Overview / Usage

We present a robust method for glaucoma screening from fundus images using an ensemble of convolutional neural networks (CNNs). The pipeline comprises of first segmenting the optic disk and optic cup from the fundus image, then extracting a patch centered around the optic disk and subsequently feeding to the classification network to differentiate the image as diseased or healthy. In the segmentation network, apart from the image, we make use of spatial coordinate (X & Y) space so as to learn the structure of interest better. The classification network is composed of a DenseNet201 and a ResNet18 which were pre-trained on a large cohort of natural images. On the REFUGE validation data (n=400), the segmentation network achieved a dice score of 0.88 and 0.64 for optic disc and optic cup respectively. For the tasking differentiating images affected with glaucoma from healthy images, the area under the ROC curve was observed to be 0.85.

Methodology / Approach

We used 57 layers of the deep convolutional neural network, for semantic segmentation. Images were cropped to nearest square size and resized to a dimension of (512, 512). The different lighting conditions and intensity variations among images across various databases were circumvented by performing normalization of the histogram using Contrast Limited Adaptive Histogram Equalization (CLAHE). The dataset was split into training, validation and testing in the ratio 70:20:10. Both the networks were trained and validated on 278 and 59 images respectively. To further address the issue of class imbalance in the network, the parameters of the network were trained by minimizing weighted cross entropy. The weight associated to each class was equivalent to the ratio of the median of the class frequency to the frequency of the class of interest. To increase the number of data points on the fly data augmentation was used with random rotation between 0-360 degree, and random flip with a probability of 0.5 was also implemented. The number of batches was set at 4, while the learning rate was initialized to 0.0001 and decayed by a factor of 10 % every-time the validation loss plateaued. The network was trained for 50 epochs using weighted cross-entropy as a cost function, weights were estimated by using entire training data as a ratio of the median of frequency

Technologies Used

Deep learning, Computer Vision, Pytorch, matplotlib, simpleITK, numpy, sklearn, and pandas

Repository

https://github.com/koriavinash1/Optic-Disk-Cup-Segmentation

Other links

• arxiv paper link