SDC-Net: Enhancing Multiclass Diabetic Retinopathy Segmentation by Shallow-Deep Collaboration Networks With...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Diabetic retinopathy (DR) is a primary cause of blindness globally.
• Accurate segmentation of fundus lesions is vital for DR screening, grading, and monitoring.
• Simultaneous segmentation of diverse DR lesions presents significant challenges due to variations in shape, size, and appearance.

Purpose of the Study:

• To develop and evaluate a novel framework for simultaneous segmentation of four types of diabetic retinopathy lesions.
• To address the challenges of segmenting multiple DR lesions with varying characteristics.

Main Methods:

• A shallow-deep collaboration network (SDC-Net) with a wavelet-guided attention mechanism was proposed.
• The framework integrates shallow and deep networks for enhanced multiscale feature extraction.
• A super-resolution auxiliary task was incorporated to improve training accuracy and robustness.

Main Results:

• SDC-Net achieved average Dice scores of 60.27, 39.98, and 42.57 on the IDRiD, DDR, and FGADR datasets, respectively.
• Average intersection over union (IoU) scores were 44.53, 25.71, and 28.58 across the datasets.
• Average area under the receiver operating characteristic curve (AUC) values reached 68.75, 57.37, and 64.21.

Conclusions:

• The dual-branch design and wavelet attention module effectively capture multiscale features and enhance lesion detail extraction.
• The super-resolution task contributed to improved accuracy and robustness in DR lesion segmentation.
• SDC-Net shows potential for enhancing clinical diagnosis of DR through precise segmentation of multiple fundus lesions.