An Intelligent Segmentation and Diagnosis Method for Diabetic Retinopathy Based on Improved U-NET Network

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Deep networks often struggle with generalization due to insufficient training samples.
• Accurate medical image segmentation is crucial for diagnosis and treatment planning.
• Existing U-net models may lose critical feature information during pooling operations.

Purpose of the Study:

• To develop an improved U-net based algorithm for automatic medical image segmentation and diagnosis.
• To enhance the generalization performance of deep networks in medical imaging tasks.
• To address the limitations of insufficient sample sizes in deep learning models.

Main Methods:

• An improved U-net architecture was proposed, replacing max-pooling with convolution to preserve feature information.
• 128x128 image regions were extracted as data samples from patient slices.
• Data augmentation techniques were applied to the training sample set.
• The model was trained using all available training samples.

Main Results:

• The proposed U-net model demonstrated superior performance compared to Fully Convolutional Network (FCN) and standard U-net.
• Dice Similarity Coefficient (DSC) and Coefficient of쾰 (CR) values were highest for the proposed method.
• While the proposed method showed a slight increase in Average Symmetric Surface Distance (0.004) and a decrease in PM coefficient (2.55%), overall segmentation and diagnosis results were significantly improved.

Conclusions:

• The improved U-net model effectively enhances deep network generalization for medical image segmentation with limited data.
• Replacing max-pooling with convolution preserves vital feature information, leading to better diagnostic accuracy.
• The proposed algorithm offers a promising solution for automated segmentation and diagnosis in medical imaging.