심층 신경망을 이용한 흉부X선사진에서 비결핵 항산균 폐질환의 예후 예측 = Deep learning-based prediction model using chest X-ray in nontuberculous mycobacterial pulmonary disease

국문 초록 (Abstract)

배경: 비결핵 항산균 폐질환은 만성적으로 폐 실질을 파괴하는 질환으로 임상 정보와 영상학적 소견을 같이 고려하더라도 예후를 정확히 예측하는데 제한이 있다. 딥러닝 기법을 이용한 흉부X선사진 판독은 사람이 인지하기 어려운 신호를 발견하여 임상 의사의 의사결정에 도움을 주고 있다. 본 연구는 딥러닝 기법을 사용하여 비결핵 항산균 폐질환 환자의 예후를 예측하는 모델을 개발하고 검증해보고자 한다.
방법: 2000년 1월부터 2015년 12월까지 서울대학교병원과 2006년 1월부터 2015년 12월까지 보라매병원에서 비결핵 항산균 폐질환으로 진단받은 환자를 대상으로 하였다. 서울대학교병원에서 진단 시 시행된 흉부X선사진을 사용하여 3년, 5년, 그리고 10년 사망을 예측하도록 딥러닝 모델을 훈련했다. 1) 딥러닝 모델에서 도출된 CXR 점수, 2) 임상 정보 (나이, 성별, 체질량 지수, 및 비결핵 항산균 종)를 이용한 모델을 만들고, 3) CXR 점수와 임상 정보를 통합한 모델을 만들어 사망 예측의 정확도를 비교했다. 서울대학교병원 환자의 흉부X선사진으로 만들어진 딥러닝 모델을 보라매병원에서 진단 시 시행된 흉부X선사진을 이용하여 검증하였다. 또한, 하위 분석을 통해 예측 모델의 정확도가 항생제 사용 여부, 균 음전 여부, 또는 균 종에 따라서 달라지는지를 확인하였다.
결과: 서울대학교병원에서 진단된 1,034명의 환자에서 시행된 1,638개의 흉부X선사진을 사용하여 예측 모델을 학습시키고 튜닝하였고, 보라매병원에서 진단된 200명의 환자에서 시행된 566개의 흉부X선사진을 사용하여 검증을 하였다. 딥러닝 모델에서 도출된 흉부X선사진 점수의 3년, 5년, 그리고 10년 사망에 대한 수신자 조작 특성 곡선 아래 면적은 각각 0.792, 0.781, 그리고 0.844로 확인되었다. 흉부X선사진 점수와 임상 정보를 통합한 모델의 3년, 5년, 그리고 10년 사망에 대한 수신자 조작 특성 곡선 아래 면적은 각각 0.865, 0.942, 0.922으로 개선되었다. 흉부X선사진 점수와 임상 정보를 통한 예측 모델의 정확도는 항생제 치료, 음성 배양 전환 여부, 그리고 비결핵 항산균 종에 따라 차이를 보였다.
결론: 진단 당시 시행된 흉부X선사진을 사용한 딥러닝 예측 모델은 비결핵 항산균 폐질환 환자의 중장기 사망률을 예측할 수 있었고, 임상 정보를 추가함으로써 예측의 정확도가 개선되었다.

다국어 초록 (Multilingual Abstract)

