Detection of Microaneurysms in Fundus Images Based on an Attention Mechanism

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Microaneurysms (MAs) are the earliest detectable lesions in DR, making their early detection crucial for timely intervention.
• Accurate and reliable automated MA detection in retinal fundus images is challenging due to image complexity and lesion size.

Purpose of the Study:

• To develop a novel deep neural network-based method for accurate and reliable detection of microaneurysms (MAs) in retinal fundus images.
• To improve the early diagnosis of diabetic retinopathy (DR) through enhanced MA detection.
• To leverage attention mechanisms and spatial relationships for improved detection performance.

Main Methods:

• Image preprocessing including equalization operations to enhance retinal fundus image quality.
• A deep neural network incorporating a multilayer attention mechanism for fusing relevant feature layers.
• Utilizing spatial relationships between MAs and blood vessels for secondary screening and refinement of detection results.

Main Results:

• The proposed method demonstrated effective improvement in the average accuracy and sensitivity for microaneurysm detection.
• The attention mechanism facilitated the fusion of informative feature layers for preliminary MA identification.
• Secondary screening based on spatial context refined the detection, leading to more reliable results.

Conclusions:

• The developed deep learning method with a multilayer attention mechanism offers a promising approach for automated microaneurysm detection.
• This technique has the potential to significantly aid in the early diagnosis and management of diabetic retinopathy.
• Further validation on diverse datasets is warranted to confirm generalizability.