Diverse Data Generation for Retinal Layer Segmentation With Potential Structure Modeling.

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Accurate retinal layer segmentation in optical coherence tomography (OCT) is crucial for diagnosing eye diseases.
• Collecting diverse and balanced OCT datasets, especially with various pathologies, is a significant challenge.
• Existing generative models struggle with data diversity due to imbalanced training data.

Purpose of the Study:

• To develop a framework for generating diverse and balanced OCT image-label pairs from imbalanced real-world data.
• To improve the performance of retinal layer segmentation by augmenting training datasets with synthetic data.
• To address the limitations of current generative models in capturing data diversity.

Main Methods:

• A novel image-label pair generation framework utilizing two customized diffusion probabilistic models.
• Generation of diverse layer masks followed by plausible OCT image synthesis.
• Introduction of pathological-related conditions and a potential structure modeling technique to guide generation and enhance diversity.

Main Results:

• The proposed method generates OCT images with superior quality and diversity compared to existing generative approaches.
• Extensive training with the generated data significantly improved downstream retinal layer segmentation performance.
• Experimental validation on two public datasets confirmed the effectiveness of the framework.

Conclusions:

• The developed framework effectively generates diverse and balanced OCT data, overcoming limitations of imbalanced datasets.
• The approach enhances the utility of generative models for medical image analysis and segmentation tasks.
• This work provides a valuable tool for improving the accuracy and robustness of retinal layer segmentation.