Bone morphology as a significant risk factor for anterior cruciate ligament injury Authors: Dingyu Wang, Chao Liu, Xinwei Du, Xueqing Yu, Jiaxin Liu, Shanggui Liu, Songlin Zhang, Tingting Jiang, Dong Jiang Published: 2026 ABSTRACT AbstractAnterior cruciate ligament (ACL) injury is a prevalent sports-related trauma with significant impact on both professional athletes and young individuals. While often viewed as accidental events, high recurrence rates suggest the presence of underlying structural risk factors. This study investigated the association between 3D bone morphology and ACL injury susceptibility. Leveraging a multicenter dataset of 5000 MRIs, we developed ACL-injury Patterning (ACL-P), a deep learning model that utilizes 3D bone morphology point clouds for injury classification. Given the short-term stability of bone geometry following trauma, this classification serves as a proxy task to quantify the correlation between pre-existing morphological traits and ACL injury through discriminative performance. The model’s discriminative performance was validated across an external civilian test set (n = 7797) and a professional athlete test set (n = 269), yielding an Area Under the Curve (AUC) of 0.887 (95% CI: 0.879–0.895) and 0.907 (95% CI: 0.872–0.941), respectively. Robustness checks confirmed that the model’s performance remained consistent across both acute and chronic injury phases (0–365 days) and was independent of secondary joint dislocation. These findings establish 3D bone morphology as a significant anatomical factor linked to injury susceptibility, enhancing our understanding of the structural correlates of knee trauma.Subject terms: Anatomy, Computational biology and bioinformatics, Diseases, Medical research, Risk factors Introduction Anterior cruciate ligament (ACL) injury is a significant and debilitating trauma that affects both professional athletes and young individuals. With an incidence rate ranging from 53 to 71 per 100,000 person-years1,2, ACL injuries impact millions globally, imposing a substantial financial burden on healthcare systems3. These injuries result in acute knee disability, chronic physical limitations, and an elevated risk of post-traumatic osteoarthritis, with an incidence as high as 87%4–7. For athletes, ACL injuries are among the leading causes of diminished athletic performance and early retirement5,7. Consequently, the prevention of ACL injuries is of paramount importance. ACL injuries predominantly occur during activities that involve cutting, pivoting, and landing, such as in basketball, football, skiing, and soccer, often through low-energy, non-contact, or indirect contact incidents8–12. Despite advancements in understanding the mechanism of ACL injuries, the incidence rate continues to rise, even among professional athletes5,6,13–15. It is estimated that each team of 25 players may experience one ACL injury every two years5,13. The reinjury rate among athletes under 25 is a staggering 23%, indicating a 30 to 40 times higher risk of sustaining an ACL injury compared to their uninjured peers16. For non-athletes, the recurrence rate is a concerning 17.6%, with contralateral and revision rates at 11.8% and 5.8%, respectively, which is unusually high for a traumatic injury17. According to Meeuwisse’s epidemiological model18, ACL injuries, though often triggered by a singular event, may result from a complex interplay of internal and external risk factors (Fig. 1). The high recurrence rate suggests that persistent risk factors predispose individuals to initial injury. Identifying these predispositional risk factors and mechanisms is crucial for effective prevention, clinical decision-making, surgical techniques, and targeted rehabilitation programs. The bone morphology as a potential risk factor for ACL injury. a Meeuwisse’s epidemiological model: the chain of ACL injury; b, c the knee bone morphology from patients with and without ACL injury. The blue dotted line indicated the well-known anatomic risk factors of ACL injury, such as intercondylar eminence of the tibia, intercondylar fossa, lateral condyle of the femur and lateral tibial slope. While the ACL injury mechanisms are well-studied8,9,11,12, the predispositional risk factors remain largely unclear. Commonly cited risk factors include female gender and young age, particularly among adolescents2, which are non-modifiable. Other proposed intrinsic and extrinsic factors include anatomical variations, neuromuscular imbalances, hormonal influences, and environmental conditions19. However, the impact of most of these factors is either negligible or debated, making it challenging to accurately assess an individual’s susceptibility to ACL injury and implement targeted prevention strategies. Mechanically, ACL load increases are caused by a combination of muscular resistance to large knee flexion moments, external forces generating knee compression, internal tibial torque, and knee abduction20. Biomechanical studies have demonstrated that certain bone morphologies, such as the tibial plateau slope, contribute to anterior tibial translation and internal tibial rotation, thereby increasing ACL loading21,22. Bone morphology is thus considered an intrinsic risk factor for ACL injury. However, despite numerous indices created to describe bone morphology, traditional 2D measurements on X-ray or singe MRI slice are inadequate for capturing the complexity of 3D bone morphology, limiting our understanding of its relationship with ACL injury23. The extent to which bone morphology influences susceptibility to ACL rupture remains controversial. The advent of machine learning offers a transformative approach to analyzing non-linear characteristics, providing a sophisticated understanding of bone morphology patterns. In this study, we employ a machine learning approach to capture the non-linear characteristics of bone morphology and establish its correlation with ACL injury. We introduce a novel algorithmic measure of ACL injury risk, named ACL-P (ACL-Injury Patterning), based solely on bone shape. Through two large external test datasets and systematic analysis, we demonstrate that bone shape alone can classify the majority of ACL injury cases, underscoring the strong correlation between bone morphology and ACL injuries. Results Overview The primary aim of this study is not to develop a tool for diagnosing ACL injury, but to understand bone morphology’s association with ACL