CADA: Multi-scale Collaborative Adversarial Domain Adaptation for unsupervised optic disc and cup segmentation

Recently, deep neural networks have demonstrated comparable and even better performance than board-certified ophthalmologists in well-annotated datasets. However, the diversity of retinal imaging devices poses a significant challenge: domain shift, which leads to degradation when applying learning models trained on one new testing domains. In this paper, we propose adaptation framework comprising multi-scale inputs along with multiple adaptors applied hierarchically both feature output spaces. The proposed training strategy novel unsupervised framework, called Collaborative Adversarial Domain Adaptation (CADA), can effectively overcome shift challenge. Multi-scale reduce information loss due pooling layers used network for extraction, while our CADA is an interactive paradigm that presents exquisite collaborative through adversarial ensembling weights at different layers. particular, produce prediction unlabeled target data, simultaneously achieve invariance model generalizability via outputs from levels maintaining exponential moving average (EMA) historical during training. Without annotating any sample domain, losses encoder decoder guide extraction domain-invariant features confuse classifier. Meanwhile, EMA reduces uncertainty adapting discriminator learning. Comprehensive experimental results demonstrate incorporating input outperform state-of-the-art methods segmenting optic disc cup fundus images stemming REFUGE, Drishti-GS, Rim-One-r3 code available https://github.com/cswin/CADA.