Skip to content

RelatedWorks.MD

Latest commit

 

History

History
56 lines (48 loc) · 2.96 KB

RelatedWorks.MD

File metadata and controls

56 lines (48 loc) · 2.96 KB

Related Works

Inspired by the following papers:

  1. Full-Resolution Network and Dual-Threshold Iteration for Retinal Vessel and Coronary Angiograph Segmentation - Liu et al. (IEEE 2022)
  2. RV-GAN: Segmenting Retinal Vascular Structure in Fundus Photographs using a Novel Multi-scale Generative Adversarial Network - Kamran et al. (MICCAI 2021)

Works of each Paper

  1. Liu et al. (IEEE 2022)

    • Contributions
      • Full-resolution UNet (FR-UNet) for vessel segmentation
      • Embed feature aggregation module before each convolution block to aggregate feature maps from up-sampling and down-sampling
      • Dual-threshold iterative algorithm to improve the connectivity of vessels that gradually extracts weak vessel pixels from the probability map
      • Introduces a quantitative evaluation method for vessel connectivity by calculating the number of connected components.
    • Parameters
      • Adam Optimizer; Weight Decay of 1e-5; 40 Epochs
      • Initial learning rate of 1e-4 which is gradually reduced by Cosine Annealing algorithm.
    • 48 × 48 sliding window with a stride of 6 is used to extract patches from vessel images to increase the quantity of training data.
      • Whole images are used for testing data
    • Data Augmentation: Random Horizontal flipping, Vertical flipping, and [90,180,270] degree rotation

  2. Kamran et al. (MICCAI 2021)

    • Contributions
      • 2 generators and 2 multi-scale autoencoding discriminators for better microvessel localization and segmentation
      • Introduced a new weighted feature matching loss to avoid fidelity suffered by traditional GAN-based segmentation systems
    • Parameters
      • Adam Optimizer; Batch-size of 24; 100 Epochs
      • Learning rate of 2e-4
      • First moment estimate (mean): β1 = 0.5
      • Second moment estimate (uncentered variance): β2 = 0.999
    • 128 x 128 sliding window with stride of 32 was used to extract patches for training and validation data
      • 4,200 training/validation images in total
      • Overlapping image patches with stride of 3 were extracted and averaged from the 20 test images.

Experiment Details

Performance Metrics

  • AUC (Area under Curve)
  • Accuracy
  • Sensitivity
  • Specificity
  • F1 Score
  • IoU (Intersection of Union)

Results on DRIVE (Digital Retinal Images for Vessel Extraction) dataset

Papers Parameters (M) Accuracy Sensitivity Specificity AUC F1 Score IoU
U-Net 7.76 0.9678 0.8057 0.9833 0.9825 0.8141 0.6864
Paper 1 5.72 0.9705 0.8356 0.9837 0.9889 0.8316 0.7120
Paper 2 14.81 0.9790 0.7927 0.9969 0.9887 0.8690 0.9762