injury. Initially, we constructed the Training Set (2500 non-contact ACL injury and 2500 normal ACL cases) to develop the bone segmentation model and convert bone shapes into point clouds. Subsequently, we trained a classification model to correlate bone shape with ACL diagnosis. The output probability of the classification model, termed ACL-P (ACL-Injury Patterning), represents the correlation between ACL injury and bone shape. Since bone is considered rigid and its shape changes minimally after an ACL injury in a short time, this probability can be viewed as a predispositional risk factor for ACL injury due to bone shape. We further evaluated the discriminative power of ACL-P on external Civilian Test (n = 7797) and Athlete Test Set (n = 269) and conducted a systematic robustness check to ensure the algorithm’s effectiveness was not primarily reliant on post-traumatic changes. Finally, we employed two strategies: contrastive matching and counterfactual explanations, to enhance the transparency of the classification decision-making process. The overview of this work is illustrated in Fig. 2. Schematic of study design. Dataset Summary statistics for the 5000 MRIs in the Training Set, 7797 MRIs in the Civilian Test Set, and 269 MRIs in the Athlete Test Set are presented in Table 1. Notably, the Athlete Test Set population is entirely independent of the Training Set and Civilian Set, as the latter group exclusively comprise non-athlete populations with distinct injury patterns and activity levels. The athletes represent 11 sports disciplines: athletics, swimming, gymnastics, basketball, football (soccer), table tennis, judo, fencing, skiing, and skating. These multicenter datasets span diverse demographic and clinical characteristics, including age (children to elderly), BMI range, balanced sex distribution, common knee pathologies, and varied injury mechanisms. Additionally, the data incorporate heterogeneity in MRI hardware and scanning parameters (Table 1). Table 1Summary statistics of the training and test setsTraining setCivilian Test SetAthlete Test SetNumber of cases5000 civilians7797 civilians269 athletesNumber of centers6 centers3 centers1 centerPatient characteristic Age (range)30.6 (12–50)29.8 (8–50)20.6 (13–45) Sex (male:female)60%:40%65%:35%57%:43% BMI25.1 (11.2–61.5)4.6 (12.7–54.4)23.3 (15.9–37.0) Injury to MRI period (range)a82 days (0–365)75 days (0–362)64 days (0–340)Disease distribution ACL injury50%30%61% Meniscus injury56%38%70% PCL injury9%5%7.4% Cartilage lesions18%13%22% Patella dislocation8%3%1% Collateral ligament injury10%4%10% Structurally normal20%41%0%Injury causes Sports trauma49%61%97% Daily work/life injury35%27%0% Traffic accident0%7%0% Overuse injury12%3%3% No causes4%2%0%Mechanism of injury Non-contact100%35%78% Direct contact0%23%15% Unknown0%42%7%Scanner manufacturer GE48%44%29% Siemens24%28%38% Philips10%0%0% United Images18%28%33%Magnetic field 3.0 T70%75%100% 1.5 T30%25%0%Scanning parameters Repetition Time (ms, range)2460 (1487–4015)2490 (1844–3217)2481 (1844–3100) Time to Echo (ms, range)35.0 (23.0–64.1)35.5 (23.0–60.0)35.8 (22.8–60.5) Slice thickness 3.0 mm4%–– 3.5 mm70%85%96% 4.0 mm26%15%4%aThe participant with structurally normal knee MRI were not calculated into this. Summary statistics of the training and test sets aThe participant with structurally normal knee MRI were not calculated into this. Validation of ACL-P After training and validating the bone shape segmentation model with a DICE coefficient of 0.968 for femur segmentation and 0.943 for tibia segmentation. Then we extracted bone shapes from the MRI scans and converted them into point clouds (see supplementary materials). We trained a Point Cloud Transformer (PCT) network to classify ACL injury based on 5000 bone shape point clouds. The output ACL injury risk, referred to as ACL-P (ACL-Injury Patterning), captures the objective features in the bone shape that indicate ACL injury. The network achieved an AUC of 0.887 for ACL injury in the Civilian Test Set, with sensitivity at 75.6% and specificity at 85.4% (Table 2). For athletes, who are more frequently exposed to sports injuries, the discriminative power of ACL-P was higher, with an AUC of 0.907 and sensitivity and specificity of 75.1% and 84.6%, respectively (Table 2). We further performed a calibration analysis of the ACL-P score using the Civilian Test Set (Fig. 3). The initial assessment using the Hosmer-Lemeshow (H-L) test indicated that the raw model output tended to overestimate the risk of ACL injury (χ2 = 268.44, P < 0.001). To address this, we applied isotonic regression to refine the probability estimates. The bias-corrected ACL-P score now demonstrates excellent agreement between the predicted and observed injury probabilities, with the H-L test showing no significant departure from perfect calibration (χ2 = 5.82, P = 0.823). Table 2The discriminative power of the ACL-P in the Civilian Test Set and Athlete Test SetaAUC of ROCSensitivitySpecificityCivilian Test Set0.887 (0.879–0.895)75.6% (73.8–77.4%)85.4% (84.4–86.3%)Athlete Test Set0.907 (0.872–0.941)75.1% (67.8–81.5%)84.6% (76.2–90.9%)Subgroup analysisSex Male0.890 (0.880–0.899)74.9% (72.7–77.0%)85.8% (84.5–87.0%) Female0.883 (0.869–0.897)77.2% (74.0–80.2%)84.8% (83.3–86.3%) P value0.434––Age 8–16 y0.907 (0.883–0.932)79.2% (72.6–84.9%)88.2% (85.8–90.3%) 17–30 y0.903 (0.892–0.914)76.4% (73.8–78.9%)87.3% (85.9–88.6%) 31–40 y0.872 (0.857–0.887)74.1% (71.0–77.1%)83.9% (81.9–85.7%) 41–50 y0.837 (0.810–0.865)74.7% (69.3–79.7%)77.6% (74.1–80.9%) P value<0.001––Injury causesb,c Sports trauma0.862 (0.847–0.876)74.8% (72.7–76.8%)81.5% (78.8–84.0%) Daily work/life trauma0.867 (0.845–0.889)77.6% (74.1–80.8%)80.0% (75.8–83.7%) Traffic accident0.835 (0.788–0.882)72.4% (65.6–80.0%)80.7% (74.2–86.1%) P value0.492–-Injury mechanism Non-contact0.927 (0.913–0.941)94.4% (92.7–95.9%)82.3% (79.3–84.9%) Direct contact0.738 (0.707–0.770)50.4% (46.5–54.3%)75.1% (70.6–79.3%) P value<0.001––Injury types Structurally normal knee vs ACL injury0.905 (0.898–0.913)75.6% (73.9–77.4%)88.8% (87.7–89.9%) P value<0.001–– ACL injury vs other knee injuries0.858 (0.848–0.869)80.3% (78.5–81.9%)75.6% (73.9–77.3%) P value<0.001––Data are presented as mean values with 95% confidence intervals.aThe subgroup analysis did not include athlete data due to the limited sample size.bThe participant with a structurally normal knee MRI was not included in the analysis.cOveruse injuries and no