Weakly Supervised Lesion Detection From Fundus Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer-Aided Detection

Background:

• Early diagnosis and monitoring of eye diseases are crucial.
• Computer-aided detection of retinal lesions has advanced but struggles with diverse lesion types.
• Automatic detection of unknown retinal lesion types remains challenging.

Purpose of the Study:

• To propose a weakly supervised method for automatic detection of diverse retinal lesions.
• To develop a model that can distinguish lesions from normal structures without specific annotations.

Main Methods:

• A weakly supervised approach using normal and abnormal retinal images.
• Fundus images modeled as background, blood vessels, and noise (lesions).
• Background modeled as low-rank structure; noise modeled using Gaussian distributions.

Main Results:

• The proposed method accurately detects retinal lesions.
• It outperforms previous related methods in lesion detection.
• The model effectively distinguishes lesions from background noise.

Conclusions:

• The weakly supervised method offers a promising solution for detecting diverse retinal lesions.
• This technique can aid in early diagnosis and continuous monitoring of eye diseases.
• The approach advances computer-aided detection in retinal imaging.