以產生一全新的客觀的心臟超音波品質指標，解決現今心臟超音波在臨床使用上主觀成分居多之困境。 欲開發出上述理想之電腦視覺軟體，吾等必須先滿足兩項必要條件。其一，必須先獲得一組高品質之心超影像資料集，而”Check-up Your Heart Program”在2017台北世大運所存取的資料十分適合，因為這些來自世界各國的年輕運動員均為健康之大學學生，因體型造成擷像品質下滑的情形較少發生。其二，吾等所採用的深度學習神經網路必須能可靠地有效地提取特徵值並避免這些特徵值在運算的過程中失真，為此吾等採用了Densenet-121卷積神經網路架構，因其具備以下優點：1.強化特徵傳遞，2.特徵重覆使用，3.大幅減少網路的參數量因而提升訓練效率。吾人主要發展出三項電腦視覺成果：1.截面辨識，2.圖像分割，3.分類激活圖。其中分類激活圖為可解釋人工智慧的一種表現方式，透過分類激活圖，吾人可以了解參數調整之後的模型效果，提升模型辨識上的正確率，確認分類器使用的分類信心足以代表心臟超音波擷像品質。 為驗證電腦視覺之分類信心是否足以作為心臟超音波擷像品質指標，吾人回溯性分析一台大醫院乳癌病患族群。此族群由五個國際大規模乳癌治療隨機對照試驗之成員所組成。納入條件主要有兩點：1.追蹤期間共測量心臟超音波8次以上且其中連續2年內反覆測量達5次以上。2.該乳癌患者追蹤期間，由臨床試驗正式回覆國際PI的報告之中不具任何達臨床意義之心衰竭或心室功能。吾人假設沒有發生癌症治療相關的心功能障礙之乳癌病患，在化療期間與之後追蹤期間之各項心臟功能參數之真值為定值，因此測量值之偏差可全數歸因於擷像品質之差異。藉此吾人可以探討分類信心在擷像品質的代表性，找出其閾值並更進一步探討擷像品質對左心室整體縱向形變信度效度之影響。吾人發現心尖四腔室擷像之分類信心高於900時可視為擷像品質良好，並且，在分類信心高於900的影像中，其左心室整體縱向形變的分析時的複本信度、評分者間信度、重測信度以及測量效度上都明顯優於分類信心低於900之影像。在縱貫性研究設計中，分類信心高於900可有效降低亞臨床心功能障礙之偽陽性偵測。 吾等之研究最主要的意義在未來可以建立高精度的心臟超音波大數據。自從心臟超音波之父Harvey Feigenbaum提出影像數位化之後，心臟超音波影像便開始累積海量的資料，並包含各式各樣的擷像角度與測量值，新的測量科技與衍生參數如雨後春筍般地持續產生。因此，心臟超音波是最符合大數據定義的心臟影像模式，然而數據並不僅僅是數量龐大就好，同時也要具備正確性，吾人之影像品質評估工具便是提升心臟超音波大數據正確性的第一步。臨床專家在執行醫療決策時往往會參考記憶中的相似個案，思考病程的未來走向，並避免過去的錯誤，減少可能產生的併發症。在未來評估新個案之心超影像時，吾人從高精度資料庫中尋找數據資料「全等」 之既往過案作為新個案的數據攣生子，為新個案進行精準的醫療規劃。如此一來，吾等便能將過往心臟超音波專家的經驗，以人工智慧的形式傳承，加速醫療的發展並造福更多的病患。
並列摘要
Cancers and cardiovascular diseases are both life threatening chronic non-communicable diseases. Cardiovascular diseases are not only the most common adverse events of cancer treatments, but also the major cause of early death in cancer survivors. In 2016, the Task Force for cancer treatments and cardiovascular toxicity of the European Society of Cardiology (ESC) announced the position paper. Cancer therapeutics–related cardiac dysfunction (CTRCD) is defined as a decrease in the left ventricular ejection fraction (LVEF) of >10% points, to a value below the lower limit of normal (53%). This decrease should be confirmed by repeated cardiac imaging done 2–3 weeks after the baseline diagnostic study showing the initial decrease in LVEF. Besides, a novel echocardiographic parameter – left ventricular global systolic longitudinal strain (LVGLS) has been reported to accurately predict a subsequent decrease in LVEF. A relative percentage reduction of GLS of >15% from baseline is considered abnormal and a marker of early LV subclinical dysfunction. In the very beginning, there was remarkable inter-vendor variance in strain measurements that can lead to deviation of LVGLS value and change clinical decision-making. When LVEF did not provide suitable risk stratification and LVGLS had inadequate reliability and validity, we focused on right ventricular functional analysis. We found that right ventricular ejection fraction (RVEF) via three-dimensional echocardiography is of the strongest prognostic value in patients under VA-ECMO support. However, RVEF is a composite result that reflects left heart function; it still needs pressure-volume analysis for disease specific decision-making. We have conducted a small population, proof-of-concept study about non-invasive, echo-based pressure-volume analysis in 2017. At the same time, EACVI/ASE/Industry task force had standardized deformation imaging, and inter-vendor agreement in LVGLS significantly improved. We retrospectively investigated a population of familial amyloid polyneuropathy at national Taiwan university hospital. We found that LVGLS does have better prognostic value than LVEF; however, for images with poor quality, strain analyses are infeasible. Since the inevitably subjective nature of image quality assessment, the concept of echocardiographic core laboratory (ECL) was proposed. Although there is robust reproducibility within ECL, there are still inter-laboratory differences in LVGLS measurements. On the other hand, the revolutionary