Rethinking Domain Adaptive Optic Disc and Cup Segmentation in Fundus Image through Dynamic Diffusion Flow


Canran Li (University of Sydney), Dongnan Liu (University of Sydney), Weidong Cai (The University of Sydney)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Impacted by the domain shift issue across varying fundus image datasets collected from different medical centres and devices, the performance of a well-trained optic segmentation network is usually affected when applied to another dataset with different distributions. To handle this issue, the unsupervised domain adaptation (UDA) strategy is widely used to improve the generalization ability of deep learning networks by using unlabeled data. However, existing UDA approaches for optic segmentation tasks are mostly adversarial learning-based, which heavily rely on the balance between the source and target datasets to align the features. In this regard, we propose a diffusion-based framework, named Dynamic Diffusion Flow Unsupervised Domain Adaptation (termed DDF-UDA), for the cross-domain optic disc (OD) and optic cup (OC) segmentation in fundus images. Specifically, we propose an adaption module based on diffusion procedure at both feature and pixel levels to alleviate the cross-domain gaps. In order to modify the domain information of the source image while minimizing changes to its content, we further propose an adjustment strategy based on Nash equilibrium, which could dynamically modify the diffusion steps. Experimental results on public datasets demonstrate that our DDF-UDA can effectively leverage unlabeled data to achieve state-of-the-art performance in OD/OC segmentation.

Citation

@inproceedings{Li_2024_BMVC,
author    = {Canran Li and Dongnan Liu and Weidong Cai},
title     = {Rethinking Domain Adaptive Optic Disc and Cup Segmentation in Fundus Image through Dynamic Diffusion Flow},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0421.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection