Search for a command to run...
Purpose To examine Generative Visual Rationales (GVRs) as a tool for visualizing neural network learning of chest radiograph features in congestive heart failure (CHF). Materials and Methods A total of 103 489 frontal chest radiographs in 46 712 patients acquired from January 1, 2007, to December 31, 2016, were divided into a labeled data set (with B-type natriuretic peptide [BNP] result as a marker of CHF) and unlabeled data set (without BNP result). A generative model was trained on the unlabeled data set, and a neural network was trained on the encoded representations of the labeled data set to estimate BNP. The model was used to visualize how a radiograph with high estimated BNP would look without disease (a "healthy" radiograph). An overfitted model was developed for comparison, and 100 GVRs were blindly assessed by two experts for features of CHF. Area under the receiver operating characteristic curve (AUC), κ coefficient, and mixed-effects logistic regression were used for statistical analyses. Results At a cutoff BNP of 100 ng/L as a marker of CHF, the correctly trained model achieved an AUC of 0.82. Assessment of GVRs revealed that the correctly trained model highlighted conventional radiographic features of CHF as reasons for an elevated BNP prediction more frequently than the overfitted model, including cardiomegaly (153 [76.5%] of 200 vs 64 [32%] of 200, respectively; P < .001) and pleural effusions (47 [23.5%] of 200 vs 16 [8%] of 200, respectively; P = .003). Conclusion Features of congestive heart failure on chest radiographs learned by neural networks can be identified using Generative Visual Rationales, enabling detection of bias and overfitted models. © RSNA, 2018 See also the editorial by Ngo in this issue.