causes were not analyzed due to the insufficient number of ACL cases in this category. The discriminative power of the ACL-P in the Civilian Test Set and Athlete Test Seta Data are presented as mean values with 95% confidence intervals. aThe subgroup analysis did not include athlete data due to the limited sample size. bThe participant with a structurally normal knee MRI was not included in the analysis. cOveruse injuries and no causes were not analyzed due to the insufficient number of ACL cases in this category. Calibration analysis of the ACL-P on Civilian Test Set. The calibration curve demonstrated original ACL-P tended to overestimated the risk of ACL injury (Hosmer-Lemeshow χ2 = 268.44, P < 0.001). After isotonic regression, excellent agreement between the bias-corrected ACL-P and the observed ACL injury probability (Hosmer-Lemeshow χ2 = 5.82, P = 0.823). The curve was created by Python (version 3.9.19) with matplotlib (version 3.8.4). The discriminative power of ACL-P varied across different subgroups. As shown in Table 2, the model more effectively distinguished between structurally normal knees and ACL injuries, with an AUC of 0.905, while its discriminative power decreased in patients with other knee injuries, showing an AUC of 0.858. This suggests that other knee injuries, such as meniscus and PCL injuries, share some common bone shape susceptibilities with ACL injury. The discriminative power of ACL-P was associated with age, showing a higher AUC in younger subgroups. We also found that the model’s performance was lower in cases of traffic accidents and direct contact mechanism, which are more extrinsic in nature, while it was higher in sports trauma cases and non-contact mechanism, which are more intrinsic. No significant difference in discriminative power was observed between male and female subgroups. Robustness check ACL-P is not primarily learning to classify from post-traumatic bone shape change: The model might have achieved its performance by merely identifying changes in bone shape following an ACL injury. However, when the Civilian Test Set was divided into five groups based on the injury period, the discriminative power of ACL-P remained consistently high from the acute phase (0–30 days, AUC 0.841) to the chronic phase (181–365 days, AUC 0.887) (Table 3). Furthermore, a regression analysis between ACL-P scores and the injury period showed no significant correlation (R2 < 0.001, p = 0.553). These findings suggest that the morphological features captured by ACL-P are largely related to intrinsic skeletal anatomy rather than being primarily driven by post-traumatic alterations Table 3The difference of the ACL-P discriminative power in subgroups with different injury to MRI periodsInjury to MRI periodaAUC of ROCSensitivitySpecificity0–30 days (36%)0.841 (0.822–0.861)73.7% (70.9–76.4%)78.7% (75.2–81.9%)31–60 days (20%)0.866 (0.837–0.894)75.4% (71.3–79.2%)82.5% (78.6–86.0%)61–90 days (14%)0.856 (0.829–0.884)81.5% (76.1–86.1%)78.3% (73.8–82.4%)91–180 days (22%)0.858 (0.834–0.881)74.2% (69.4–78.6%)81.3% (78.0–84.2%)181–365 days (9%)0.887 (0.853–0.922)80.0% (74.6–84.7%)80.6% (72.9–86.9%)P value0.209––aThe participant with structurally normal knee MRI were not calculated into this. The proportion of the subgroups was showed in parentheses. The difference of the ACL-P discriminative power in subgroups with different injury to MRI periods aThe participant with structurally normal knee MRI were not calculated into this. The proportion of the subgroups was showed in parentheses. ACL-P is not simply learning to classify from the femur and tibia relative dislocation: The model could have achieved its performance by merely identifying post-ACL injury tibial anterior dislocation. However, after randomly selecting 700 MRIs from the Civilian Test Set and manually measuring anterior tibial translation (ATT), we found that using only ATT to classify ACL injury resulted in an AUC of 0.759 (95% CI: 0.727 to 0.790). This is considerably lower than the ACL-P’s AUC of 0.925 (95% CI: 0.904 to 0.945) within the same subgroup. Regressing ACL-P against ATT yielded an R2 of 0.217, indicating that the model’s explanatory power for ACL injury is not fully explained by tibial anterior dislocation. We further developed a classification model using separate inputs for the femur and tibia, effectively eliminating the influence of their relative positions, including tibial anterior dislocation, knee varus, and valgus. The model still demonstrated discriminative ability for ACL injury, achieving an AUC of 0.816 and 0.719 in the Civilian and Athlete Test Sets, respectively (Table 4). We further conducted a subgroup analysis of the Civilian Test Set and found that removing the relative position of the femur and tibia mainly affected the subgroup with knee injuries, reducing the AUC of the ROC by 14.4%. However, the model’s ability to distinguish ACL injuries from structurally normal knees remained largely intact. Table 4Impact of femur and tibia relative location information on ACL-P discriminative powerAUC of ROC joint input of the femur and tibiaAUC of ROC separate input of the femur and tibiaReduced byP valueCivilian Test Set0.887 (0.879–0.895)0.816 (0.806–0.825)8.0%<0.001 Subgroup: structurally normal vs ACL injury0.905 (0.898–0.913)0.870 (0.861–0.879)3.8%<0.001 Subgroup: ACL injury vs other knee injuries0.858 (0.848–0.869)0.734 (0.720–0.748)14.4%<0.001Athlete Test Set0.907 (0.872–0.941)0.719 (0.660–0.778)20.7%<0.001Although the discriminative power of ACL-P decreased when we separately input the femur and tibia and eliminated the influence of their relative positions, the model still demonstrated discriminative ability for ACL injury, achieving an AUC of 0.816 and 0.719 in the Civilian and Athlete Test Sets, respectively. Impact of femur and tibia relative location information on ACL-P discriminative power Although the discriminative power of ACL-P decreased when we separately input the femur and tibia and eliminated the influence of their relative positions, the model still demonstrated discriminative ability for ACL injury, achieving an AUC of 0.816 and 0.719 in the Civilian and Athlete Test Sets, respectively. ACL-P is not only reweighting features already known to clinician: The model could have achieved its performance by merely adjusting known clinical factors and reweighting them to generate a score. To evaluate this, we calculated the coefficient of ACL-P in a