Background: Prognostic prediction of nontuberculous mycobacteria pulmonary disease using a deep learning technique has not been tried yet. We aimed to develop a model predicting the prognosis of nontuberculous mycobacteria pulmonary disease using a deep learning technique.
Methods: Patients diagnosed with nontuberculous mycobacteria pulmonary disease at Seoul National University Hospital (train/validation dataset) between January 2000 and December 2015 and at Seoul Metropolitan Government-Boramae Medical Center (test dataset) between January 2006 and December 2015 were included. We trained deep learning models to predict the 3-, 5-, and 10-year overall mortality using baseline chest radiographs at diagnosis. We tested the predictability for the corresponding mortality using only deep learning-driven radiographic scores and using both radiographic scores and clinical information (age, sex, body mass index, and mycobacterial species). The prediction model was externally validated. In addition, we figure out whether the performance of the prediction is different according to the various clinical features.
Results: The datasets comprised 1,638 (train/validation set) and 566 (test set) chest radiographs from 1,034 and 200 patients, respectively. The deep learning-driven radiographic score provided areas under the receiver operating characteristic curve of 0.844, 0.781, and 0.792 for the 10-, 5-, and 3-year mortality, respectively. The logistic regression model using both the radiographic score and clinical information provided areas under the receiver operating characteristic curves of 0.922, 0.942, and 0.865 for the 10-, 5, and 3-year mortality, respectively. The accuracy of our prediction model with radiographic score and clinical information was different according to antibiotic treatment, negative culture conversion, and the species of nontuberculous mycobacteria.
Conclusions: The deep learning model we developed could predict the mid- to-long-term mortality of patients with nontuberculous mycobacteria pulmonary disease using a baseline radiograph at diagnosis, and the predictability increased with clinical information.

목차 (Table of Contents)

• 제 1 장 서론 6
• 제 1 절 연구의 배경 6
• 제 2 절 연구의 목적 17
• 제 2 장 연 구 방 법 19
• 제 1 절 연구 설계와 대상자 19
• 제 2 절 데이터 수집 19
• 제 3 절 딥러닝 모델의 개발 20
• 제 4 절 예측 모델의 성능 평가 22
• 제 5 절 통계 분석 24
• 제 6 절 연구 윤리 24
• 재 3 장 연 구 결 과 26
• 제 1 절 기본적인 특성 26
• 제 2 절 기관 별 환자의 임상적 차이 27
• 제 3 절 생존 여부에 따른 환자의 임상적 차이 28
• 제 4 절 딥러닝 예측 모델의 성능 29
• 제 5 절 항생제 치료 여부에 따른 하위 그룹 분석 31
• 제 6 절 음성 배양 전환 여부에 따른 하위 그룹 분석 32
• 제 7 절 비결핵 항산균 종에 따른 층화 분석 33
• 제 4 장 고 찰 35
• 참고문헌 66
• 표 목차
• [표 1] 84
• [표 2] 85
• [표 3] 87
• [표 4] 89
• [표 5] 90
• [표 6] 91
• [표 7] 92
• [표 8] 93
• [표 9] 94
• 그림 목차
• [그림 1] 95
• [그림 2] 96
• [그림 3] 97
• [그림 4] 98
• [그림 5] 99
• Abstract in English 100

참고문헌 (Reference)

1. Anonymous mycobacteria in pulmonary disease, Runyon EH, 43(1):273-90, , 1959

2. Mycobacterium abscessus: a new antibiotic nightmare, Murray A, Reyrat JM, Gicquel B., Nessar R, Cambau E, 67(4):810-8, , 2012

3. Mycobacterium abscessus: challenges in diagnosis and treatment, Benwill JL, Wallace RJ Jr., 27(6):506-10, , 2014

4. Update on pulmonary disease due to non-tuberculous mycobacteria, Yew WW, Stout JE, Koh WJ, 45:123-34, , 2016

5. Cardiovascular Disease Risk Assessment: Insights from Framingham, Massaro JM, Pencina MJ, D'Agostino RB, Sr., Coady S., 8(1):11-23, , 2013

6. Deep Learning to Assess Long-term Mortality From Chest Radiographs, Lu MT, Ivanov A, Mayrhofer T, Hosny A, Aerts H, Hoffmann U, 2(7):e197416, , 2019

7. Reduced lung-cancer mortality with low-dose computed tomographic screening, Clapp JD, Berg CD, Aberle DR, Fagerstrom RM,, Black WC, Adams AM, 365(5):395-409, , 2011

8. Occurrence and clinical relevance of Mycobacterium chimaera sp. nov., Germany, Petrich A, Göbel UB, Buchholz P, Goldenberg O, Schweickert B, Richter E, 14(9):1443-6, , 2008

9. Epidemiology of Nontuberculous Mycobacterial Infection, South Korea, 2007-2016, Koh WJ, Lee H, Myung W, Jhun BW, Moon SM, 25(3):569-72, , 2019

