Preoperative planning for THA has evolved from the widely used 2D digital template to 3D software planning, with a gradual improvement in the accuracy of prediction. However, most of the software operation procedures of 3D planning are still complicated and time-consuming at present. In this study, the AI HIP was used for preoperative planning of patients undergoing primary THA. The accuracy of acetabular cup and femoral stem prediction was 74.58% (44 of 59) and 71.19% (42 of 59), respectively, and the average operation time was 3.91±0.64 min. Compared with the 3D mimics, the accuracy was similar, but the operation time was greatly shortened, making the operation more convenient. Compared with the traditional 2D digital template, its accuracy was much higher, with a slightly longer operation time than that of the digital template, but it can provide a 3D perspective that cannot be provided by the 2D digital template. In this study, AI technology was used for preoperative planning of THA for the first time, and was compared with the accuracy and time efficiency of 3D and 2D planning. Combined with the current era of big data, the clinical application value of this technology in preoperative planning was established.
In order to highlight the objectivity of AI HIP, 3D mimics and 2D digital templates were planned and compared. In previous studies on 3D planning, Inoue et al.  conducted a retrospective study on 65 hips of 57 patients with a total of 65 hips using 3D planning software-zed-hip. The results showed that the complete accuracy of the femoral stem was 65%, 98% in one size, and 92% in the acetabular cup, and 100% in one size. Similarly, using the 3D planning software-zed-hip, Erik et al. conducted a retrospective preoperative planning study on 116 patients. The results showed that the complete accuracy of the femoral stem was 58.6%, 94% in one size, and 56.9% in the acetabular cup and 86.2% in one size . The accuracy rate of 3D planning in the preoperative planning of complex hip joint diseases is also high. For example, Zeng et al. 3 D mimics was used to study the preoperative planning of 20 cases of DDH patients with high dislocation. The results showed that the complete accuracy rate of acetabular cup was 70%, and that of one size was 100%, while the results of the traditional 2D template showed that the complete accuracy rate of the acetabular cup was 25%, and that of one size was 45% . Similarly, peihui et al.  also used mimics to plan the acetabular cup in 41 patients with 49 hip DDH. The results showed that the complete accuracy was 71%, and 100% within a size. It can be seen that the above 3D preoperative planning related research results show a high accuracy rate, and has a great advantage in the planning of complex hip diseases. This is because 3D planning software provides different 3D perspectives. In our study, the results of 3D mimics planning showed that the complete accuracy of acetabular cup was 71.19%, and that of one size was 93.22%. The complete accuracy of the femoral stem was 76.27%, and that of one size was 93.22%, which was similar to the previous 3D planning results. The accuracy of the AI HIP also reached the prediction level of this 3D software (cup: complete accuracy was 74.58%, in one size was 94.92%; stem complete accuracy was 71.19%, in one size 93.22%). In previous studies on 2D digital templates, the accuracy of prosthesis prediction was low due to the influence of many factors. Strøm et al.  used a double-blind method to perform preoperative planning for 41 patients. The results showed that the complete accuracy rate of acetabular cup was 7.3%, that of one size was 41%, that of femoral stem was 34%, and that of one size was 76%. Shaarani et al.  carried out preoperative planning for 100 patients with the same digital template. The results showed that the complete accuracy rate of the acetabular cup was 38%, one size was 80%, and the complete accuracy rate of the femoral stem was 36%. Since the adjacent model of the accolade stem was 0.5, the internal size of one was 75%. Schiffner et al.  compared 3D and 2D planning methods, and the results showed that the complete accuracy rate of 2D planning acetabular cup was 44.8%, 45.7% for femoral stem, 56.9% for 3D planning of acetabular cup, and 58.6% for femoral stem. The above studies on 3D and 2D showed that the accuracy of 3D planning was significantly higher than that of 2D planning. In our 2D digital template planning, the results showed that the complete accuracy of the acetabular cup was 42%, including one size was 75%, the femoral stem was 50%, and including one size was 77%. The results of this study were similar to those of previous relevant literature. All plannings were checked by two experienced surgeons. If there was a disagreement, the planning was repeated by all the three surveyors agreed upon. Therefore, our results are both effective and reliable.
