Low accuracy and dice values - Executing inference

[vlgonz]

I'm running inference using only the IPGN checkpoint (.pt, converted from .pkl), which includes frozen weights for Graph, Point, and Fusion.

I tested a small batch of airway data, and results are very low:

filename,accuracy,dice score
pulse_00007_airway.npz,0.025,0.003
pulse_00016_airway.npz,0.032,0.003
pulse_00019_airway.npz,0.034,0.004
pulse_00020_airway.npz,0.043,0.005

Also, when loading pulse_00007_airway.nii.gz in 3D Slicer, the segmentation appears as a single green segment with label 1.

Am I missing any step?
Thanks in advance!

[cwaid3]

I was having a hard time loading the saved checkpoints, they are not in .pt format. I had to re-train the models using
python train_network.py -d $dataset_specs_json_file
as specified in the README. Training went faster than expected though. I did not adjust the format that the models were saved in, but for some reason after re-training, I was able to run the inference.
python reconstruction_inference.py -d $dataset_specs_json_file
After doing this I got better dice values than what you reported, and they were in line with the results from the paper:
filename,accuracy,dice score
pulse_00007_airway.npz,0.907,0.802
pulse_00016_airway.npz,0.933,0.85
pulse_00019_airway.npz,0.854,0.786
pulse_00020_airway.npz,0.879,0.849