Automatic retinal structure segmentation via U-Net framework and quantitative analysis with ImageJ

Retina is a window reflecting the state of the general system and the condition of the eye. The traditional heavy-manual quantification of the morphological characteristics of retina is in need of improving. Therefore, we aim to automatically segment the retinal vessels and lesions and further analyze segmented results. Three public datasets including DRIVE, DIARETDB, IDRID were selected, and the images from them were preprocessed and augmented with a series of enhancements, random rotation and gamma transformation. Attention gate (AG) U-Net framework was used to segment retinal vessels, OD and exudates respectively. The performance of AG U-Net model was evaluated. Furthermore, the segmented results were analyzed with different shape descriptors using ImageJ. The geometric features such as width, length, area, circumference, and roundness were extracted and statistically analyzed in grade 2 and 3 DR images. The results approved the superiority of Attention-U-Net in retinal structure segmentation and the geometric features were successfully extracted and analyzed. In conclusion, the proposed AG U-Net empowered automatic segmentation and ImageJ based analytic framework are worthy of applications on retinal images, thus fostering the clinical science investigations.