Pull Request Overview

This PR intends to remove pinning monai<=1.5.0, after having identified, analyzed and found resolutions to the issues in MONAI (more precisely errors in the older MONAI models' inference config for mixing up patient anatomical coordinate system, RAS vs LPS)

• The PR #8473 to address the orientation issue is included in MONAI v1.5.1, and it enables the proper handling of image/Metatensor using LPS+ anatomical coordinate system, e.g. DICOM Radiology images. The App SDK had been setting this metadata space correctly, so no changes needed except for moving it to the DICOM series to volume operator.
• The example apps fail to get good segmentation because the older MONAI models' configs have orientation transform to RAS in the pre-processing transforms. They can get correct seg result after correcting the model's pre-processing transforms. Detailed analysis in [BUG] Examples using older MONAI Bundles not compatible with MONAI v1.5.1 (which reveals errors in older MONAI bundles) #557.
• MONAI v1.5.1 supports torch 2.8 which has embedded CUDA Runtime that is compatible with Holoscan SDK v3.6, so now technically MD App SDK is compatible with HSDK v3.6.0.
• Tested example apps that could and did change inference config to target orientation LPS instead of RAS, and obtained correct inference results.

Note

• Even though technically the App SDK works with MONAI 1.5.1 and Holoscan 3.6.0, documentation build fails if requiring Holoscan 3.6.0 because the former needs Python 3.9 while the latter requires min Python be 3.10. So, the SDK still pins Holoscan SDK <=3.5.0.
• Example apps using MONAI Bundle (without means to alter its inference config for the pre-processing orientation to target LPS) would fail because the models in MONAI Model Zoo and Huggingface still have RAS orientation in their inference configs.