7951 anchors are not centered on grid cells

[hachoj]

This pull request resolves issue #7951 by correctly centering anchor boxes within their grid cells.

Description

Changes Made:

• Modified the anchor generation logic in monai/apps/detection/utils/anchor_utils.py to add a stride // 2 offset.
• Refactored the corresponding unit test in tests/apps/detection/utils/test_anchor_box.py to validate this new, correct behavior against the torchvision baseline by accounting for the offset.

Fixes #7951

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests --disttests.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.
• [ ]

[ericspod]

Hi @hachoj thanks for the contribution, it looks fine to me as far as I can tell. Could you please check that this works with the luna16 tutorial example? If so I think we're good. @Can-Zhao I think you were the original author of this file, any thoughts?

[Cado87]

Hi @ericspod

I ran a smoke test of the LUNA16 detection tutorial pipeline with this patch applied (MONAI commit 8f6e2c7, PyTorch
2.10.0, Tesla T4 GPU).

What was tested:

• Verified the anchor fix is active: with a 2×2×2 grid and stride=4, the first anchor center is now at (2.0, 2.0,
  2.0) instead of (0.0, 0.0, 0.0)
• Ran luna16_training.py (fold 0, 3 epochs, synthetic volumes matching the fold-0 datasplit paths) — training
  completed without errors, loss was finite and decreasing
• Ran luna16_testing.py on the saved checkpoint — inference processed 67 images, produced 94 predicted boxes with
  scores in a valid range

The full pipeline (data loading → ATSS matching → training → sliding window inference → box post-processing) runs
without issues with the centered anchors.

Note: since real LUNA16 data was not used, detection quality metrics are not meaningful here. The test only
validates that the fix does not break the tutorial pipeline. With real data, a full training run would be needed to
confirm the expected improvement in detection accuracy from properly centered anchors.