Microcirculation

Segmentation of human eye capillaries based on ophthalmic slit lamp images using UNet++

• Anatoly Medvedev
• ID: 1603212269

Table of Contents

1. Environment
2. Training
3. Usage
4. Reference

Environment

Training

Dataset Structure:

data/
    |
    train_dataset/
    |            |
    |            1.png
    |            1.geojson
    |            ...
    eye_test/
            |
            784.png
            ...

Model was trained in parallel on 2 GPUs Tesla V100 32GB. To train model, change the training flag in training.py to True and run it in the background:

$ nohup python training.py > log.txt &

Training results:

Model	Backbone	F1 Score
UNet++	ResNet-50	0.512379
	ResNet-101	0.525109

Usage

Follow steps in demo.ipynb to learn more about the model, image preparation, and model validation.

Reference

• Zhou, Z. et al. (2018) "UNet++: A Nested U-Net Architecture for Medical Image Segmentation". arXiv. doi: 10.48550/ARXIV.1807.10165