Artificial intelligence in foot and ankle surgery: current concepts

With many patients seeing nonorthopedic care providers for foot and ankle radiograph interpretation, DL and AI can be important in getting patients accurately and quickly diagnosed and referred to more specialized providers. Convoluted neural networks (CNNs), a form of DL, recognize visual patterns from raw image pixels which makes them potentially useful for medical imaging. While CNNs developed for radiographic images demonstrate high fracture detection, they are ultimately limited in that radiographs provide only a 2D representation of 3D joints. To address this, AI for ankle and foot fracture detection expands beyond radiographs to computed tomography (CT) imaging as well. Through de novo and pretrained CNNs, DL has been found to successfully detect and accurately classify 92–98% of Sanders calcaneal fracture types [7].

The field of robotics for use in surgery is not limited to robotic arm applications for intraoperative assistance. AI has been also advocated for imaging analysis, patient-specific instrumentation in preoperative planning, and robotics-aided rehabilitation [8‐10].

Several authors have described the use of AI in the diagnosis and treatment of ankle and foot fractures. Ashkani-Esfahani et al. internally validated two deep convolutional neural networks (DCNN) for identifying ankle fractures from radiographs and achieved a near-perfect area under the curve (AUC) of 0.99 [11]. Kitamura et al. internally validated 5 separate CNNs for detecting ankle fractures from plain radiographs and achieved a fair fracture detection accuracy of 81% [12]. Prijs et al. internally and externally validated a DL model for detecting, classifying, and localizing ankle fractures from plain radiographs and achieved an excellent AUC of 0.92 and an accuracy of 99 % on external validation [13]. Guermazi et al. internally validated a DL model for detecting fractures from foot and ankle plain radiographs, which performed excellently with an AUC of 0.97, sensitivity per patient of 93%, and specificity per patient of 93% [14]. Olczak et al. internally validated neural network models for classifying ankle fractures from radiographs according to the AO Foundation/Orthopaedic Trauma Association (AO/OTA) 2018 classification, which showed fair to excellent performance with AUCs ranging from 0.79 to 0.99 in classifying AO types [15]. Pinto Dos Santos et al. internally validated a CNN for detecting fractures in anteroposterior ankle radiographs, which performed well with an AUC of 0.85 [16]. Ashkani-Esfahani et al. internally validated 2 DCNN models for detecting Lisfranc instability from single-view (anteroposterior) and 3‑view radiographs (anteroposterior, lateral, oblique), which performed excellently with AUCs ranging from 0.925 to 0.994 [11]. Aghnia Farda et al. internally validated a CNN model for classifying calcaneal fractures on CT images into the Sanders system, which performed well with a classification accuracy of nearly 72% after augmenting the data 17. Pranata et al. internally validated 2 separate DCNN models for detecting the presence or absence of calcaneal fractures on CT images and achieved an excellent accuracy of 98% [17]. Hendrickx et al. internally validated 4 ML and DL models for predicting patients with tibial shaft fractures and associated occult posterior malleolar fractures. The models performed well with AUCs ranging from 0.81 to 0.89 [18]. Oosterhoff et al. internally validated 5 models for predicting posterior malleolar involvement in distal tibial shaft fractures using the same data set as that in the previously described study by Hendrickx et al [19]. Oosterhoff et al. found that all the models performed well with AUCs  0.80 (highest 0.89) and 4 of 5 having a Brier score of 0.11 [19].

Wang et al. internally validated several radiomics-based ML models for diagnosing Achilles tendinopathy from ultrasonographic images in skiers and achieved an excellent AUC of 0.99, 90% sensitivity, and 100% specificity [20]. Kapiński et al. internally validated several DL models to classify Achilles tendons injuries on MRI scans, achieving a maximum accuracy of 97.6%, a sensitivity of 98.3%, and a specificity of 99.45% [21]. Merrill et al. internally validated a logistic regression and gradient boosting model for predicting short-term complications, including mortality and readmissions, in patients who have undergone open reduction and internal fixation (ORIF) in acute ankle fractures. Both models performed similarly, with AUCs for gradient boosting ranging from 0.6979 to 0.7580 and AUCs for logistic regression ranging from 0.7101 to 0.7583 [22].

