AI-based biplane X-ray image-guided method for distal radius fracture reduction.

Background: In the course of manual reduction of distal radius fractures, many doctors rely on tactile perception to assess the displacement of the fracture. However, a more accurate determination of the severity of the fracture and the success of the reduction requires measurement software to annotate the fracture images, which is difficult to achieve real-timely in actual procedure of reduction. Which may lead to misdiagnosis when experienced doctors rely on their intuition. Therefore, it is necessary to develop an AI-based method for calculating fracture parameters to provide real-time display, particularly in fracture reduction machines. Methods: An AI-based method for automatically calculating of radiographic parameters in distal radius fractures (DRF) was developed. Initially, anteroposterior (AP) and lateral (LAT) X-ray images of patients with distal radius fractures were collected from three hospitals and preprocessed. Subsequently, several neural network structures, UNet, DeeplabV3+, PSPNet, and TransUNet, are compared in terms of utility and accuracy, and finally, the models obtained from the UNet image segmentation algorithm are used for semantic segmentation of the radius and ulna. Following this, the contours of the radius and ulna were extracted using OpenCV, key points were detected, and the principal axes were calculated. Finally, the computed parameters including radial angle (RA), radial length (RL), ulnar variance (UV), and palmar tilt (PT) were calculated and displayed on the image. Results: The advantages and disadvantages of several models were considered, and finally the UNet neural network model was used as the core algorithm of the image segmentation model in this study. The segmentation accuracy for the radius and ulna in the AP and LAT X-ray images reached 91.31% and 88.63%, respectively. The average errors between the automated calculations of parameters RA, RL, UV, and PT and the manually annotated results by physicians were -1.36°, -1.7 mm, 0.66 mm, and -1.06°, respectively. The system has been initially deployed on the same computer that operates the radial fracture fracture repositioning robot. Conclusion: The automated parameter calculation method developed in this study accurately computes diagnostic parameters for assessing distal radius fractures and can be utilized in the image-guided reduction process of fracture rehabilitation robots. This method has the potential to evolve into an intelligent diagnostic tool for physicians, thereby enhancing the accuracy of distal radius fracture diagnosis.