A Deep Learning Framework for Interpretable Diagnoses of Pulmonary Diseases through Odds-Ratio Analysis in Radiographic Findings
- Nov. 2020
- by Jaemin Son et. al.
To propose and evaluate a framework that quantifies the relationships between radiographic findings and pulmonary diseases using a deep neural network (DNN)
METHOD AND MATERIALS
A DNN model with ResNet-50 architecture was first trained to identify nodules, consolidation, interstitial opacity, pleural effusion, and pneumothorax in 7,104 normal and 7,821 abnormal chest X-ray images (CXRs). Layers following the last feature extraction layer were then replaced with randomly initialized logistic regression layers. These logistic regression layers were then trained without modifying other layers’ weights on public datasets to identify the onset of pulmonary diseases: tuberculosis (Shenzen and Montgomery), edema (CheXpert), and pneumonia (Shenzen, Montogomery, and RSNA pneumonia challenge). Shenzen, Montgomery, and RSNA datasets were split in half to training and validation sets. Random augmentation (shift, rotation, and resize) was employed throughout the training phase. To quantify the relationships between radiographic findings and pulmonary diseases, odds-ratios were computed for every finding-disease pair using the cosine distance between the logistic models’ weights. Heatmaps were displayed to analyze findings that attribute the most in diagnosing individual diseases.
The logistic regression models performed reliably in diagnosing pulmonary diseases on both cross-validation (91.9% and 85.0% AUROC for Shenzhen to Montgomery and Montgomery to Shenzhen) and reserved validation datasets (95.9%, 92.2%, 93.3%, 88.4% AUROC for CheXpert, Shenzhen, Montgomery, and RSNA). More importantly, the odds ratio demonstrated concordant relationships between pulmonary diseases and radiographic findings as perceived in clinical practice.
An interpretable deep learning framework that quantifies odds-ratio between radiographic findings and pulmonary diseases was developed without compromising the accuracy. The proposed framework enables quantifying the attribution of each finding in making diagnostic decisions as commonly done by radiologists.
By quantifying the odds-ratio between radiographic findings and pulmonary diseases, a DNN can assist physicians to diagnose diseases while providing interpretable information on predominant findings