Test-Time Augmentation Consistently Improves U-Net-Family Brain Tumor Segmentation: A Controlled Study on BraTS 2020

Abstract

Test-time augmentation (TTA) averages a model's predictions over label-preserving input transformations and is a common, training-free addition to segmentation pipelines. Its isolated effect, however, is rarely quantified under matched conditions across a family of models. In this work, TTA is evaluated as a standalone intervention on sixteen previously trained U-Net-family configurations for brain tumor segmentation on BraTS 2020 (four decoders × four losses, sharing a ResNet34 encoder). Each model is evaluated on 2,242 test slices with and without four-flip TTA, and the difference is assessed by paired bootstrap. TTA improves mean per-slice Dice for every one of the sixteen models, with all sixteen improvements statistically significant. Pooled across models, TTA raises mean Dice from 0.813 to 0.826 (Δ = +0.0129, 95% CI [+0.0119, +0.0140], p < 0.001). The gain is largest for the architectures that were weakest without augmentation, so that TTA also narrows the performance gap between decoder families. Because TTA requires no additional training and only a fixed multiple of inference cost, it is a reliable and low-cost improvement for U-Net-family brain tumor segmentation.