Deep Learning Retinal Vessel Segmentation From a Single Annotated - - PowerPoint PPT Presentation

Deep Learning Retinal Vessel Segmentation From a Single Annotated Example

Praneeth Sadda1, John A. Onofrey1, Xenophon Papademetris1,2

1 Department of Radiology and Biomedical Imaging, Yale University 2 Department of Biomedical Engineering, Yale University

Semantic Segmentation

Bagci et al. 2014 Garcia-Peraza-Herrera et al. 2014 Stahl et al. 2004

FCNN

Fundamental Issue of Supervised Learning

• Data is easy to acquire
• Data is difficult to label

Many Datasets are Partially Labeled

Fully Labeled Dataset Partially Labeled Dataset

FCNN

Style Transfer

Zhu et al. 2017

Synthetic Image Generation

Unlabeled Image Labeled Synthetic Image Labeled Image

Proposed Workflow

Partially Labeled Dataset Generation of Synthetic Examples Training with Fully-Labeled Data

FCNN

CycleGAN

CycleGANs

!"#"$% &, ( = *+ ~ -(+) ( & 0 − 0 2 + *# ~ - # & ( 4 − 4 2 ( & 0 ≈ 0 & ( 4 ≈ 4

Methods

• Provide a manual ground truth segmentation for a single “template”

image.

• Divide images into patches.
• Train one CycleGAN for each unlabeled image (~10 hours per image)

using a patchwise approach.

• Transfer the style of the template image using a patchwise approach.
• Train FCNN on the style transferred images using a patchwise

approach.

Synthetic Image Generation

Phyisican Rater FCNN trained with one labeled example FCNN trained with one labeled and 19 unlabeled examples FCNN trained with 20 labeled examples

Results

Results

Training Dataset Sensitivity Specificity Accuracy 20 Labeled 0.60 ± 0.10 0.98 ± 0.01 0.94 ± 0.01 1 Labeled + 19 Unlabeled 0.62 ± 0.10 0.95 ± 0.01 0.93 ± 0.02 1 Labeled 0.86 ± 0.04 0.53 ± 0.06 0.56 ± 0.05

Conclusion

• Style transfer can be used to exploit the information in unlabeled

examples for supervised learning of semantic segmentation.

• For segmenting