Multi-task deep learning for glaucoma detection from color fundus images | JoVE Visualize

Area of Science:

Ophthalmology
Medical Imaging
Artificial Intelligence

Background:

Glaucoma diagnosis relies on expert analysis of subtle optic disc changes in retinal images, a process challenged by time constraints and the need for extensive expert annotations for deep learning models.
Current deep learning models for retinal image analysis require significant annotated data, which is costly and time-consuming to obtain from medical experts.

Purpose of the Study:

To design and train a novel multi-task deep learning model for glaucoma diagnosis by leveraging similarities between related eye-fundus tasks.
To improve the efficiency and accuracy of glaucoma diagnosis using deep learning by enabling simultaneous learning of segmentation and classification tasks.

Main Methods:

Developed a novel multi-task deep learning architecture capable of performing various segmentation and classification tasks related to glaucoma diagnosis concurrently.
Trained and evaluated the model on a diverse dataset of 1200 retinal fundus images from multiple sources, comparing its performance against single-task models and human experts.

Main Results:

The multi-task model achieved a [Formula: see text] AUC performance, outperforming a single-task model with the same backbone trained solely for glaucoma detection ([Formula: see text] AUC).
The proposed approach demonstrated superior performance compared to other multi-task learning models and matched expert-level accuracy.
The model achieved expert-level performance using significantly fewer parameters ( [Formula: see text] times fewer) than training individual tasks separately.

Conclusions:

The novel multi-task deep learning model offers an efficient and effective approach for glaucoma diagnosis from retinal fundus images.
This method reduces the need for extensive annotations and computational resources while achieving high diagnostic accuracy comparable to human experts.
Publicly available data and code facilitate reproducibility and further research in AI-driven ophthalmology.