TestFit: A plug-and-play one-pass test time method for medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) excels in medical image segmentation, but research primarily focuses on network design and training.
• Limited work exists on improving DL models during the testing phase (test time).

Purpose of the Study:

• To investigate the feasibility of stably improving off-the-shelf DL segmentation networks on-the-fly during test time.
• To introduce TestFit, a novel online test-time method for enhancing segmentation performance by fitting test data distributions.

Main Methods:

• TestFit utilizes a supplementary network (SuppNet) derived from the original network (OGNet).
• OGNet preserves original training data insights, while SuppNet adapts to test data.
• Segmentation results and supervision signals are generated by combining OGNet and SuppNet outputs dynamically.
• The method requires a single pass per test sample and no prior training preparation.

Main Results:

• TestFit demonstrates stable and effective online test-time improvement for DL segmentation models.
• The method shows capability across diverse 2D and 3D medical imaging datasets.
• Experiments confirm the effectiveness of technical treatments for stability and performance.

Conclusions:

• TestFit offers a plug-and-play solution for enhancing existing DL segmentation models at test time.
• The method requires minimal hyperparameter tuning and is practical for real-world applications.
• TestFit successfully improves segmentation accuracy by adapting to test data distributions without compromising original model integrity.