The effect sizes reported in the activation maps changed as well. With 3 mm voxels, the peak t-statistic was moderate, and the cluster spanned several cubic centimeters. With 1.5 mm voxels, the peak was sharper, and the cluster was smaller. The researchers noted that the finer voxels reduced partial volume effects, revealing a more precise location of the response. However, the smaller voxels also produced more noise, necessitating additional smoothing to achieve a comparable signal-to-noise ratio.

This protocol change was not a trivial tweak. It required re-optimizing the acquisition parameters, including repetition time (TR), echo time (TE), and flip angle. The lab also had to update their preprocessing pipeline to handle the higher resolution data, including motion correction and spatial normalization. The effort paid off in terms of spatial specificity, but it also introduced a new variable that complicated comparisons with their earlier studies.

The Oslo lab's experience highlights the sensitivity of fMRI findings to methodological details. A change in voxel size can alter the location, extent, and statistical significance of activation clusters. This is particularly problematic in small structures like the amygdala, where even a few millimeters of shift can change the interpretation. The study serves as a concrete example for researchers who assume that identical scanning parameters across studies guarantee comparable results.

Why Smaller Voxels Aren't Always Better

While smaller voxels offer improved spatial resolution, they come with significant drawbacks. The most obvious is the increase in the number of voxels, which inflates the number of statistical comparisons. In a typical whole-brain analysis, a 3 mm voxel grid yields about 50,000 voxels; a 1.5 mm grid yields about 400,000. This necessitates stricter multiple comparison corrections, such as the Bonferroni or false discovery rate (FDR) methods, which can reduce statistical power.

Another issue is the reduction in signal per voxel. The BOLD signal depends on the volume of tissue sampled, and smaller voxels contain fewer spins, leading to a lower SNR. To compensate, researchers often increase the number of repetitions or use larger smoothing kernels, which partially negate the benefits of high resolution. The trade-off between spatial and temporal resolution is a limitation in fMRI.

Motion artifacts become more problematic with smaller voxels. Head movement of a few millimeters, which is common in human subjects, can cause voxel misalignment and signal dropout. At 1.5 mm, even small movements can shift the voxel grid relative to the brain, introducing errors that are harder to correct. Motion correction algorithms exist, but they are imperfect and can introduce their own biases.

Finally, smaller voxels can lead to overfitting in statistical models. With more voxels, there are more opportunities for random noise to produce spurious clusters. This is particularly concerning in studies with modest sample sizes, where the statistical threshold may not be stringent enough to control for false positives. Researchers must balance the desire for detail against the risk of unreliable results.

What This Means for Reproducibility in Neuroimaging

The neuroimaging field has faced a reproducibility crisis, with many published findings failing to replicate in subsequent studies. While many factors contribute, methodological variability, including voxel size, is a key factor. A 2012 study by the Open Science Collaboration found that only about a third of psychology studies replicated (Open Science Collaboration, 2015), and fMRI studies are difficult to reproduce. Voxel size is one of many parameters that can vary across labs, making direct comparisons challenging.

Efforts to improve transparency, such as the COBIDAS (Checklist for Brain Imaging Data and Structure) guidelines, encourage researchers to report detailed acquisition parameters, including voxel size, echo time, and flip angle. The goal is to make methods sections comprehensive enough that other labs could replicate the study exactly. However, even with such guidelines, subtle differences in preprocessing and analysis can still lead to divergent results.

The Oslo lab's experience suggests that voxel size should be reported explicitly and justified in every fMRI paper. A reader should be able to assess whether the chosen resolution is adequate for the structure of interest. For small subcortical regions like the amygdala, 1.5 mm voxels may be necessary, but for larger cortical areas, 3 mm might suffice. The choice should be driven by the research question, not by convention.

Cross-lab comparisons would benefit from standardized protocols, but such standardization is difficult to achieve in practice. Different scanners, sequences, and analysis pipelines introduce variability that cannot be fully controlled. The field is moving toward multi-site studies and open data sharing, which can help identify which findings are robust across variations. However, until voxel size is treated as a critical variable, the risk of contradictory results remains.

Broader Implications for the Field

The Oslo lab's findings are not just a cautionary tale; they point to a systemic issue in neuroimaging. Voxel size is often chosen based on convention or convenience, not on the specific demands of the research question. This can lead to results that are not directly comparable across studies, undermining the cumulative nature of science. The field needs to move toward more standardized reporting and, where possible, standardized protocols.

One potential solution is the adoption of common data elements for fMRI, including voxel size. Organizations like the International Neuroinformatics Coordinating Facility (INCF) have proposed such standards, but adoption has been slow. Another approach is to encourage multi-site studies that use identical protocols, as seen in projects like the Human Connectome Project. These efforts can help identify which findings are robust across different voxel sizes and scanner types.

There is also a need for more research on the effects of voxel size itself. Systematic studies that vary voxel size while keeping other parameters constant can provide guidance on when finer resolution is worth the cost. For example, a study might compare amygdala activation maps obtained with 1.5 mm, 2 mm, and 3 mm voxels in the same participants. Such data would help researchers make informed decisions about their own protocols.

Ultimately, the choice of voxel size is a trade-off that should be made explicit and justified. The Oslo lab's switch from 3 mm to 1.5 mm voxels was driven by a specific hypothesis about amygdala subregions, and it yielded new insights. But it also made their results harder to compare with earlier studies. As the field moves forward, embracing methodological transparency and variability will be essential for building a robust understanding of the brain.