Overcoming the sensor delta for semantic segmentation in OCT images

Abstract

The performance of a segmentation network optimized on data from a specific type of OCT sensor will decrease when applied to data from a different sensor. In this work, we deal with the research question of adapting models to data from an unlabeled new sensor with new properties in an unsupervised way. This challenge is known as unsupervised domain adaptation and can alleviate the need for costly manual annotation by radiologists. We show that one can strongly improve a model's result that was trained in a supervised way on the source OCT sensor domain on the target sensor domain. We do this by aligning the source and target domain distributions in the feature space through a semantic clustering method. Apart from the unsupervised domain adaptation, we improved even the supervised training compared to the results in the RETOUCH challenge by employing a sophisticated training strategy. The RETOUCH challenge contains three different types of OCT scanners and provides annotations for the task of disease-related fluid classes.

Reducing the impact of domain shift in deep learning for OCT segmentation using image manipulations

“…done by alignment of the feature space distributions of source and target domain by clustering as e.g. in Niemeijer et al 4 or through adversarial training as in Hoffman et al, 10 Li et al 11 or Wang et al 12 Output space adaptation usually consists of self training on the target domain as e.g. applied in Zheng et al 13 or adversarial training on the predicted outputs as in Tsai et al 14 But since we are studying the manipulation of the input space, the input space adaptation is the most interesting to our work.…”

“…This phenomenon is called domain shift and an actively studied topic. [3][4][5][6] We study the domain shift problem under the assumption that labels are only given for a training (source) domain but not the test (target) domain. Domain adaptation can be categorized depending on the adaptation spaces, which are the input space, the feature space, and the output space of a DNN.…”