CNNs Enable Accurate and Fast Segmentation of Drusen in Optical Coherence Tomography

Shekoufeh Gorgi Zadeh, Maximilian Wintergerst, Vitalis Wiens, Sarah Thiele, Frank Holz, Robert Finger und Thomas Schultz

In proceedings of Deep Learning in Medical Image Analysis (DLMIA), Springer, 2017

Abstract

Optical coherence tomography (OCT) is used to diagnose and track progression of age-related macular degeneration (AMD). Drusen, which appear as bumps between Bruch's membrane (BM) and the retinal pigment epithelium (RPE) layer, are among the most important biomarkers for staging AMD. In this work, we develop and compare three automated methods for Drusen segmentation based on the U-Net convolutional neural network architecture. By cross-validating on more than 50,000 annotated images, we demonstrate that all three approaches achieve much better accuracy than a current state-of-the-art method. Highest accuracy is achieved when the CNN is trained to segment the BM and RPE, and the drusen are detected by combining shortest path finding with polynomial fitting in a post-process.

Bilder

Paper herunterladen

Bibtex

@INPROCEEDINGS{GorgiZadeh:DLMIA17,
     author = {Gorgi Zadeh, Shekoufeh and Wintergerst, Maximilian and Wiens, Vitalis and Thiele, Sarah and Holz,
               Frank and Finger, Robert and Schultz, Thomas},
      title = {CNNs Enable Accurate and Fast Segmentation of Drusen in Optical Coherence Tomography},
  booktitle = {Deep Learning in Medical Image Analysis (DLMIA)},
     series = {LNCS},
       year = {2017},
  publisher = {Springer},
   abstract = {Optical coherence tomography (OCT) is used to diagnose and track progression of age-related macular
               degeneration (AMD). Drusen, which appear as bumps between Bruch's membrane (BM) and the retinal
               pigment epithelium (RPE) layer, are among the most important biomarkers for staging AMD. In this
               work, we develop and compare three automated methods for Drusen segmentation based on the U-Net
               convolutional neural network architecture. By cross-validating on more than 50,000 annotated images,
               we demonstrate that all three approaches achieve much better accuracy than a current
               state-of-the-art method. Highest accuracy is achieved when the CNN is trained to segment the BM and
               RPE, and the drusen are detected by combining shortest path finding with polynomial fitting in a
               post-process.}
}