Deep Learning-Based Full-Process Automatic CPAK Classification System and Its Application in the Analysis of...

Background/Objectives: Coronal Plane Alignment of the Knee (CPAK) classification enables individualized alignment assessment in total knee arthroplasty (TKA), yet manual evaluation is time-consuming and lacks preoperative-to-postoperative transition analysis. Methods: This retrospective, single-center study aimed to develop and validate a fully automated deep learning-based CPAK classification system using internal validation on a held-out test set (n = 92) and to investigate individual-level transition patterns and their association with short-term clinical outcomes using paired radiographic data from a large Chinese cohort. A total of 919 KOA patients undergoing TKA were analyzed. A keypoint detection model (HRNet-W32) was developed to automatically calculate the medial proximal tibial angle, lateral distal femoral angle, arithmetic hip-knee-ankle angle, and joint line obliquity, from which CPAK types were derived. Results: On the validation set (92 cases), the model achieved a Mean Radial Error of 1.22 ± 0.43 mm for keypoint detection; mean absolute errors for MPTA and LDFA were ≤0.74°, while for aHKA and JLO they were 0.91° and 1.12°, respectively, with intraclass correlation coefficients ≥0.96 compared to manual annotations. Automatic CPAK classification accuracy was 80.98% (kappa = 0.767). Transition matrix analysis showed that only 9.36% of all patients maintained their original type postoperatively, with most shifting to types IV, V, or VII. After inverse probability weighting, no significant differences in clinical outcomes were observed among transition groups (all adjusted p > 0.05). Conclusions: These results demonstrate that the proposed automated system enables efficient CPAK assessment, revealing substantial postoperative alignment transitions that were not associated with differential short-term outcomes, thereby supporting AI-assisted individualized alignment planning in TKA.