The AIHIP planning system included preoperative assessment and postoperative outcome simulations. Image segmentation was required to differentiate the femur from the pelvis. Featured anatomic landmarks were identified to serve as references. Then, the pelvis was corrected to a neutral position in the sagittal and coronal planes according to the identified landmarks. Assessment of preoperative deformity was also conducted referring to the identified landmarks. Postoperative outcome simulations showed the planned implant size, implant coverage and to what degree the preoperative deformity could be corrected. The comprehensive workflow was integrated into the AIHIP software. The accuracy of the AIHIP was validated through a clinical study. The research workflow is shown in Figure 1.

Over 1.2 million qualified CT images from 3,000 anonymized patients were included in this study. All patients were scheduled to receive total hip arthroplasty. The primary diagnosis included osteonecrosis, osteoarthritis, rheumatoid arthritis and developmental dysplasia of the hip. Standardized pelvic CTs were conducted prior to the operation. The range of each CT scan began from the highest point of the pelvis to 15 cm below the lesser trochanter at 1 mm intervals. All CTs were stored according to the DICOM protocol.

CT images were manually segmented by a group of engineers and three orthopedic surgeons using Mimics Software (Materialise NV, Leuven, Belgium). All engineers and surgeons had performed manual segmentation for at least 50 cases prior to this study. The contours of the femur and pelvis were manually annotated. Featured anatomic landmarks were manually marked, which included the anterior superior iliac spine (ASIS), pubic symphysis, center of the femoral head, medial edge of the lesser trochanter, and anatomic axis of the femur.

The complete dataset was randomly assigned to a training set, validation set and testing set at a ratio of 6:2:2. All images were resized to 512 × 512 pixels. The neural network structure was developed based on the attention U-Net with a point rend module. The U-Net convolutional neural network can automatically segment CT images and has achieved high accuracy in recognizing abdominal organs and tissues (16). The point rend module was used to provide point-based predictions to further enhance segmentation performance (17). An attention U-Net was developed based on a U-Net with added skip connections and attention gates (18). The use of an attention gate enables the network to automatically focus on target structures without requiring large computational power and model parameters. A skip connection was conducted between the corresponding encoder and decoder layers. Implementing skip connections provided segmentation results with a higher level of accuracy because more graphic features from the basic level were preserved and integrated into the output feature. The increased number of decoders also provides a more detailed segmentation (Figure 2A). The Dice similarity coefficient (DSC) and loss was used to assess the model performance of the AIHIP in the segmentation of CT images. DSC and loss were calculated for every 100 iterations.

An example of manual identification of feature anatomic landmarks and correction of pelvis is shown in Figure 3A. Based on the segmented pelvis, the automatic recognition of featured anatomic landmarks was conducted with a stacked hourglass network, which has been used in human pose estimation (19). Repeated bottom-up and top-down inference is beneficial in predicting the coordinates of the featured anatomic structure. The stacked hourglass network automatically recognized featured anatomic structures, including the bilateral anterior superior iliac spine (ASIS), the pubic symphysis and the center of the femoral head (Figure 3B). The least square method was used to determine the anatomic axis of the femur. The pelvis could be adjusted, and the preoperative deformity could be measured based on these identified points (Figure 3C).

Preoperative planning was conducted by two orthopedic surgeons using AIHIP software (Version 3.0, Longwood Valley Technology, China). Planning was carried out by first determining the position and size of the acetabular component. Inclination, anteversion, and coverage of the acetabular component were planned as well. Then, the position and size of the femoral component were determined, and the level of femoral resection was determined. Different types of acetabular components, femoral components, and femoral heads could be chosen. A simulation of the postoperative effect was generated, which showed the postoperative leg length, offset and coverage of the acetabular component. Leg length and offset of the contralateral side were also shown so that changes in the surgical plan could be made accordingly. The distance between the tip of the lesser trochanter and the tip of the femoral stem (neck length) and the distance between the tip of the lesser trochanter and the resection line of the femoral neck (calcar length) were measured in preoperative planning as references. Then, the operating surgeon measured the neck length and calcar length intraoperatively for verification (Figures 4A–C).

Approval from the Institutional Review Board and written informed consent was acquired to conduct a prospective clinical study from October 2019 to February 2021. Patients were included if they 1. were diagnosed with osteonecrosis, osteoarthritis and developmental dysplasia of the hip (Crowe I) and received THA; 2. provided written informed consent to participate in the study. Patients were excluded if 1. the preoperative or postoperative radiographs were not standardized or 2. different types of prostheses were used during surgery. A total of 120 cases were included. X-ray-based planning was conducted in 60 cases (control group) according to the method described by Della Valle et al. (3), where planning was completed directly over the printed X-ray using templates. AIHIP planning was conducted in 60 cases (AIHIP group).

Leg length discrepancy, offset, neck length, and calcar length were measured on postoperative radiographs with the patients’ names concealed. Each measurement was made and recorded by the two observers at least 4 weeks after the operation to avoid any recollection bias. The mean value of the two measurements was used for statistical analysis. Inter-observer reliability of radiographic measurement was assessed with intraclass correlation coefficient (ICC).

Functional outcome was assessed by Hip Disability and Osteoarthritis Score Joint Replacement (HOOS JR) (20) and EuroQol 5 Dimensions Questionnaire (EQ-5D) (21, 22). Patients were followed up until 12 weeks postoperatively. Surgical time and blood loss were also recorded.

