Image segmentation enhanced by heuristic assistance for retinal vessels case

Binary segmentation is one of the basic issues of image processing. It also has a wide range of applications. One of them is the use of segmentation tools to support disease prevention and diagnosis. This makes it possible to automate this task and perform analyses regularly and for a large number of people, compared to manual analysis. However, segmentation tools must be particularly accurate for this to be possible. In this paper, we focus on improving retinal vessel segmentation methods. First of all, the U-Net network architecture has been expanded for even more accurate detection of the thinnest vessels. This is the most difficult part of the issue - spotting features in great detail. In addition, we introduced a strategy for selecting images to expand the training set, assisted by heuristic algorithms. Thanks to this, the set is expanded only to include those samples containing the most information about thin vessels. Moreover, we have also modified the method of binary pixel classification, introducing the possibility of capturing image parts with a lower probability of having a vessel, using local information. As a result of our work, we checked the effectiveness of the model with such metrics as accuracy, sensitivity, specificity, AUC, F1 score, and Intersection over Union using classic and modified thresholding. An improvement in the sensitivity value from 0.7562 to 0.7822 and an improvement in AUC from 0.8681 to 0.8776 was achieved with minor losses in other metrics.