Description
This relates to the lesion screening experiments done in the "Benchmarking off-the-shelf statistical shape modeling tools in clinical applications" paper.
Given a shape model trained on control subjects, a pathological sample can be represented in the context of the controls population using its closed-form, orthogonal projection onto the PCA subspace of controls. The lesion can then be detected by quantifying the deviation of the pathological shape from the shape reconstructed based on the model of controls. In lesion screening, lesions are not known a priori, and hence representing a pathologic sample with respect to the controls’ statistics should down-weight the lesion in the projection of the pathologic shape to reduce false positives in the lesion identification process.
A slack variable approach was formulated to identify the lesions and estimate the offsets in the lesions: Appendix B. Slack variables-based optimization
This should be included in studio to analyze models.
We have the python code available for this task. This can be included in ShapeWorks as an example or as a consolidated API for studio and python.