Collaborative feature ensembling adaptation for domain adaptation in unsupervised optic disc and cup segmentation

A system comprising:

• one or more processors;
• and
• memory storing computer-executable instructions that, when executed by the one or more processors, cause performance of operations comprising:
• for a plurality of training iterations, drawing a mini-batch of labeled source domain samples from a source domain and unlabeled target samples from a target domain, wherein a domain shift exists between the source domain and the target domain;
• initiating training of a first network to learn detection of OC or OD regions within a labeled source sample from the source domain, wherein training weights of the first network are adapted based on a loss calculated from an output of the first network and a ground truth for a same source sample;
• sharing training weights of the first network with a second network;
• initiating training of the second network to learn detection of OC or OD regions within an unlabeled sample from the target domain, wherein training weights of the second network are adapted based on an adversarial loss calculated from an output of the second network and the output of the first network for the same sample across the source and target domains;
• adjusting the training weights of the first network and the second network based on the calculated adversarial loss;
• transferring average training weights of the second network to a third network;
• initiating training of the third network to learn detection of OC and OD regions within an unlabeled sample from the target domain;
• computing a mean square error loss between an output of the third network and the output of the second network for a same target sample;
• adjusting the training weights of the second network based on the mean square error loss computation;
• and
• proceeding with a next iteration of the plurality of training iterations until each iteration of the plurality of training iterations has been completed.