Deep Learning for Automated Measurement of Patellofemoral Anatomic Landmarks

Background: Patellofemoral anatomy has not been well characterized. Applying deep learning to automatically measure knee can provide a better understanding of anatomy, which be key factor in improving outcomes. Methods: 483 total patients with CT imaging (April 2017–May 2022) from 6 centers were selected cohort scheduled for arthroplasty and healthy anatomy. A 7 patellofemoral landmarks annotated on 14,652 images approved by senior musculoskeletal radiologist. two-stage model was trained predict landmark coordinates using modified ResNet50 architecture initialized self-supervised pretrained weights RadImageNet. Landmark predictions evaluated mean absolute error, derived measurements analyzed Bland–Altman plots. Statistical significance assessed paired t-tests. Results: Mean error between predicted ground truth 0.20/0.26 cm the healthy/arthroplasty cohort. Four parameters calculated, including transepicondylar axis length, transepicondylar-posterior femur angle, trochlear medial asymmetry, sulcus angle. There no statistically significant parameter differences (p > 0.05) both cohorts, except Conclusion: Our accurately identifies ~0.20–0.26 accuracy produces human-comparable pathological knees. This work represents first regression automated annotation physiologic pathologic at this scale. novel enhance our ability analyze compartment