Compound Scaling Encoder-Decoder (CoSED) Network for Diabetic Retinopathy Related Bio-marker Detection.

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Diabetic Retinopathy (DR) screening relies on manual detection of microvascular lesions, which is time-consuming and prone to errors.
• The increasing prevalence of diabetes necessitates efficient and accurate automated methods for DR screening.

Purpose of the Study:

• To develop a novel compound scaling encoder-decoder network (CoSED-Net) for enhanced accuracy and efficiency in microvascular lesion segmentation.
• To improve the automated detection and classification of DR biomarkers.

Main Methods:

• Proposed a lightweight encoder scaled in depth, width, and resolution for faster training.
• Incorporated an attention mechanism using Concurrent Spatial and Channel Squeeze and Channel Excitation (scSE) blocks in the decoder.
• Utilized a compound loss function with transfer learning to address imbalanced data.

Main Results:

• The proposed CoSED-Net demonstrated superior performance in microvascular lesion segmentation compared to existing methods.
• Evaluated on the large-scale DDR and FGADR datasets, confirming the method's effectiveness.
• Achieved improved accuracy and running efficiency in segmenting DR-related biomarkers.

Conclusions:

• The developed CoSED-Net offers a promising approach for automated diabetic retinopathy screening.
• The compound scaling and attention mechanisms significantly enhance segmentation performance.
• This method can aid in timely diagnosis and management of diabetic retinopathy.