In general, I would probably try GATK before DeepVariant:

C: How Much Does It Cost To Align A Flowcell In The Cloud?

GATK performance also seems similar (or slightly better) than DeepVariant in this paper (with BWA alignments):

https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-019-2928-9

In the specific case of targeted gene panels, I have previously found indications that there can be a relatively greater need for modification for targeted gene panels than Exome sequencing (where I was using VarScan with a set of non-default parameters for targeted gene panels):

https://peerj.com/articles/600/

[please note that the y-axis is different in the panels from the figure in the Warden et al. 2014 paper]

So, without directly testing DeepVariant, I would definitely recommend checking for your variants (and I think it is possible you may want to use an alternative model, from a large panel of samples with your own target design).

I don't think there would be a problem using DeepVariant for targeted gene panels. make_examples accepts a bed file defining your regions with the --regions flag; you would have to decide whether to use the WGS model or the WES model for call_variants. I'd recommend using the WES model, assuming your targeted data has the kind of non-uniform coverage patterns seen in WES data.