볼륨-플로우 그래프 기반 폐질환 분류를 위한앙상블 딥러닝 모델 = CNN-based Ensemble Deep Learning Model for Pulmonary Diseases Classification using Flow Volume Loops

다국어 초록 (Multilingual Abstract)

Chronic Obstructive Pulmonary Disease (COPD) is a respiratory disease characterized bychronic airway obstruction. COPD often progresses to a severe stage, since thereare few noticeable symptoms in the early stages. Also regression equations involingvarious factors such as race, gender, height, and weight to determine whether ornot there is pulmonary disease is complex and needs to be updated periodically.
Therefore, there is a demand for a system that can easily analze the presence orabsence of the pulmonary disease, even for non-experts. In this paper, aCNN-based flow volume loops classification model using ensemble learning andappropriate data pre-processing algorithms was proposed and validated to diagnosepulmonary disease in the early stages. The ensemble model was organized by fourCNN models based on VGG16, VGG19, Resnet50, and MobileNet and used transferlearning and fine-tuning for each pre-trained model. Specifically, to overcome asmall amount of data, several data augmentation techniques that took into accountthe characteristics of flow volume loops were used, and soft voting was employedfor the ensemble model. The proposed ensemble model not only could diagnose thepresence or absence of pulmonary disease but could also classify into a total of fourcategories: normal, restrictive, obstructive, and combined pulmonary diseases. As aresult of the experiment, the performance of the proposed ensemble model showedan accuracy of 90.91%, precision of 91.11%, and recall of 90.91%.

국문 초록 (Abstract)

만성 폐쇄성 폐질환은 만성적인 기도 폐쇄를 특징으로 하는 호흡기 질환이다. 만성 폐쇄성폐질환은 초기에 자각 증상이 거의 없어, 대부분 중증 상태로 악화된다. 또한, 인종, 성별, 키,몸무게 등 다양한 요인을 포함한 폐 질환 분류 회귀식은 복잡하고, 정확한 판별을 위해서는지속적인 갱신을 필요로 한다. 따라서 폐질환의 초기 진단이 용이하도록 간편한 휴대형 페기능 검사기를 통해 산출 가능한 볼륨-플로우 그래프 이미지 기반 분류 모델이 요구된다.
본 논문에서는 폐질환 조기 진단을 위해 볼륨-플로우 그래프 이미지의 전처리 및 합성곱 신경망 기반 앙상블 딥러닝 모델을 구현하였고, 이를 검증했다. 합성곱 신경망 기반 앙상블 딥러닝 모델은 VGG16, VGG19, ResNet50, 그리고 MobileNet 구조 기반 4개의 모델로 구성되며, 전부 전이학습 및 미세조정하여 사용하였다. 세부적으로는 부족한 수의 학습 데이터를볼륨-플로우 그래프 이미지의 특성을 고려하여 적합한 데이터 증강기법을 적용하였고, 4개의 모델들은 가중치 기반의 간접투표 방식을 사용했다. 최종 앙상블 모델은 단순히 폐질환유무를 판별하는 것이 아닌 정상, 제한성 폐질환, 폐쇄성 폐질환, 그리고 혼합성 폐질환과 같이 총 4개의 클래스로 분류하는 모델임에도 불구하고, 테스트 데이터를 통한 성능은 정확도90.91%, 가중치 평균 정밀도 91.11%, 가중치 평균 재현율 90.91%로 높은 수치를 보였다.

참고문헌 (Reference)

• 1 K. Simonyan, "Very Deep Convolutional Networks for Large-Scale Image Recognition"
• 2 Pedro Marcelino, "Transfer learning from pre-trained models"
• 3 Ibrahem Kandel ; Mauro Castelli, "The effect of batch size on the generalizability of the convolutional neural networks on a histopathology dataset" 6 (6): 312-315, 2020
• 4 H. Bass, "The Flow Volume Loop : Normal Standards and Abnormalities in Chronic Obstructive Pulmonary Disease" 63 (63): 172-176, 1973
• 5 Brian L. Graham, "Standardization of Spirometry 2019 Update. An Official American Thoracic Society and European Respiratory Society Technical Statement" 200 (200): 70-88, 2019
• 6 J. F. Morris, "Spirometric standards for healthy nonsmoking adults" 103 (103): 57-67, 1971
• 7 A. Majid, "Pulmonary function and flow-volume loop patterns in patients with tracheobronchomalacia" 58 (58): 1521-1526, 2013
• 8 김지항, "Pulmonary Emphysema : Visual Interpretation and Quantitative Analysis" 82 (82): 808-816, 2021
• 9 최정근 ; 백도명 ; 이정오, "Normal Predictive Values of Spirometry in Korean Population" 58 (58): 230-242, 2005
• 10 N. G. Csikesz, "New developments in the assessment of COPD : early diagnosis is key" 9 : 277-286, 2022
• 11 A. G. Howard, "MobileNets : Efficient Convolutional Neural Networks for Mobile Vision Applications"
• 12 E. Doria, "Exploring Current Concepts and Challenges in the Identification and Management of Early-Stage COPD" 12 (12): 2023
• 13 N. Srivastava, "Dropout : A Simple Way to Prevent Neural Networks from Overfitting" 15 (15): 1929-1958, 2014
• 14 Sang Chul Lee, "Development and validation of a nomogram to predict pulmonary function and the presence of chronic obstructive pulmonary disease in a Korean population" 21 (21): 2021
• 15 H. Kaiming, "Deep Residual Learning for Image Recognition"
• 16 A. D. Lopez, "Chronic obstructive pulmonary disease, current burden and future projections" 27 (27): 397-412, 2006
• 17 S. Kumari, "An ensemble approach for classification and prediction of diabetes mellitus using soft voting classifier" 2 : 40-46, 2021