ForensicSexEstimation

A deep learning framework for forensic sex estimation from 3D CT

Environment setup

Setup conda environment Python3.9

Step 1:

git clone https://github.com/aehrc/ForensicSexEstimation.git
cd ForensicSexEstimation
pip install -r requirements.txt

Step 2: Manually install TotalSegmentator. please use the v2.0.5 release:

https://github.com/wasserth/TotalSegmentator/releases/tag/v2.0.5

Please place your raw dicom CT data in the folder ./Cranial CT data/CT dicom data

Data preprocessing

Converting dcm file into nifti, isotropic resampling to 1X1X1mm, skull segmentation via TotalSegmentator, skull region extraction

cd ./run
python3.9 run_preprocessing.py

• The preprocessed images will be save into ./Cranial CT data/Cranial CT nifti isotropic crop, and the preprocessed skull masks will be saved into ./Cranial CT data/Cranial CT isotropic segmentations crop

• The training/validation(5-fold cross-validation) and testing partition need to be specified in the ./data_partition/Case_partition.xlsx file.

Training

python3.9 train.py --mdl_type [model_type] --input_mode [input_mode]

The model_type and input_mode variables can be selected as shown in the table below. After training, the models can be found under folder ForensicSexEstimation.

Model Type	Model Type Explanation	Input Mode	Input Mode Explanation
ResNet_two_output	Model generates Walker scores and forensic sex sequentially.	image	Use image as input.
		image_skull	Use both image and the skull mask as input in 2 channels.
		skull	Use the intersection of image and skull mask (image ∩ skull mask) as input.
ResNet_auxiliary	Model architecture with auxiliary tasks to estimate both Walker scores and forensic sex.	image
		image_skull
		skull
ResNet_single_output	Model only generates forensic sex estimation.	image
		image_skull
		skull

• For information on the Walker scores, please refer to:

@article{walker2008sexing, 
  title={Sexing skulls using discriminant function analysis of visually assessed traits}, 
  author={Walker, Phillip L}, 
  journal={American Journal of Physical Anthropology: The Official Publication of the American Association of Physical Anthropologists}, 
  volume={136}, 
  number={1}, 
  pages={39--50}, 
  year={2008}, 
  publisher={Wiley Online Library} 
}

Evaluate

After training completed and before testing, the probability thresholds need to be determinded on the validation set.

cd ./run
python3.9 evaluate.py --mdl_type [model_type] --input_mode [input_mode]

Testing

• Generating test outcomes

cd ./run
python3.9 test.py --mdl_type [model_type] --input_mode [input_mode]

• Calculate testing performance

python3.9 calculate_test_performance.py --mdl_type [model_type] --input_mode [input_mode]

The sex estimation results for the testing data can be found under the 'test' folder in the model folders.

Weights

Trained weights (ResNet_model_weights.zip ) are available for image ∩ skull mask as input at: https://github.com/aehrc/ForensicSexEstimation/releases/tag/Weights.

Please place the model folders under folder ForensicSexEstimation.

Testing on sample data

Two sample images can be found in /Cranial CT data/sample_data, which have been preprocessed and only include the skull region. The forensic sex for these two cases are case 1: F(female) and case 2: M(male).

To test sex estimation prediction on these two sample images, please use:

cd ./run
python3.9 Test_sample_data.py --mdl_type [model_type]