The AI HIP was found in 10 consecutive planned hip joints that the prediction of acetabular cup and femoral stem fluctuated in one size. In the first 10 patients planning, the accuracy of the acetabular cup and femoral stem was 70%. In the planning of 10-20 patients, the accuracy rates of acetabular cup and femoral stem were 80% and 70%, respectively. It shows that AI HIP was relatively stable, and there was no obvious deviation in the automatic learning in the early stage, so the planning could continue without elimination. In the whole study, there were 11 hips in one size, 2 hips in two sizes, and 1 hip in more than three sizes. In two hips with a difference of 2 sizes, one was osteoarthritis due to malunion of a chronic femoral neck fracture. It became the shallow acetabulum fossa edge on the part of the deformation, when the acetabulum cup matches 54 mm bone coverage rate of 87%, while the actual intraoperative application of 50 mm acetabulum cup, the reason may be that AI planned the position and size of acetabular cups in the ideal state. However, when the acetabular cup was ground to 50mm during the operation, it was found that the coverage rate of the acetabular cup was sufficient and the stability of the cup was satisfactory. So, the grinding was not deepened. The other was the same AI software that planned a 54mm acetabular cup, and a 50mm acetabular cup was used in the operation. The possible reason was that in patients with ankylosing spondylitis because of ankylosis of the hip joint, the acetabulum was fused with the femoral head, and the software could not accurately segment the acetabulum and femoral head, resulting in a large deviation. Thus, the application of AI HIP in preoperative planning for serious AS needs further investigation. There was one case with a difference of more than 3 sizes. This patient had severe hip osteoarthritis with osteophytes around the hip joint. However, the software matched a 52mm acetabular cup according to the true anterior and posterior diameter of the acetabulum, including the upper and lower diameter of the acetabulum. Perhaps the surrounding osteophytes interfered with the software in recognizing the acetabulum morphology. This led to a large error in the planning results. In the planning of the femoral stem, there were 10 hips in one size, 3 hips in two sizes, and 1 hip in three sizes. In 3 hips with a difference of 2 sizes, 2 hips were DDH, which may be due to: the variation of femoral neck anteversion angle and the increase of neck stem angle, the enlargement and upward movement of the femoral anterior arch, the enlargement of the anteroposterior diameter of the proximal femoral medullary cavity, and the reduction of transverse diameter. This may lead to the deviation of AI in identifying: the epiphysis of the femoral shaft, the medial and lateral cortex of the distal femur, resulting in large planning errors, which requires AI HIP to make sufficient planning after a large number of DDH cases. It can determine the femoral variation of DDH, so that it can be more accurate in DDH. In the case of 3 sizes, the anterior femoral arch of the patient was significantly enlarged, which may have led to a bias towards matching the smaller femur stem during AI planning.
Our study also compared the operation time of the three planning methods. The results showed that the operation time of AI HIP was shorter than that of 3D mimics. The study of preoperative planning of the operation time has also been reported before, Erik Schiffner et al.  compared the operation time of 3D and 2D for the first time. The results showed that the average operation time of the 2D template was 12 minutes (range: 8-23 min), while that of the 3D template was 17 minutes (range: 10-25 min). The operation time of 2D was faster than that of 3D, with an average of 5 minutes interval. When sariali et al. used hip-plan 3D planning software, the operation time was 10-15minutes. Similarly, Inoue  et al. used a zed hip for preoperative planning, and the operation time was also 10-15 minutes. The reason for which their 3D planning time was faster than our 3D mimics planning time was because the algorithms of the pelvic plane, femoral morphology, and anatomical markers in their 3D software were automatic, while our 3D mimics was manual in pelvic segmentation, modeling, prosthesis implantation, and other operations. However, the AI HIP in our study was 3-5 times faster than the 3D software planning reported before, and 10 times faster than 3D mimics. This is because the software uses 3D segmentation neural network and 3D anatomical recognition neural network technology, which can quickly identify, segment, correct, and measurement by artificial intelligence, greatly shortens the planning time. Petretta  compared the time efficiency of the acetate template and digital template, and the results obtained showed that the operation time of the digital template was 154 s (range: 73-343 s). In our study, the operation time of the 2D digital template was 2.96±0.48 minutes, which was similar to the results of petretta.
