Show abstract - PDF - Reviews

Miscalibration of deep neural networks (DNNs) can lead to unreliable predictions and hinder their use in clinical decision-making. This miscalibration is often caused by overconfident probability estimates. Calibration techniques such as model ensembles, regularization terms, and post-hoc scaling of the predictions can be employed to improve the calibration performance of DNNs. In contrast to DNNs, graph neural networks (GNNs) tend to exhibit underconfidence. In this study, we investigate the efficacy of calibration techniques developed for DNNs when applied to GNNs trained on medical image data, and compare the calibration performance of binary and multiclass node classification on a benchmark dataset and a medical image dataset. We find that post-hoc methods using Platt scaling or Temperature scaling, or methods that add a regularization term to the loss function during training are most effective to improve calibration. Our results further indicate that these calibration techniques are more effective for multiclass classification tasks compared to binary classification tasks.
Hide abstract