Automated Femoral, Acetabular, and Global Offset Measurements on Pelvis Radiograph Using Deep Learning Authors: Jack C. Casey, Seong J. Jang, Billy I. Kim, Kyle N. Kunze, Chris Anderson, David J. Mayman, Seth A. Jerabek, Jonathan M. Vigdorchik, Peter K. Sculco Published: 2026 ABSTRACT PurposeOffset measurement is critical in total hip arthroplasty (THA) for guiding restoration of native anatomy. However, measurements are time-consuming and measurer-dependent, creating obstacles for large cohort analyses. We aim to create an objective and reliable offset measurement algorithm using deep learning.Materials and MethodsFive hundred radiographs from the Osteoarthritis Initiative (OAI) were segmented with identification of the teardrop, femoral head, implant head, and femoral diaphysis. A U-Net convolutional neural network was trained to identify these landmarks and optimized using the multi-class Dice coefficient metric. Femoral axis and femoral/implant head center of rotation were calculated with the model predictions, and measurements of offset were compared against two trained readers on an independent testing cohort.ResultsThe optimized model had a Dice coefficient of 0.96 and a foreground mask accuracy of 96.2%. The model measured femoral, acetabular, and global offset on both limbs at a rate of 1.67 sec/image. On an independent cohort (n=90), the intraclass correlation coefficient between readers and the algorithm was 0.86 (95% confidence interval [CI] 0.80-0.91) for femoral offset, 0.87 (95% CI 0.78-0.91) for acetabular offset, and 0.94 (95% CI 0.91-0.96) for global offset. When applied to the entire OAI cohort (n=4,188), all relevant anatomical features (femoral axis, implant/femoral center of rotation, inter-teardrop line) were correctly calculated in 83.0% of images.ConclusionWe report the development of an accurate and rapid offset measurement model using deep learning that can be applied before and after THA. Future work will involve external model validation. INTRODUCTION Total hip arthroplasty (THA) has proven to be an effective procedure for providing pain relief, improving range of motion, and enhancing quality of life1). When performing THA, restoring native biomechanics of the hip results in improved long-term clinical outcomes and increased implant survivorship2). Offset, which can be assessed using three measurements—femoral offset (FO), acetabular offset (AO), and global offset (GO), is an important anatomic variable that should be closely restored at the time of THA as it is related to hip stability, range of motion, and optimized abductor function. Accurate and efficient measurement of offset is also important for large patient cohort studies for examining the interaction between restoration of radiographic offset and postoperative outcomes3-5). Preoperative templating provides the information needed to prepare the proper instrumentation, implants, and prepare for any anticipated surgical challenges6). Offset is traditionally measured on anteroposterior (AP) radiographs, however different imaging techniques have also been employed7,8). Offset is influenced by femoral anteversion, rotation of the pelvis, and measurement can be difficult, particularly in large cohort analyses, because the long axis of the femur must be precisely bisected by the measurer9). These challenges create variability among clinicians measuring offset, as there is no reported gold standard method. This is important to note as assessing offset restoration is important when studying outcomes after THA; failure to restore offset may result in gait abnormalities, bone impingement, instability, patient dissatisfaction, and lower quality of life10). Deep learning algorithms can be leveraged for automating measurements with high accuracy and precision which may facilitate the accuracy and efficiency of performing offset measurements on AP radiographs11). Specifically, convolutional neural networks (CNNs), which are largely used for image analysis, can perform segmentations on radiographic images and precisely identify anatomic landmarks for calculation of pertinent preoperative measurements such as offset12). This method of measuring is ideal as it may minimize or prevent user error and early studies have suggested that it may be more efficacious than traditional methods used by radiologists and surgeons13,14). Therefore, the purpose of this study is to (1) create a rapid, objective, and validated deep learning algorithm for offset measurement and to (2) leverage the algorithm for use in large cohort analysis of offset on standard AP radiographs. We hypothesize that our CNN will be able to identify the necessary anatomical landmarks for measurement of offset with high accuracy and show excellent correlation to human raters. MATERIALS AND METHODS AP radiographs of the pelvis from the Osteoarthritis Initiative (OAI) database were included. The OAI is a publicly available database consisting of images and prospective data on 4,796 participants aged 45-79 years from 2004 to 2015. This study was deemed exempt from review by Hospital for Special Surgery’s Institutional Review Board given the use of a de-identified public database. 1. Method of Radiographic Offset Measurements for Human Graders Anonymized DICOM (Digital Imaging and Communications in Medicine) images were uploaded to a public medical image viewer, HorosTM (https://horosproject.org/), for measurement of FO, AO, and GO using the software’s preestablished analytical tools. A standard method for measuring FO was employed4). First, using a best fit circle around the femoral head, the center of rotation (COR) of the femoral head was identified. Perpendicular lines were drawn connecting the long axis of the femur to the COR of the femoral head for measurement of FO. For measurement of AO, a separate perpendicular line was used to measure the distance from the femoral COR to a vertical line that closely parallels the ipsilateral teardrop. Finally, GO was generated by the sum of the FO and the AO. All measurements were performed after leveling the pelvis based on the superior portion of the obturator foramen. Two orthopedic residents who were trained to take measurements were blinded and measured each variable on 90 femurs. Assessment of reliability was performed using intraclass correlation coefficient (ICC) (two-way mixed, single score, ICC3) between the two human raters and between each rater and the deep learning algorithm. 