Open lassoan opened 6 months ago
Thanks for pointing this, @lassoan This https://github.com/lassoan/SlicerMONAIAuto3DSeg/pull/39 should solve the issue.
UPDATE: This is still and issue. Removed the label arg in this PR: https://github.com/lassoan/SlicerMONAIAuto3DSeg/pull/60
MONAI Auto3DSeg prostate model only works well if the IJK axes of the input are approximately aligned with the patient RAS axes. This is a bug, because the orientation of the IJK coordinate system can be freely chosen by applications. All that an application guarantees is that the image is oriented correctly in physical (LPS) coordinate system.
It seems that the prostate model (but maybe others, too?) does not appropriately reorder slices, rows, or columns, but simply ignores the IJK to LPS transform.
This is proven by the test images created from ProstateX-0000 T2 and ADC images: https://github.com/lassoan/PublicTestingData/releases/download/data/prostate-orientation.zip The images are in the same physical position but with the IJK axes oriented in different ways. Only the images in
Good-straightened-flipped
folder provide correct segmentation results.