Various fixes to PIN integration example

[CPBridge]

Hi @aihsani

Thanks again for this nice integration example and sorry for taking so long to get around to looking over it properly. I just played around with it for a while and it seems to be working quite nicely. During this process, I made a bunch of small changes, which I've collected in this pull request:

• A few fixes to the documentation (small typos and things)
• Despite the docs saying that a GPU is not required, the device for the inference part is actually hard coded as "cuda", meaning it won't work on CPU and will just crash. I fixed this. Unfortunately this does affect me as I cannot figure out how to get GPUs active within a Docker container when using the version of Docker compose that I have (if you have a quick pointer to a fix I'd appreciate it but I don't necessarily need a solution).
• I did make a fairly substantial change to the way the GSPS output works. Even though this is not the point of the integration, I think it's important that this follows best practices for GSPS creation. Generally speaking, a single GSPS should relate to an entire series, not one GSPS object per image or detection (as currently implemented). Generally speaking, if there are multiple GSPS objects for a series, a viewer (including eUnity) will only apply a single, manually-selected GSPS at a time. Therefore the current way things work would give rise to a pretty bad UI in the viewer. Therefore, I tweaked this to created a single GSPS for the entire series. This also has the effect of applying the windowing to the whole series.

[MMelQin]

• Despite the docs saying that a GPU is not required, the device for the inference part is actually hard coded as "cuda", meaning it won't work on CPU and will just crash. I fixed this. Unfortunately this does affect me as I cannot figure out how to get GPUs active within a Docker container when using the version of Docker compose that I have (if you have a quick pointer to a fix I'd appreciate it but I don't necessarily need a solution).

This may not be exactly what you are seeking for, but to achieve this we use NV container base, and run docker with GPU options (--gpus or --runtime=nvidia), and for Docker Compose by defining driver

[CPBridge]

This may not be exactly what you are seeking for, but to achieve this we use NV container base, and run docker with GPU options (--gpus or --runtime=nvidia), and for Docker Compose by defining driver

Thanks Min, but this hasn't been working for me. It's ok, no need to worry about his here.

Please rebase, as the inference.py has been updated in a previous PR by me to make the App SDK compositable with MONAI Core 0.9.1 and newer; for this one there is a small change in the pre-processing to remove the use of ToTensord

Rebased, and fixed some style issues (that bizarrely were not caught before)

[aihsani]

Ready to merge?

[MMelQin]

LGTM

[CPBridge]

@aihsani I made a couple of small tweaks to the README, as we discussed on slack. I think we are ready to merge

[sonarqubecloud]

Kudos, SonarCloud Quality Gate passed!   

0 Bugs
0 Vulnerabilities
0 Security Hotspots
0 Code Smells

No Coverage information
0.0% Duplication

[MMelQin]

LGTM