Automatic Coronary Calcium Scoring from Gated Coronary CT Using DL-based FP detection model

CAC Scoring from NCCT Using DL With External Validation

This work introduces an automatic CAC scoring method that uses multi-atlas segmentation for whole heart segmentation (WHS) and a DL model as a supervised classifier for correcting false positives (FP).

Descriptions

• The repository provides a DL-based FP (of CAC) model.
• The model is developed by using the Stanford AIMI COCA dataset which is publicly available for research purpose
• We used the multi-atlas segmentation pipeline implemented by the Biomedical Imaging Group Rotterdam (BIGR)
• Our work was externally validated on the Rotterdam Study

Run

Generate labeled patches with annotated images

python3 patch_prep.py -patch_size 45

Split the patch data into non-overlapping 5 folds w.r.t subjects

python3 k-fold_prep.py -normalize

Evaluate binary classification performance and save the trained models

python3 fp_classifier_train_subject_fold.py -batch_size 32 -n_epochs 100 -lr 1e-4

Compute CAC scores

python3 coca_internal_eval.py -trained_model 'fp_vgg_trained_model_3.pth'

Assess the agreement between computed scores and reference scores

python3 coca_score_agreement.py

References

TBD

Bibtex entry ready to be cited

TBD