Multi-source Unsupervised Domain Adaptation Fundus Lesion Segmentation of Various OCT Devices with Moment Consistency

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of fundus lesions in optical coherence tomography (OCT) images is crucial for diagnosing retinopathy and choroidopathy.
• Domain shift caused by OCT images from various manufacturers poses a significant challenge for traditional segmentation models.

Purpose of the Study:

• To develop a novel multi-source domain adaptation framework for robust fundus lesion segmentation in OCT images from diverse manufacturers.
• To address the challenge of domain shift in OCT image analysis for improved ophthalmological diagnostics.

Main Methods:

• A multi-order moment consistency approach using moment generating function (MGF) to align feature distributions across domains.
• A perturbation-based feature consistency strategy linking semantic consistency with feature distribution alignment.
• Population stability whitening to automatically separate style-related and content-related features.

Main Results:

• The proposed framework demonstrated significant superiority over state-of-the-art approaches on datasets with diverse OCT device domains.
• The multi-order moment consistency and perturbation-based strategies effectively aligned feature distributions and improved model robustness.
• Population stability whitening successfully separated style and content features, enhancing segmentation performance.

Conclusions:

• The novel multi-source domain adaptation framework effectively overcomes domain shift challenges in OCT image segmentation.
• The proposed methodological innovations lead to superior performance in segmenting fundus lesions across different OCT devices.
• This work advances automated analysis of OCT images, aiding ophthalmologists in diagnosing eye conditions.