In general, I would recommend to ask images from your clinical collaborators. They can provide not just images but also help interpreting them and ensure you are developing clinically useful solutions.
If you just want to find some random images then you can browse large image databases, such as TCIA. Since it is a database for cancer cases, you won’t find a lot of knee and shoulder MRIs, but there should be some.
You can also check out the visible human project data sets. The image acquisition protocols are quite old and do not reflect current musculoskeletal imaging protocols used for shoulder/knee MRI nowadays, but may be OK for quick initial tests.