Hi everyone,

I am analyzing a Xenium spatial transcriptomics dataset from tumor tissues and would appreciate advice on how to select biologically and statistically reasonable radii for a dual-scale spatial niche analysis.

My dataset contains approximately 1.2 million cells from multiple tumor samples. The cells have already been annotated into major cell types, including tumor cells, T-cell subsets, B cells, macrophages, CAFs, pericytes, and endothelial-cell subtypes.

I am currently testing a dual-scale neighborhood framework. For each anchor cell, I define:

Inner neighborhood: cells located within an inner radius, representing the local or direct microenvironment. Outer neighborhood: cells located between the inner and outer radii, representing the broader or indirect microenvironment.

The neighborhood features are mainly the fractions of different cell types within each spatial scale. I then plan to use these features for Leiden clustering to identify spatial niche subtypes.

At present, I am considering the following candidate radii:

Inner radius: 20, 30, 40, or 50 µm Outer radius: 80, 100, 150, or 200 µm

However, I am concerned that selecting one radius combination based only on visual inspection or the biological size of a few cell types may be too arbitrary.

My main questions are:

What criteria are commonly used to select inner and outer radii in spatial transcriptomics niche analysis? Should the radii be selected according to physical cell-to-cell interaction distances, neighborhood cell counts, clustering stability, or sample-level reproducibility? Is it reasonable to compare multiple radius combinations using metrics such as silhouette score, Leiden cluster stability, adjusted Rand index, or patient-level reproducibility? How can I avoid choosing radii that mainly reflect differences in cell density among samples? Should one global radius combination be used for all samples, or should the radius be adapted according to sample-specific cell density? Are there recommended sensitivity analyses for demonstrating that the identified spatial niches are not dependent on one arbitrarily selected radius combination?

One possible strategy I am considering is to run all candidate inner–outer radius combinations and select a combination that satisfies the following conditions:

sufficient neighborhood cell counts for most anchor cells; stable niche assignments across random seeds and nearby radius combinations; reproducibility across patients; clear but not overly fragmented cell-type composition patterns; limited association between niche assignment and technical factors such as cell density or total detected transcripts.

Would this be an acceptable strategy? Are there published methods or practical guidelines for choosing spatial scales in this type of analysis?

Any suggestions or references would be greatly appreciated. Thank you!