Keywords: generalization, deep learning, unets, visual representations
TL;DR: We propose new techniques for visualizing interior features of UNets and new metrics that exploit those features to predict generalization (test) performance on previously-unseen data.
Abstract: Fully-convolutional neural networks, such as the 2D or 3D UNet, are now pervasive in medical imaging for semantic segmentation, classification, image denoising, domain translation, and reconstruction. However, evaluation of UNet performance, as with most CNNs, has mostly been relegated to evaluation of a few performance metrics (e.g. accuracy, IoU, SSIM, etc.) using the network's final predictions, which provides little insight into important issues such as dataset shift that occur in clinical application. In this paper, we propose techniques for understanding and visualizing the generalization performance of UNets in image classification and regression tasks, giving rise to metrics that are indicative of performance on a withheld test-set without the need for groundtruth annotations.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: methodological development
Primary Subject Area: Uncertainty Estimation
Secondary Subject Area: Interpretability and Explainable AI
Source Code Url: https://gitlab.com/abhe/UNet-Generalization_MIDL2021
Source Latex: zip