GVRheart

The study aimed to evaluate Generative Visual Rationales (GVRs) as a tool for visualizing how neural networks learn features of congestive heart failure (CHF) from chest radiographs. Using a dataset of 103,489 chest radiographs, the researchers trained a generative model on unlabeled images and a neural network on labeled images to estimate B-type natriuretic peptide (BNP) levels, then visualized what high BNP radiographs would look like if they were “healthy.” The findings showed that the correctly trained model more accurately identified CHF features, such as cardiomegaly and pleural effusions, compared to an overfitted model, demonstrating the potential of GVRs to detect biases and improve model transparency.