DDFU-Net: A Deep Decoder-Focused U-Net Model for Retinal Lesion Segmentation.

Area of Science:

• Ophthalmology
• Medical Image Analysis
• Deep Learning

Background:

• Early detection of retinal lesions is crucial for preventing vision loss.
• Common lesions include soft exudates, hard exudates, microaneurysms, and hemorrhages.
• Accurate segmentation of these lesions is challenging due to variations in size, contrast, and inter-class similarity.

Purpose of the Study:

• To develop an automated method for accurate multi-lesion segmentation in fundus images.
• To introduce the Deep Decoder-Focused U-Net (DDFU-Net) model for enhanced retinal lesion detection.
• To evaluate the effectiveness of multi-task learning for simultaneous segmentation of multiple lesion types.

Main Methods:

• Developed DDFU-Net, an asymmetric dense U-Net architecture with more decoder layers for refined boundary reconstruction.
• Employed multi-task learning to simultaneously segment four types of retinal lesions.
• Conducted experiments on the IDRiD and DDR datasets to validate the model's performance.

Main Results:

• DDFU-Net demonstrated superior performance compared to state-of-the-art methods on both IDRiD and DDR datasets.
• Achieved high performance metrics, including mean Area Under the Precision-Recall Curve, mean Intersection Over Union, and mean Dice scores.
• The asymmetric design effectively captured detailed features and improved segmentation accuracy for complex structures.

Conclusions:

• The proposed DDFU-Net model offers an accurate and automated solution for multi-lesion segmentation in fundus images.
• This approach can significantly aid in the early diagnosis of eye diseases, reduce specialists' workload, and improve patient care.
• The asymmetric U-Net architecture with an enhanced decoder proves effective for fine-grained feature preservation in medical image segmentation.