logistic regression with ACL injury as the dependent variable, initially without controlling for other known features (9.74; 95% CI, 8.44–11.0, Nagelkerke R2 = 0.660), and then compared it to the coefficient when controlling variables were added (10.8; 95% CI, 8.64–13.0, Nagelkerke R2 = 0.854). The small change in the coefficient suggests that the model’s explanatory power for ACL injury was not entirely captured by currently measured features. We found that ACL-P had barely correlation with most features (see supplemental material: Manually measured bone shape, Table S4), with ATT (R2 = 0.220) and lateral tibial slope (LTS) (R2 = 0.104) explaining the largest fraction of variance in ACL-P. Furthermore, ACL-P could not be fully explained by all reweighted bone shape features combined (R2 = 0.360). Explanation of ACL-P After ruling out these explanations, we attempted to understand how ACL-P make the classification based on bone shape. The counterfactual explanation method mainly modified the femoral posterior condyle, femoral trochlear, intercondylar notch, intercondylar eminence, tibia plateau and tibia tuberosity (Fig. 4a). While most features were difficult to be summarized as verbalizable features, we could still recognize some well-known factors associated with ACL injury such as NW (notch width), NH (notch height), LCFCR (lateral condylar flexion circle radius), ATT (anterior tibia translation). Besides, we found that counterfactual explanation model often modified the angle between the femur axis and long axis of the fitting ellipse of femoral condyle, thus we designed a new measurement named LFCA (lateral femoral condylar angle), and manually measured LFCA on 700 MRIs. We proved that larger LFCA was significantly associated with ACL injury (OR: 1.241, 95% CI 1.150–1.339, p < 0.001). Specifically, we did not observe that the counterfactual explanation incorporated chronic ACL injury-related bone shape features, such as bone spurs around the notch, into the cases. Representative examples of the counterfactual explanations and contrastive matching. a The generated cases are from counterfactual explanations, while real cases are from actual patients. b The matched normal and injured ACL cases are from actual patients with high bone morphology similarity. Using contrastive matching, we built 500 pairs of normal/injured ACL cases with most similar bone shape. By comparing the bone shape, we found strong heterogeneity between two groups. Besides the commonly reported posterior tibial slope and anterior tibia translation, we found that the bone shape of two groups differed in the posterior femur condyle, femur trochlear, tibial intercondylar eminence, posterior tibial plateau (Fig. 4b). Discussion In summary, we employed a machine learning algorithm to demonstrate that bone morphology as a strong risk factor for ACL injuries. Given the excellent discriminative capability of our new algorithmic measure, ACL-P, we propose that ACL-P could potentially assess the risk of ACL injury. This tool could be valuable for evaluating injury risk and aiding in injury prevention for elite athletes. The risk factors for ACL injury are multifaceted, with bone morphology long considered integral to its pathogenesis. Over the past decades, clinicians have identified several bone morphology characteristics as risk factors, such as intercondylar notch width, posterior tibial slope, tibial eminence size, and Q angle23. Numerous indices have been developed to describe bone morphology and establish thresholds for high risk24–27. However, the extent to which bone morphology contributes to ACL rupture susceptibility remains controversial. Micicoi measured four risk factors in 382 healthy participants, namely, notch width index (NWI), lateral femoral condylar index (LFCI), medial posterior plateau tibial angle (MPPTA), and lateral posterior plateau tibial angle (LPPTA), and found that only 14.4% exhibited no ACL bony risk factors23. The high proportion of healthy participants deemed at risk raises questions about the specificity and sensitivity of these indices. In this study, we employed a machine learning algorithm to correlate 3D bone morphology with ACL injury. Using solely bone shape data, ACL-P distinguished ACL injury patients with an AUC of 0.887 in the Civilian Test Set and 0.907 in the Athlete Test Set. By leveraging the power of machine learning in non-linear regression, we can analyze bone morphology from a higher-dimensional perspective. We demonstrated that ACL-P is not merely a reweighted combination of 2D features or a reflection of post-injury bone changes. The robustness of this method, combined with extensive test sets, reinforces the conclusion that bone morphology is a significant risk factor for ACL injury. ACL injury is a career-threatening trauma for athletes, leading to prolonged absence from sports, a high risk of subsequent knee injuries, and long-term consequences such as early-onset osteoarthritis. Preventing ACL injury is thus a priority in sports medicine and physiotherapy research28. Prevention strategies encompass a wide range of interventions, including neuromuscular training, proprioceptive exercises, stretching, plyometrics, movement training, core strengthening, balance training, resistance training, and speed training29. Despite these efforts, preventing ACL injuries remains challenging. Over the past 20 years, the overall ACL injury rate among professional athletes has not declined29, in contrast to the significant decrease in medial collateral ligament injuries and ankle sprains30,31. For athletes consistently exposed to inciting events during training or competition, the strong association between bone morphology and ACL injury may explain the difficulty in preventing and the ease of recurrence of ACL injuries. As a significant intrinsic risk factor, bone morphology is not easily modifiable; however, this tool allows us to identify individuals at increased risk of ACL injury or recurrence. A deeper understanding of predisposing anatomical factors can inform athletes about their risks and enhance their motivation for injury prevention32. For patients with high ACL-P scores specialized rehabilitation and modified sports programs should be considered to reduce the risk of contralateral or graft rupture. Another intriguing finding of this study is the association between the discriminative power of the ACL-P and age. The model demonstrated greater