Kurtosis Transformation for Multi-Source Domain Generalization Segmentation of OCT Images From Multi-Manufacturers' Devices.

Abstract

Objective: Optical coherence tomography (OCT) images can visualize retinal layers and fundus lesions. Retinal structure segmentation is of great significance in early lesion detection and treatment guidance. However, devices from different OCT manufacturers are largely different from each other, which often leads to degraded results in image segmentation.

Methods: To enrich the diversity of multi-source domains in intensity distributions and image contrasts of multi-manufacturers' OCT images, a kurtosis transformation method is proposed to generate a kurtosis-transformed image. To make the generated potential T-styles as different as possible from the source domain styles, a mean style contrastive learning method is proposed to maximize the distance between style features of a kurtosis-transformed image and multi-source domain images. To improve the diversity and independence of potential T-styles in a high-level orthogonal space, variance style orthogonalization is proposed to impose an orthogonal constraint on the reparametrized variance styles. Mean style features and variance style features are combined to modulate an input image for the training of the segmentation network.

Results: Comprehensive experiments have been performed on two OCT image datasets. Compared to state-of-the-art methods, the proposed method can achieve better segmentation.

Conclusion: The proposed segmentation method can be trained on labeled OCT images from multi-manufacturers' devices and can be tested on unseen manufacturer's device, and has good domain generalization performance in both retinal layer and lesions segmentation tasks.

Significance: The proposed method can be used in routine clinical settings, when OCT images from multi-manufacturers' devices are available.