Request to include pixel size/resolution in the spec

[petebankhead]

The topic of including the requested pixel size (resolution) in the model spec came up at the recent HT hackathon.

I wrote a little rant on Zulip about why I think it's important, and at @ctrueden's suggested wanted to include it here.

In summary:

1. It's extremely common in pathology (and potentially useful elsewhere) to train models from images from a variety of sources, often acquired using different scanners that use different pixel sizes. Rescaling images later to have a standardized pixel size is an effective way to reduce at least one source of variation, and have a model that can potentially generalize well.
2. It's also extremely common to train models that should be run at different resolutions for different purposes (e.g. 0.5µm/px for nucleus detection, 5µm/px for tumor recognition)

QuPath definitely needs resolution information, but other software does as well - e.g. see

• [ENH] look into bioimage spec SBU-BMI/wsinfer#68

Having the intended resolution as an optional entry in the spec would help avoid more reliance on custom configs.

[FynnBe]

we have discussed quite a bit already and this is the current proposal: #290
@petebankhead it would be great to get your feedback on that!

[petebankhead]

Sorry, hadn't seen that - I'll think some more, but first impression is that the approach looks good to me!

[FynnBe]

unit and scale were added to the axis descriptions in model 0.5