Explore: Quantify error when the two x-rays are misaligned

[msrepo]

Description

See Limitations section of 3D reconstruction of proximal femoral fracture from biplanar radiographs with fractural representative learning

• second view needs to be precisely 90 degree from the first view or else the error increases(obvious, because the model was only trained on perfectly aligned inputs)
  This is quantified here
  

they give reasons why it may not always be possible to get perfectly aligned views

• some view such as judet view are hard to obtain biplanar xray than standard AP/LAT views
• if the x-ray machine does not support lateral head tilt, the patient has to lie down in 45 degree slope to obtain judet view (more here Selecting the correct X-ray tube tilt angle and roof pillar rotation for bedside radiography with combined cranio-caudal and lateral cassette tilt

Tasks

• Recreate the above diagram for our benchmarked architectures
• Train in a more robust way by including these misaligned inputs during training, and see if things improve

[msrepo]

Interesting followup: Representing Input Transformations by Low-dimensional parameter subspaces

[msrepo]

Can we run a quick experiment to learn the desired robustness via data augmentation?

[msrepo]

• Generate data for perturbation by 1,2.5,5,10 degrees
• Create csv list for testpaths
• see Domain shift evaluation workflow for example

[msrepo]

upto +/- 10 degree perturbation of lat view

[msrepo]

Use pandas pivot() to format and work with the perturbation csv