Our pipeline (Figure 6) is made of four main steps, and aims at diagnosing glaucoma based on the vCDR of the eye, which is obtained from OD and OC segmentations. The four steps are described below.


**Pre-processing:** Input images are of high resolution, which would considerably increase the processing time of a deep convolutional network. We thus systematically resize (without crop) the input data to 256*256.

**Deep CNN:** We use a UNet [[4]](https://towardsdatascience.com/understanding-semantic-segmentation-with-unet-6be4f42d4b47) , which is a CNN encoder-decoder designed for image segmentation, and well known in medical imaging. Its main specificity is to combine the encoder feature maps with those of the same scale in the decoder, in order to restitute more accurate spatial information (it creates a sort of "U" shape of the network). The network outputs a prediction maps for both OD and OC segmentation, of the size of its input (256*256 in our case), and is trained with the segmentation ground truth. We monitor the [Dice Coefficient](https://towardsdatascience.com/metrics-to-evaluate-your-semantic-segmentation-model-6bcb99639aa2) during training to monitor the segmentation network performances.


**Post-processing:** In order to accurately determine the OD and OC heights, we only retain the largest connected component of the prediction. This will filter potential noise in the prediction.


**Linear classifier:** We use a simple logistic regression model, which predicts the final classification probability from the vertical Cup-to-Disk Ratio (vCDR). The classifier is trained with the classification labels.

![Pipeline](https://i.postimg.cc/02xPZhYt/pipeline.png)
                                                        Figure 6: Baseline solution pipeline


Although this baseline produces good performance, there are many ways you could think of to improve it, either by modifying some of the four processing steps, or by improving/replacing the pipeline itself.