Purchase this article with an account.
Meng Wang, Kai Yu, Weifang Zhu, Fei Shi, Xinjian Chen; Multi-Strategy Deep Learning Method for Glaucoma Screening on Fundus Image. Invest. Ophthalmol. Vis. Sci. 2019;60(9):6148.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To improve the performance of glaucoma screening, we proposed a multi-strategy deep learning algorithm, which not only focus on the features of the optic disc region and the CDR, but also take the global information of fundus into account.
As shown in fig.1, Our multi-strategy deep learning method for glaucoma screening on fundus image consists of three steps:1) the glaucoma screening results are directly produced by classification network, which adopts the ResNet  as the backbone model to learn the global information on the whole fundus image. And the region of optic disc is detected by the detection network directly adopts the Faster-RCNN.2) This stage consists of two classification networks and one segmentation network(SegNet), whose input data are cropped from the whole fundus image based on the detection optic disc region result in stage one. The calssification network is same as the stage-one’s calsfficaton network; The segmentation network is utilized to segment the optic disc and optic cup for calculating the vertical cup-to-disc ratio(CDR). The main architecture of the segmentation network adopts the U-shape convolutional network(U-Net) , which is widely used in biomedical image segmentation. In our segmentation network, the dense block is used as the convolutional blocks for obtaining more context information. Moreover, the pixel deconvolutional layer(PixelDCL) , which has been proven that can establish direct relationships among adjacent pixels on the up-sampled feature map and address the checkerboard problem in segmentation results . The encoder path of segmentation network is adopted as the backbone network of the other classification network, which followed two fully connected layers to produce the results of glaucoma screening.3) The glaucoma probability map is obtained by combined the results of all classification networks and the CDR produced from segmentation network.
Dice similarity coefficient for optic disc and cup of our segmentation network are 0.9675 and 0.896. The glaucoma screening precision and recall are 0.9512 and 0.975, respectively.
Different from the previous method that only focus on the region of optic disc, our proposed method not only focus on the features of the optic disc region and the CDR,but also take the global information of fundus into account.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Fig.1: Multi-strategy deep learning method for glaucoma screening on fundus image.
This PDF is available to Subscribers Only