cwmok / DIRAC

This is the official Pytorch implementation of "Unsupervised Deformable Image Registration with Absent Correspondences in Pre-operative and Post-Recurrence Brain Tumor MRI Scans" (MICCAI 2022), written by Tony C. W. Mok and Albert C. S. Chung.
MIT License
36 stars 1 forks source link

why alpha #9

Closed 18370029656 closed 1 year ago

18370029656 commented 1 year ago
image

In this step, how exactly does alpha=0.015 get it? For a new data set, how should the corresponding alpha value be calculated? Looking forward to your answer

cwmok commented 1 year ago

Hi @18370029656,

We selected the alpha value based on the forward and backward error maps of the validation set, and we found it generalized well to the test set. An ideal alpha value should be able to highlight regions with absent correspondence, e.g., tumor and resected tumor region, while excluding the regions with valid correspondence. A more rigorous way to choose the alpha is to measure the overlap (e.g. Dice score) of thresholded region and the tumor region of the validation set.