Multi-scale Multi-target Domain Adaptation for Angle Closure Classification

Deep learning (DL) has made significant progress in angle closure classification with anterior segment optical coherence tomography (AS-OCT) images. These AS-OCT images are often acquired by different imaging devices/conditions, which results a vast change of underlying data distributions (called “data domains”). Moreover, due to practical labeling difficulties, some domains (e.g., devices) may not have any labels. As result, deep models trained on one specific domain device) difficult adapt and thus perform poorly other devices). To address this issue, we present multi-target adaptation paradigm transfer model labeled source multiple unlabeled target domains. Specifically, propose novel Multi-scale Multi-target Domain Adversarial Network (M2DAN) for classification. M2DAN conducts multi-domain adversarial extracting domain-invariant features develops multi-scale module capturing local global information Based these at scales, the is able classify even without annotations Extensive experiments real-world dataset demonstrate effectiveness proposed method.