Adapted from a U-Net implementation by zhixuhao.
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.
Please download the LIDC-IDRI database
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
shell_scripts/parallel_script_runs_wrapper.sh generates lung RA950 segmentation masks from the LIDC-IDRI CT data
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.
- Tensorflow
- Keras >= 1.0
- Matlab
- Python >= 2.7
- bash shell_scripts/prepare_script_lidc_mat.sh
- bash shell_scripts/parallel_script_runs_wrapper.sh
- bash shell_scripts/deep_learning_wrapper.sh
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.
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.
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.