predictive accuracy in younger individuals, with the under-16 subgroup achieving the highest AUC of 0.907. As adolescents did not constitute the majority of the training set, we ensured that the high discriminative power of the ACL-P was not due to data imbalance. It is well-known that younger individuals, both civilians and athletes, have a higher rate of ACL reinjury7,16. It is often theorized that younger patients face increased reinjury risk because they are more likely to return to high-risk pivoting activities33–35. The results of this study suggest an additional explanation: some young patients with immature bone morphology characteristics may have an increased risk of ACL injury. In this study, we observed consistent discriminative performance of ACL-P across both male and female subgroups. However, this stability should be interpreted as evidence of the model’s generalizability and robustness rather than proof that sex-specific morphological risk factors do not exist. While ACL-P identifies a shared skeletal feature associated with ACL injury susceptibility in both genders, our current analysis may not be sufficiently powered to detect subtle, sex-stratified anatomical interactions. Consequently, ACL-P provides a sex-agnostic framework for morphological assessment, while further research is needed to explore potential gender-specific pathways. This study has several limitations: First, the retrospective, cross-sectional design of this study identifies morphological associations rather than definitive causality. While establishing a causal trajectory would ideally require pre-injury baseline MRIs, for a cohort of this magnitude, such an approach is currently unfeasible. We utilized the relative stability of bone shape within the first-year post-injury as a pragmatic proxy for baseline anatomy. Although our robustness checks aimed to mitigate the influence of post-traumatic artifacts, the current findings primarily characterize morphological susceptibility rather than longitudinal risk. Second, in a retrospective framework, the potential influence of post-traumatic morphological remodeling cannot be entirely dismissed. While we restricted our cohort to patients within one year of injury and demonstrated through subgroup analyses that ACL-P performance remains stable across the acute-to-chronic phases (from 0 to 365 days), subtle bone remodeling or secondary changes following trauma may still occur. Given the “black box” nature of deep neural networks, it is impossible to definitively isolate these subtle secondary alterations from intrinsic skeletal traits. Consequently, while our data suggest that the model captures inherent risk factors, the impact of post-traumatic bone shape changes remains a persistent confounding possibility. Third, although our segmentation model achieved high accuracy (Dice coefficient = 0.9685), factors such as severe bone marrow edema, cartilage damage, and significant joint effusion might indirectly affect the precision of bone boundary segmentation and 3D geometry extraction. Given the scale of this dataset, manual verification of every image segmentation was unfeasible, representing a necessary trade-off between dataset breadth and individual granular precision. Fourth, the definition of ground truth was not uniform across all study cohorts. While the injury group consisted of 4545 patients with arthroscopically-confirmed diagnoses, the control group comprised individuals with structurally normal MRIs without surgical verification. Due to ethical considerations, it is unfeasible to obtain surgical confirmation for asymptomatic or non-surgical candidates. Although we implemented a dual-reading protocol involving senior musculoskeletal radiologists and strictly excluded any cases with ambiguous or suspect pathology to minimize false positives and negatives, the possibility of subtle radiological misinterpretation cannot be entirely eliminated. Fifth, the composition of our control group reflects a pragmatic clinical cohort rather than the general asymptomatic population. Our controls included symptomatic patients with structurally normal MRIs, a design that follows established risk-factor analysis methodologies by comparing injured patients to those with similar clinical exposures. While our subgroup analyses demonstrated robust discriminatory power in distinguishing ACL injuries from both structurally normal knees (AUC = 0.905) and other knee pathologies (AUC = 0.858), suggesting potential generalizability. However, we acknowledge that underlying biomechanical abnormalities in the control group could still exist, and future research involving long-term prospective tracking of asymptomatic cohorts would complement our findings on the causal pathways of bone morphology and injury risk. Sixth, the current findings were derived exclusively from Chinese populations, posing a potential demographic bias. Given that bone morphology exhibits significant variability across different ethnicities, the generalizability of our ACL-P model to non-Asian groups remains to be established. To facilitate global validation, we have open-sourced our 3D bone morphology data and classification algorithms, inviting international collaborators to recalibrate and test this framework across diverse ethnic groups. Finally, identifying the specific high-dimensional features utilized by the neural network remains an ongoing challenge. While we have leveraged counterfactual explanations to bridge the gap between AI discovery and clinical intuition, leading to the identification of reproducible biomarkers such as the Lateral Femoral Condylar Angle (LFCA), much of the model’s internal reasoning remains opaque. To address this and empower further investigation, we have publicly released our counterfactual explanation results, inviting the global communities to collaboratively identify additional morphological biomarkers and potential surgical targets. Decoding these complex features into standardized, manually measurable clinical targets is a crucial objective for our future research to enhance the translational utility of this framework. Caution is advised because, despite ACL-P’s strong correlation with ACL injury, the pathogenesis is still multifaceted, resulting from a complex interaction of internal, external risk factors and injury mechanisms. A promising option for integrating our algorithm into sports injury prevention programs is to use it as a decision