Kirill Lipatov, MD
Clinical Fellow
Mayo Clinic College of Medicine
Rochester, Minnesota, United States
Disclosure information not submitted.
Michelle Herberts, MD
Clinical Fellow
Mayo Clinic, United States
Disclosure information not submitted.
Brian Pickering, MD, MB, BCh, BAO
Professor
Mayo Clinic - College of Medicine, United States
Disclosure information not submitted.
Vitaly Herasevich, MD, PhD, FCCM
Professor of Anesthesiology and Medicine
Mayo Foundation
Rochester, Minnesota
Disclosure information not submitted.
Title: Classification of Idiopathic Pulmonary Fibrosis Using Chest X-rays AND Deep Learning Approach
Introduction: Idiopathic pulmonary fibrosis (IPF) is the most common type of PF. Approximately 50,000 new cases of IPF are diagnosed yearly in the US. This chronic lung condition is irreversible and progressive over time with 3 to 5 years average survival time. Accurate and early diagnosis of IPF is essential but complicated and may require a multidisciplinary and multidimensional approach covering clinical, pathologic, laboratory and radiologic factors. The first aim of this project is to recognize signs of IPF using deep learning technology on chest X-rays.
Methods: A retrospective analysis of front-chest and lateral X-rays was conducted by convolutional neural network (CNN) models. The front-view model was trained with 1,196 front-chest X-rays (596 IPF and 600 Normal). Lateral model was trained with 1,189 lateral X-rays (589 IPF and 600 Normal). These models classified images into 2 different categories: images with IPF vs. images without IPF (Normal). 80% of the images were randomly split for training and 20% were used for validation. Data augmentation methods were employed to increase the models’ performance and generalization ability. Regularization and dropout helped prevent overfitting. The models included the classifier following by 6 feature extractors. Each feature extractor had a convolutional layer, an activation layer (ReLU) and a max pooling layer. The classifier included a flattened layer, two dense layers with ReLU layer in between and Softmax activation function for outputting the probability prediction scores. The outcome of study is diagnostic performance of CNN model compared to Gold Standard of clinical assessment of chest X-Rays.
Results: In validation cohort the front-view model had better classification accuracy rate (94%) compared to 90% accuracy rate of the lateral model. The front-view model had the specificity of 92%, sensitivity of 96%, PPV of 92%, NPV of 97%, and AUC of 94%. The lateral model had the specificity of 86%, sensitivity of 95%, PPV of 84%, NPV of 96%, and AUC of 90%.
Conclusions: Identifying IPF from X-rays by CNN can yield highly accurate results. For future work, a combined model which includes both lateral and front-view predictions will be investigated. Furthermore, other features such as clinical, pathologic, and lab results may also be included for modeling.