8000 GitHub - aguron/EmphysemaScoring: Train and evaluate U-Net models in a lung emphysemic region segmentation task
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

aguron/EmphysemaScoring

Repository files navigation

EmphysemaScoring

Adapted from a U-Net implementation by zhixuhao.


Overview

Emphysema

Emphysema is the most common form of chronic obstructive lung disease, which is the third leading cause of the death in the United States. Emphysema is often assessed with the RA950 score, which segments out dead lung tissue in a CT scan. In this project, we train a neural network to perform the segmentations.

Dataset

Please download the LIDC-IDRI database

Organizing Data for Training and Evaluation

Please run shell_scripts/prepare_script_lidc_mat.sh and shell_scripts/parallel_script_runs_wrapper.sh in that order. LIDC-IDRI (1010 patients) is partitioned into 10 parts (101 patients each) and each part, in turn, is partitioned into sections a (80 patients) and b (21 patients). In the training-evaluation scheme, 9 parts are used for training (section a) and validation (section b) while the 10th part (sections a and b) is used for evaluation.

  • bash shell_scripts/prepare_script_lidc_mat.sh
  • bash shell_scripts/parallel_script_runs_wrapper.sh

Data Preprocessing

shell_scripts/parallel_script_runs_wrapper.sh generates lung RA950 segmentation masks from the LIDC-IDRI CT data

U-Net Model

We modified the U-Net implementation by zhixuhao, which is based on U-Net: Convolutional Networks for Biomedical Image Segmentation, so that the training loss for different samples (e.g. foreground versus background pixels) can be weighted differently.

Usage

Dependencies

  • Tensorflow
  • Keras >= 1.0
  • Matlab
  • Python >= 2.7

Preprocessing, Training, and Evaluation

  • bash shell_scripts/prepare_script_lidc_mat.sh
  • bash shell_scripts/parallel_script_runs_wrapper.sh
  • bash shell_scripts/deep_learning_wrapper.sh

Results

Slice of Lung CT Scan

example_images/CT_slice.png

Lung CT RA950 Segmentation Mask

example_images/CT_slice_mask.png

Lung CT RA950 Predictions

The following image is representative of when the relative segmentation foreground-background sample (pixel) training loss weighting is just about right. The average dice coefficient based on images selected for evaluation was 0.946.

example_images/CT_slice_predict.png

The following image is representative of when the relative segmentation foreground-background sample (pixel) training loss weighting could be increased. The average dice coefficient based on images selected for evaluation was 0.378.

example_images/CT_slice_predict_f_w_l.png

The following image is representative of when the relative segmentation foreground-background sample (pixel) training loss weighting could be decreased. The average dice coefficient based on images selected for evaluation was 0.121.

example_images/CT_slice_predict_f_w_m.png

About

Train and evaluate U-Net models in a lung emphysemic region segmentation task

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published
0