SSMD-UNet: semi-supervised multi-task decoders network for diabetic retinopathy segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss, necessitating early detection through retinal screening.
• Automated deep learning models for DR segmentation from retinal fundus images are crucial for timely diagnosis.
• Current models face limitations due to insufficient annotated training data.

Purpose of the Study:

• To develop an accurate semi-supervised deep learning model for diabetic retinopathy segmentation.
• To leverage large unlabelled datasets to overcome data scarcity issues in DR segmentation.
• To improve the generalisation and robustness of DR segmentation models.

Main Methods:

• A semi-supervised multitask learning approach utilizing a novel multi-decoder architecture.
• Incorporation of both unsupervised and supervised learning phases.
• Training an unsupervised auxiliary task to enhance the primary DR segmentation task using unlabelled data.

Main Results:

• The proposed model significantly improves DR segmentation performance.
• Outperforms existing state-of-the-art techniques on public datasets (FGADR and IDRiD).
• Demonstrates enhanced generalisation and robustness in cross-dataset evaluations.

Conclusions:

• Semi-supervised learning effectively utilizes unlabelled data to boost DR segmentation accuracy.
• The proposed multi-decoder architecture offers a robust solution for DR screening.
• This approach aids ophthalmologists in earlier and more accurate DR diagnosis, potentially preventing vision loss.