eduzhai > Applied Sciences > Engineering >

The Little W-Net That Could State-of-the-Art Retinal Vessel Segmentation with Minimalistic Models

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 13 pages

Abstract: The segmentation of the retinal vasculature from eye fundus images representsone of the most fundamental tasks in retinal image analysis. Over recent years,increasingly complex approaches based on sophisticated Convolutional NeuralNetwork architectures have been slowly pushing performance on well-establishedbenchmark datasets. In this paper, we take a step back and analyze the realneed of such complexity. Specifically, we demonstrate that a minimalisticversion of a standard U-Net with several orders of magnitude less parameters,carefully trained and rigorously evaluated, closely approximates theperformance of current best techniques. In addition, we propose a simpleextension, dubbed W-Net, which reaches outstanding performance on severalpopular datasets, still using orders of magnitude less learnable weights thanany previously published approach. Furthermore, we provide the mostcomprehensive cross-dataset performance analysis to date, involving up to 10different databases. Our analysis demonstrates that the retinal vesselsegmentation problem is far from solved when considering test images thatdiffer substantially from the training data, and that this task represents anideal scenario for the exploration of domain adaptation techniques. In thiscontext, we experiment with a simple self-labeling strategy that allows us tomoderately enhance cross-dataset performance, indicating that there is stillmuch room for improvement in this area. Finally, we also test our approach onthe Artery Vein segmentation problem, where we again achieve resultswell-aligned with the state-of-the-art, at a fraction of the model complexityin recent literature. All the code to reproduce the results in this paper isreleased.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...