Automated CPAK Alignment Planning: What AI Does

Salnus runs its preoperative planning entirely client-side and implant-agnostic: DICOM images are parsed and analyzed on your own machine, and they never leave it. That matters most for a measurement task that surgeons repeat on nearly every knee: CPAK classification. This pillar page explains what CPAK is, how the four main alignment philosophies relate to it, and how AI can automate CPAK measurement from imaging. Salnus offers automated CPAK as a distinctive Research Use Only capability, one that most planning tools compute but rarely name.

What Is CPAK (Coronal Plane Alignment of the Knee)?

The Coronal Plane Alignment of the Knee (CPAK) classification was introduced by Samuel MacDessi and colleagues in 2021 in The Bone & Joint Journal. It is a pragmatic system for describing knee phenotypes in the coronal plane, and it works for both healthy and arthritic knees. The original study analyzed 500 healthy and 500 osteoarthritic knees to validate the framework.

CPAK organizes every knee along two independent variables:

• Arithmetic hip-knee-ankle angle (aHKA): an estimate of a patient's constitutional (pre-arthritic) limb alignment, calculated as aHKA = MPTA minus LDFA. It is grouped into varus, neutral, and valgus.
• Joint line obliquity (JLO): the orientation of the joint line relative to the floor in double-leg stance, calculated as JLO = MPTA plus LDFA. It is grouped into apex distal, neutral, and apex proximal.

Both variables derive from the same two coronal-plane angles: the medial proximal tibial angle (MPTA) and the lateral distal femoral angle (LDFA). If you want a refresher on how those landmark angles are defined and measured, see our guide to mechanical axis alignment with LDFA, MPTA, and HKA.

For JLO, a sum near 180 degrees indicates a roughly neutral joint line. A sum greater than 180 degrees indicates an apex proximal joint line, while a sum less than 180 degrees indicates an apex distal joint line.

The Nine CPAK Phenotypes

Placing the three aHKA subgroups against the three JLO subgroups produces a 3-by-3 matrix of nine knee phenotypes.

The distribution is not uniform. Type I (varus, apex distal) is the most common phenotype in many populations. In the original healthy versus arthritic comparison, Type I accounted for roughly a quarter of normal knees and became substantially more prevalent in arthritic knees, and it is especially common in Asian cohorts where constitutional tibial varus is more frequent. Reporting a knee's CPAK type gives surgeons a shared vocabulary for how a given knee is built, before any decision is made about how to reconstruct it.

CPAK and the Four Alignment Philosophies

CPAK describes anatomy. Alignment strategy is a separate choice about how to position implants relative to that anatomy. Knowing a knee's CPAK type helps a surgeon reason about which strategy fits, and what soft-tissue behavior to expect.

Mechanical alignment aims for a neutral hip-knee-ankle angle (roughly 180 degrees, plus or minus 1.5 degrees) with orthogonal cuts, followed by soft-tissue release to balance the gaps. Kinematic alignment instead tries to restore the patient's pre-arthritic anatomy so implants track the native kinematic axes. Restricted kinematic alignment keeps that anatomic intent but caps varus-valgus so extreme alignments are avoided. Functional alignment adjusts bone resections within limits to achieve balanced gaps and equal soft-tissue tension, minimizing releases, and is commonly executed with robotic assistance and intraoperative gap measurement.

Randomized and systematic comparisons of these strategies show statistically detectable differences in some patient-reported scores, but many between-group differences fall below accepted minimal clinically important difference thresholds. In other words, the alignment debate is active and unresolved. CPAK does not settle it; it gives every camp a common, measurable starting point.

How AI Automates CPAK Measurement and Classification

Traditionally, CPAK requires a surgeon to place landmarks on a long-leg standing radiograph, measure LDFA and MPTA, then compute aHKA and JLO and read off the phenotype. Done manually, this is time-consuming and prone to inter-observer variability.

AI changes the workflow. Published work has demonstrated fully automated CPAK pipelines. For example, a 2026 deep-learning study built a keypoint-detection model that automatically located the anatomical landmarks, computed MPTA, LDFA, aHKA, and JLO, and derived the CPAK type. On a held-out validation set of 92 cases it reported mean absolute errors at or below roughly 0.74 degrees for MPTA and LDFA and near 0.9 to 1.1 degrees for aHKA and JLO, with intraclass correlation coefficients at or above 0.96 against manual annotation. Separate studies have shown that automated software can measure full-leg standing radiographs with high reproducibility.

The AI pipeline for automated CPAK generally follows these steps:

1. Landmark or keypoint detection: a model localizes the hip center, knee landmarks, and ankle center on the image.
2. Angle computation: MPTA and LDFA are calculated from those landmarks, then aHKA (MPTA minus LDFA) and JLO (MPTA plus LDFA) are derived.
3. Phenotype classification: the aHKA and JLO values are binned into varus/neutral/valgus and apex distal/neutral/proximal to assign one of the nine CPAK types.

The clinical value is consistency and speed. Automated measurement removes the manual protractor step, standardizes how angles are computed, and makes it practical to report a CPAK type on every case rather than only on complex ones. The value of any such system still depends on validation against expert readers on the imaging and population it is used with.