change of artificial intelligence (AI) has begun. In 2006, Geoffrey Everest Hinton and colleagues published the article “Reducing the dimensionality of data with neural network” in Science, which provides evidence of possibility of unsupervised learning. In 2012, Alex Krizhevsky won the first price of ILSVRC with the first attempt of deep learning supported by GPU. This opened a new era of deep learning based object detection. In 2018, Jeffrey Zhang proposed the computer vision based, fully automated, scalable analysis pipeline for echocardiogram interpretation, including (1) view identification, (2) image segmentation, (3) quantification of structure and function, and (4) disease detection. However, the source quality control of the echocardiography was still not emphasized at that time. Explainable artificial intelligence (XAI) helps the investigators better trust and understand how does the model come to the decision. On the other hand, automated software can reduce the inter-observer variability during strain analysis. As a result, the major source of error during strain analysis should remain to the image quality. Through XAI, we realized that feature detection is the basis of view classification. Since the process during feature detection was not subject to experts or sonographer, the detected feature should be an objective quality marker of echocardiography. We firstly collected a high quality echocardiographic dataset from 2017 summer universiade “Check-up your heart program” to saturate the deep learning process of feature detection. Secondly, we adopted the Densenet-121 convolutional neural network for its strengths in feature propagation, feature reuse, and substantially reducing the number of parameters. We completed the following computer vision tasks: view classification, chamber segmentation and activation mapping. And we used the classification confidence (CC) to stand for objective image quality. We then validated our model in a breast cancer population recruited from five international randomized control trials about trastuzumab based targeted therapy. We hypothesized that for patients without CTRCD, the value of LVGLS in serial echocardiograms should be a constant. As a result, the fluctuation of measured value of LVGLS should be contributed to image quality. Through the inter-modality consistency with the magnetic resonance imaging feature tracking, we determined the CC of apical four chamber view (A4C) above 900 as adequate image quality for improving the precision of echocardiographic strain measurements. The strain analysis showed higher parallel forms, inter-rater, and test-retest reliabilities in patients with CC of A4C > 900. During sequential comparisons of automated LVGLS in individual patients, those with CC of A4C > 900 had a lower false-positive detection rate of CTRCD. The potential of our current study is to build echocardiographic big data with high precision through objective quality control. Since Harvey Feigenbaum proposed digitalization of echocardiography, there is an exponential growth of echocardiographic data including emerging parameters and focused acquisition views. However, due to the susceptibility to inter-observer variability, big data application of echocardiographic datasets is lacking, while our quality control tool might be a potential solution. In the process of clinical decision-making, clinicians usually consider patients’ disease progression and prognosis basing on previous cases with similar presentation. If we can build a precise echocardiographic big data, AI can help identify the “digital twin” of an individual patient. As a result, AI can pass on and perpetuate the legacy of previous cardiologists and thus accelerate the improvement of cardiovascular care for more patients.