Dynamic Gaussian Convolution for Retinal Optical Coherence Tomography Segmentation

Ziyun Yang      Xiaorui Peng      Zhanghao Yang

zy104@duke.edu     xp24@duke.edu zy105 @duke.edu

Paper PDF

Increasing the imaging equality of spectral-domain optical coherence tomography (SD-OCT) is a key factor to expand OCT-relevant application, such as OCT segmentation and further diagnosis. In this paper, the method of denoising the noisy detected image by adding a trainable Gaussian kernel as a denoising physical layer(named dynamic Gaussian convolution) is investigated. Combined with traditional segmentation network, we showed that using our dynamic Gaussian convolution as physical layer is better than traditional convolution layer in both segmentation and denoising task. Multiple factors like noise levels, role of physical layer, training strategy are also discussed.

Paper: Paper PDF