Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologists

Abstract

Background: Chest radiograph interpretation is critical for the detection of thoracic diseases, including tuberculosis and lung cancer, which affect millions of people worldwide each year. This time-consuming task typically requires expert radiologists to read the images, leading to fatigue-based diagnostic error and lack of diagnostic expertise in areas of the world where radiologists are not available. Recently, deep learning approaches have been able to achieve expert-level performance in medical image interpretation tasks, powered by large network architectures and fueled by the emergence of large labeled datasets. The purpose of this study is to investigate the performance of a deep learning algorithm on the detection of pathologies in chest radiographs compared with practicing radiologists.

Methods and findings: We developed CheXNeXt, a convolutional neural network to concurrently detect the presence of 14 different pathologies, including pneumonia, pleural effusion, pulmonary masses, and nodules in frontal-view chest radiographs. CheXNeXt was trained and internally validated on the ChestX-ray8 dataset, with a held-out validation set consisting of 420 images, sampled to contain at least 50 cases of each of the original pathology labels. On this validation set, the majority vote of a panel of 3 board-certified cardiothoracic specialist radiologists served as reference standard. We compared CheXNeXt's discriminative performance on the validation set to the performance of 9 radiologists using the area under the receiver operating characteristic curve (AUC). The radiologists included 6 board-certified radiologists (average experience 12 years, range 4-28 years) and 3 senior radiology residents, from 3 academic institutions. We found that CheXNeXt achieved radiologist-level performance on 11 pathologies and did not achieve radiologist-level performance on 3 pathologies. The radiologists achieved statistically significantly higher AUC performance on cardiomegaly, emphysema, and hiatal hernia, with AUCs of 0.888 (95% confidence interval [CI] 0.863-0.910), 0.911 (95% CI 0.866-0.947), and 0.985 (95% CI 0.974-0.991), respectively, whereas CheXNeXt's AUCs were 0.831 (95% CI 0.790-0.870), 0.704 (95% CI 0.567-0.833), and 0.851 (95% CI 0.785-0.909), respectively. CheXNeXt performed better than radiologists in detecting atelectasis, with an AUC of 0.862 (95% CI 0.825-0.895), statistically significantly higher than radiologists' AUC of 0.808 (95% CI 0.777-0.838); there were no statistically significant differences in AUCs for the other 10 pathologies. The average time to interpret the 420 images in the validation set was substantially longer for the radiologists (240 minutes) than for CheXNeXt (1.5 minutes). The main limitations of our study are that neither CheXNeXt nor the radiologists were permitted to use patient history or review prior examinations and that evaluation was limited to a dataset from a single institution.

Conclusions: In this study, we developed and validated a deep learning algorithm that classified clinically important abnormalities in chest radiographs at a performance level comparable to practicing radiologists. Once tested prospectively in clinical settings, the algorithm could have the potential to expand patient access to chest radiograph diagnostics.

Fig 1. ROC curves of radiologists and algorithm for each pathology on the validation set.

Each plot illustrates the ROC curve of the deep learning algorithm (purple) and practicing radiologists (green) on the validation set, on which the majority vote of 3 cardiothoracic subspecialty radiologists served as ground truth. Individual radiologist (specificity, sensitivity) points are also plotted, where the unfilled triangles represent radiology resident performances and the filled triangles represent BC radiologist performances. The ROC curve of the algorithm is generated by varying the discrimination threshold (used to convert the output probabilities to binary predictions). The radiologist ROC curve is estimated by fitting an increasing concave curve to the radiologist operating points (see S1 Appendix). BC, board-certified; ROC, receiver operating characteristic.

Fig 2. Performance measures of the algorithm and radiologists on the validation set for mass, nodule, consolidation, and effusion.

Each plot shows the diagnostic measures of the algorithm (purple diamond), micro-average resident radiologist (unfilled orange diamond), micro-average BC radiologist (filled orange diamond), individual resident radiologists (unfilled green diamond), individual BC radiologists (filled green diamond). Each diamond has a vertical bar denoting the 95% CI of each estimate, computed using 10,000 bootstrap replicates. The ground truth values used to compute each metric were the majority vote of 3 cardiothoracic specialty radiologists on each image in the validation set. Kappa refers to Cohen's Kappa, and F1 denotes the F1 score. BC, board-certified; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value.

Fig 3. Interpreting network predictions using CAMs.

In the normal chest radiograph images (left), the pink arrows and circles highlight the locations of the abnormalities; these indicators were not present when the image was input to the algorithm. (a) Frontal chest radiograph (left) demonstrates 2 upper-lobe pulmonary masses in a patient with both right- and left-sided central venous catheter. The algorithm correctly classified and localized both masses as indicated by the heat maps. (b) Frontal chest radiograph demonstrates airspace opacity in the right lower lobe consistent with pneumonia. The algorithm correctly classified and localized the abnormality. More examples can be found in S2 Fig. CAM, class activation mapping.