| Abstract [eng] |
This paper serves as a master thesis titled "Fusion of medical images and preclinical data in ophthalmology using deep learning". It explores data fusion and compares it to simple segmentation based on deep learning. At first, an overview of relevant literature is provided. It gives knowledge about deep learning, CNV pathology, data fusion, as well as achievements in the field. Then in order to do the task, ophthalmology data was prepared: preclinical data was collected that would tell various information about animals, as well as two sets of images (OCT b-scans and FA images) were annotated. Later on, three algorithms are compared. The first one is image segmentation using the trained neural net model. Then early data fusion algorithm, and the late data fusion algorithm both follow different approaches to the task. Results are calculated: image segmentation is compared using the Dice coefficient and methods are compared to each other using the Wilcoxon rank-sum test. Results show that data fusion using given methods didn’t improve image segmentation: the only combination where the difference of Dice coefficients wasn’t significant was between initial segmentation and late data fusion algorithms for FA images, but that model had other problems. Discussions and conclusions are formed explaining why data fusion wasn’t as good performing as expected. |
