Tackling domain generalization for out-of-distribution endoscopic imaging

TL;DR

This paper proposes a novel approach for improving domain generalization in endoscopic image segmentation by combining style-content feature preservation and a restitution module, significantly outperforming existing methods on multiple datasets.

Contribution

It introduces a new method that leverages style and content information with instance normalization, feature covariance mapping, and a restitution module to enhance generalization in endoscopic imaging.

Findings

13.7% improvement over DeepLabv3+ on EndoUDA polyp dataset

19% improvement over baseline on EndoUDA BE dataset

Significant performance gains over state-of-the-art methods

Abstract

While recent advances in deep learning (DL) for surgical scene segmentation have yielded promising results on single-center and single-imaging modality data, these methods usually do not generalize well to unseen distributions or modalities. Even though human experts can identify visual appearances, DL methods often fail to do so when data samples do not follow a similar distribution. Current literature addressing domain gaps in modality changes has focused primarily on natural scene data. However, these methods cannot be directly applied to endoscopic data, as visual cues in such data are more limited compared to natural scenes. In this work, we exploit both style and content information in images by performing instance normalization and feature covariance mapping techniques to preserve robust and generalizable feature representations. Additionally, to avoid the risk of removing salient feature representations associated with objects of interest, we introduce a restitution module within the feature-learning ResNet backbone that retains useful task-relevant features. Our proposed method shows a 13.7% improvement over the baseline DeepLabv3+ and nearly an 8% improvement over recent state-of-the-art (SOTA) methods for the target (different modality) set of the EndoUDA polyp dataset. Similarly, our method achieved a 19% improvement over the baseline and 6% over the best-performing SOTA method on the EndoUDA Barrett's esophagus (BE) dataset.