Bin size of bin20 or bin50 at Stereo-seq is 10um x 10um and 25um x 25um capture, while the former is expected to be single cell resolution it might be a bit small to consider an entire large cell like neurons. The latter will also be close to single cell resolution but might give you transcripts from a few cells at once (3/4 cells)

Stereo-seq is a non targeted spot based method, there is no cell segmentation, therefore, as small as your bins can be, you can never be sure that you are looking at only one cell. The spot might be sitting exactly at the cell border.

One method to annotate spots or here bins, is to run a transfer annotation from a single cell dataset from the literature (even better if you have your own one on the same cohort, and even better if you have a consecutive slide to run single cell on). Use the predictive scores for each bin to consider your cell population in each of them. Other methods such as RTCD or cell2location for example. See STOmics docs.

On the UMAP, your collagen genes are predominant in clusters 2 and 5, with some sparse expression in the other clusters, while your dot plot show a more homogeneous expression across clusters.

Could you double check the normalization of the data on your dot plot, as "mean expression in group" scaled from 0 to 1, does not fit with "mean expression".

Could you plot the raw expression of the collagen genes spatially to check their spatial expression patterns ?

Please give more details on your experiment and code.