iVesseger Framework

This repository contains the code for the paper "Deep Leaning-Based Interactive Segmentation of Three-Dimensional Blood Vessel Images". As a deep leaning-based interactive vascular segmentation framework, the iVesseger employs mouse-click operations to select regions in 3D space, guiding neural networks to correct any erroneous segmentation results. iVesseger comprises four steps: vessel pre-segmentation, interaction point selection, vascular image enhancement, and fine iterative segmentation.

Procedure

The specific steps for using iVesseger are as follows:

step 1. Make sure the Pytorch, PyQt, and VTK are included in the coding environment，the package version is:

PyQt5 - 5.15.5
PyTorch - 1.12.1
VTK - 9.2.6

step 2. Implementation Platform & External Tools:

PyCharm - 2022
pyUIC & QT Desinger

step 3. Run the pyqt_vtk.py to start the program.Four parameters need to be set before running:

Two parameters are set in thick_detect.py:
  # Pre-segmentation with 3D U-Net by default
  net = UNet3D(1, 1, 64, layer_order='cbr')
  # Import model parameters (We provide the training parameters of U-Net, called "U_Net.pth")
  model_path = 'logs/thick_seg/U_Net.pth'

Two parameters are set in refine_detect.py:
  # Fine-segmentation with HCS-Net by default
  net = HCS_Net(2, 1, image_size)
  # Import model parameters (We provide the training parameters of HCS-Net, called "HCS_Net.pth")
  model_path = 'logs/refine_seg/HCS_Net.pth'

step 4. Click on the Load image button in the left interface to import data. We provide a data block for testing(named "test_image.tif").

step 5. After data importing, use the Max slider and Min slider to adjust the brightness of the original image. The initial values of the Min-slider and Max-slider default to the maximum and minimum gray values in the original image.

step 6. Click on the PreSeg button in the left interface to generate the pre-segmentation result.

step 7. Interaction point selection process:

In single-ray mode:
  Left-click to adjust the viewing angle
  double left-click to generate interaction points
  right-click to eliminate erroneous interaction points

In cross-ray mode:
  Left-click to adjust the viewing angle
  double left-click to generate interaction rays
  right-click to eliminate erroneous interaction rays

step 8. After point selection, click on the Seg button in the left interface to generate the fine segmentation result.

step 9. Repeat the point selection and fine segmentation process to optimize the segmentation result.

step 10. Click on the Save Label button in the left interface to save the segmentation result.

step 11. The checkbox at the bottom is used to select single-ray or cross-ray mode.

NOTE The training parameters should be downloaded and replace separately to ensure data consistency. The .pth file called "HCS_Net.pth" is stored in the path "/framework/logs/refine_seg"， and the .pth file called "U_Net.pth" is stored in the path "/framework/logs/thick_seg".