C-Arm Guidance

Official implementation of the paper "C-Arm Guidance: a Self-Supervised Approach to Automated Positioning During Stroke Thrombectomy" (ISBI 2025)

Prerequisites

Training & Testing

• Pytorch
• torchvision
• pillow
• tqdm

GUI

in progress

Training

For training you'll need to provide a dataset composed of X-ray images. Feel free to either use our provided GUI or your own data (make sure to adapt the code, if you face any problems don't hesitate to submit an issue)

For the best experience, make sure your data repository tree is as follows:

regression
├── case_00000/
   └── *.png
├── case_00001/
   └── *.png
├── .
├── .
├── case_N/
   └── *.png

Landmarks
├── 20/
   └── *.png
├── 19/
   └── *.png
├── .
├── .
├── 1/
   └── *.png 

Also, the dataset class assumes to have an annotation .csv files with the following columns:

Regression

column	example	Description
case_number	case-10065	uniques case ID
filename	root/case_10065/5.png	file path for the Xray .png image
x	-232.5	x-position of the image in the CT
y	-34.5	y-position of the image in the CT
z	208.2	z-position of the image in the CT
part	upper	Xray belongs to the 'upper' or 'lower' CT
age_years	66	patient age
sex_code	Male	patient sex
cadaver_weight	73.0	cadaver weight (kg)
cadaver_length	174.0	cadaver legnth (m)
mode	train	'train' or 'test'

Classifier

column	example	Description
case_number	case-10065	uniques case ID
filename	root/20/10065.png	file path for the Xray .png image
x	-232.5	x-position of the image in the CT
y	-34.5	y-position of the image in the CT
z	208.2	z-position of the image in the CT
part	upper	Xray belongs to the 'upper' or 'lower' CT
age_years	66	patient age
sex_code	Male	patient sex
cadaver_weight	73.0	cadaver weight (kg)
cadaver_length	174.0	cadaver legnth (m)
mode	train	'train' or 'test'
landmark	20	landmark label [1-20]

Pretext Task (Regression)

To train the first task (regression) navigate to the root directory and run the following script

bash ./scripts/train_self_supervised.sh

Note that you can customize the training experiment using the arguments provided in ./src/train_self_supervised.py by simply passing them to ./scripts/train_self_supervised.sh

This script will save checkpoints at every epoch in ./LOG_DIR/EXP_NAME/checkpoints/ (all of these variables should be provided in train_self_supervised.sh)

Downstream Task (Classification)

To finetune your model on the classification task, navigate to the root directory and run the following script

bash ./scripts/train_classifier.sh

Note that you can customize the training experiment using the arguments provided in ./src/train_classifier.py by simply passing them to ./scripts/train_classifier.sh (make sure to locate the checkpoint from the pretraining phase)

This script will save checkpoints at every epoch in ./LOG_DIR/EXP_NAME/checkpoints/ (all of these variables should be provided in train_classifier.sh)

Training Experiments

You can customize the training experiment using the arguments provided in each script by simply passing them to ./scripts/train_classifier.sh or ./scripts/train_self_supervised.sh, for example:

--pretrained_weights="imagenet"

The following are the main experiments ran in the paper:

Argument	Experiment
--pretrained_weights="position"	Determine the classifier pretraining
--pretrained_weights="imagenet"	Determine the classifier pretraining
--pretrained_weights="none"	Determine the classifier pretraining
--linear_probing	Implement linear probing and freeze all layers except the last linear ones
--remove_patient_stats	Ablation study to not include the patient demographics

GUI & Dataset

In Progress