BPPR

This repo provides the code related to the paper titled "BPPR: A Framework for Content Navigation in Multi-Contrast Body CT Images using Deep Regression Models".

Usage

(1) Clone this repo and install the requirements.txt (python >= 3.10).

(2) Create a folder named checkpoints inside the repo, download model checkpoints bpr_model.pth and epr_model.pth from https://drive.google.com/drive/folders/1hDUEp47wdzPDElf7bVHQNoHyYFqeLmTL?usp=sharing and place them in the checkpoints folder.

(3) Prepare the test body CT image in nifti format.

(4) Use the following command to perform inference:

python inference_args.py \
    --input_path /path/to/image.nii.gz \
    --output_path /path/to/output_folder \
    --case_name case_name \
    --device cpu

Argument	Description
--input_path	Full path to the input body CT image in nifti format.
--output_path	Folder where the results are saved.
--case_name	Name used to identify the case in the outputs.
--device	cpu or cuda.

Output

(1) A png visually illustrating the prediction results.

(2) A json containing the scores for each axial slice and the phase for both slice and volume. The volume phase label is inferred from the slice phase labels using the positional weighting mechanism explained in the paper. The json contains the fields scores, phase_label, x1x2 and pred_x1x2:

{
  "scores": [1.5509, 1.5229, 1.5083, "...", -1.5172, -1.5237],
  "phase_label": 0,
  "pred_x1x2": [-0.1390, 0.9903],
  "x1x2": [[0.3788, -0.4238], [0.3814, -0.4120], "...", [-0.2020, 1.5327]].
}

• scores: the body part regression score of each axial slice, ordered along the slice axis of the input volume. See Body part score below.
• phase_label: descrete phase label of the whole volume.
• pred_x1x2: continuous phase label of the whole volume.
• x1x2: continuous phase label of each slice, from which the pred_x1x2 and phase_label is derived using positional weighting mechanism.

Body part score

The score is a continuous coordinate describing the longitudinal position of an axial slice within the body. It is defined by seven anatomical landmarks, which are assigned fixed reference scores spanning the range from the head to the femur:

#	Landmark	Definition	Score
1	Head end	Base of the skull	-1
2	Neck end	Slice at the C7 cervical vertebra	-0.64
3	Aorta arc	Slice through the center of the aortic arch	-0.42
4	Liver start	Apex of the dome of the liver	-0.04
5	Pancreas mid	Middle slice of the pancreas	0.21
6	Kidney end	Most inferior slice of the kidney	0.53
7	Femur end	Most superior slice of the femur ball	1

The score therefore increases from the head towards the feet, and the spacing between the reference values reflects the relative anatomical distance between the landmarks. The predicted scores for slices between two landmarks vary continuously between the corresponding reference values, so a score can be read directly as a body region, and slices from different volumes can be matched by comparing their scores. Slices lying beyond the first or the last landmark can be assigned scores outside the [-1, 1] range.

Phase label

The phase_label field encodes the contrast enhancement phase of the volume:

Value	Phase
0	Non-contrast
1	Arterial phase
2	Portal venous phase
3	Delayed phase