Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study

Abstract

Background: There is interest in using convolutional neural networks (CNNs) to analyze medical imaging to provide computer-aided diagnosis (CAD). Recent work has suggested that image classification CNNs may not generalize to new data as well as previously believed. We assessed how well CNNs generalized across three hospital systems for a simulated pneumonia screening task.

Methods and findings: A cross-sectional design with multiple model training cohorts was used to evaluate model generalizability to external sites using split-sample validation. A total of 158,323 chest radiographs were drawn from three institutions: National Institutes of Health Clinical Center (NIH; 112,120 from 30,805 patients), Mount Sinai Hospital (MSH; 42,396 from 12,904 patients), and Indiana University Network for Patient Care (IU; 3,807 from 3,683 patients). These patient populations had an age mean (SD) of 46.9 years (16.6), 63.2 years (16.5), and 49.6 years (17) with a female percentage of 43.5%, 44.8%, and 57.3%, respectively. We assessed individual models using the area under the receiver operating characteristic curve (AUC) for radiographic findings consistent with pneumonia and compared performance on different test sets with DeLong's test. The prevalence of pneumonia was high enough at MSH (34.2%) relative to NIH and IU (1.2% and 1.0%) that merely sorting by hospital system achieved an AUC of 0.861 (95% CI 0.855-0.866) on the joint MSH-NIH dataset. Models trained on data from either NIH or MSH had equivalent performance on IU (P values 0.580 and 0.273, respectively) and inferior performance on data from each other relative to an internal test set (i.e., new data from within the hospital system used for training data; P values both <0.001). The highest internal performance was achieved by combining training and test data from MSH and NIH (AUC 0.931, 95% CI 0.927-0.936), but this model demonstrated significantly lower external performance at IU (AUC 0.815, 95% CI 0.745-0.885, P = 0.001). To test the effect of pooling data from sites with disparate pneumonia prevalence, we used stratified subsampling to generate MSH-NIH cohorts that only differed in disease prevalence between training data sites. When both training data sites had the same pneumonia prevalence, the model performed consistently on external IU data (P = 0.88). When a 10-fold difference in pneumonia rate was introduced between sites, internal test performance improved compared to the balanced model (10× MSH risk P < 0.001; 10× NIH P = 0.002), but this outperformance failed to generalize to IU (MSH 10× P < 0.001; NIH 10× P = 0.027). CNNs were able to directly detect hospital system of a radiograph for 99.95% NIH (22,050/22,062) and 99.98% MSH (8,386/8,388) radiographs. The primary limitation of our approach and the available public data is that we cannot fully assess what other factors might be contributing to hospital system-specific biases.

Conclusion: Pneumonia-screening CNNs achieved better internal than external performance in 3 out of 5 natural comparisons. When models were trained on pooled data from sites with different pneumonia prevalence, they performed better on new pooled data from these sites but not on external data. CNNs robustly identified hospital system and department within a hospital, which can have large differences in disease burden and may confound predictions.

Conflict of interest statement

I have read the journal's policy and the authors of this manuscript have the following competing interests: MAB and ML are currently employees at Verily Life Sciences, which played no role in the research and has no commercial interest in it. EKO and ABC receive funding from Intel for unrelated work.

Figures

Fig 1. Pneumonia models evaluated on internal and external test sets.

A model trained using both MSH and NIH data (MSH + NIH) had higher performance on the combined MSH + NIH test set than on either subset individually or on fully external IU data. IU, Indiana University Network for Patient Care; MSH, Mount Sinai Hospital; NIH, National Institutes of Health Clinical Center.

Fig 2. CNN to predict hospital system detects both general and specific image features.

(A) We obtained activation heatmaps from our trained model and averaged over a sample of images to reveal which subregions tended to contribute to a hospital system classification decision. Many different subregions strongly predicted the correct hospital system, with especially strong contributions from image corners. (B-C) On individual images, which have been normalized to highlight only the most influential regions and not all those that contributed to a positive classification, we note that the CNN has learned to detect a metal token that radiology technicians place on the patient in the corner of the image field of view at the time they capture the image. When these strong features are correlated with disease prevalence, models can leverage them to indirectly predict disease. CNN, convolutional neural network.

Fig 3. Assessing how prevalence differences in aggregated datasets encouraged confounder exploitation.

(A) Five cohorts of 20,000 patients were systematically subsampled to differ only in relative pneumonia risk based on the clinical training data sites. Model performance was assessed on test data from the internal hospital systems (MSH, NIH) and from an external hospital system (IU). (B) Although models perform better in internal testing in the presence of extreme prevalence differences, this benefit is not seen when applied to data from new hospital systems. The natural relative risk of disease at MSH, indicated by a vertical line, is quite imbalanced. IU, Indiana University Network for Patient Care; MSH, Mount Sinai Hospital; NIH, National Institutes of Health Clinical Center; ROC, receiver operating characteristic; RR, relative risk.