NAGIOS: RODERIC FUNCIONANDO

Deep Learning Architectures for Diagnosis of Diabetic Retinopathy

Repositori DSpace/Manakin

Valencià Castellano

Inici
Investigació
e-prints
22 - Física
Visualitza element

IMPORTANT: Aquest repositori està en una versió antiga des del 3/12/2023. La nova instal.lació está en https://roderic.uv.es/

Deep Learning Architectures for Diagnosis of Diabetic Retinopathy

Mostra el registre parcial de l'element

dc.contributor.author	Martínez Sober, Marcelino
dc.contributor.author	Solano, Alberto
dc.contributor.author	Dietrich, Kevin N.
dc.contributor.author	Barranquero Cardeñosa, Regino
dc.contributor.author	Vila Tomás, Jorge
dc.contributor.author	Hernández Cámara, Pablo
dc.date.accessioned	2023-05-23T07:30:02Z
dc.date.available	2023-05-23T07:30:02Z
dc.date.issued	2023
dc.identifier.citation	Martínez Sober, Marcelino Solano, Alberto Dietrich, Kevin N. Barranquero Cardeñosa, Regino Vila Tomás, Jorge Hernández Cámara, Pablo 2023 Deep Learning Architectures for Diagnosis of Diabetic Retinopathy Applied Sciences 13 7 1 14
dc.identifier.uri	https://hdl.handle.net/10550/86831
dc.description.abstract	For many years, convolutional neural networks dominated the field of computer vision, not least in the medical field, where problems such as image segmentation were addressed by such networks as the U-Net. The arrival of self-attention-based networks to the field of computer vision through ViTs seems to have changed the trend of using standard convolutions. Throughout this work, we apply different architectures such as U-Net, ViTs and ConvMixer, to compare their performance on a medical semantic segmentation problem. All the models have been trained from scratch on the DRIVE dataset and evaluated on their private counterparts to assess which of the models performed better in the segmentation problem. Our major contribution is showing that the best-performing model (ConvMixer) is the one that shares the approach from the ViT (processing images as patches) while maintaining the foundational blocks (convolutions) from the U-Net. This mixture does not only produce better results (DICE=0.83) than both ViTs (0.80/0.077 for UNETR/SWIN-Unet) and the U-Net (0.82) on their own but reduces considerably the number of parameters (2.97M against 104M/27M and 31M, respectively), showing that there is no need to systematically use large models for solving image problems where smaller architectures with the optimal pieces can get better results.
dc.relation.ispartof	Applied Sciences, 2023, vol. 13, num. 7, p. 1-14
dc.subject	Física
dc.subject	Enginyeria biomèdica
dc.subject	Enginyeria
dc.title	Deep Learning Architectures for Diagnosis of Diabetic Retinopathy
dc.type	journal article
dc.date.updated	2023-05-23T07:30:03Z
dc.identifier.doi	10.3390/app13074445
dc.identifier.idgrec	158967
dc.rights.accessRights	open access