Evaluating Spatial Correspondence of Multimodal MRI via Anatomical Landmarks

1833 

Abstract Submission 

Joy Zhao¹, Alaa Taha¹, Mohamed Abbass², Greydon Gilmore³, Chris Zajner³, Violet Liu⁴, Homa Vahidi¹, Arun Thurairajah⁵, Ali Khan¹, Jonathan Lau⁶

¹University of Western Ontario, London, Ontario, ²Western Univeristy, London, Ontario, ³Department of Clinical Neurological Sciences, Division of Neurosurgery, London, Ontario, ⁴Western Univeristy, London, MT, ⁵Department of Clinical Neurological Sciences, Division of Neurosurgery, Western University, London, ON, ⁶Department of Clinical Neurological Sciences, Division of Neurosurgery, London, ON

Joy Zhao  
University of Western Ontario
London, Ontario

Alaa Taha  
University of Western Ontario
London, Ontario

Mohamed Abbass, MD  
Western Univeristy
London, Ontario

Greydon Gilmore  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, Ontario

Chris Zajner  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, Ontario

Violet Liu  
Western Univeristy
London, MT

Homa Vahidi  
University of Western Ontario
London, Ontario

Arun Thurairajah  
Department of Clinical Neurological Sciences, Division of Neurosurgery, Western University
London, ON

Ali Khan  
University of Western Ontario
London, Ontario

Jonathan Lau  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, ON

Multimodal fusion involves the combination of two or more datasets acquired using different techniques and parameters to enhance our understanding of brain structure and function [Zhang et al. 2020, 149-187]. It is often common practice to align multimodal images to each other and to a standard neuroimaging template space. Voxel-based overlap metrics, commonly used in neuroimaging analysis, are less sensitive to focal misregistration [Rohlfing 2012, 153-163], compared to a point-based millimetric framework [Lau et al. 2019, 4163-4179].
We validated and released data [Taha et al. 2023] of over 5,000 salient anatomical fiducials (AFIDs) in the human brain across T1w scans with varying MRI field strengths and neurodegenerative diseases (Figure 1). We previously demonstrated the high sensitivity of AFIDs during point-based evaluation of image registration [Abbass et al. 2022, 393-405] which has broad implications in the field and clinical contexts (e.g., pooling deep brain stimulation electrodes for group-level analysis).
In this study, we aim to 1) provide an overview of the AFIDs framework and associated data, and 2) extend the AFIDs framework to T2w MRI datasets, enabling evaluation of biases in registration accuracy when using standard neuroimaging software.

Imaging Datasets. We curated across a number of different neuroimaging modalities and resolutions: 1) a previously published dataset [Chen et al. 2023] consistent with healthy paired imaging acquired at 7-T and 3-T MRI employing both T1w and T2w imaging sequences and 2) a local neurosurgical dataset with patients who underwent: a) T1- and T2- weighted MRI (at 1.5- and 7-Tesla), b) contrast-enhanced MRI at 1.5-Tesla, c) contrast- and non-contrast-enhanced CT (peri- and post-operatively), and d) diffusion-weighted imaging (DWI) with quantitative derivatives including fractional anisotropy.
Anatomical Landmarks. We made our curated placement protocols publicly available [Lau et al. 2023]. In total, 34 AFIDs (10 midline and 12 bilateral) that survey the brain were applied. To curate AFID ground-truth data, we recruited a total of 20 human raters over the past 4 years, constituting a wide spectrum of expert and novice raters. Novice raters went through the open-access protocol and leveraged our online web app platform "AFIDs Validator" [Kai et al. 2023] during training.
Localization Assessment (Figure 2A). We computed the anatomical fiducial localization error (AFLE) of these 34 landmarks across rater applications (i.e., mean AFLE) and also across MRI field-strengths (i.e., inter-scan AFLE) of paired imaging.
Registration Assessment (Figure 2B). We quantified the anatomical fiducial registration error (AFRE) when registering a participant's scans to a common standard template, using presets from a nonlinear deformation framework validated via 11,000 non-linear warps across 100 subjects [Ewert et al. 2019].
We focus our reporting here on new findings validating AFIDs for T2w imaging.

Rater demographics. Six human raters were trained and subsequently recruited to place the AFIDs protocol on T2w imaging.
AFLE. The mean AFLE across all scans and AFIDs was 1.03 +/- 0.55 mm. Furthermore, the error between AFID localization on 3T and 7T MRI scans was 0.65 +/- 0.35 mm (7 out of 34 AFIDs were localized more accurately on 7T MRI).
AFRE. The mean AFRE was statistically higher (p < 0.001) on 3T (3.12 +/- 1.93 mm) when compared to 7T scans (2.86 +/- 1.94 mm).

This study validates AFIDs as a reliable and sensitive tool for T2w image registration and strives towards better standardization of intermediate steps taken in a variety of neuroimaging and clinical workflows. Using the curated datasets from this study, future work will expand this protocol to other commonly used imaging modalities (e.g., DWI and CT) and explore approaches that automate AFID localization.

Image Registration and Computational Anatomy ¹

Neuroanatomy Other

Databasing and Data Sharing ²

Informatics Other

Computational Neuroscience

Data Registration

HIGH FIELD MR

MRI

Open-Source Code

Spatial Normalization

Spatial Warping

Other - multimodal imaging, anatomical landmarks, non-linear registration