ANTs SyN Normalization with Brain Tumor Lesions: Patient-to-MNI (Moving Mask) vs. MNI-to-Patient (Fixed Mask)?

[cxstevenddd]

I am working on a pipeline for the Lesion Quantification Toolkit (LQT) and need to normalize Brain Tumor patient T1 images (with tumors) to MNI space. I plan to use ANTs SyN with Cost Function Masking. My core question is about the direction of registration and how the mask is handled during optimization: Option 1: Patient (Moving) to MNI (Fixed). I would input the tumor lesion mask as a "Moving Mask" to exclude the lesion from the metric. Option 2: MNI (Moving) to Patient (Fixed). I would input the tumor lesion mask as a "Fixed Mask," and then apply the calculated inverse warp to the patient image/mask later. Intuitively, Option 2 seems safer because the patient image serves as the fixed domain, ensuring the tumor lesion mask remains static and perfectly aligned with the anatomy throughout the optimization process. I am concerned that in Option 1, as the patient image deforms iteratively, the "Moving Mask" might not deform perfectly in sync, potentially failing to mask the lesion artifacts effectively. Is my concern about the "Moving Mask" valid? Does ANTs handle robust resampling of moving masks across iterations, or is the "Fixed Patient + Inverse Warp" strategy generally preferred for brain tumor lesion normalization?

[ntustison]

The mask specifies where the similarity metric and gradient is calculated so you'd need to invert the tumor mask. Also, you'd have to go with Option 2 as one can specify just a fixed mask, both a fixed mask and a moving mask, but not just a moving mask.

[cxstevenddd]

Thank you for the quick response! However, I believe there might be a misunderstanding.

Option 1 (Patient→MNI with Moving Mask) is actually working well in my pipeline:

antsRegistration \
  --masks [NULL,${MOVING_MASK}] \
  ...

I've successfully processed a couple of brain tumor cases with this approach. The moving mask correctly excludes the lesion from the cost function during optimization.

My specific question is about the theoretical concern: Does ANTs internally update the moving mask position synchronously with the moving image during iterative deformation? Or could there be desynchronization issues?

Based on my empirical results, Option 1 seems robust, but I'd appreciate clarification on the algorithm's internal handling of moving masks across iterations.

[cookpa]

Internally, neither the fixed nor the moving images actually get deformed during registration, only the transforms get updated. To evaluate the metrics, points from the virtual space are mapped to the fixed and moving images, in order to evaluate the interpolated intensity.

The decision about whether a point is inside the masks is made as part of that process. If the current transform maps a point to a position outside of the mask, it is not used to compute the metric. So the masks cannot get misaligned from their images during pairwise registration.

[cxstevenddd]

Thank you so much for taking the time to provide such a detailed and authoritative explanation!

As someone working on brain tumor registration, I feel incredibly fortunate to receive clarification directly from one of ANTs' core developers. Your explanation has completely resolved my concern about mask-image alignment during iterative optimization.

To confirm my understanding: since ANTs updates only the transformation parameters (not the images themselves) and evaluates mask membership at each iteration using the current transform, the moving mask maintains perfect anatomical correspondence with the moving image throughout the registration process. This internal mechanism ensures synchronous mask-image mapping, eliminating the theoretical concern about "desynchronization" that initially motivated my question.

Follow-up question: Does this mean I can confidently choose the forward direction (brain tumor patient T1 as moving image, with lesion mask as moving mask) when designing my registration pipeline? This direction feels more intuitive conceptually (patient space → standard space), and based on your explanation, there appears to be no technical disadvantage compared to the reverse approach in terms of mask handling reliability.