Uncertainty-guided cross-level fusion network for retinal OCT image segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning excels at optical coherence tomography (OCT) segmentation but struggles with uncertainty due to data distribution and network limitations.
• Accurate uncertainty estimation is crucial for reliable confidence assessments and improved OCT segmentation predictions.

Purpose of the Study:

• Propose a novel uncertainty-guided cross-layer fusion network (UGCFNet) for enhanced retinal OCT segmentation.
• Integrate uncertainty quantification into deep neural network training to improve segmentation accuracy.

Main Methods:

• Utilize an encoder-decoder architecture that quantifies uncertainty across multiple stages to focus on high-uncertainty regions.
• Implement cross-layer feature fusion for a comprehensive understanding of semantic and morphological details.
• Incorporate improved Bayesian neural network and uncertainty-aware loss functions for effective uncertainty modeling.

Main Results:

• Achieved state-of-the-art performance on AI-Challenger and OIMHS OCT segmentation datasets.
• Demonstrated high accuracy with average Dice similarity coefficients of 79.47% and 93.22% on the respective datasets.
• Extensive experiments validated the model's effectiveness on large-scale OCT datasets.

Conclusions:

• The proposed UGCFNet significantly advances retinal OCT segmentation.
• Integrating uncertainty guidance and cross-level feature fusion leads to more reliable and accurate segmentation outcomes.