| Title | Investigation of eye fundus image quality on vascular segmentation using deep neural networks / |
|---|---|
| Translation of Title | Akies dugno vaizdų kokybės poveikis kraujagyslių segmentavimui naudojant giliuosius neuroninius tinklus. |
| Authors | Domarkaitė, Julija |
| Full Text |
|
| Pages | 52 |
| Keywords [eng] | Segmentation of retinal vessels, Deep neural networks, Quality of eye fundus image, ResUnet, BCDU-Net, Image size, DRIVE, CHASE_DB1, STARE, HRF, IOSTAR, DRHAGIS, ARVDB, Input size, Nearest neighbour interpolation, Bilinear interpolation, Bicubic interpolation |
| Abstract [eng] | Master thesis “Investigation of eye fundus image quality on vascular segmentation using deep neural networks” was conducted at Vilnius university by the student of Systems biology master program Julija Domarkaitė. This master thesis investigates the impact of fundus image quality on deep neural networks performance on retinal vessel segmentation. Automatic segmentation of retinal vessels from fundus images provides information about vessels characteristics but models are unable to handle diverse data. This problem requires a deep investigation on the impact of image quality on deep neural network performance. In order to give insights, this master thesis analyses seven datasets of different properties and quality and their influence on the performance of two deep neural networks. The effects of image size, resizing technique and input size on the model performance are observed. This master thesis shows that models perform on diverse data differently and modifications of image size alter provided information to the models leading to changes in their performance. Different interpolation techniques create different pixel representations which affects segmentation results. Changes of input size provide different contextual information to the model which also affects the segmentation of retinal vessels. In conclusion, this master thesis evaluates the effects of quality of the images and various components on the segmentation of fundus images and provides valuable information about the best approaches to deal with diverse data. |
| Dissertation Institution | Vilniaus universitetas. |
| Type | Master thesis |
| Language | English |
| Publication date | 2022 |