aid to screen at-risk athletes. Additionally, this algorithm could be potentially applied in clinical practice to identify patients at high risk of reinjury, allowing for specialized rehabilitation programs and adjustments in sports and lifestyle for high-risk patients. Such collaboration between humans and algorithms might decrease ACL incidence in athletes and improve clinical outcomes for ACL injury patients. More broadly, our results illustrate how deep learning algorithms can be employed to mine data and analyze complex relationships in healthcare. Methods This study was a multicenter, retrospective study designed to investigate the association between 3D bone morphology and ACL injury. The study protocol was approved by the Institutional Review Board of Peking University Third Hospital (PUTH) Groups (IRB00006761-M2024043), and the requirement for informed consent was waived due to the retrospective nature of the data analysis. Study design The study comprised three phases: (1) development of a bone segmentation model to extract femur and tibia morphology from MRI scans and convert them into point cloud representations; (2) training and validation of a classification model (Point Cloud Transformer) to establish the correlation between bone shape and ACL injury status, generating an algorithmic risk score termed ACL-P (ACL-Injury Patterning); and (3) external validation of ACL-P’s discriminative performance across two large, independent test sets: a civilian population (n = 7797) and an athlete population (n = 269) followed by systematic robustness checks and model interpretability analyses. Dataset construction The flow diagram for inclusion and exclusion of Training and Test Sets are summarized in Fig. 5. Flow diagram for inclusion and exclusion of Training and Test Sets. a Flow diagram of Training Set; b flow diagram of Civilian Test Set; c flow diagram of Athlete Test Set. Training set The training set was extracted from our continuously updated multicenter MRI database, which has cataloged 41,636 knee MRIs from six centers since 2023. These centers include PUTH-Main Center, the Second Affiliated Hospital of Xi’an Jiaotong University, Shanxi Jincheng Hospital, Inner Mongolia Wuhai Hospital, the Second Hospital of Inner Mongolia Medical University, and Qilu Hospital of Shandong University-Dezhou Center. The MRIs were consecutively collected in each center. The inclusion criteria of the training set were: (1) non-athlete patients with non-contact mechanism injury; (2) complete patient demographic information and injury time. MRIs with an injury period exceeding one year were excluded. We then stratified and sampled 5000 MRIs (2500 with ACL injury and 2500 without) to construct the training set for this study. If a patient had multiple knee MRIs from one or both sides, only one MRI was randomly selected. Civilian Test Set This test set included both injured and structurally normal knees. For the injured knee MRIs, we searched the hospital information system (HIS) of the PUTH Groups, encompassing three centers: PUTH main center, Beijing Capital International Airport Center, and Northern Center. The inclusion criteria were: (1) all patients who underwent knee arthroscopy surgery from January 1, 2019, to December 31, 2022; (2) patients aged 8 to 50 years old, with prospectively collected demographic information, injury history, and ipsilateral knee MRI; (3) patients without knee fractures. The exclusion criteria were: (1) patients duplicated with the Training Set; (2) patients with a history of previous knee surgery; (3) patients certified as athletes by the General Administration of Sport of China; (4) the period from knee injury to MRI exceeding 12 months (to avoid chronic bone shape changes); (5) MRIs lacking sagittal fat suppressed PD or T2 series; (6) additional MRIs from the same knee; (7) the period from knee MRI to surgery exceeding 6 months (to prevent inconsistency between MRI and surgical diagnosis); (8) patients without a definite knee injury history or exact date (to avoid chronic bone shape changes). For the structurally normal knee MRIs, we searched the picture archiving and communication system (PACS) of the PUTH Groups. These MRIs typically came from individuals experiencing discomfort but without positive structural findings. The inclusion criteria were: (1) individuals who received knee MRIs from January 1, 2016, to December 31, 2022; (2) MRI reports diagnosed as “normal knee.” The exclusion criteria were: (1) patients without demographic information; (2) patients duplicated with the Training Set or already included in the injury subgroup of the Civilian Test Set; (3) patients certified as athletes by the General Administration of Sport of China; (4) MRIs lacking sagittal fat suppressed PD or T2 series; (5) additional MRIs from the same knee. Athletes Test Set the inclusion criteria were: (1) athletes certified by the General Administration of Sport of China, with certification levels including international master of sports, master of sports, and national first, second, or third levels; (2) only active athletes were included, while retired athletes were considered civilians; (3) athletes who underwent knee arthroscopy surgery in the Department of Sports Medicine of PUTH main center from January 1, 2016, to December 31, 2022; (4) injuries were restricted to sports trauma and training overuse injuries. Other inclusion and exclusion criteria were the same as those for the Civilian Test Set. Ground truth The ground truth for injured knee MRIs in Civilian Test Set and Athletes Test Set was determined by surgical diagnosis. For the structurally normal knee MRIs in Civilian Test Set, the ground truth is from MRI reports, which were conducted by one junior and one senior musculoskeletal radiologist at PUTH. Any MRIs with suspect pathology of ACL and other structures are directly excluded from the test set. Image processing and bone morphology extraction We initially trained a segmentation model to segment the femur and tibia on sagittal MRI sequences, then converted the bone segmentation into point cloud to extract the bone morphology. Training data for segmentation A total of 3812 sagittal images from 200 knee MRIs were selected for training the segmentation model. Each image was accompanied by manually annotated femur and tibia boundary masks, verified by one junior and one senior musculoskeletal radiologist at PUTH. All images were normalized