Fundus Unimodal and Late Fusion Multimodal Diabetic Retinopathy Grading

Sara El-Ateif, Ali Idri, Ali Idri

2023

Abstract

Diabetic Retinopathy (DR) is an eye disease with complications, if left untreated grow, split into four grades: mild, moderate, severe, and proliferative. We propose to (1) compare and evaluate three different recently used deep learning models: EfficientNet-B5, Swin Transformer, and Hybrid-EfficientNetB0-SwinTF (HES) on the APTOS 2019 dataset’s fundus and early fused (EF) weighted gaussian blur fundus. (2) Evaluate three fine-tuning and pre-processing schemes on the best model. And (3) choose the best model-scheme per modality and perform late fusion on them to get the final DR grade. Results show that our best method, late fusion HES model, results in F1-socre of 81.21%, accuracy of 81.83%, and AUC of 96.30%. We propose using late fusion HES model in population-wide diagnosis to assist doctors in Morocco to reduce DR burden.

Paper Citation

in Harvard Style

El-Ateif S. and Idri A. (2023). Fundus Unimodal and Late Fusion Multimodal Diabetic Retinopathy Grading. In Proceedings of the 12th International Conference on Data Science, Technology and Applications - Volume 1: DATA; ISBN 978-989-758-664-4, SciTePress, pages 219-226. DOI: 10.5220/0012048700003541

in Bibtex Style

@conference{data23,
author={Sara El-Ateif and Ali Idri},
title={Fundus Unimodal and Late Fusion Multimodal Diabetic Retinopathy Grading},
booktitle={Proceedings of the 12th International Conference on Data Science, Technology and Applications - Volume 1: DATA},
year={2023},
pages={219-226},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012048700003541},
isbn={978-989-758-664-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 12th International Conference on Data Science, Technology and Applications - Volume 1: DATA
TI - Fundus Unimodal and Late Fusion Multimodal Diabetic Retinopathy Grading
SN - 978-989-758-664-4
AU - El-Ateif S.
AU - Idri A.
PY - 2023
SP - 219
EP - 226
DO - 10.5220/0012048700003541
PB - SciTePress