What do the 3 input channels of Pillar0-BreastMRI represent?
Hi, thanks for releasing Pillar0-BreastMRI.
I am trying to use the pretrained vision encoder for 3D breast MRI feature extraction.
The released config specifies in_chans=3 and an input size of [384, 384, 192]. I also noticed that the pretraining code applies high_contrast MR windowing for breast_mr.
Could you clarify what the three input channels represent before windowing?
Specifically, are they:
three DCE-MRI phases/sequences,
three derived intensity representations/windows of one MRI volume,
or another three-channel composition?
I currently have a single 3D DCE subtraction volume per patient, so I would like to reproduce the pretrained Breast MRI input preprocessing as faithfully as possible.
Thank you.
Hello and thank you for your question!
The model input is the following 3 different breast MRI series:
- T1 fat-saturation
- T2 fat-saturation
- Peak contrast-enhanced series
Would you be able to procure these?
@longchaoliu Thank you again for the clarification. I have a few follow-up questions about reproducing the preprocessing for Pillar0-BreastMRI:
For exams with multiple post-contrast DCE phases, how is the peak contrast-enhanced series selected? Is it a fixed post-contrast phase, based on acquisition time, or determined according to the degree of enhancement?
Are the T1 fat-saturated, T2 fat-saturated, and peak contrast-enhanced volumes registered to a common reference space before being stacked? If so, which sequence is used as the reference, and what registration method is used?
After preprocessing, are the three volumes stacked directly as three channels in the order [T1 fat-saturation, T2 fat-saturation, peak contrast-enhanced] before being passed to the model?
We checked the public Pillar pretraining, RAVE, and RATE-Evals code. RATE-Evals appears to apply MR high-contrast windowing independently to each input channel, but we could not find the public implementation for peak-phase selection or the three-series registration/stacking procedure.
@longchaoliu Thank you very much for your help!