Abstract: BACKGROUND: Surgical decision-making for posterior cruciate ligament avulsion fractures is highly dependent on imaging evaluation. Traditional methods rely on subjective interpretation of CT images, which suffer from limitations such as difficulties in quantifying three-dimensional spatial displacement parameters and insufficient precision in assessing rotational angles. Given the advancements in artificial intelligence (AI) technology, there is a need to develop automated, intelligent image recognition software based on AI algorithms.
OBJECTIVE: To investigate the intelligent diagnostic capabilities of AI algorithms for posterior cruciate ligament tibial avulsion fractures in 3D CT images and their effectiveness in accurately assessing 3D parameters of fracture fragments.
METHODS: Knee CT data from 24 patients with posterior cruciate ligament tibial avulsion fractures who were treated at the Wangjing Hospital of the China Academy of Chinese Medical Sciences between December 1, 2022, and August 30, 2024, were retrospectively collected. Three-dimensional reconstruction, intelligent fracture point recognition, and simulated reduction were performed using self-developed AI image recognition software. Translational and rotational parameters of the fracture fragments along the X, Y, and Z axes were obtained. The results were compared with those obtained using picture archiving and communication system (PACS system). The consistency of the two methods was assessed using the rank sum test, Bland-Altman analysis, and linear regression model. The coefficient of variation was calculated to assess software stability.
RESULTS AND CONCLUSION: (1) No statistically significant differences were observed in fracture fragment displacement (translation/rotation along X/Y/Z axes) between AI and traditional methods (P > 0.05). (2) Bland-Altman analysis indicated good consistency between the two methods, with no significant differences (P > 0.05). (3) The R² values of the linear regression models for X, Y, and Z axis displacement and angle fitting were all > 0.99. (4) The coefficient of variation of the AI software for three repeated fracture point identifications showed that the coefficient of variation was < 20% for 21 imaging samples for total fracture identification and < 20% for 18 imaging samples for articular surface fracture identification. (5) This indicates that the AI image recognition software can accurately quantify the three-dimensional parameters of posterior cruciate ligament avulsion fracture fragments. Its measurement results are consistent with those of traditional methods and are highly stable. It can assist physicians in determining the degree of displacement and provide accurate data support for preoperative planning. This software has promising application prospects in the treatment of avulsion fractures. Future studies should expand the sample size and validate its impact on surgical outcomes.

Key words: posterior cruciate ligament avulsion fracture, artificial intelligence, CT image recognition, three-dimensional measurement, preoperative planning