Li et al. aimed to internally validate a DL model to detect 18 anatomical landmarks from weight-bearing radiographs, including the hallux valgus angle (HVA), hallux interphalangeal angle (HIA), first-second intermetatarsal angle (IMA), and distal metatarsal articular angle (DMAA). The observed (manual by a radiologist) and predicted (model) values of the 4 angles correlated well (intraclass correlation: 0.89–0.96, r = 0.81–0.97) [23]. Day et al. aimed to assess the performance of an AI-based software that automatically measures the M1–M2 IMA from weight-bearing cone beam computed tomography (WBCT) scans in patients with hallux valgus. The AI-based software was faster than manual measurements, correlated well with manual measurements, and had higher and nearly perfect test-retest reliability (0.99 intrasoftware intraclass correlation coefficient for both 3D and 2D IMA) [24]. Wang et al. validated a support vector machine model to classify patients with symptomatic hallux valgus using HVA, IMA, and DMAA, with a fair accuracy of 76.4% [20].

Wang et al. internally and externally tested a DL system for detecting and grading fatigue fractures (a type of stress fracture) from plain radiographs, which performed excellent (AUC 0.911, sensitivity 90.8%) in the detection of fatigue fractures for the foot images and good (AUC 0.877, sensitivity 85.5%) for the tibiofibula images. External validity for grading of fatigue fractures was not demonstrated as the DL system performed poorly with an overall accuracy of 62.9% for the tibiofibular images and an accuracy of 61.1% for the foot images [20].

Diniz et al. internally validated one ML model for predicting whether soccer players would return to similar performances after Achilles tendon rupture, achieving a good AUC of 0.81 and a Brier score loss of 0.12 [25]. Lu et al. internally validated many ML models for predicting the occurrence of a lower extremity muscle strain (quadriceps, calf, hamstring, groin) in elite basketball players [26]. Among them, the XGBoost model achieved the highest AUC of 0.840, representing the best-performing model if the Brier score and calibration were also considered [25]. Jauhiainen et al. internally validated 2 ML models for predicting moderate and severe knee and ankle injuries in young basketball and floorball players (age ≤ 21 years), which performed poorly with an AUC of 0.63 for the random forest model and 0.65 for the logistic regression model [27]. Ruiz-Pérez et al. internally validated many ML models to predict lower extremity non-contact soft tissue injury in professional futsal players, which generally performed fairly, with the best model achieving an AUC of 0.767, a sensitivity of 85.1%, and a specificity of 62.1% [28]. Suda et al. internally validated several support vector machine models for classifying running experience levels based on foot-ankle kinematic and kinetic patterns to potentially assist with running rehabilitation and training. The models performed well with classification accuracies of 88.5% for less experienced runners, 87.2% for moderately experienced runners, and 84.6% for experienced runners [29].

Yin et al. internally validated a neural network model for predicting patients that would achieve the minimum clinically successful therapy (decrease in the visual analogue score, VAS, by 60% or more from baseline) at 6 months following extracorporeal shock wave therapy (ESWT) in patients with chronic plantar fasciitis. The model performed well, with an overall accuracy of 92.5%, a sensitivity of 95.0%, and a specificity of 90.0% [30]. Keijsers et al. internally validated a neural network model for differentiating patients who have forefoot pain and those that do not use plantar pressure data, which performed satisfactorily with an accuracy of 70.4% [31]. Zhu et al. investigated whether AI-assisted ultrasonography-guided needle knife therapy improves the outcomes of patients treated for chronic plantar fasciitis. Patients who were allocated to the AI-assisted group evidenced statistically significant higher American Orthopaedic Foot & Ankle Society (AOFAS), lower plantar fascia elasticity scores and plantar fascia thickness at 2, 4, and 8 weeks of follow-up [32].

Hernigou et al. applied AI and ML to assist in conducting their study for developing a method of defining the ideal and patient-specific motion axes of the tibiotalar joint, intending to improve robotic-assisted total ankle arthroplasty (TAA) [33].

Ardhianto et al. applied DL to help with the automated measurement of the foot progression angle (FPA) from plantar pressure images, helping clinicians in assessing gait abnormalities [34].

