View abstract

Cet abstract a été assigné à session AOB Free Papers

Type de présentation	Oral presentation
Titre	LUNet: Deep Learning for the Segmentation of Arterioles and Venules in Fundus Images

But	Automatic segmentation of the retinal arterioles and the venules (A/V) is essential in building an automatic pipeline to study the patient’s microvasculature. Nevertheless, there is a lack of a data-driven algorithm for A/V segmentation that may enable analyzing the patient’s vasculature on a large scale. This research aims to develop a novel robust, i.e. with high performance and generalizable, data-driven algorithm for A/V segmentation.
Méthodes	We created a new digital fundus image (DFI) dataset, named UZF, which contains a total of 240 crowd-sourced A/V segmentations made by fifteen residents and reviewed by a senior oph- thalmologist from the Universitair Ziekenhuis (UZ) Leuven Hospital. To develop LUNet we have adapted the U-Net architecture, redefining the A/V segmentation problem as a multilabel binary segmentation and designed a custom loss function based on a weighted average of the most commonly used function for binary segmentation while taking into account the continuous tubular structure of the blood vessels. We evaluate the generalization performance of LUNet on two external datasets from the United Sates and Paraguay respectively. We compare LUNet’s performances against an open source state-of-the-art (SOTA) algorithms.
Résultats	LUNet achieved an average dice score of 84.56 on the arterioles and 82.43 on the veins on the UZF dataset and on the external test sets. This was significantly (p<0.5) better than the SOTA benchmark. Conclusion: We introduce LUNet a robust and generalizable deep learning algorithms for A/V segmentation in DFI.
Conclusion	LUNet outperforms existing algorithms to segment the retinal vasculature.
Conflit d'intérêt	Non

Auteur 1

Nom	VAN EIJGEN
Initiales	J
Institut	KU Leuven
Ville	Leuven

Auteur 2

Nom	Fhima
Initiales	J
Institut	Technion
Ville	Haifa

Auteur 3

Nom	Behar
Initiales	J
Institut	Technion
Ville	Haifa

Auteur 4

Nom	Christinaki
Initiales	E
Institut	KU Leuven
Ville	Leuven

Auteur 5

Nom	Stalmans
Initiales	I
Institut	KU Leuven
Ville	Leuven