3D topographic acquisitions to predict spinal curvature in adolescent idiopathic scoliosis

Abstract

Aims

This study aims to determine the reliability, accuracy, and usability of a new health application that uses AI to estimate major coronal curve magnitude in patients with adolescent idiopathic scoliosis (AIS) from 3D surface topography (ST) captured on a smartphone video scan.

Methods

This is a prospective validation study. AIS patients, aged ten to 18 years, with coronal curve magnitudes ≤ 45° were recruited at a tertiary care spine clinic. A single trained researcher performed scans twice, six months apart, during participants’ routine clinical and radiological assessment. Participants were asked to complete a scan once a month between clinic visits, starting the day of recruitment. Agreement was calculated by comparing scan curve magnitude predictions to the reference standard: a three-foot standing spine radiograph measured by blinded spine clinicians. Inter-rater reliability was assessed by comparing in-clinic to home scan predictions. Measures of diagnostic accuracy to determine the app’s ability to screen for coronal deformity > 25° and its ability to detect progression > 5° over a six-month period were determined. Successful compared with failed scans were recorded.

Results

Among participants (n = 63), 59 patients (94%) had at least one successful in-clinic scan and 32 patients (51%) had at least one successful home scan. Agreement with the reference standard was moderate for in-clinic scans (intraclass correlation coefficient (ICC) 0.535) and poor for home scans (ICC 0.402). Inter-rater reliability between in-clinic and home scans was poor (ICC 0.168). The app had an accuracy of 70% when discriminating between curve magnitudes ± 25° and detecting curve progression > 5°. A larger proportion of scans failed at-home (30%) compared with in-clinic (16%).

Conclusion

Conceptually, the app shows potential as an accessible screening tool for scoliosis. However, the accuracy and reliability suggest it is not yet a reasonable replacement for radiographs and in-person clinical evaluation.

Cite this article: Bone Jt Open 2026;7(4):473–481.

Introduction

Standard diagnosis and management of adolescent idiopathic scoliosis (AIS) rely on full length spine radiographs performed at least every six months to evaluate curve magnitude and detect meaningful curve progression during skeletal growth.^1,2 It has been estimated that AIS patients will have an average of 9 to 14 radiographs over three-years depending on curve severity and the management plan.³

There is growing concern over the cumulative radiation exposure and associated health consequences in children with AIS.^2,4 Approximately 630 mGy of ionizing radiation is absorbed per radiograph,⁵ and subsequently, several years of repeat radiographs increase the cumulative risk of malignancy.^4,6 Young females with AIS have nearly double the risk of developing breast cancer compared with the general female population.^7,8

Surface topography (ST) offers an alternate approach to assessing spine deformity by evaluating 3D trunk shape. Historically, ST techniques such as scoliometry⁹ and Moire topography¹⁰ have aimed to diagnose scoliosis without radiation but have had limited success with high variability in measurements.^11-13 More advanced systems (i.e. ISIS, Fornometric) have used ST with better ability to detect the presence of scoliosis, particularly a curve magnitude above 20°, but lack accessibility and have meaningful errors in curve magnitude predictions.^14-18

Recently, AI and mobile device applications have emerged as accessible means for applying ST in scoliosis management. Mohamed et al¹⁸ used torso ST captured from an iPad camera fitted with a Structure sensor to distinguish between AIS patients (curve magnitude > 25°) and controls, achieving 95% accuracy. Similarly, Rohde et al¹⁹ evaluated an app that analyzed 3D ST torso scans to estimate the probability of scoliosis > 20°. Using a ≥ 50% probability threshold to classify positive cases, the app achieved an accuracy of 91%. While these apps aim to reduce radiation exposure, they are limited to scoliosis screening without providing information on curve severity.

A new AI-driven health app has been developed that is designed to predict spinal curve magnitude from 3D ST captured on a smartphone video (ST scan). The app aims to reduce radiation exposure through remote patient monitoring, recommending that families perform ST scans once a month at home to monitor curve progression between clinic appointments.

This is the first prospective validation study using this health app. The primary objective of this study was to determine the agreement between the app-predicted curve magnitude and radiological measurements, and inter-rater reliability between scans performed in clinic compared with at home. The secondary objectives were to determine accuracy for identifying moderate curves (≥ 25°) and detecting change in curve magnitude (> 5°), evaluate the app’s usability, and identify factors associated with increased error.

Methods

Study design

This is a prospective validation study comparing curve magnitude predicted by a mobile health app with the reference standard. Using AI, the app analyzes ST modelled from two smartphone videos (ST scan) to predict the magnitude of spinal curvature (°). The reference standard was three-foot standing spine radiographs evaluated by spine clinicians prior to the clinic ST scan being performed. No study team member has a proprietary relationship with the application and/or the development team.

All patients had repeat ST scans and three-foot standing EOS radiographs during two consecutive clinic appointments (initial and six-month follow-up). Between appointments, participants were asked to perform a ST scan at home on the same day of their initial clinic appointment and monthly until follow-up (Figure 1).

Fig. 1.

Study schematic illustrating the outcomes measured. Mo, months; ST, surface topography.

All clinic ST scans were performed by a single trained research assistant (EN) using the patient’s smartphone while teaching the ST scan procedure to the patient and family. The research assistant was blinded to reference standard measures.

