shenshuming / artificial-contrast

artificial-contrast

Deep learning for cerebral angiography segmentation from non-contrast computed tomography

Repository

Requirements: python 3.6+

Modules

1. ./src/ contains artificial-contrast library used later in scripts
2. ./scripts/ contains python scripts needed to reproduce results of our study
3. ./data/ contains sample random data required to run the code.
4. ./examples/ contains example scripts with all environmental variables set

Installation

1. create new virtual environment
2. install library using command: $ pip install ./src

Data structure

For each step we require one directory train - for cross validation and test for final test of the best model chosen in cross validation process. Each patient should have BC directory with non-contrast examinations and label with corresponding masks.

Steps to reproduce

1. Generate experiments: $ ./examples/generate_experiments.sh. Results can be found in ./results/config/
2. Run CV (results are always stored by patient and by fold):
   1. $ ./examples/cv_simple_window.sh Preprocessing approach: simple, Configuration: Radiodensity range from -100 HU to 300 HU on all 3 channels. Results in:
      • ./results/simple_window_-100_300_result.csv
      • ./results/simple_window_-100_300_by_patient_result.csv
   2. $ ./examples/cv_simple_multiple_windows.sh Preprocessing approach: simple, Configuration: Radiodensity range: (-40, 120), (-100, 300), (300, 2000) HU. Results in:
      • ./results/simple_multiple_windows_result.csv
      • ./results/simple_multiple_windows_by_patient_result.csv
   3. $ ./examples/cv_uniform_no_limit.sh Preprocessing approach: Uniform, Configuration: Without clipping radiodensity range. Results in:
      • ./results/freqs_no_limit_window_result.csv
      • ./results/freqs_no_limit_window_by_patient_result.csv
   4. $ ./examples/cv_uniform_window.sh Preprocessing approach: Uniform, Configuration: Radiodensity range from-100 HU to 300 HU
      • ./results/freqs_window_-100_300_result.csv
      • ./results/freqs_window_-100_300_by_patient_result.csv
3. Train model chosen using CV using whole training set and evaluate its performance using test set. (In our case the best model was: Preprocessing approach: simple, Configuration: Radiodensity range: (-40, 120), (-100, 300), (300, 2000) HU.): $ ./examples/test.sh.
   NOTE we used: export DCM_CONF="{\"windows\": [[-40, 120], [-100, 300], [300, 2000]], \"norm_stats\": [[0.191989466547966, 0.1603623628616333, 0.02605995163321495], [0.3100860118865967, 0.2717258334159851, 0.1233396977186203]]}" (this is one of parameters created in first step).
   Final result can be found in: ./results/simple_multiple_windows_testset_result.csv

Final model weights

Model weights are available here