2. Method of Algorithm Training A U-Net CNN was created for pixel identification of bony landmarks. A random sample of 500 patients with baseline and 96-month follow-up pelvis radiographs from the OAI were used for training (80.0%) and validation (20.0%). The adequacy of this sample size for orthopaedic image deep learning analysis has been demonstrated11) and we previously reported on the same techniques for measuring and assessing leg-length discrepancy15). After training for identification of landmarks, the U-Net CNN was optimized using the multi-class Dice coefficient metric. The model segmentation predictions were used for calculating femoral axis and femoral/implant head COR. Measurements of offset were compared against two trained readers on an independent testing cohort. The algorithm was then deployed to measure offset for a large cohort of images (n=3,477) from the OAI database. RESULTS 1. U-Net Convolutional Neural Network Performance After training, the optimized model had a multiclass dice coefficient of 0.96 with a foreground mask accuracy of 96.2%. This result demonstrates acceptable performance by the U-Net model for predicting landmarks, which was completed at a rate of 1.67 seconds per image. An example of landmark prediction, the results of pixel analyses for identification of points of interest, and the final output are shown on an AP pelvic radiograph image with an implant in Fig. 1, and on an image without an implant in Fig. 2. For assessment of model validity, the model predictions were compared against human measurers on an independent cohort of 90 hips for measurements of FO (ICC 0.86, 95% confidence interval [CI] 0.80-0.91), AO (ICC 0.87, 95% CI 0.78-0.91), and GO (ICC 0.94, 95% CI 0.91-0.96) (Fig. 3). FO and AO measurements showed good-to-excellent agreement, while GO showed excellent agreement. 2. Offset Measurements in the OAI Database Out of the 4,188 AP pelvic radiographs in the OAI cohort, anatomical landmarks were correctly identified by the algorithm in 83.0% (n=3,477). In these 3,477 images, landmarks were predicted by the algorithm and bilateral measurements of three types of offset were made in under 108 minutes. No significant differences were observed between left and right offset measurements (P>0.05), while FO, AO, and GO were greater in males (P<0.01). The offset discrepancy correlations among FO, AO, and GO are shown in Table 1. DISCUSSION The main findings of the current study were as follows: (1) a deep-learning algorithm can perform accurate automatic measurements of FO, AO, and GO bilaterally from standard AP pelvic radiographs in under two seconds; (2) excellent correlation was observed between the algorithm and human readers; and (3) when this tool was applied to a publicly available patient cohort, no significant differences were observed between left and right offset measurements, greater offset values were observed in males compared to females, and GO discrepancies between hips in a single patient were found to be driven by differences in FO. Surgeons aim to restore FO, AO, and GO, clinically important parameters, during THA. Poor restoration of FO has been associated with worse postoperative range of motion16). High postoperative FO has been associated with elevated serum cobalt levels in patients with metal-on-polyethylene implants17), and worse scores for postoperative pain18). Meanwhile, low post-THA FO has been associated with increased linear polyethylene wear19) and adverse alterations in gait and an increased risk for post-operative dislocation20). Offset determines the lever arm for muscles of hip abduction; higher FO enables greater abductor muscle strength3,21). Restoring AO affects both lateralization of the femur and the COR of the hip and therefore must also be considered by the surgeon. Increases in AO reduce bone-bone impingement, thus AO is an important parameter under the surgeon’s control that can minimize the risk of dislocation22). Intraoperatively, AO is controlled by surgeons, who choose the degree of cup medialization and whether to use an offset liner. During THA, AO is frequently reduced with compensatory increases in FO to maintain GO, although this has negative effects on patient reported outcomes compared to when FO and AO are preserved23). Offset restoration plays a central role in promoting optimal outcomes following THA; therefore, surgeons often utilize preoperative planning to assist in restoring these parameters. Intraoperatively, manual measurement jigs have been developed that allow for measurement of GO. Computer and robot-assisted surgical systems enable surgeons to assess changes of offset intraoperatively and compare to a preoperative plan24). The preoperative planning process is therefore imperative for surgeons performing both manual procedures as well as robot-assisted surgeries to ensure proper restoration of offset. Multiple imaging options are available for use in preoperative planning. Measurements of offset can be performed using AP X-ray films of the hip or pelvis, computed tomography (CT) scans, or biplanar low dose radiographs (EOS; ATEC Spine Group). CT scan planning is superior to conventional pelvic radiographs, which tend to underestimate FO25-27), although AP radiographs of the hip are efficacious28). The EOS system is less commonly used but does provide an accurate and reliable medium for measuring offset29). AP radiographs of the pelvis, although