A PyTorch implementation of Hybrid-FPN-AACNet for granular Abdominal Aortic Calcification (AAC) detection in lateral view DXA images.
Custom dataset was used for the training of this model which has not been shared due to data ethics limitations. For training, testing, and prediction using Hybrid-FPN-AACNet, make sure the directory like this:
|-- data
|-- Manitoba
|-- DE
|-- mat
SBH-01022013_144654.mat
SBH-01022016_105551.mat
...
|-- npy_cropped
SBH-01022013_144654.npy
SBH-01022016_105551.npy
...
|-- SE
|-- mat
SBH-01022013_144654.mat
SBH-01022016_105551.mat
...
|-- npy_cropped
SBH-01022013_144654.npy
SBH-01022016_105551.npy
...
|-- US_Model_Combined
|-- SE
|-- dcm
PNC-1.dcm
MODEL CK1_CK_VFA_BL_19FEB2020.dcm
...
|-- npy_cropped
PNC-1.npy
MODEL CK1_CK_VFA_BL_19FEB2020.npy
...
For data conversion from .dcm format to .npy format, navigate to ./utilities and use the following command:
python dcm_to_npy.py --src_folder_path ../data/US_Model_Combined/SE/dcm --dst_folder_path ../data/US_Model_Combined/SE/npy_cropped --perc_crop_from_top 0.50 --normalization_max_value 4096 --normalization_min_value 0
The images from US and Model datasets are cropped 50% from top and are normalized with '0' and '4096' as minimum and maximum values.
For data conversion from .mat format to .npy format, navigate to ./utilities and use the following commands:
python mat_to_npy.py --src_folder_path ../data/Manitoba/SE/mat --dst_folder_path ../data/Manitoba/SE/npy_cropped --perc_crop_from_top 0.50 --normalization_max_value 32767 --normalization_min_value -32767 --modality SE
python mat_to_npy.py --src_folder_path ../data/Manitoba/DE/mat --dst_folder_path ../data/Manitoba/DE/npy_cropped --perc_crop_from_top 0.50 --normalization_max_value 32767 --normalization_min_value -32767 --modality DE
The images from Manitoba datasets are cropped 50% from top and are normalized with '32767' and '-32767' as minimum and maximum values.
For Manitoba SE dataset training using 10 fold stratified cross validation approach, run the following commands sequentially:
python Step_1_Pretraining_Classification_Manitoba_SE.py
python Step_2_Training_Regression_Manitoba_SE.py
For Manitoba DE dataset training using 10 fold stratified cross validation approach, run the following commands sequentially:
python Step_1_Pretraining_Classification_Manitoba_DE.py
python Step_2_Training_Regression_Manitoba_DE.py
For US-Model combined SE dataset training using 10 fold stratified cross validation approach, run the following commands sequentially:
python Step_1_Pretraining_Classification_us_model_combined_SE.py
python Step_2_Training_Regression_us_model_combined_SE.py
The trained models and the corresponding results for each fold would be saved in the following folders:
|-- outputs
|-- Step_1_DE_Manitoba
|-- Fold_0_Step_1_DE_Manitoba.csv
|-- Fold_1_Step_1_DE_Manitoba.csv
...
|-- Fold_0_Step_1_DE_Manitoba_MSE_xxxacc_0th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
|-- Fold_1_Step_1_DE_Manitoba_MSE_xxxacc_1th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
...
|-- Step_2_DE_Manitoba
|-- Fold_0_Step_2_DE_Manitoba.csv
|-- Fold_1_Step_2_DE_Manitoba.csv
...
|-- Fold_0_Step_2_DE_Manitoba_MSE_xxxacc_0th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
|-- Fold_1_Step_2_DE_Manitoba_MSE_xxxacc_1th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
...
|-- Step_1_SE_Manitoba
|-- Fold_0_Step_1_SE_Manitoba.csv
|-- Fold_1_Step_1_SE_Manitoba.csv
...
|-- Fold_0_Step_1_SE_Manitoba_MSE_xxxacc_0th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
|-- Fold_1_Step_1_SE_Manitoba_MSE_xxxacc_1th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
...
|-- Step_2_SE_Manitoba
|-- Fold_0_Step_1_SE_US_Model_Combined.csv
|-- Fold_1_Step_1_SE_US_Model_Combined.csv
...
|-- Fold_0_Step_1_SE_US_Model_Combined_MSE_xxxacc_0th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
|-- Fold_1_Step_1_SE_US_Model_Combined_MSE_xxxacc_1th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
...