10. Changing epidemiology of nontuberculous mycobacterial lung disease in South Korea, Jeong SH, Kim SK, Lee SK, Lee EJ, Kang YA, Chang J, 44(10):733- 8, , 2012

11. M ycobacterium abscessus pulmonary disease: individual patient data meta-analysis, Dalcolmo MP, Gayoso R, Kwak N, Daley CL, Hasegawa N, Eather G, 54(1), , 2019

12. Deep-COVID: Predicting COVID-19 from chest Xray images using deep transfer learning, Sonka M, Kafieh R, Minaee S, Jamalipour Soufi G., Yazdani S, 65:101794, , 2020

13. Epidemiology of human pulmonary infection with nontuberculous mycobacteria: a review, Prevots DR, Marras TK, 36(1):13-34, , 2015

14. Nontuberculous mycobacteria in bronchiectasis: Prevalence and patient characteristics, Screaton NJ, Condliffe A, Fowler SJ, Foweraker J, French J, Haworth CS,, 28(6):1204-10, , 2006

15. Understanding nontuberculous mycobacterial lung disease: it's been a long time coming, Griffith DE, Aksamit TR., 5:2797., , 2016

16. Deep Learning to Determine the Activity of Pulmonary Tuberculosis on Chest Radiographs, Lee JK, Kwak N, Lee YJ, Lee S, Yim JJ, Lee JY,, 301(2):435-42, , 2021

17. Nontuberculous mycobacterial disease prevalence and risk factors: a changing epidemiology, McNelly E, Hedberg K, Cassidy PM, Winthrop KL, Saulson A, 49(12):e124-9, , 2009

18. Predictors of radiographic progression for NTMpulmonary disease diagnosed by bronchoscopy, Huang HL, Liu CJ, Lee MR, Cheng MH, Lu PL, Wang JY, 161:105847, , 2020

19. Inhaled amikacin for treatment of refractory pulmonary nontuberculous mycobacterial disease, Glaser TS, Fleshner M, Bhattacharyya D, Olivier KN, Shaw PA, Brewer CC, 11(1):30-5, , 2014

20. Prognostic factors of 634 HIVnegative patients with Mycobacterium avium complex lung disease, Yanagisawa T, Miyahara Y, Hayashi M, Takayanagi N, Sugita Y., Kanauchi T, 185(5):575-83, , 2012

21. Comparison of Clinical Features, Virulence, and Relapse among Mycobacterium avium Complex Species, Qi C., Reddy S, Zembower TR, Boyle DP, 191(11):1310-7, , 2015

22. Outcome of patients with and poor prognostic factors for Mycobacterium kansasii-pulmonary disease, Liu CJ, Wang JY, Lu PL, Shu CC, Cheng MH, Huang HL, 151:19-26, , 2019

23. Population-based Incidence of Pulmonary Nontuberculous Mycobacterial Disease in Oregon 2007 to 2012, Hedberg K, Schafer S, Winthrop KL, Novosad S, Henkle E, 12(5):642-7, , 2015

24. The natural history of non-cavitary nodular bronchiectatic Mycobacterium avium complex lung disease, Kim WS, Oh YM, Lee JH, Song JW, Kwon BS, Koh Y, 150:45-50, , 2019

25. Natural history of Mycobacterium avium complex lung disease in untreated patients with stable course, Jo KW, Hwang JA, Shim TS, Kim S, 49(3), , 2017

26. Clinical characteristics and treatment outcomes of pulmonary disease caused by Mycobacterium chimaera, Jhun BW, Park HY, Moon SM, Kim SY, Lee H, Jeon K, 86(4):382-4, , 2016

27. Prognostic nutritional index as a predictor of mortality in nontuberculous mycobacterial lung disease, Murata K, Takei K, Koga Y, Tsuchiya T, Hachisu Y, Tsurumaki H, 12(6):3101-9, , 2020

28. Long-Term Outcomes in a Population-based Cohort with Respiratory Nontuberculous Mycobacteria Isolation, Hedberg K, Novosad SA, Ku J, Siegel SAR, Shafer S, Henkle E, 14(7):1120-8, , 2017

