UNISELF: A unified network with instance normalization and self-ensembled lesion fusion for multiple sclerosis lesion...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Neuroimaging

Background:

• Manual segmentation of multiple sclerosis (MS) lesions from MRI is time-consuming and prone to variability.
• Deep learning (DL) methods offer state-of-the-art performance for automated MS lesion segmentation but struggle with generalization across different datasets.
• Existing DL models often fail to maintain high accuracy and generalize well when trained on limited single-source data, especially with variations in imaging protocols or missing data contrasts.

Purpose of the Study:

• To develop a novel deep learning (DL) method, UNISELF, for accurate and generalizable automated segmentation of multiple sclerosis (MS) lesions.
• To improve both in-domain accuracy and out-of-domain generalization for MS lesion segmentation, overcoming limitations of current DL approaches.
• To address challenges posed by domain shifts and missing contrasts in multi-center and multi-protocol MRI datasets.

Main Methods:

• UNISELF utilizes test-time self-ensembled lesion fusion to enhance segmentation accuracy.
• The method incorporates test-time instance normalization (TTIN) of latent features to mitigate domain shifts and handle missing input contrasts.
• The model was trained on the ISBI 2015 longitudinal MS segmentation challenge training dataset.

Main Results:

• UNISELF achieved top-tier performance on the ISBI 2015 challenge test dataset.
• The method demonstrated superior performance compared to benchmark methods across diverse out-of-domain datasets (MICCAI 2016, UMCL, private multisite).
• UNISELF effectively handled domain shifts and missing contrasts arising from variations in acquisition protocols, scanner types, and imaging artifacts.

Conclusions:

• UNISELF provides a robust solution for automated MS lesion segmentation, achieving high accuracy within a training domain and strong generalizability across varied datasets.
• The proposed method, incorporating test-time self-ensembling and TTIN, effectively addresses domain shifts and missing contrasts in real-world clinical scenarios.
• UNISELF represents a significant advancement in automated MS lesion segmentation, offering improved efficiency and reproducibility for clinical applications.