search

xmed-lab / DIF-Gaussian

MICCAI 2024: Learning 3D Gaussians for Extremely Sparse-View Cone-Beam CT Reconstruction
40 stars 3 forks source link

Can results be stored and visualised? #1

Open dfdu233 opened 1 month ago

dfdu233 commented 1 month ago

I trained LUNA16 data with your code but found the results can't be visualised.Will you consider accomplish that function?

lyqun commented 1 month ago

Hi,

For the training, the visualization is not supported yet as we want to simplify the evaluation during training. Note that the evaluation has been simplified during training with a lower output resolution (0.5x).

For the testing, you can add '--save_results' (https://github.com/xmed-lab/DIF-Gaussian/blob/main/code/evaluate.py#L91) and then the results will be saved in a specific directory (see https://github.com/xmed-lab/DIF-Gaussian/blob/main/code/evaluate.py#L129)

dfdu233 commented 3 weeks ago

Hi, thanx for the reply.I test the results on brain CT dataset,but results seem so strange image Is there any way to prove that?

lyqun commented 3 weeks ago

What dataset was the model trained on? If you trained the model on LUNA16, it may not be appropriate to be applied to brain CT due to the domain gap.

dfdu233 commented 2 weeks ago

Yes，I am working with a head CT dataset called CQ500, but the results are unsatisfactory. The images appear blurred and exhibit artifacts. Are there any methods to improve this? many thanks

dfdu233 commented 2 weeks ago

BTW, it's trained on head CT too, so I wonder what causes this

lyqun commented 1 week ago

The results look wired. Can you check the processed data -- the CT and rendered X-ray projections? If the dataset is very different from the lung/dental CT, the preprocessing may need to be modified, e.g., the value range.