not the best, are the most prevalent imaging modality used for preoperative planning. It is therefore imperative to optimize the functionality of this method by ensuring high quality radiographic images are taken6). Measurement accuracy requires a consistent method for taking measurements. Deep learning automation is one solution as it standardizes the measurement method and thus minimizes human variability30). A deep learning algorithm has previously been used to enhance preoperative planning for THA, in which artificial intelligence performed a preoperative assessment for identifying anatomic landmarks and then simulated postoperative outcomes to show the surgeon implant sizes and possibilities for correction of deformity31,32). Compared to traditional manual X-ray based planning, preoperative planning involving artificial intelligence performing measurements on CT images enabled more accurate restoration of FO postoperatively, although there was no difference in AO or GO31). Because AP pelvic radiographs are more common than CT scans preoperatively, a deep learning tool that makes accurate offset measurements on radiographs would be very useful. To the best of our knowledge, this tool is the first application of deep learning in performance of clinically relevant offset measurements on AP pelvic radiographs. Artificial intelligence (AI) can potentially benefit the clinical, preoperative, and postoperative settings by reliably producing rapid and consistent measurements of offset. Most importantly, this tool enables completion of these measurements using AP radiographs of the hip, which most surgeons already obtain regularly for patients for which THA is being considered. Measuring offset by hand is tedious, user-dependent, and sometimes deemed unnecessary if the implants needed to accommodate for a patient who appears to have high or low offset can be visually predicted by a surgeon. AI automates measurements and reduces burden on the surgeon to identify landmarks and calculate distances for generation of offset values. As such, surgeons could easily obtain a quantitative assessment of offset restoration during postoperative visits when obtaining routine postoperative AP radiographs. The utility of this tool extends beyond the direct patient setting, with many potential benefits in orthopaedic research. As discussed previously, offset restoration during THA directly affects postoperative complications and patient reported outcomes. This tool would enable nearly effortless, high throughput automation of radiographic measurements on large databases enabling further investigation into the importance of FO, AO, and GO. Of particular importance, when compared against human readers and deployed in a large population, the current algorithm produced accurate results aligning with findings from previous literature. The results of this study demonstrate that deep learning measurements of FO, AO, and GO from the OAI database were greater in males than females, which aligns with findings regarding FO in previous studies26,32). No significant difference was observed between left and right side offset measurements, again aligning with previous studies32). Further examination of differences between offset measurements in a single patient, as shown in Table 1, shows that GO discrepancies on standard radiographs are driven by FO discrepancies, which may be in part due to femoral and pelvic rotation, or differences in OA progression among patients with bilateral OA pathology. This study has several limitations. A systematic analysis of the causes of failure in prediction of algorithm landmarks was outside the scope of this study and would be an important addition to future studies aimed at validating the algorithm for clinical use. To strengthen generalizability, future studies should also include a larger and more diverse pool of human measurers, including radiologists and surgeons at different training levels. The OAI cohort is comprised of patients with or at risk of OA. The OA grade was not available at the timepoint at which measurements of radiographs were performed, and therefore sub-analyses on the variability in offset measurements according to OA grade could not be performed. Rigorous external validation of this tool will be necessary prior to clinical application. Another limitation is that radiographs do not account for pelvic rotation, femoral rotation, or patient positioning, which all affect offset measurements. Future studies should examine the question of whether radiographic features of the pelvis can be used for detecting femoral rotation. If femoral rotation is the main factor driving errors in offset measurement, predicting rotation could potentially enable correction of offset values. This study did not use three-dimensional images as comparisons with the two-dimensional calibrated measurements, which is important in order to account for how measurements are currently scaled at the level of the receiver beam, leading to overestimations. CONCLUSION In this study a novel deep learning algorithm was successfully applied to a heterogeneous patient cohort, collecting rapid, accurate, and objective measurements for the calculation of three types of offset from AP pelvic radiographs. This algorithm can be applied preoperatively and postoperatively in patients undergoing THA for measurement of FO, AO, and GO. The algorithm produces these measurements much more rapidly than a human, at an average of 1.67 seconds per patient. Of particular importance, these measurements are taken rapidly without sacrificing accuracy. On an independent cohort of 90 hips, the ICC between human readers and the deep learning algorithm was high for all types of offset. While external validation is still needed, this tool has the potential to streamline and enhance preoperative planning, postoperative assessment, and facilitate large scale research studies in patients undergoing THA.