Tackling Domain Generalization for Out-of-Distribution Endoscopic Imaging

Abstract

While recent advances in deep learning (DL) for surgical scene segmentation have yielded promising results on single-centre and single-imaging modality data, these methods usually do not generalise to unseen distribution or unseen modalities. Even though human experts can identify visual appearances, DL methods often fail to do so if data samples do not follow the similar data distribution. Current literature for tackling domain gaps in modality changes has been done mostly for natural scene data. However, these methods cannot be directly applied to the endoscopic data as the visual cues are very limited compared to the natural scene data. In this work, we exploit the style and content information in the image by performing instance normalization and feature covariance mapping techniques for preserving robust and generalizable feature representations. Further, to eliminate the risk of removing salient feature representation associated with the objects of interest, we introduce a restitution module within the feature learning ResNet backbone that allows the retention of useful task-relevant features. Our proposed method obtained 13.7% improvement over the baseline DeepLabv3+ and nearly 8% improvement on recent state-of-the-art (SOTA) methods for the target (different modality) set of EndoUDA polyp dataset. Similarly, our method achieved 19% improvement over the baseline and 6% over best performing SOTA on EndoUDA Barrett’s esophagus (BE) data.