to a size of 512 × 512 pixels. The training set was partitioned with an 8:2 ratio for training and internal validation. Segmentation model architecture The segmentation model utilized U2Net as its backbone, renowned for its ability to capture boundary details with high performance36. The U2Net architecture is a variant of UNet37 following an Encode-Decode framework (Fig. 6a). It incorporates multiple RSU (Residual Symmetric U-Net) modules within a Feature Pyramid Network (FPN) structure at various levels to enhance segmentation results. Unlike conventional approaches in salient object detection (SOD) that rely on feature extraction from established backbones like ResNet38 and DenseNet39, the RSU module in U2Net is specifically designed to capture comprehensive contextual information. This design enables seamless training from scratch using in-house datasets, producing high-resolution segmented images without significantly increasing memory usage and computational complexity. The architecture of the models of this study. a Segmentation model and b classification model. Segmentation model validation we constructed an extra external test set of 100 MRIs with 2076 sagittal images. The Dice coefficient was calculated to evaluate the segmentation model (see supplementary materials: Comparison of 2D and 3D Bone Segmentation, Table S1 and Fig. S1). Point cloud conversion the bone shape derived from the segmentation model was transformed into a point cloud format (see supplementary materials: Point cloud transformation and Table S2). This format preserves only the spatial information of the bone structure, eliminating redundant data compared to using masks for classification, thus providing a more precise representation of bone morphology and better elucidating the relationship between bone shape and ACL tears. Development of ACL injury classification model We developed a classification model to determine ACL status based on bone morphology. If a correlation between bone shape and ACL injury exists, this task will quantify it in terms of classification accuracy. Given that point cloud data is represented as a sequence of three-dimensional coordinates, we employed the Point Cloud Transformer (PCT) method as the feature extractor40. The structural diagram of the classification network is depicted in Fig. 6. Compared to the basic Transformer architecture41, PCT introduces enhancements such as coordinate-based feature encoding, an adaptable offset-attention module, and a localized information extraction module specifically designed for point cloud data. Within the PCT framework, after the initial one-dimensional convolutional feature extraction, multi-step feature extraction is performed through successive attention mechanisms, and multi-scale feature fusion is achieved using LBR (Linear, Batchnorm, ReLU) operations40. The resulting features are then subjected to maximum and average pooling before being processed through linear layers to produce the final classification output. The output of the ACL classification model gives probability of the ACL injury according to the bone shape. Given that bone shape barely changes after an ACL injury, this probability can be considered a predispositional risk factor attributable to bone shape. We have named this new algorithmic measure of ACL injury risk ACL-P (ACL-Injury Patterning). Model evaluation and statistical analysis Following the extraction of 3D bone morphology as point clouds, the ACL classification model was applied to the Civilian and Athlete Test Sets for validation. We evaluated the model’s performance using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy, with 95% confidence intervals (CIs) calculated for all metrics. Subgroup analyses were performed across various categories, including sex, age, injury cause, injury type, and injury mechanism. All primary statistical analyses were conducted in R software (version 4.5.1) using the pROC package (version 1.19.0.1). Pairwise comparisons of AUCs between subgroups were performed using DeLong’s non-parametric test, while the overall equality of AUCs across multiple categories was assessed using an omnibus Wald χ2 test based on DeLong variance estimates. To assess the agreement between predicted probabilities and observed outcomes, calibration was performed in Python (version 3.9.19) using Isotonic Regression via scikit-learn (version 1.6.1). The calibration quality was quantitatively evaluated with the Hosmer-Lemeshow (HL) goodness-of-fit test and visually inspected through calibration curves generated using matplotlib (version 3.8.4). To further investigate whether the model’s performance was driven by intrinsic skeletal anatomy or confounding factors, we conducted three robustness checks: Robustness Check 1 ACL-P is not primarily classifying based on post-traumatic bone shape changes. The model might achieve its predictive performance by identifying changes such as osteophytes after an ACL injury. To test this, we calculated the correlation between ACL-P and the time from injury to MRI using Pearson’s R2 using Python (version 3.9.19). We also divided the test set by injury duration to see if ACL-P’s discriminative ability increased with longer post-traumatic periods. Robustness Check 2 ACL-P does not rely solely on classifying from tibial anterior dislocation caused by ACL injury. The model might achieve its performance by recognizing post-injury tibial dislocation. To assess this, we randomly selected 700 MRIs from the test set and manually measured anterior tibial translation. The correlation between ACL-P and tibial anterior dislocation was evaluated using Pearson’s R2. We further modified the ACL classification model to input separate femur and tibia data, removing all location information, including tibial anterior dislocation. The discriminative ability of ACL-P was recalculated for the two test sets. Robustness Check 3 ACL-P is not merely reweighting features already known to clinicians. The model might perform by emphasizing known factors, such as medial tibial slope, over others like notch width. To evaluate this, we randomly extracted 700 MRIs from the test set and manually measured 36 bone shape parameters noted in previous studies (see supplementary materials: Manually measured bone shape and measurement of the 36 commonly reported bone shape parameters, Figs. S3–S7). Two sports medicine clinicians performed these measurements. The intraclass correlation