Clarification needed regarding tissuenet_model_0019999.pth and Whole-cell vs. Nuclear Segmentation

[Zhiyuan-Weng]

Hi authors,

Thanks for your great work on CelloType! I've been exploring the repository and trying to reproduce the results on the TissueNet dataset.

I noticed that the paper reports performance for both Whole-cell Segmentation and Nuclear Segmentation, but I have a few questions regarding the provided pretrained model and preprocessing scripts:

Model Ambiguity: The download.sh script provides a single weight file tissuenet_model_0019999.pth. Could you please clarify if this model was trained on nuclear masks or whole-cell masks?
Preprocessing Logic: In [prep_tissuenet.py], the code explicitly extracts the second channel ([mask = data_y[i,:,:,1]]), which corresponds to the nuclear mask in TissueNet, and saves the file as dataset_dicts_nuclear_{}.npy. However, the training script [train_tissuenet.py] attempts to load dataset_dicts_cell_{}.npy.
Does this imply the provided model is strictly a nuclear segmentation model?
If I want to perform whole-cell segmentation, is it sufficient to just change the channel index to 0 in [prep_tissuenet.py] and retrain, or are there other configuration changes needed?
It would be very helpful if you could clarify which task the provided .pth file corresponds to, or if there might be a missing weight file for the other task.

Thank you for your help!