Adaptive wavelet-VNet for single-sample test time adaptation in medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Domain shift in medical imaging (e.g., different scanners/protocols) degrades deep learning segmentation model performance.
• Manual data labeling is costly and time-consuming; privacy concerns limit data collection.
• Test-time adaptation (TTA) using unlabeled target domain data offers a practical solution for unseen datasets.

Purpose of the Study:

• To enhance deep learning segmentation accuracy on unseen datasets.
• To improve the efficiency and stability of TTA for individual samples from heterogeneous medical imaging domains.

Main Methods:

• Proposed a dynamically adaptive wavelet-VNet (WaVNet) for TTA using unlabeled test samples.
• Embedded multiscale wavelet coefficients into a V-Net encoder for adaptive spatial and spectral feature adjustment.
• Optimized model parameters using a hybrid objective function with shape-aware, Refine, and entropy loss functions.

Main Results:

• Achieved superior segmentation accuracy on liver and prostate datasets compared to baseline methods.
• Reported mean Dice coefficients (DSC) of 78.10% (liver) and 80.02% (prostate).
• Demonstrated significant improvement over the no-adaptation baseline, with DSC increases of 11.67% (liver) and 15.27% (prostate).

Conclusions:

• The adaptive WaVNet effectively enhances image segmentation accuracy in unseen domains during test time via unsupervised learning and multi-objective optimization.
• The method is beneficial for clinical applications with scarce data or changing data distributions, such as online adaptive radiotherapy.
• Open-source code is available for WaVNet implementation.