Effects of thresholding on correlation-based image similarity metrics

Area of Science:

• Neuroimaging
• Data Analysis
• Computational Neuroscience

Background:

• Image similarity computation is crucial for neuroimaging analyses such as decoding and meta-analysis.
• The impact of empty voxels on image similarity metrics is not well understood.
• Understanding how thresholding affects these metrics is essential for accurate comparisons.

Purpose of the Study:

• To investigate the influence of varying degrees of image thresholding on pairwise image comparison outcomes in neuroimaging.
• To determine optimal thresholding strategies for enhancing the accuracy of image similarity computations.
• To assess the robustness of meta-analytic comparisons using thresholded brain maps.

Main Methods:

• Systematic analysis of pairwise brain map comparisons with varying thresholding levels.
• Evaluation of image similarity metrics based on the intersection of non-zero voxels.
• Comparison of results from thresholded versus unthresholded brain maps.

Main Results:

• Thresholding at Z = ±1.0 maximizes accuracy for retrieving maps of the same contrast.
• Thresholding up to Z = ±2.0 can enhance accuracy compared to using unthresholded maps.
• Maps thresholded up to Z = ±3.0 (25% non-empty voxels) maintain over 90% accuracy.

Conclusions:

• A moderate degree of thresholding can significantly improve the accuracy of image similarity computations in neuroimaging.
• Robust meta-analytic comparisons are achievable using thresholded brain images.
• The findings provide practical guidelines for optimizing image similarity analyses in neuroimaging research.