Pancreas Tumor segmentation CT dataset

[sovanlal]

Hi,

Thanks for your outstanding work and sharing the code. I was trying to test this model on an internal pancreas tumor (PDA) segmentation dataset to see how the model behaves for these complex CT images.

I trained the GenSeg3D model using >1000 CT abdominal CT images with PDA and tested on ~241 CT images. I organized the data as per the instructions i.e. creating a 'train" and "val" folder and within each folder, I uploaded the ground-truth PDA segmentation inside "t1" and raw CT images inside "t2" sub-folder. I also manually changed the all the paths inside as per my dataset. After, inference on validation subset, the DSC was around ~0.27.

1. Do you have any suggestions for why this performance is so low?
2. Do I need to add any other steps in addition to what I have done?
3. Has this model ever been tested on complex problem like PDA segmentation?
4. Does this model also save the individual predicted segmentations in .nii.gz format or just provide the mean test dice?

Thank you for your help.