Deep learning network with differentiable dynamic programming for retina OCT surface segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Multiple-surface segmentation in OCT images is challenging due to weak boundaries.
• Deep learning methods show promise but struggle with limited medical training data, hindering global structure learning like surface smoothness.

Purpose of the Study:

• To develop an end-to-end deep learning method for retina OCT surface segmentation.
• To explicitly enforce surface smoothness by unifying a U-Net with a constrained differentiable dynamic programming module.

Main Methods:

• A U-Net architecture was integrated with a differentiable dynamic programming module for feature learning and segmentation.
• The method utilizes feedback from downstream model optimization to guide feature learning and enforce global surface structures.

Main Results:

• The proposed method achieved subvoxel accuracy in retinal layer segmentation on Duke AMD and JHU MS OCT datasets.
• Mean Absolute Surface Distance (MASD) errors were 1.88 ± 1.96 μm and 2.75 ± 0.94 μm, respectively.

Conclusions:

• The unified deep learning and dynamic programming approach effectively segments retina OCT surfaces.
• The method demonstrates superior enforcement of global surface structures and smoothness, leading to high accuracy.