The optic disc is a critical area for detecting glaucoma, as this condition primarily affects the optic disc and its surrounding areas. To improve the efficiency of glaucoma detection, I developed a simple yet effective threshold-based algorithm for automatically detecting the optic disc in fundus images.
I hate manually labeling images to train a deep learning model, and I’m that kind of developer who happily spends a couple of days automating a task that takes a few hours! So, I spent around a week on this project reading state-of-the-art papers regarding optic disc detection. I realized that a threshold method can be used to extract the optic-disc areas, but it’s almost impossible to find the right threshold. However, we know a couple of facts about the optic disc:
- it’s circular.
- it has a fixed aspect ratio to the entire eye area.
So, I wrote a Python code with the help of OpenCV that starts to extract optic disc areas from images and changes the threshold values until it finds an area that checks all of the criteria (check the code if you are interested in this part!).
Then, I used this threshold-based method to label all of the dataset and train a deep neural network for the final optic disc localization. I used ResNet34 with SmoothL1Loss to train a the deep learning model to predict the top-left and bottom-right corner of optic discs. This approach was possible because I normalized all of the images to a fixed size.
Results of the deep learning model:
GitHub link: https://github.com/mh-salari/self_training_optic_disc_localization/blob/main/threshold_localization.ipynb