Salnus and Automated CPAK: An RUO Capability

Salnus is building automated CPAK classification into its browser-based, client-side preoperative planning workflow as a Research Use Only capability. From a CT-derived model, Salnus derives the coronal-plane angles that CPAK depends on, then reports the aHKA, the JLO, and the resulting phenotype. Because the same CT bone segmentation that produces the 3D model also anchors the anatomical landmarks, the alignment read-out is a natural extension of the planning pipeline rather than a separate manual task.

Several design choices make this distinctive:

• Named, not just computed. Many planning tools calculate coronal angles internally but never surface a CPAK phenotype. Salnus reports the CPAK type explicitly, so the surgeon gets a shared classification vocabulary, not just raw numbers.
• Client-side. DICOM images are parsed and the analysis runs on-device (via Cornerstone3D and ONNX Runtime Web). Images never leave the machine, which means scan to plan in minutes without a cloud upload step. For a deeper look at that trade-off, see our comparison of cloud versus client-side medical AI.
• Implant-agnostic and vendor-neutral. Salnus does not favor a particular implant system, which is why the same alignment read-out is useful to surgeons and to OEM partners evaluating fit across catalogs.

Salnus's peer-reviewed foundation is a controlled laboratory study, not a clinical outcomes trial. Cirdi and colleagues (including B. Serteser and U. Akgun) showed that 3D-printed patient-specific guides reduced femoral tunnel convergence in anatomic multiligament knee reconstruction, published in the Orthopaedic Journal of Sports Medicine in 2026. That study validates the accuracy of Salnus's patient-specific guide workflow on the bench; it does not, on its own, establish clinical outcomes for CPAK-based planning. You can read more about that work in our summary of the OJSM patient-specific instrumentation study, and about our related work on soft-tissue-aware planning in the PCL as a forgotten ligament in AI research.

To be clear about status: Salnus tools are RUO and pilot-stage. They are not cleared or approved medical devices, and nothing here should be read as a regulatory or clinical-equivalence claim.

FAQ

What does CPAK stand for?

CPAK stands for Coronal Plane Alignment of the Knee. It is a classification system introduced by MacDessi and colleagues in 2021 that sorts knees into nine phenotypes based on constitutional alignment (aHKA) and joint line obliquity (JLO).

How are aHKA and JLO calculated?

Both come from two coronal-plane angles, the medial proximal tibial angle (MPTA) and the lateral distal femoral angle (LDFA). The arithmetic HKA is MPTA minus LDFA, and JLO is MPTA plus LDFA. A JLO sum near 180 degrees is a roughly neutral joint line; greater than 180 is apex proximal, and less than 180 is apex distal.

Is CPAK the same as an alignment strategy like kinematic alignment?

No. CPAK describes a knee's anatomy. Mechanical, kinematic, restricted kinematic, and functional alignment are strategies for positioning implants relative to that anatomy. Knowing the CPAK type informs the strategy choice, but it does not dictate it.

Can AI classify CPAK automatically?

Yes. Published deep-learning pipelines detect anatomical landmarks, compute MPTA, LDFA, aHKA, and JLO, and assign the CPAK phenotype automatically, with reported agreement close to expert manual measurement. Accuracy still depends on validation against expert readers for the specific imaging and population.

Does Salnus offer automated CPAK, and is it a cleared device?

Salnus provides automated CPAK classification as a Research Use Only capability inside its client-side, implant-agnostic planning workflow. It is pilot-stage and not a cleared or approved medical device. Regulatory status should be verified independently for your jurisdiction.

Why does client-side processing matter for CPAK planning?

Because DICOM images are parsed and analyzed on your own device, they never leave the machine, and there is no cloud upload step. That supports a scan-to-plan-in-minutes workflow while keeping imaging local.

Explore the Salnus Surgeon Portal →

---

Disclaimer: This article is for educational and research purposes only. Salnus tools are designated for Research Use Only (RUO) and are not cleared medical devices. Mention of third-party products is for educational context only and does not imply endorsement or comparison of clinical equivalence. Clinical decisions should be made by qualified physicians, and regulatory status should be independently verified for your jurisdiction.

References:

• MacDessi SJ, Griffiths-Jones W, Harris IA, Bellemans J, Chen DB. Coronal Plane Alignment of the Knee (CPAK) classification: a new system for describing knee phenotypes. The Bone & Joint Journal. 2021;103-B(2):329-337. DOI 10.1302/0301-620X.103B2.BJJ-2020-1050.R1.
• Cirdi YU, Serteser B, Mavi A, Ergun S, Akgun U. 3D-Printed Patient-Specific Guides Reduce Femoral Tunnel Convergence in Anatomic Knee Multiligament Reconstruction: Controlled Laboratory Study. Orthopaedic Journal of Sports Medicine. 2026. DOI 10.1177/23259671261417360.
• Deep Learning-Based Full-Process Automatic CPAK Classification System and Its Application in the Analysis of Alignment Outcomes Before and After Knee Arthroplasty. Diagnostics. 2026. DOI 10.3390/diagnostics16091389.
• Current concept of kinematic alignment total knee arthroplasty and its derivatives. Bone & Joint Open. 2022. Overview of mechanical, kinematic, restricted kinematic, and functional alignment.
• Evaluation of a deep learning software for automated measurements on full-leg standing radiographs. PMC. On the reproducibility of automated long-leg radiograph angle measurement.
• A systematic review of geographic differences in knee phenotypes based on the CPAK classification. PMC. On CPAK phenotype distribution across populations.