Adaptive Feature Fusion Neural Network for Glaucoma Segmentation on Unseen Fundus Images

TL;DR

This work tackles glaucoma fundus segmentation under unseen-domain, low-data conditions. It introduces AFNN, an adaptive architecture consisting of a domain adaptor, a feature-fusion network, and self-supervised multi-task learning, augmented by a two-stage optimization and a weighted-dice-loss to balance optic-disk and optic-cup segmentation. Empirical results on four public glaucoma datasets show AFNN achieving state-of-the-art generalization with improved edge accuracy (HD/ASD) and balanced segmentation performance across OC and OD. The proposed framework offers a practical pathway to robust medical image segmentation when annotated data are scarce and domain shifts are prevalent, with potential applicability to other data-limited medical imaging tasks.

Abstract

Fundus image segmentation on unseen domains is challenging, especially for the over-parameterized deep models trained on the small medical datasets. To address this challenge, we propose a method named Adaptive Feature-fusion Neural Network (AFNN) for glaucoma segmentation on unseen domains, which mainly consists of three modules: domain adaptor, feature-fusion network, and self-supervised multi-task learning. Specifically, the domain adaptor helps the pretrained-model fast adapt from other image domains to the medical fundus image domain. Feature-fusion network and self-supervised multi-task learning for the encoder and decoder are introduced to improve the domain generalization ability. In addition, we also design the weighted-dice-loss to improve model performance on complex optic-cup segmentation tasks. Our proposed method achieves a competitive performance over existing fundus segmentation methods on four public glaucoma datasets.

Figures (12)

• Figure 1: Overview of adaptive fusion neural network (AFNN). AFNN mainly contains three modules: the domain adaptor, the feature-fusion network, and the self-supervised multi-task learning module. In particular, the domain adaptor maps a variety of domain distributions into a normalized general distribution. Then with the help of feature-fusion network, AFNN improves its feature learning ability. The last self-supervised multi-task learning further improves model representation ability by learning from the limited training data.
• Figure 2: Domain-adaptive learning in AFNN. The domain adaptor consists of convolution layers and normalization layers, which maps the raw data to a common distribution and adapts the inputs to the pretrained backbone.
• Figure 3: The network architectures comparisons of DeepLab, UNet and our feature-fusion network. UNet conducts feature fusion with multiple layers, and DeepLab conducts feature fusion with multiple scales. Our feature-fusion network is a blend of UNet and DeepLab.
• Figure 4: Comparison of MSE loss and weighted dice loss."OD" and "OC" denote the optic-disk and optic-cup correspondingly. "OD" and "OC" have different intersection areas.
• Figure 5: Comparisons of different glaucoma datasets. Dataset distribution visualizations are listed on the top-left of each dataset, and the domain gaps of different datasets are shown on the right sub-figure.