郭秀鏐 丁鶯 徐秋萍
[摘要] 人工智能是計算機科學的一個分支,是一門新的技術科學。以強大的計算和學習能力而廣泛應用于臨床實踐的各個領域。本文回顧了人工智能在胰腺疾病診斷及治療中的應用,特別是在急性胰腺炎的嚴重程度及預后評估、胰腺癌的診斷和預后等方面。然而人工智能是以“大數據”為基礎的,多中心數據庫的建立仍需要我們進一步努力。此外,隨著人工胰腺在糖尿病應用中的普及,人機關系在醫療實踐中占的比重也會越來越大。人工智能技術將會給臨床診療活動帶來更多的便利。
[關鍵詞] 人工智能;胰腺疾病;人工神經網絡;胰腺炎;胰腺癌;機器學習
[中圖分類號] R57;R-05? ? ? ? ? [文獻標識碼] A? ? ? ? ? [文章編號] 1673-9701(2020)17-0188-05
Application and prospect of artificial intelligence in diagnosis and treatment of pancreatic diseases
GUO Xiuliu? ? DING Ying? ? XU Qiuping
Zhejiang University School of Medicine,Hangzhou? ?310020,China
[Abstract] Artificial intelligence is a branch of computer science and a new technical science. It is widely used in various fields of clinical practice with strong computing and learning capabilities. This article reviews the application of artificial intelligence in the diagnosis and treatment of pancreatic diseases, especially in the assessment of the severity and prognosis of acute pancreatitis, and the diagnosis and prognosis of pancreatic cancer. However, artificial intelligence is based on "big data", and the establishment of a multi-center database still requires our further efforts. In addition, with the popularization of artificial pancreas in the application of diabetes, human-machine relationship will also become more and more important in medical practice. Artificial intelligence technology will bring more convenience to clinical diagnosis and treatment activities.
[Key words] Artificial intelligence;Pancreatic diseases;Artificial neural network;Pancreatitis; Pancreatic cancer;Machine learning
人工智能(Artificial intelligence,AI)是一門新的技術科學,主要研究開發能夠模擬、延伸和擴展人類智能的理論、方法、技術和應用系統。21世紀人工智能得到了飛速發展,在醫療、軍事、化學工業、地質勘探等各個領域都取得了驚人的成果。早在20世紀50年代后期,人工智能就在醫療領域有了研究,在醫學診斷中有了初步的探索[1]。70余年來,在經歷了曲折的發展之后,目前人工智能在我國臨床診斷、治療,病原學檢測,疾病預后預測及醫療影像等方面應用廣泛,為我國的醫療事業做出巨大貢獻[2]。其中在胰腺疾病的診斷及治療方面,已經有多種人工智能技術在應用,如胰腺炎的診斷及預測、胰腺惡性腫瘤的診斷及鑒別診斷、人工胰島的應用等。本文就人工智能在胰腺疾病診斷及治療中的應用及展望作出綜述。現報道如下。
1 人工智能在醫學研究中的主要方法
1956年,在由一些心理學、神經生理學、計算機學等學科參加的達特茅斯會議中,“人工智能”的概念首次被提出,并希望可以用計算機來構造擁有與人類智慧同樣本質特性的機器。人工智能研究領域范圍很廣,包括專家系統、機器學習、進化計算、模糊邏輯、計算機視覺、自然語言處理、推薦系統等。其中機器學習與醫學研究關系最為密切。機器學習(Machine learning,ML)其實是一種實現人工智能的方法[3]。簡而言之就是使用算法來解析已有的臨床數據,從中學習,然后對臨床事件做出決策和預測。與傳統的為解決特定任務、硬編碼的軟件程序不同,機器學習是用大量的數據來“訓練”,通過各種算法從數據中學習如何完成任務。傳統的機器學習算法包括決策樹、聚類、貝葉斯分類、支持向量機、EM、Adaboost等。從學習方法上來分,機器學習算法可以分為監督學習(如分類問題)、無監督學習(如聚類問題)、半監督學習、集成學習、深度學習和強化學習等。目前,深度學習(Deep learning,DL)[4]方法在醫學研究中應用最為廣泛。由于醫療數據具有龐大、復雜、無序的特殊性,傳統的機器學習方法并不能勝任這樣繁雜的任務。而深度學習采用了深度神經網絡(DNN)、卷積神經網絡(CNN)等方法,與傳統計算機回歸分析的單層結構不同,神經網絡是復雜的多層感知模型,包括了輸入層、模擬神經元層、輸出層三個部分。其在數據處理能力上可以分析傳統的回歸分析所無法處理的非線性數據。只要選擇合適的輸入層與輸出層,通過網絡模型對大量臨床數據的學習和調試,就能找到一個輸入層與輸出層的函數關系,一個無限靠近現實真相的關聯關系。使用訓練成功的網絡模型,對臨床工作具有巨大的推動作用。
2 人工智能在胰腺疾病診斷及治療中的應用
2.1人工智能與胰腺炎
急性胰腺炎是一種常見的急腹癥,其發病率與死亡率均較高[5-6]。自從人工智能發展以來,其在急性胰腺炎預測方面就有了不少研究與探索。上世紀九十年代,Pofahl WE等[7]對神經網絡在預測急性胰腺炎患者住院時間(Length of stay,LOS)中的作用展開了研究。他們建立了一種反向傳播神經網絡,并對195例急性胰腺炎患者的病例資料進行回顧,其中156例用于對神經網絡模型的訓練,在剩余39例中進行測試。結果表明,相比于其他方法,神經網絡模型在預測LOS>7 d中具有最高的靈敏度(75%)。盡管該研究并未涉及急性胰腺炎發病早期的預測,但也證實了人工智能在急性胰腺炎領域擁有廣闊的研究前景。之后Keogan MT[8]團隊也利用人工智能對急性胰腺炎患者的預后進行預測。他們建立的人工神經網絡模型(ANN)利用CT和實驗室數據對92例急性胰腺炎患者的預后進行預測。輸入節點為CT、實驗室數據,輸出節點為患者住院時間。最后ANN成功地預測了患者有無超過平均住院時間(Az=0.83±0.05)。相比于Ranson分級(Az=0.68±0.06,P<0.02)和Balthazar分級(Az=0.62±0.06,P<0.003),他們建立的人工神經網絡模型有著明顯優勢。但與線性判別分析(Az=0.82±0.05,P=0.53)相比,其結果不具有差異。此外,對于急性胰腺炎嚴重程度的預測,有研究者建立了一個神經網絡預后模型[9]。該模型經增強CT掃描證實其敏感性為100%,入院時特異性為70%。Pearce CB等[10]利用機器學習來提高APACHEⅡ評分和C反應蛋白的入院值對急性胰腺炎嚴重程度的預測作用。選取了256例患者作為研究對象,采用年齡、CRP、呼吸頻率、空氣中PO2、動脈pH、血肌酐、白細胞計數和GCS評分這8個項目作為輸入節點,其受試者-操作特征曲線(AUC)下的面積為0.82(SD 0.01),預測嚴重程度的最佳臨界值為0.87,特異度為0.71。預測結果明顯優于入院APACHE Ⅱ評分(AUC 0.74)和歷史入院APACHE Ⅱ數據(AUC 0.68~0.75)(P=0.0036)。表明機器學習技術顯著改善了入院后首次觀察值的預測性能,并減少了預測因素的數量。