Ahh, I'm with you now! I didn't catch that you were talking about the choices sheet.
This is because image, big-image and label have special meaning to the XLSForm converter when used on the choices tab. Here's the XML that's generated for the "Pomme" choice:
<text id="internal_choices-0">
<value>Pomme</value>
<value form="big-image">jr://images/Image1.jpg</value>
<value form="image">jr://images/Image1_SM.jpg</value>
</text>
...
<item>
<itextId>internal_choices-0</itextId>
<emoji>🍎</emoji>
<name>apple</name>
<notanimage>notanimage1.jpg</notanimage>
<shape>round</shape>
</item>
You can see that under the hood choice images use the same mechanism as translations. If you remove the image and big-image columns, the following XML will be generated for "Pomme":
<item>
<label>Pomme</label>
<emoji>🍎</emoji>
<name>apple</name>
<notanimage>notanimage1.jpg</notanimage>
<shape>round</shape>
</item>
The itext/translation indirection will be removed and you'll be able to get the label directly.
This is implicit and very annoying, I'm sorry you've run into it. I'm not sure what we can really do about it without rethinking the way translations and images are defined, though.