A Novel Weakly Supervised Multitask Architecture for Retinal Lesions Segmentation on Fundus Images

Area of Science:

• Medical image analysis
• Computer vision
• Ophthalmology

Background:

• Automated diagnosis of retinopathy requires accurate retinal lesion segmentation.
• Limited pixel-level ground truth data hinders deep learning model generalization.
• Existing methods struggle with the variability of retinal lesion appearances.

Purpose of the Study:

• To develop a novel deep learning architecture for precise retinal lesion segmentation.
• To improve the interpretability and generalization of automated retinopathy diagnosis systems.
• To address the challenge of limited ground truth data using a hybrid learning approach.

Main Methods:

• A convolutional multi-task architecture trained with supervised and weakly supervised learning.
• Simultaneous training for segmentation of red and bright lesions, alongside lesion detection.
• A new preprocessing method ensuring color consistency in retinal images.

Main Results:

• The system accurately segments both red and bright retinal lesions.
• Validation across four databases using pixel-level and per-image metrics.
• Achieved an Area Under the ROC Curve of 0.839 for lesion screening on the Messidor dataset, comparable to state-of-the-art.

Conclusions:

• The proposed hybrid learning approach effectively segments retinal lesions despite data limitations.
• The system demonstrates strong performance in automated screening for retinopathy.
• This method advances the development of interpretable and generalizable automated diagnostic tools for eye diseases.