Dilated residual FPN-based segmentation for mouse retinal images

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Diabetes can cause irreversible blindness through diabetic retinopathy (DR).
• Early diagnosis of DR is crucial for effective treatment.
• Analysis of mouse retinal images aids in understanding DR progression.

Purpose of the Study:

• To develop an effective method for segmenting mouse retinal images.
• To improve the early diagnosis of diabetic retinopathy (DR).
• To enhance the segmentation of cells and layers in retinal images.

Main Methods:

• Designed a dilated residual network based on a feature pyramid network (FPN).
• Utilized dilated convolution and residual blocks for enhanced feature extraction.
• Integrated a squeeze-and-excitation (SE) attention module for detailed object recognition.
• Employed transposed convolution for improved upsampling in the decoding pathway.

Main Results:

• Validated the model on ganglion cell and mouse retinal cell/layer segmentation tasks.
• Achieved superior precision in both binary and multiclass semantic segmentation compared to existing deep learning methods.
• Demonstrated the robustness of the dilated residual FPN for retinal image analysis.

Conclusions:

• The dilated residual FPN is a highly effective method for mouse retinal image segmentation.
• This approach can significantly assist in the early diagnosis of diabetic retinopathy (DR).
• Accurate segmentation contributes to preventing vision loss associated with DR.