另一項研究[11]將神經網絡預測急性胰腺炎嚴重程度的準確性與APACHE Ⅱ和GCS評分進行比較,結果顯示ANN在預測嚴重病程進展(P<0.05和P< 0.01)、預測多器官功能障礙綜合征的發展(P<0.05和P<0.01)以及預測AP死亡(P<0.05)方面優于APACHE Ⅱ或GS評分系統,其靈敏度和特異度分別達到89%、96%。急性胰腺炎是一種很復雜的疾病,根據之前的研究可以得出,想要利用人工智能預測急性胰腺炎的嚴重程度,危險因素的選擇十分關鍵[12]。Andersson B等[13]設計的人工神經網絡模型,首次將疼痛持續時間作為危險變量提出。然而,Hong WD等[14]指出該研究的幾個局限性:樣本量小,缺少數據點,急性胰腺炎發病和數據收集之間的時間間隔不清楚,所以該研究結果有待進一步闡明。急性胰腺炎癥狀出現后第一周內持續的器官衰竭一個致命結局的標志,Hong WD等[14]認為,這可以作為使用人工神經網絡分析急性胰腺炎患者持續性器官衰竭的預測因素之一。同時,他們也提到了該研究的一些局限性,如數據是回顧性的,樣本量較小,可能會造成結果的一些偏差。另有一篇綜述[15]表示,與當前的評分系統相比,神經網絡預測疾病嚴重程度的準確性更高,需要的變量更少,并且能更早地作出評估。但是van den Heever M等[15]也發現,現有的大部分研究,其數據來源的數據庫大多是為管理目的而設計的,對臨床或研究人員價值有限,希望未來能建立智能數據庫,促進多中心數據收集。
急性胰腺炎本身病程十分復雜,在疾病發展過程中會出現各種各樣的并發癥[16]。Fei Y等[17]的一項研究利用人工神經網絡模型來預測門脾靜脈血栓形成的能力,并與傳統Logistic回歸進行比較。結果顯示所建立的人工神經網絡模型靈敏度為80%,特異度為85.7%,陽性預測值為77.6%,陰性預測值為90.7%。準確率為83.3%。綜合性能優于Logistic回歸模型。如果能加入更多的臨床因素或生物標志,該模型的預測能力也許會進一步提高。于是Fei Y等[18]改進了研究方法,采用徑向基函數(RBF)人工神經網絡(ANN)模型預測AP誘發PVT的風險,結果得出RBF神經網絡模型預測PVT的敏感性、特異性和準確性分別為76.2%、92.0%和88.1%。該研究證明RBF神經網絡模型是預測AP后PVT風險的有效工具,并且提出AMY、D-二聚體、PT和HCT是AP誘發PVT的重要預測因子。以同樣的方法,Fei Y等[19]人對重癥急性胰腺炎(SAP)并發急性肺損傷(ALI)的危險性也做了相關探索,并得到陽性結果。最近還有一項研究[20]表明基于CECT的放射組學模型在預測AP復發方面效果良好。這可能為一些復發患者就預防措施方面提供重要幫助。
慢性胰腺炎是各種病因引起胰腺組織和功能不可逆改變的慢性炎癥性疾病,終末期有嚴重的并發癥,包括內外分泌功能不全和胰管腺癌。慢性胰腺炎是胰管腺癌的危險因素之一[21],人工智能在慢性胰腺炎領域尚未做深入研究,現有研究主要利用人工智能相關算法鑒別診斷胰腺癌與慢性胰腺炎[22-25]。目前主要采用的方法是利用實時內鏡超聲(EUS)彈性成像提供關于胰腺病變特征的附加信息,再通過人工神經網絡分析,最后使用計算機輔助診斷來評估實時EUS彈性成像在胰腺局灶性病變中的準確性。其中一項研究[26]中神經網絡計算方法的敏感性為87.59%,特異性為82.94%,陽性預測值為96.25%,陰性預測值為57.22%,說明使用人工智能方法可以提供快速準確的診斷。自身免疫性胰腺炎(AIP)是慢性胰腺炎的一個獨特亞型,其臨床表現與胰腺導管腺癌(PDA)有許多相似之處。Zhang Y等[27]利用多種特征提取算法對CT和PET圖像進行紋理特征提取,結果顯示病灶紋理分析有助于準確區分AIP和PDA。
[5] 急性胰腺炎協作組.中國6223例急性胰腺炎病因及病死率分析[J]. 胰腺病學2006,6(6):321-325.
[6] Koutroumpakis E,Slivka A,Furlan A,et al. Management and outcomes of acute pancreatitis patients over the last decade: A US tertiary-center experience[J]. Pancreatology:official journal of the International Association of Pancreatology (IAP),2017,17(1): 32-40.
