Developing Patient-Specific Functional Atlases with Inverse Distance Weighting of MR Images 📝

Author: Chibawanye I. Ene, Sherise D. Ferguson, Ping Hou, Vinodh A. Kumar, Ho-Ling Anthony Liu, Kyle R. Noll, Sujit S. Prabhu, Jian Ming Teo, Max Wintermark 👨‍🔬

Affiliation: Department of Neuro-Oncology, The University of Texas MD Anderson Cancer Center, Department of Neuroradiology, The University of Texas MD Anderson Cancer Center, Department of Neurosurgery, The University of Texas MD Anderson Cancer Center, Department of Imaging Physics, The University of Texas MD Anderson Cancer Center 🌍

Abstract:

Purpose:
Functional brain atlases are used to guide clinical functional MRI (fMRI) analyses. Imprecise assertions may introduce the ecological fallacy as atlases are reflective of the constituent cohort and not the subject. This study presents a first approach to introduce a weighting scheme to bring functional atlases in closer alignment with the subject.
Methods:
Presurgical sentence completion task-based (tb-) fMRI, T1-weighted and T2-weighted FLAIR images from an atlas dataset and testing dataset of 270 and 79 brain tumor patients, respectively, were analyzed using AFNI and SPM12. Pearson distance metric was calculated between normalized T1-weighted images to obtain nT1-distance, and between normalized FLAIR images to obtain nFLAIR-distance in MNI space. For testing dataset patients, the fractional inverse of the L2-norm (with weight order κ=0,1) was used to construct a weighted probabilistic overlap map (POM) from tb-fMRI of atlas dataset patients. κ=0 is the unweighted POM. For each κ, 10 atlases were generated by randomly leaving 10% of atlas dataset patients out. Atlas agreement with testing dataset patients was determined by highest dice coefficient across atlas thresholds with the detected language network within left hemisphere primary language areas (PLA) defined by the AAL3 atlas. Wilcoxon signed-rank test and paired Cohen’s d were calculated.
Results:
For left hemisphere PLA, mean highest dice for κ=1 atlases (0.456±0.166) was found to be significantly higher (p<0.001) than κ=0 atlases (0.454±0.166) with small effect size (d=0.449). For left hemisphere posterior PLA, mean highest dice for κ=1 atlases (0.417±0.197) was found to be significantly higher (p<0.001) than κ=0 atlases (0.412±0.195) with moderate effect size (d=0.574).
Conclusion:
This work indicates the potential of a weighting scheme in developing functional atlases personalized to patients. Future work will expand on the types of images and information used in calculating total-distance and weight order optimization.