Knock-knee diagnosis in Chinese adolescents: Expert evaluation and defensive strategies in image analysis - A population study.

Abstract

Background: Knock-knee, a prevalent postural deformity problem among adolescents, poses significant challenges to traditional diagnostic methods in terms of complexity, high cost, and radiation risk. Therefore, there is a demand for diagnostic techniques that are more accessible, safe, and non-invasive for knock-knee.

Methods: We collected 1519 clear whole-body images from 1689 Chinese adolescents aged 10-19 years as image data, and obtained expert annotations on the presence or absence of knock-knee from three orthopedic surgeons. Utilizing Real-Time Multi-Person Pose Estimation (RTMpose), we manually extracted ten features related Knock-knee to construct the dataset. Regard to model, we employed a defense strategy called BitSqueezing.

Results: The proposed model achieved an accuracy of 72.81%, a recall of 62.12%, and an AUC of 76.12%, outperforming the benchmark model that achieved an accuracy of 62.45%, a recall of 43.35%, and an AUC of 76.17%.

Conclusion: The proposed model is a promising non-contact, high-performance knock-knee detection method that can overcome the limitations of traditional diagnostic methods. The proposed model can facilitate more accurate and efficient deformity detection and postural correction in adolescents. The proposed model also demonstrates the effectiveness of adversarial defense in improving the reliability and accuracy of pose estimation tasks. Future work should validate the proposed model in larger and more diverse populations, and explore other applications of pose estimation and adversarial defense in deformity detection.