Training Deep Learning Models to Work on Multiple Devices by Cross-Domain Learning with No Additional Annotations.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of retinal layers and intraretinal fluid in Optical Coherence Tomography (OCT) images is crucial for diagnosing and monitoring macular diseases.
• Current deep learning segmentation models often require extensive labeled data, limiting their generalizability across different OCT devices and manufacturers.

Purpose of the Study:

• To develop an unsupervised cross-domain segmentation algorithm capable of segmenting intraretinal fluid and retinal layers on OCT images from various manufacturers and devices.
• To leverage generative adversarial networks (GANs) for domain generalization in OCT image segmentation without requiring labeled data from the target domain.

Main Methods:

• An unsupervised GAN model, GANSeg, was developed to segment retinal layers and intraretinal fluid.
• The model was trained on labeled Heidelberg Spectralis OCT images (domain A) and applied to unlabeled Topcon 1000 OCT images (domain B).
• Performance was validated against manual segmentations by masked graders and compared to a baseline U-Net model trained on the same labeled data.

Main Results:

• GANSeg achieved comparable Dice scores to human experts for specific retinal layers and intraretinal fluid segmentation on unseen OCT devices.
• The algorithm significantly outperformed the baseline U-Net model in cross-domain segmentation tasks.
• GANSeg demonstrated successful generalization to OCT devices (Zeiss and Topcon Maestro2) not included during its training.

Conclusions:

• GANSeg facilitates the transfer of supervised deep learning segmentation models across different OCT devices without the need for target domain labeled data.
• This unsupervised approach significantly enhances the applicability and utility of deep learning algorithms in diverse OCT imaging settings.