29. An official ATS/IDSA statement: diagnosis, treatment, and prevention of nontuberculous mycobacterial diseases, Griffith DE, Aksamit T, Catanzaro A, Gordin F, Daley C, Brown-Elliott BA, 175(4):367-416, , 2007

30. Basics of Deep Learning: A Radiologist's Guide to Understanding Published Radiology Articles on Deep Learning, Do S, Song KD, Chung JW, 21(1):33- 41, , 2020

31. Increasing trend of isolation of non-tuberculous mycobacteria in a tertiary university hospital in South Korea, Yoo JW, Kim DS,, Lee SD, Kim WS, Kim MN, Jo KW, 72(5):409-15, , 2012

32. Long-term natural history of non-cavitary nodular bronchiectatic nontuberculous mycobacterial pulmonary disease, Ko RE,, Kim S, Baek SY, Moon SM, Jhun BW, Jeon K, 151:1-7, , 2019

33. Impact of prognostic nutritional index on outcomes in patients with Mycobacterium avium complex pulmonary disease, Lee EH, Leem AY, Moon SW, Lee SH, Choi JS, Lee SH, 15(5):e0232714, , 2020

34. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach, DeLong ER, DeLong DM, Clarke-Pearson DL., 44(3):837-45., , 1988

35. Incidence and Prevalence of Nontuberculous Mycobacterial Lung Disease in a Large U. S. Managed Care Health Plan, 2008-2015, Zhang Q, Marras TK, Winthrop KL, Wang P, Zhang H, Adjemian J, 17(2):178-85, , 2020

36. Screening by chest radiograph and lung cancer mortality: the Prostate, Lung, Colorectal, and Ovarian (PLCO) randomized trial, Hocking WG, Oken MM, Andriole GL, Kvale PA, Buys SS, Church TR,, 306(17):1865-73, , 2011

37. Deep learning at chest radiography: automated classification of pulmonary tuberculosis by using convolutional neural networks, Sundaram B, Lakhani P, 284(2):574-82, , 2017

38. Increased risk of nontuberculous mycobacterial infection in asthmatic patients using long-term inhaled corticosteroid therapy, Hojo M, Sugiyama H, Kobayashi N, Hirano S, Kudo K., Iikura M, 17(1):185-90, , 2012

39. Relative risk of all-cause mortality in patients with nontuberculous mycobacterial lung disease in a US managed care population, Marras TK, Vinnard C Zhang Q, Hamilton K, Adjemian J, Eagle G,, 145:80- 8, , 2018

40. Risk factors for mortality in patients with pulmonary infections with non-tuberculous mycobacteria: a retrospective cohort study, van Dessel H, Linssen CF, van Haren E, Even P, Gommans EP, de Vries GJ,, 109(1):137-45, , 2015

41. Artificial intelligence solution for chest radiographs in respiratory outpatient clinics: Multicenter Prospective Randomized Study, Oh S, Jeong IB, Kang S-Y, Jin KN, Lee SM, Lee HW, In Press, , 2022

42. Patient-Centered Research Priorities for Pulmonary Nontuberculous Mycobacteria (NTM) Infection. An NTM Research Consortium Workshop Report, Aksamit T, Leitman P, Griffith D, Henkle E, Daley CL, Barker A, 13(9):S379-84, , 2016

43. Prognostic factors associated with long-term mortality in 1445 patients with nontuberculous mycobacterial pulmonary disease: a 15-year follow-up study, Jeon K, Yoo H, Jhun BW, Kwon OJ, Moon SM, Carriere KC, 55(1), , 2020

44. Training and Validating a Deep Convolutional Neural Network for Computer-Aided Detection and Classification of Abnormalities on Frontal Chest Radiographs, Cicero M, Bilbily A, Colak E, Dowdell T, Gray B, Perampaladas K,, 52(5):281-7, , 2017

45. Identification of potentially undiagnosed patients with nontuberculous mycobacterial lung disease using machine learning applied to primary care data in the UK, van der Laan R, Doyle OM, McMahon P, Pitcher A, Obradovic M, Daniels F, 56(4), , 2020