Hugging Face's logo

Datasets:
BodyMaps
/
CancerVerse

Tasks:
Image Classification
Image Segmentation
Image-to-Text
Size:
10K - 100K
Tags:
medical-imaging
computed-tomography
oncology
cancer
radiology-reports
longitudinal
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
10

Recommended Pipeline for Longitudinal CT Registration in CancerVerse

#10
by Wei1998ustc - opened
Discussion
Wei1998ustc

Hi, thank you for releasing this valuable dataset!

I'm currently working with longitudinal CT scans from CancerVerse (multiple timepoints from the same patient) and have encountered some challenges when aligning the volumes for downstream longitudinal analysis.

  1. Inconsistent image geometry: CT volumes acquired at different timepoints often have different dimensions, voxel spacings, origins, and fields of view (FOVs). For example, two scans from the same patient have shapes of 512 × 512 × 620 (0.79 × 0.79 × 1.0 mm) and 512 × 512 × 641 (0.82 × 0.82 × 1.0 mm), respectively.

  2. Suboptimal affine registration: I attempted to register follow-up scans to their corresponding baseline scans using ANTs (antsRegistrationSyNQuick.sh -t a, affine registration with mutual information). However, the resulting alignment was still insufficient, particularly around organ boundaries, making voxel-wise longitudinal comparisons unreliable.

I would greatly appreciate your insights on the following questions:

  1. Recommended pipeline: Is there an official or recommended preprocessing/registration pipeline for aligning longitudinal CT scans in CancerVerse?
  2. Pre-registered data: Are there any plans to release pre-registered longitudinal CT volumes or corresponding transformation fields?
  3. Handling geometric differences: Do you have any suggestions for handling substantial differences in FOV, image origins, and anatomical coverage before registration?

Any guidance or recommendations would be extremely helpful for researchers interested in longitudinal modeling and disease progression analysis using CancerVerse.

Thank you again for making this dataset publicly available and for your valuable work!
BodyMaps
 Owner

Hi @Wei1998ustc ,

Thanks for the detailed questions, and for the kind words.

  1. Recommended pipeline: we do not have an official registration pipeline. The scans are released exactly as acquired in the clinic, so different timepoints naturally differ in scanner, protocol, spacing, and field of view.
  2. Pre-registered data: we have no plans to release pre-registered volumes or transformation fields. The right registration depends on the downstream task, so we leave that choice to the user.
  3. Handling geometric differences: for a voxel-wise comparison, I would recommend uniGradICON (https://github.com/uncbiag/uniGradICON), a foundation model for medical image registration that handles varying geometry well and gives a better deformable alignment than affine alone. Before running it, resample both scans to a common spacing and crop them to the same anatomical range, for example by using an organ segmentation to find the overlapping region, so the registration is not fighting the FOV mismatch.

Hope this helps, and we would be glad to hear what works for you.

Thanks,
Wenxuan

Sign up or log in to comment