Exploring Test Time Adaptation for Subcortical Segmentation of the Fetal Brain in 3D Ultrasound

TL;DR

This paper tackles the challenge of domain shifts in 3D fetal brain ultrasound for subcortical segmentation by applying test time adaptation (TTA) to a pretrained 3D UNet. It introduces two novel TTA strategies, LayerInspect and EntropyKL, the latter incorporating a normative atlas as a prior on tissue proportions to guide updates without labels. Across simulated shifts, gestational age ranges, and unseen scanners, all TTA methods improve segmentation performance, with EntropyKL showing the strongest robustness, especially for unseen sites; single-volume adaptation often yields larger gains. The work advances automated fetal brain monitoring by enabling more reliable segmentation across diverse US acquisitions and weeks of gestation, and provides code for replication.

Abstract

Monitoring the growth of subcortical regions of the fetal brain in ultrasound (US) images can help identify the presence of abnormal development. Manually segmenting these regions is a challenging task, but recent work has shown that it can be automated using deep learning. However, applying pretrained models to unseen freehand US volumes often leads to a degradation of performance due to the vast differences in acquisition and alignment. In this work, we first demonstrate that test time adaptation (TTA) can be used to improve model performance in the presence of both real and simulated domain shifts. We further propose a novel TTA method by incorporating a normative atlas as a prior for anatomy. In the presence of various types of domain shifts, we benchmark the performance of different TTA methods and demonstrate the improvements brought by our proposed approach, which may further facilitate automated monitoring of fetal brain development. Our code is available at https://github.com/joshuaomolegan/TTA-for-3D-Fetal-Subcortical-Segmentation.

Figures (5)

• Figure 1: TTA framework schematic. The source model is the pretrained 3D UNet from Hesse2022, which is adapted on the target data and then used on the same data to make a prediction.
• Figure 2: Results from adapting to simulated augmentations. Dice scores are averaged across all four subcortical regions, and the shaded region indicates the standard deviation bounds.
• Figure 3: EntropyKL with varying $\lambda$ (a). LayerInspect performance for various rotations, varying the number of layers updated (b) and the learning rate (c).
• Figure 4: LHS: Average Dice score across all structures at each gestational age. Differences between methods are negligible and so overlap on the graph. RHS: The same plane from the US atlas Namburete2023 at different ages, showing the development of the four subcortical structures.
• Figure 5: An example slice from a scan from a) a GE scanner and b) a Canon scanner, comparing the performance of the different methods, with the reference atlas slice to show expected anatomy. Red=LPVH, Green=CP, Yellow=CB.