Running inference using sample mpMRI data - questions about parameters in config file

[deepakri201]

Hi,

First, thank you for making this interesting work and repository public!

I am running the inference code on the sample mpMRI dataset provided, and have noticed a few small issues:

1. The name of the config file specified is incorrect, and should be/configs/config_infer_MRI.yaml

2. I can successfully run the code using the T2, ADC, DWI, gland, and cancer nii files as written here: configs/config_infer_MRI.yaml. However, in inference_mpMRI.py, it seems that the cancer nii file is optional:

I tried removing this from the yaml file. However, it does not seem like this case is handled:

3. It is my understanding that for inference, only the T2, DWI, and ADC images are necessary, as written in the paper:

"The model processes three MRI sequences (T2-weighted, DWI, and ADC) independently through sequence-specific decoders that generate pixel-level probability maps for prostate gland segmentation and cancer classification"

If so, it might be useful to handle the cases where the gland/cancer nii files are not provided in the config file for inference.

Thanks!

[so-sal]

Hi,
Thank you for your comment!

This issue is due to the current implementation of the data loader.
You can fix it by replacing Dataset['Cancer'] with Dataset['Gland'].

If you don't have the gland segmentation data, we’d be happy to provide T2 based gland-segmentation model.
Feel free to reach out via email, tkdlvk27@gmail.com

Thanks again!

[deepakri201]

Hi,

Thank you for your response, I will try that.

So for my previous question:

3. It is my understanding that for inference, only the T2, DWI, and ADC images are necessary, as written in the paper:

"The model processes three MRI sequences (T2-weighted, DWI, and ADC) independently through sequence-specific decoders that generate pixel-level probability maps for prostate gland segmentation and cancer classification"

My understanding now is that a T2 gland segmentation is additionally required, and therefore the above line in your paper is false?

Thanks!

[so-sal]

1. ProViCNet model predicts four pixel-level classes: background, prostate gland, non-csPCa, and csPCa.
2. We used prostate gland segmentation data for normalization, but not as input to the model.
3. We already have a highly accurate prostate segmentation model that has been externally validated (DOI link), so manual segmentation isn’t necessary. If needed, we’d be happy to share this model with you.

Thank you.

[deepakri201]

Thank you very much for the clarification!