A retinal segmentation model without learning algorithm
This projects aims at developping a model to perform vessel segmentation on retinal images (SLO) without using any learning algorithm, but only basic image processing techniques such as filters and morphological operations.
The data for this projet come from IOSTAR dataset.
The performance criteria are the accuracy, recall, and F1-score between the segmented images and the ground truth ones.
First, we implement a sequence of image processing operations that gives quite good results on one image.
Then, we use Optuna library in order to optimize the parameters selection over the whole dataset (maximization of mean F1-score).
And we obtain the following results:
Best mean F1-score: 0.821