Unsupervised domain adaptation model for lesion detection in retinal OCT images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical Coherence Tomography (OCT) is crucial for retinal imaging, generating high-resolution anatomical data.
• Deep learning models are increasingly used for automated lesion detection in OCT images to reduce clinician workload.
• Domain discrepancies in OCT images from different devices pose a significant challenge for deep learning model performance.

Purpose of the Study:

• To develop an unsupervised domain adaptation model for robust lesion detection in cross-device retinal OCT images.
• To address the challenge of domain shift in deep learning for medical image analysis, specifically for OCT lesion detection.

Main Methods:

• Proposed a faster-RCNN based unsupervised domain adaptation model.
• Minimized domain shift by simultaneously reducing image-level and instance-level discrepancies.
• Employed a domain classifier and Wasserstein distance critic for shift alignment.

Main Results:

• Achieved an average accuracy improvement of over 8% compared to models without domain adaptation.
• Demonstrated superior performance over other comparable domain adaptation methods on cross-device OCT data.
• Successfully reduced domain shift in retinal OCT lesion detection tasks.

Conclusions:

• The proposed unsupervised domain adaptation model is effective in mitigating domain shift for OCT lesion detection.
• The method offers improved accuracy and robustness for analyzing retinal OCT images from diverse devices.
• Highlights the potential of domain adaptation techniques in medical deep learning applications.