All patients underwent total hip arthroplasty through a posterior approach by one experienced orthopedic surgeon in one facility. The prostheses used were Pinnacle Cup (DePuyOrthopaedics, Warsaw, IN, United States), Corail Stem (DePuyOrthopaedics, Warsaw, IN, United States) and Trilock stem (DePuyOrthopaedics, Warsaw, IN, United States). Standard perioperative care and patient education were administered to all patients.

Statistical analysis was performed with SPSS version 25 (IBM, New York, NY, United States) and GraphPad Prism version 8 (GraphPad Software, San Diego, CA, United States). According to previous literatures (23–25), accurate prediction was defined as the predicted size to be within ± 1 size from the implanted size. Absolute error was defined as the absolute difference between planned and implanted component size and the difference between plan and postoperative radiographic measurement. Mean error was defined as the average value of the planned component size minus the implanted component size. A p-value less than 0.05 was considered statistically significant. Discontinuous variables were recorded as incidence and rate. The chi-square test was used to compare the discontinuous variables between groups. Continuous variables were recorded as the means and standard deviation. A general linear model was used to test whether there was a statistically significant difference between the two groups considering confounding factors, including age, sex, BMI, and primary diagnosis.

The effect of manual segmentation and AI segmentation from four common primary diagnoses are shown in Figure 2B. The DSC curves and loss from the training set and validation set are shown in Figure 2C. Both curves reached convergence by 15,800 iterations, which indicated optimal DSC and loss. At 15,800 iterations, the DSC of the training set was 0.983, and the DSC of the validation set was 0.987. The losses were 0.008 and 0.013 for the training set and validation set, respectively (Figure 2C). The testing set was used to validate algorithm performance. The testing set achieved a DSC of 0.973 and loss of 0.012, which was comparable to that of the training set and validation set.

The average time consumption was 0.99 ± 0.94 min for AIHIP segmentation and 0.87 ± 0.07 min for AIHIP correction and deformity assessment. The average time was 124.55 ± 16.87 min for manual segmentation and 60.85 ± 11.11 min for manual correction and deformity assessment (Figure 2D). The total time required for the AIHIP algorithm to process the CT for one case was 1.86 ± 0.12 min on average, compared with 185.40 ± 21.76 min for the manual workflow (P < 0.001).

A total of 120 cases were included in the study. The demographic characteristics are listed in Table 1, which include age, sex, height, weight, BMI, and primary diagnosis. The difference in age between the control group (mean: 53.75, range: 24–78) and the AIHIP group (mean: 47.62, range: 23–78) was statistically significant (P = 0.033). There were no statistically significant differences in terms of sex, weight, height, or BMI. Demographic characteristics were considered as confounding variables. Their influences on the prediction accuracy were assessed and adjusted with generalized linear models.

The predicted cup size and implanted cup size were exactly the same in 66.67% of the AIHIP cases and 20% of the control cases (P < 0.001). For femoral stem, the exact same size was achieved in 55% of the AIHIP cases and 31.67% of the control cases. The cup size was accurately predicted to within ± 1 size in 55.00 and 96.67% of patients in the control group and AIHIP group, respectively (P < 0.001). Stem size was accurately predicted to within ± 1 size in 65.00 and 96.67% of the control group and AIHIP group, respectively (P < 0.001). Further analysis showed more detailed comparison in Figures 5A, B. The tendency toward overestimation or underestimation of component size was assessed by mean error. Comparing with AIHIP planning, acetate templating tended to underestimate cup size by 2.13 ± 2.11 (P < 0.001) and underestimate the stem size by 0.53 ± 1.28 (P = 0.013). The comparison of the mean absolute error between the two groups is shown in Table 2. Compared with the control group, high offset/varus stems were more commonly used in the AIHIP group (P = 0.004) (Figure 5C). The average time it took to conduct X-ray planning was 7.37 ± 1.32 min and the average time it took to conduct AIHIP planning was 8.11 ± 0.98 min (P = 0.001).

Neck length, calcar length, LLD, and offset were measured on radiographs. The difference between preoperative planning and postoperative radiographic measurements was recorded as the mean absolute error and is shown in Table 2. The ICCs for all radiographic measurement were above 0.9, which indicated substantial inter-observer agreement (Table 3). The mean absolute error for neck length was 6.13 ± 3.16 mm in the control group and 5.49 ± 4.40 mm in the AIHIP group (P = 0.813). The mean absolute error of calcar length was 4.51 ± 2.96 mm in the control group and 3.92 ± 2.79 mm in the AIHIP group (P = 0.249). The average postoperative LLD for the control group and AIHIP group was 5.68 ± 4.06 mm and 5.03 ± 3.67 mm, respectively (P = 0.360) (Figure 5D). Although a trend was observed in the above radiographic outcomes, the differences were not statistically significant. Changes in acetabular offset and global offset were not significantly different between the two groups. Femoral offset was more accurately restored in the AIHIP group than in the control group (P = 0.001) (Figure 5E).

The average operation time was 106.83 ± 18.20 min in the AIHIP group and 109.58 ± 21.98 min in the control group (P = 0.457) (Figure 5F). The average blood loos was 285.00 ± 127.33 ml in the AIHIP group and 315.67 ± 164.68 ml in the control group (P = 0.256). There were no statistically significant differences in HOOS score preoperatively (P = 0.605) and 12 weeks postoperatively (P = 0.22) between the two groups. There were no statistically significant differences in EQ5D index preoperatively (P = 0.846) and 12 weeks postoperatively (P = 0.203) between the two groups.