coefficient (ICC) assessed intraobserver and interobserver agreement. We calculated the coefficient of ACL-P in a logistic regression model, first without controlling for known features, using Nagelkerke’s R2, and then compared it to the coefficient when controlling variables were added. If the coefficient remained unchanged, it suggests that the model’s explanatory power was not fully captured by currently measured features. We also calculated the correlation between ACL-P and manually measured features using Pearson’s R2 (see supplementary materials: Correlation of the measured parameters with ACL-P, Table S4). Explanation of ACL-P To tackle the challenge of understanding the decision-making process of the classification model, we employed two strategies: counterfactual explanations42 and contrastive matching. Counterfactual explanations By introducing slight modifications to the input data, the classification model can produce an opposite output. This process highlights the regions that significantly impact classification, offering valuable insights into the model’s internal workings and the factors it deems most important. A limitation of this strategy is that the model might generate a bone shape that does not exist in reality. In this study, we designed a 3D-counterfactual method based on Generative Adversarial Networks (GANs). We represent the 3D shape of bone morphology using point clouds. To elucidate the classification model, we train a point cloud generative network that takes point clouds as input and outputs modified versions. The output should be minimally altered and conform to the distribution of real bone morphology. Additionally, the output by the classification model should be reversed (from normal to ACL injury or vice versa). Thus, the modified region can be identified as the critical area in predicting ACL rupture. By generating these counterfactual examples, we can pinpoint the specific features or regions of bone morphology crucial for the model’s decision-making process (Details see supplementary materials: Counterfactual explanation, Fig. S2 and Table S3). We released representative cases of counterfactual examples online (see Data availability section) to help researchers to identify reproducible morphological biomarkers or clinically measurable features. Contrastive matching The concept of contrastive matching is similar to counterfactual explanations. By comparing the differences in bone shapes with high and low ACL-P scores, we can infer which bone areas the model emphasizes. Due to individual variations in bone shape and size, we pair bone shapes with opposite ACL-P classification in a case-control manner. In this study, we selected 500 MRIs with the lowest ACL-P scores to form the “algorithmic normal group”, and 500 MRIs with the highest scores to form the “algorithmic injury group” from the Civilian Test Set. Each bone shape in the “algorithmic normal group” was paired with the most similar one in the “algorithmic injury group.” The similarity of the two bone shapes, in terms of the point cloud, was calculated using chamfer distance. It operates by computing the average of the minimum distances from each point in one set to its nearest neighbor in the other set, and vice versa. We opted for this approach primarily due to its exemplary computational efficiency and its notable robustness in the presence of noise and outliers. Ultimately, we obtained 500 pairs of bone shapes and analyzed the differences between them. Supplementary information Supplementary Information Acknowledgements This work was supported in part by the Beijing Natural Science Foundation under Grant L242104, in part by National Natural Science Foundation of China under Grant 82441025 and in part by Beijing Municipal Science and Technology Program under Grant Z251100006025005. We also acknowledge High-Performance Computing Platform of Peking University for providing computational resources. Author contributions D.Y.W, C.L., T.T.J., and D.J. designed the study. D.Y.W. built the datasets, conducted the data analysis and wrote the primary draft of the manuscript and all revisions. C.L. built the bone shape segmentation model and ACL classification model. X.W.D. and S.L.Z. determined the measurement methods of the bone shape parameters and conducted the manual measurement. X.Q.Y. designed and built the counterfactual explanations model. J.X.L. and S.G.L assisted with study coordination and database management. T.T.J. and D.J. guided the study conduct and were responsible for study results and interpretations. All authors had full access to all the data in the study, approved the final manuscript, and had the final responsibility for the decision to submit for publication. Data availability The 3D bone morphology point cloud datasets and counterfactual explanation examples are available at [https://drive.google.com/drive/folders/12cKL5yRYuvV83bchv35_C4AEI7yYgA6o?usp=drive_link]. Code availability The code used in the current study for developing the algorithm is provided at [https://github.com/RunningPill/ACL_P]. Python version 3.8.18 was used for all statistical analyses in this study. The following third-party Python packages were used: Pytorch version 1.13.0 was used to build the DL models. Open3d version 0.17.0 and matplotlib version 3.7.4 was used for reconstructing the point cloud. Pointnet2-ops version 3.0.0 was used for unifying sampling point cloud. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Dingyu Wang, Chao Liu. Contributor Information Tingting Jiang, Email: ttjiang@pku.edu.cn. Dong Jiang, Email: bysyjiangdong@126.com. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41746-026-02903-z. References Associated Data Supplementary Materials Supplementary Information Data Availability Statement The 3D bone morphology point cloud datasets and counterfactual explanation examples are available at [https://drive.google.com/drive/folders/12cKL5yRYuvV83bchv35_C4AEI7yYgA6o?usp=drive_link]. The code used in the current study for developing the algorithm is provided at [https://github.com/RunningPill/ACL_P]. Python version 3.8.18 was used for all statistical analyses in this study. The following third-party Python packages were used: Pytorch version 1.13.0 was used to build the DL models. Open3d version 0.17.0 and matplotlib version 3.7.4 was used for reconstructing the point cloud. Pointnet2-ops version 3.0.0 was used for unifying sampling point cloud.