OCT Segmentation Guide

Zhongyang Sun szy019@gmail.com

Environment

Hardware

CPU AMD Threadripper 1900X

RAM 32GB（8GB * 4）

GPU NVIDIA GeForce 1080TI & NVIDIA GeForce 1060

System

Linux Ubuntu 18.04

Python 3.6.6

Caffe 1.0.0

Cuda 9.0.176

opencv 3.4.1

cuDNN 7.0.5

NVIDIA Driver 390.116

Directory Structure

fcn
 |
data - oct - Labels - *.mat
 |              |
 |           originalImages - *.jpg
 |              |
 |           test.txt
 |              |
 |           trainval.txt
 | 
FCN_16s - models - OCT_Segmentation.caffemodel
            |         |
            |      siftflow-fcn16s-heavy.caffemodel
            |       
          results - *.jpg
            |
          infer.py
            |
          infer_collection.py
            |
          oct_layers.py
            |
          score.py
            |
          solve.py
            |
          surgery.py
            |
          solver.prototxt
            |
          train.prototxt
            |
          test.prototxt

Files

data/oct
Labels	Store label of each pixel in mat format for all images in training set
originalImages	Store data of training and test images in jpg format
test.txt	Includes all image names of images from test set
train.txt	Includes all image names of images from training set

Fcn_16s
models	Store trained models
results	Store the Segmentated OCT image in jpg format
infer.py	Used to segmentation single OCT image to estimate the performance of model
Infer_collection.py	Used to segmentation all images in tests, and store the segmentated OCT images in results folder
oct_layers.py	Used to replace caffe's default input layer
solve.py	Including the specific procedure of the training process, such as source model of migration training, or the number of training iterations, etc.
solver.prototxt	Includes mainly parameters in the training procedure, such as training rate, batch size, etc.
train.prototxt	Define network structure for training
test.prototxt	Define network structure for testing

Instructions

Training

1.Copy images in training set to originalImages folder

2.Generate label files for every image in training set, and store in Labels folder

3.Write name of images in trainset set in test.txt

4.Modify slove.py, solver.protxt, identify training processes and parameters

5.Execute python3 solve.py ，start training and store trained model in moldels folder

Testing

1.Copy images in test set to originalImage folder

2.Write name of images in test set in train.txt

3.Execute python3 infer_collection.py ，segmentation images in test set, and store the segmentation result in results folder

THOCT1800 Dataset

This dataset consists of 1800 preprocessed retinal SD-OCT B-scans (600 AMD, 600 DME, and 600 NOR), all images are intended for use in research and education situations, and every use of this dataset should include citation in their corresponding papers.