[7] Pofahl WE,Walczak SM,Rhone E,et al. Use of an artificial neural network to predict length of stay in acute pancreatitis[J]. Am Surg,1998,64(9):868-872.
[8] Keogan MT,Lo JY,Freed KS,et al. Outcome analysis of patients with acute pancreatitis by using an artificial neural network[J]. Acad Radiol,2002,9(4):410-419.
[9] Nazarenko GI,Sidorenko VI,Lebedev DS. Prognosis of the severity of acute pancreatitis by the neural network method[J]. Vestn Khir Im I I Grek,2005,164(1):50-54.
[10] Pearce CB,Gunn SR,Ahmed A,et al. Machine learning can improve prediction of severity in acute pancreatitis using admission values of APACHE Ⅱ score and C-reactive protein[J]. Pancreatology,2006,6(1-2):123-131.
[11] Mofidi R,Duff MD,Madhavan KK,et al. Identification of severe acute pancreatitis using an artificial neural network[J]. Surgery,2007,141(1):59-66.
[12] Bartosch-Harlid A,Andersson B,Aho U,et al. Artificial neural networks in pancreatic disease[J]. Br J Surg,2008, 95(7):817-826.
[13] Andersson B,Andersson R,Ohlsson M,et al. Prediction of severe acute pancreatitis at admission to hospital using artificial neural networks[J]. Pancreatology,2011,11(3):328-335.
[14] Hong WD,Chen XR,Jin SQ,et al. Use of an artificial neural network to predict persistent organ failure in patients with acute pancreatitis[J]. Clinics(Sao Paulo),2013, 68(1):27-31.
[15] Yan den Heever M,Mittal A,Haydock M,et al. The use of intelligent database systems in acute pancreatitis--a systematic review[J]. Pancreatology,2014,14(1):9-16.
[16] Hines OJ,Pandol SJ. Management of severe acute pancreatitis[J]. BMJ (Clinical research ed),2019,367:6227.
[17] Fei Y,Hu J,Li WQ,et al. Artificial neural networks predict the incidence of portosplenomesenteric venous thrombosis in patients with acute pancreatitis[J]. J Thromb Haemost,2017,15(3):439-445.
[18] Fei Y,Hu J,Gao K,et al. Risk prediction for portal vein thrombosis in acute pancreatitis using radial basis function[J]. Ann Vasc Surg,2018,47:78-84.
[19] Fei Y,Gao K,Li WQ. Artificial neural network algorithm model as powerful tool to predict acute lung injury following to severe acute pancreatitis[J]. Pancreatology,2018, 18(8): 892-899.
[20] Chen Y,Chen TW,Wu CQ,et al. Radiomics model of contrast-enhanced computed tomography for predicting the recurrence of acute pancreatitis[J]. Eur Radiol,2019, 29(8): 4408-4417.
[21] Huang CT,Lin CK,Lee TH,et al. Pancreatic fibrosis and chronic pancreatitis: Mini-review of non-histologic diagnosis for clinical applications[J]. Diagnostics (Basel,Switzerland),2020,10(2):87.
[22] Yeaton P,Sears RJ,Ledent T,et al. Discrimination between chronic pancreatitis and pancreatic adenocarcinoma using artificial intelligence-related algorithms based on image cytometry-generated variables[J]. Cytometry,1998,32(4):309-316.
[23] Norton ID,Zheng Y,Wiersema MS,et al. Neural network analysis of EUS images to differentiate between pancreatic malignancy and pancreatitis[J]. Gastrointest Endos,2001,54(5): 625-629.
[24] Saftoiu A,Vilmann P,Gorunescu F,et al. Neural network analysis of dynamic sequences of EUS elastography used for the differential diagnosis of chronic pancreatitis and pancreatic cancer[J]. Gastrointest Endosc,2008,68(6):1086-1094.