Pakhomov et al. applied ML to automate the identification and classification of foot examination findings from clinical notes as normal, abnormal, or not assessed, and their models performed well with overall accuracies ranging from 81% to 87% [35].

While foot and ankle surgery has lagged behind other orthopaedic specialities, employing and studying robotics more extensively in this field is necessary. CNNs can be trained for autonomous outcome prediction and are currently focused on fracture detection with projected optimization in a multitude of clinical settings. Lastly, considering post-injury and post-surgey outcomes, robotic foot braces, emulators, and assistive limb devices have a variety of adaptive functions with options for real-time patient feedback that profoundly individualize patient rehabilitation.

Advancements in robotic-assisted TKA and THA demonstrated good clinical outcomes, showing a promising future for application in TAA; however, because of the broad range of foot and ankle surgery with lower volumes in singular procedures than arthroplasty, significant cost barriers exist for the widespread adoption of these technologies. Translational cadaveric studies might help clarify the native mechanical strains and injury biomechanics of ankle joints, test the current TAA systems, and introduce novel machinery for hands-off fracture reduction. At the clinical end, robotics and computer-based systems are being employed for increased precision in TAA and trauma, but these developments are less extensive when contrasted with THA and TKA robotics. Therefore, contained air solutions (CAS) and robots with open technological capacities will likely be more widely adopted in the coming years for use in the foot and ankle; however, improving implant positioning with robotic-assisted TAA can lead to a reduction in long-term healthcare costs, especially given the high failure rates of TAA compared to other joint replacements. If open robotic systems are also developed with capabilities for other procedures that often accompany TAA, such as soft tissue manipulations, longitudinal costs and outcomes will likely be significantly improved both in the operative suite and for patient quality of life.

With future improvements in ankle prostheses, orthotics, and therapeutics on the horizon, further work would help optimize the design of these systems to create more lightweight devices to reduce mechanical work on behalf of the user and to recreate better natural motion [36]. The expansion of the ankle orthosis to a foot-ankle-knee orthosis for more debilitating pathologies has also been described in the literature [37]. Other suggestions include individualized protocols that are tailored to individual patient needs rather than a standardized, one size- fits all protocol. Ultimately, patients will benefit from these technologies through modifiable products promoting individualized recovery, lending to improved post-surgery outcomes.

Advancements in AI and DL will allow for incorporation in the primary care and acute care settings for increased efficiency and accuracy of ankle pathology diagnoses. Especially regarding scenarios in which practitioners are less familiar with complex orthopedic injuries, these systems can close a gap in knowledge in practice while decreasing the cost of care and time spent interpreting radiographic imaging for more swift referrals, treatment plans, and time to surgical intervention. Given the impressive precision and accuracy of these algorithms, another application is telehealth, allowing for remote diagnostics, potentially without a radiologist’s interpretation. Beyond fracture detection, AI systems can also be employed to inform surgeons of patient-specific projected outcomes based on prior data patterns, answering questions such as “What is my patient’s risk of reoperation or implant failure?” or “How long until this patient is back to work?”.

There exists a vast potential for the application of robotics in the realms of preclinical and translational research, clinical evaluation (e.g., with AI), preoperative planning, and CAS robotics, among others. Future research should be aimed at incorporating robotic technologies specifically into surgical procedures and clinical practice, for which cadaveric translational studies have proven to be an accurate and replicable pipeline.

In vitro and in vivo gait simulators can begin to transition to human subjects; however, less invasive versions should be first developed. Additionally, because most cadaveric models in the past have been static with one plane of motion, employing more dynamic robotic simulators with more degrees of freedom will allow for a more realistic positioning of the specimens to represent biological motion better. Moreover, these static simulators apply only one or two dimensions of action, such as torque or axial load, over fixed ranges of motion. With knowledge of the complexity of joint loading and strain, it would be of interest to apply these concepts to robotic systems to mirror joint kinematics during daily activities such as walking, lunging, and pivoting. This would also necessitate quantification of these types of loads during these activities, which has yet to be elucidated. This research will enrich our understanding of the ankle joint, which can be directly applied to surgical planning and postoperative therapy and return to motion.