Automating CPAK Knee Phenotyping

Key takeaways

The Coronal Plane Alignment of the Knee (CPAK) classification, introduced by MacDessi and colleagues, sorts knees into nine phenotypes using just two derived measures: the arithmetic hip-knee-ankle angle (constitutional alignment) and joint-line obliquity. It has become the common language for personalized and kinematic alignment because it captures a patient's native coronal anatomy. Automating CPAK from imaging is attractive because the inputs are just two angles Salnus's engine already computes, but there is a catch: the underlying angles are highly reproducible while the phenotype classification is not, because knees near a boundary flip type with tiny measurement changes. A reliable automated CPAK therefore needs accurate angle measurement plus careful handling of the boundaries, ideally anchored to imaging that removes projection error. Salnus builds automated CPAK phenotyping into its planner, currently Research Use Only (RUO).

What CPAK is

CPAK is a 3-by-3 grid built from two numbers:

• Arithmetic hip-knee-ankle angle (aHKA) = MPTA minus LDFA, estimating constitutional (pre-arthritic) limb alignment as varus, neutral or valgus.
• Joint-line obliquity (JLO) = MPTA plus LDFA, describing whether the joint line is apex-distal, neutral or apex-proximal.

The nine combinations are the CPAK types. MPTA and LDFA are the medial proximal tibial angle and lateral distal femoral angle. Because both derived measures come from just these two angles, CPAK is a near-zero-marginal extension of any engine that already measures coronal alignment.

Why automation is attractive

CPAK is the phenotyping layer that tells you which alignment philosophy a knee is eligible for. Doing it by hand on a spreadsheet is slow and observer-dependent. Automated measurement from imaging is fast, consistent and scalable, and it is where deep-learning alignment tools already perform at or above human reproducibility.

The catch: reliable angles, unreliable classification

Here is the counter-intuitive part. Studies show that even with near-perfect angle reproducibility, the CPAK phenotype disagrees between readers about one time in five, because a knee sitting close to a boundary flips type with a fraction of a degree of change. Fixing this by making the angles even more precise is not realistic. The instability lives in the classification boundaries, and it concentrates in the joint-line term.

This has two implications for automation:

1. Report the angles with their uncertainty, not just a hard type label, so a near-boundary knee is flagged rather than silently forced into one bin.
2. Anchor the measurement to imaging that reduces projection error. A 3D reconstruction from CT removes the rotational and projection artifacts that push a 2D radiograph across a boundary. For how the 3D model is built, see our note on CT bone segmentation.

What a reliable automated CPAK needs

• Accurate, validated MPTA and LDFA measurement (the non-inferiority bar is roughly 1 degree of error against expert readers).
• Explicit handling of boundary cases, with a flag rather than a forced label.
• A consistent joint-line definition, since the joint-line term is the fragile one.
• Honest reporting of the human agreement ceiling as the reference, rather than a claim of superhuman accuracy.

The honest caveat

Automated CPAK is not a new claim of clinical superiority. It is a reproducibility and workflow tool: it standardizes a phenotype that is otherwise observer-dependent, and it should present uncertainty rather than false precision. Salnus is Research Use Only today, and CPAK output is a characterization to support planning, not a diagnosis or a treatment recommendation.

Bottom line

CPAK is easy to compute and hard to classify reliably, because the phenotype boundaries are unstable near discriminatory values. A good automated CPAK measures the angles accurately, flags boundary cases, and anchors to projection-error-free imaging. Salnus builds toward exactly this, as Research Use Only software.