Retinal Vessel Segmentation by U-Net with VGG-16 Backbone on Patched Images with Smooth Blending

0585312 - ÚGN 2025 RIV CH eng C - Konferenční příspěvek (zahraniční konf.)
Hasal, M. - Pecha, Marek - Nowaková, J. - Hernández-Sosa, D. - Snášel, V. - Timkovič, J.
Retinal Vessel Segmentation by U-Net with VGG-16 Backbone on Patched Images with Smooth Blending.
Advances in Intelligent Networking and Collaborative Systems. Cham: Springer Cham, 2023 - (Barolli, L.), s. 465-474. Lecture Notes on Data Engineering and Communications Technologies, 182. ISBN 978-3-031-40970-7. ISSN 2367-4512. E-ISSN 2367-4520.
[International Conference on Intelligent Networking and Collaborative Systems (INCoS-2023). Chiang Mai (TH), 06.09.2023-08.09.2023]
Institucionální podpora: RVO:68145535
Klíčová slova: deap learning * semantic segmentation * retina vessels
Obor OECD: Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)
https://link.springer.com/chapter/10.1007/978-3-031-40971-4_44

Detecting vessels in retinal images is crucial for various medical applications, including diagnosing and monitoring eye diseases such as diabetic retinopathy, glaucoma, and macular degeneration. This paper presents a study on applying the U-Net architecture with a VGG-16 backbone for retinal vessel segmentation trained on patched images. As a source of training images, three well-labeled datasets, DRIVE, STARE, and CHASE DB1, were used for the training of the segmentation algorithm. We implemented the task-specific data class to further divide training images into patches, and the data augmentation techniques to increase the size of training set and to promote the model’s generalization ability. Additionally, a blending technique was employed to achieve smooth predictions by blending image patches. The experimental results highlight the effectiveness of the proposed approach in accurately detecting blood vessels in retinal images, providing promising prospects for improving ophthalmic diagnosis and treatment.
Trvalý link: https://hdl.handle.net/11104/0353022