In the analysis of factors influencing the accuracy of AI HIP, we found that sex had no influence on the software. Holzer et al., studied the accuracy of digital template in cementless prosthesis in 2D preoperative planning. The results obtained also showed that sex had no influence on the accuracy of digital template prediction  because of little variation and limited differences in sex in the amplification rate of markers. . Previous studies on the influencing factors of BMI on the accuracy of 3D planning also found that obese patients were not affected by the accuracy of CT navigation to locate the acetabular cup. In our study, AI HIP adds AI technology based on the CT data of patients. It is also based on the CT data for 3D planning, which is not affected by the magnification. Therefore, the accuracy of its prediction is not affected by BMI. In the analysis of hip dysplasia, we found that the accuracy of AI HIP in the DDH group was lower than that in the non-DDH group, and the difference was statistically significant. However, there was no significant difference in the accuracy between the DDH and non-DDH groups in the planning of the femoral stem. The main reason for these results was that 16 hips in the DDH group were Crowe I-II type, and the degree of femoral deformity variation was small. So, there was no significant difference in accuracy between the DDH and non-DDH groups. On the acetabular side, although the deformity variation is relatively light, which is not as severe as Crowe III-IV high dislocation deformity, AI HIP was in situ reconstruction during the planning of acetabular cup, without proper upward or inward movement according to the degree of variation of acetabular deformity, and the placement angle is 40° abduction angle and 20° anteversion angle. In the actual operation, 13 (81.25%) of the 16 hips in the DDH group had the acetabular cup appropriately moved upward or inward in order to achieve greater bone coverage. This may be the reason why the software was inaccurate in predicting acetabular cups in patients with DDH. Therefore, the software needs to expand the planning of the number of cases of each type of DDH, continuously strengthen the deep learning of the angle and position of the placement of each type of DDH acetabular cup, so as to master the placement rules of each type of DDH prosthesis, and achieve accurate planning in DDH planning.
Based on AI technology and big data, compared with previous 2D and 3D studies, AI HIP has the following innovations: (1) Speed: one-button intelligent operation, simple operation; (2) Visualization: 3D visualized image output is immersive, providing realistic 3D visual anatomy; (3) Pelvic correction: Automatically corrects the position of the pelvis and lower limbs, without constant manual adjustment; (4) automatic planning: artificial intelligence planning can avoid planning deviation caused by personal experience, while the results of other 2D template methods and 3D template measurements are related to the experience level of the surveyors, and the more skilled the operation, the more accurate the results; (5) output scheme: generate a multi-perspective observation planning scheme, which is easy to realize in operation; (6) continuing education; further support and improvement of continuing education to shorten the doctor's learning curve. At the same time, in this study, in order to avoid surgeon selection deviation, the blind method was adopted for the surgeon because if the surgeon participated in preoperative planning, the surgeon might change the grinding technique and insert a larger or smaller part by removing more or less bones when necessary, which will lead to errors.
(1) The prosthesis used in this study was limited to depuy products, and the range of prosthesis selection was small. (2) The sample size of the study was small, and the influencing factors of diseases were only divided into DDH and non-DDH groups, without further subdividing the difference in accuracy of the software in different diseases. (3) Only the accuracy of prosthesis size was evaluated, but important parameters such as: rotation center, offset, and leg length discrepancy were not further evaluated. (4) Because of the need for CT scanning, the X-ray radiation exposure of patients was large, and the economic cost was also high. (5) CT data need to be copied, whether the data transmission process can be optimized or needs no further research. However, this is not the scope of this experiment. Next, we will continue to: expand the cases, further study the accuracy of AI HIP in different diseases, evaluate its reproducibility in biomechanical planning and its impact on postoperative clinical efficacy.