|-- Step_1_SE_US_Model_Combined
|-- Fold_0_Step_2_SE_US_Model_Combined.csv
|-- Fold_1_Step_2_SE_US_Model_Combined.csv
...
|-- Fold_0_Step_1_SE_US_Model_Combined_MSE_xxxacc_0th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
|-- Fold_1_Step_1_SE_US_Model_Combined_MSE_xxxacc_1th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
...
|-- Step_2_SE_US_Model_Combined
|-- Fold_0_Step_1_SE_US_Model_Combined.csv
|-- Fold_1_Step_1_SE_US_Model_Combined.csv
...
|-- Fold_0_Step_2_SE_US_Model_Combined_MSE_xxxacc_0th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
|-- Fold_1_Step_2_SE_US_Model_Combined_MSE_xxxacc_1th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
...
Copy the corresponding file id_to_labels.csv to the output folder. For example, in case of Manitoba DE, copy the mentioned files in ./outputs/Step_2_DE_Manitoba folder
and then copy the file best_results_sorting_per_fold.py in same folder. It would look something like the following:
|-- outputs
|-- Step_2_DE_Manitoba
|-- id_to_labels.csv
|-- best_results_sorting_per_fold.py
|-- Fold_0_Step_2_DE_Manitoba.csv
|-- Fold_1_Step_2_DE_Manitoba.csv
...
|-- Fold_0_Step_2_DE_Manitoba_MSE_xxxacc_0th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
|-- Fold_1_Step_2_DE_Manitoba_MSE_xxxacc_1th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
...
Run the following command:
python best_results_sorting_per_fold.py --file id_to_labels.csv
Results would be saved in the form of a single .csv file i.e. best_results.csv and as a confusion matrix in the form of an image file as shown below:
|-- outputs
|-- Step_2_DE_Manitoba
|-- id_to_labels.csv
|-- best_results_sorting_per_fold.py
|-- Fold_0_Step_2_DE_Manitoba.csv
|-- Fold_1_Step_2_DE_Manitoba.csv
|-- best_results.csv
|-- confusion_matrix.png
...
|-- Fold_0_Step_2_DE_Manitoba_MSE_xxxacc_0th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
|-- Fold_1_Step_2_DE_Manitoba_MSE_xxxacc_1th-fold_lr-xxx_sen-xxx_spec-xxx_corr-xxx_epoch-xxx_pv-xxx_rho-xxx_prho-xxx.pt
...
For results calculation including 3-class accuracy based on AAC-24 scoring method, Pearson Correlation, Kendall Tau etc., copy the file ./scripts/results_calculation.py to ./outputs/results_calculation.py
and run the following command by adding the relevant folders in the file:
python results_calculation.py --folder Step_2_DE_Manitoba Step_2_SE_Manitoba Step_2_SE_US_Model_Combined
The results would be saved in the form of .txt files as follows:
|-- outputs
|-- results_calculation.py
|-- Step_1_DE_Manitoba
|-- Step_2_DE_Manitoba
|-- Step_1_SE_Manitoba
|-- Step_2_SE_Manitoba
|-- Step_1_SE_US_Model_Combined
|-- Step_2_SE_US_Model_Combined
|-- Step_2_DE_Manitoba_results.txt
|-- Step_2_SE_Manitoba_results.txt
|-- Step_2_SE_US_Model_Combined_results.txt
For prediction using Manitoba SE trained models, use the following command:
python predict_Manitoba_SE.py
For prediction using Manitoba DE trained models, use the following command:
python predict_Manitoba_DE.py
For prediction using US-Model combined SE trained models, use the following command:
python predict_US_Model_combined_SE.py
The predictions would be in the following directories:
|-- outputs
|-- predictions
|-- Manitoba_DE
|-- fold_0_model_predictions.csv
|-- fold_1_model_predictions.csv
...
|-- Manitoba_SE
|-- fold_0_model_predictions.csv
|-- fold_1_model_predictions.csv
...
|-- US_Model_combined_SE
|-- fold_0_model_predictions.csv
|-- fold_1_model_predictions.csv
...
|-- Step_1_DE_Manitoba
|-- Step_2_DE_Manitoba
|-- Step_1_SE_Manitoba
|-- Step_2_SE_Manitoba
|-- Step_1_SE_US_Model_Combined
|-- Step_2_SE_US_Model_Combined