[25] Zhu M,Xu C,Yu J,et al. Differentiation of pancreatic cancer and chronic pancreatitis using computer-aided diagnosis of endoscopic ultrasound (EUS) images:A diagnostic test[J]. PLoS One,2013,8(5):e63820.
[26] Saftoiu A,Vilmann P,Gorunescu F,et al. Efficacy of an artificial neural network-based approach to endoscopic ultrasound elastography in diagnosis of focal pancreatic masses[J]. Clin Gastroenterol Hepatol,2012,10(1):84-90.
[27] Zhang Y,Cheng C,Liu Z,et al. Differentiation of autoimmune pancreatitis and pancreatic ductal adenocarcinoma based on multi-modality texture features in (18)F-FDG PET/CT[J]. Sheng Wu Yi Xue Gong Cheng Xue Za Zhi,2019,36(5):755-762.
[28] Exarchakou A,Papacleovoulou G,Rous B,et al. Pancreatic cancer incidence and survival and the role of specialist centres in resection rates in England,2000 to 2014:A population-based study[J]. Pancreatology,https://doi.org/10.1016/j.pan.2020.01.012.
[29] O'Reilly D,Fou L,Hasler E,et al. Diagnosis and management of pancreatic cancer in adults: A summary of guidelines from the UK National Institute for Health and Care Excellence[J]. Pancreatology:official journal of the International Association of Pancreatology(IAP)[et al],2018,18(8): 962-970.
[30] Okon K,Tomaszewska R,Nowak K,et al. Application of neural networks to the classification of pancreatic intraductal proliferative lesions[J]. Anal Cell Pathol,2001, 23 (3-4): 129-136.
[31] Xing F,Su H,Neltner J,et al. Automatic Ki-67 counting using robust cell detection and online dictionary learning[J].IEEE Trans Biomed Eng,2014,61(3):859-870.
[32] Momeni-Boroujeni A,Yousefi E,Somma J. Computer-assisted cytologic diagnosis in pancreatic FNA:An application of neural networks to image analysis[J]. Cancer Cytopathol,2017,125(12):926-933.
[33] Kuwahara T,Hara K,Mizuno N,et al. Usefulness of deep learning analysis for the diagnosis of malignancy in intraductal papillary mucinous neoplasms of the pancreas[J].Clin Transl Gastroenterol,2019,10(5):1-8.
[34] Li S,Jiang H,Wang Z,et al. An effective computer aided diagnosis model for pancreas cancer on PET/CT images[J].Comput Methods Programs Biomed,2018,165:205-214.
[35] Chung WY,Correa E,Yoshimura K,et al. Using probe electrospray ionization mass spectrometry and machine learning for detecting pancreatic cancer with high performance[J]. American Journal of Translational Research,2020,12(1):171-179.
[36] Turksoy K,Cinar A. Adaptive control of artificial pancreas systems-a review[J]. J Healthc Eng,2014,5(1):1-22.
[37] Boughton CK,Hovorka R. Advances in artificial pancreas systems[J]. Science Translational Medicine,2019,11(484):4949.
[38] Thamprajamchit S,Ongphiphadhanakul B,Krittiyawong S,et al. A simple prediction rule and a neural network model to predict pancreatic beta-cell reserve in young adults with diabetes mellitus[J]. J Med Assoc Thai,2001, 84(3):332-338.
[39] El-Jabali AK. Neural network modeling and control of type 1 diabetes mellitus[J]. Bioprocess Biosyst Eng,2005, 27(2):75-79.
[40] Zecchin C,Facchinetti A,Sparacino G,et al. Jump neural network for real-time prediction of glucose concentration[J]. Methods Mol Biol,2015,1260:245-259.
[41] Samadi S,Turksoy K,Hajizadeh I,et al. Meal detection and carbohydrate estimation using continuous glucose sensor data[J]. IEEE J Biomed Health Inform,2017,21(3):619-627.
[42] Asadi S,Nekoukar V. Adaptive fuzzy integral sliding mode control of blood glucose level in patients with type 1 diabetes:In silico studies[J]. Math Biosci,2018,305:122-132.
(收稿日期:2020-03-03)