Choroid Segmentation Based on Dense Atrous Convolution and Coordinate Parallel Attention

Changes in the choroid are closely related to many ophthalmic diseases. Doctors often need to manually segment choroid in optical tomography image (OCT) during diagnosis, and then quantitatively analyze choroidal health. However, manual segmentation is time-consuming and laborious. The difficulty of automatic choroidal segmentation lies in the blurring of the lower choroidal boundary in the OCT images, which makes it difficult to capture context information. Morever, the choroidal structure is similar to the retina structure, which is easy to be confused. In order to solve this difficulty, this paper proposes a residual codec model that combines coordinate parallel attention module and dense hole convolution module. A bridge structure is designed, including attention mechanism and cavity convolution, which can increase the receptive field of the model while suppressing shallow noise. In order to make the model focus on choroidal structure information, a hybrid loss function containing structural similarity is introduced to train the model. It is shown from experimental results that this model can effectively improve the segmentation accuracy of choroid, and the similarity between Dice coefficient and Jaccard is 97.63% and 95.28% on the OCT choroid dataset.