Delving into transfer learning within U-Net for refined retinal vessel segmentation: An extensive hyperparameter...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Blood vessel segmentation is challenging due to low contrast, overlapping structures, and varying vessel thickness.
• Traditional Convolutional Neural Networks (CNNs) face limitations in accurately segmenting complex vascular networks.
• U-Net and its variants show promise for medical image segmentation tasks.

Purpose of the Study:

• To propose a novel approach for retinal vessel segmentation using transfer learning.
• To evaluate the performance of various pretrained models as encoders within the U-Net architecture.
• To compare the efficacy of different optimizers (Adam, SGD, RMSProp) for retinal vessel segmentation.

Main Methods:

• Implemented U-Net architectures with pretrained encoders: VGG16, VGG19, ResNet50, MobileNetV2, and DenseNet121.
• Trained and evaluated models on the DRIVE dataset for retinal vessel segmentation.
• Utilized Adam, Stochastic Gradient Descent (SGD), and RMSProp optimizers for model training.

Main Results:

• Models employing the Adam optimizer generally yielded superior results.
• ResNet50 U-Net demonstrated the highest specificity (0.9875).
• MobileNetV2 U-Net achieved the highest recall (0.8056), while DenseNet121 U-Net reached an accuracy of 0.9689.
• VGG16 U-Net and MobileNetV2 U-Net attained a Dice coefficient of 0.849.

Conclusions:

• Transfer learning with pretrained U-Net models offers a robust approach to retinal vessel segmentation.
• The choice of pretrained encoder and optimizer significantly impacts segmentation performance.
• ResNet50 U-Net, MobileNetV2 U-Net, and DenseNet121 U-Net show strong potential for clinical applications in retinal imaging.