For home ST scans, automatic reminders were programmed through the app to notify the patient/family to complete their ST scan one month following their most recent scan. Patients were contacted by a member of the research team if the ST scan was not completed within one week.

This study was conducted in accordance with The Hospital for Sick Children(Protocol no: 1000080724) Research Ethics Board-approved protocol and the manuscript was prepared in conformity with the Standards for Reporting of Diagnostic accuracy (STARD) criteria.

Participants

Participants were recruited between April and November 2024, from a single tertiary care centre. This was a consecutive series of AIS patients that were approached if they had a major coronal curve magnitude ≤ 45° and were scheduled for a follow-up appointment in six months. Patients were excluded if they were aged under ten years old, had nonidiopathic aetiology, if they had previously undergone corrective scoliosis surgery (i.e. posterior spinal fusion), or if they did not have access to the following phone models compatible with the app: iPhone 11 (Apple, USA), Samsung S23 (Samsung Electronics, South Korea), Google Pixel 6 (Google, USA), and Galaxy Note 5 or newer (Samsung Electronics). All patients provided written informed consent. A sample size of convenience was used for study feasibility, and recruitment ended once 50 patients had a successful ST scan completed at the initial clinic visit (Figure 1). This pragmatic target would allow preliminary estimation of agreement and accuracy metrics. At the time of study design, no prior data were available on the expected diagnostic performance of this specific STscan algorithm, making meaningful assumptions for a conventional power calculation unfeasible. Participants were assisted in downloading the app onto a personal smartphone and were logged into individual de-identified accounts.

Participant characteristics

A total of 63 patients consented to participate in the study and had an initial in-clinic ST scan. Patients had a mean age of 13 years (SD 1), and were mostly female (n = 48, 76%), with a mean major curve magnitude of 29° (SD 11°) (Table I).

Table I.

Sample characteristics at initial clinic visit.

Variable	Data
Biological sex at birth, n (%)
Male	15 (23.8)
Female	48 (76.2)
Mean age, yrs (SD)	13.45 (1.29)
Mean height, cm (SD)	153.51 (29.50)
Mean weight, kg (SD)	48.33 (13.01)
Mean BMI, kg/m2 (SD)	19.01 (3.87)
Mean axial trunk rotation (SD)	10.6 (4.6)
Major curve profile, n (%)
Thoracic	40 (63.5)
Thoracolumbar	4 (6.3)
Lumbar	19 (30.2)
Mean major coronal curve, ° (SD)	28.6 (10.6)
Risser score, n (%)
0	23 (36.5)
1	4 (6.3)
2	7 (11.1)
3	13 (20.6)
4	14 (22.2)
5	2 (3.2)
Triradiate cartilage fused, n (%)	50 (79.4)
Brace, n (%)*	59 (93.7)
Mean months bracing (SD)*	20.27 (17.84)

^*Including those prescribed a brace at initial visit.

Surface topography scans

During a ST scan, participants were shirtless or wore a bra if female, with pants lowered below their waist and long hair tied up. A ST scan consists of two 30-second videos of the patient in a standing and forward bend position. In each video, the scanner circles the patient three times, with each circle capturing the patient’s body at a specific angle (i.e. thorax, abdomen, and lower limbs while standing and torso, mid torso cross-section, and lower limbs during forward bend). The videos are then uploaded and processed by the app. Processing takes between 15 and 30 minutes and once complete, two 3D representations of the patient’s body (face excluded) in both positions (standing and forward bend) along with the AI-predicted major curve magnitude are displayed on both the patient-facing app and clinician-facing platform. ST scans may fail if the video image quality is inadequate to generate a 3D model and/or AI-predicted curve magnitude.

AI-predicted curve magnitudes were recorded after each clinic and home ST scan was completed. Throughout the study period, the app underwent periodic software updates affecting the AI-algorithm and altering previous curve magnitude predictions. All ST scan predictions used in this study were made on the most up-to-date algorithm, as of 21 June 2025.

Data collection

Electronic patient health charts were reviewed following each clinic visit. Collected clinical variables included biological sex at birth, age, height, weight, BMI, axial trunk rotation, apex of the major curve, major coronal curve magnitude (reference standard), Risser score, triradiate cartilage fusion, and prescribed treatment.

Outcomes

The primary outcomes were to determine: 1) the agreement in major curve magnitude between the app’s predictions and the reference standard measurements (reference standard vs clinic and home ST scan predictions); and 2) the inter-rater reliability of the app’s predictions (clinic vs home ST scan predictions).

The secondary outcomes included the accuracy of the app to identify moderate sized curves (≥ 25°) and change in curve magnitude (> 5°) between two clinic appointments, determine the app’s usability, measured by the number of successful and failed home ST scan attempts, and identify any factors associated with increased error between reference standard and ST scan predictions. Failed ST-scan attempts were defined as ST scans that failed to produce a curve magnitude prediction.

Work flow

All participants were retained in the study, regardless of whether their initial in-clinic ST scan was successful. If the initial clinic ST scan failed to generate a curve magnitude prediction, the ST scan performed at the patient’s six month follow-up visit, when successful, was used in the agreement and accuracy analyses. This ensured that each participant contributed data wherever a successful ST scan and corresponding radiograph were available.

For at-home ST scans, the scan attempted on the same day as the initial clinic visit was used for reliability analysis when successful. If that scan failed or was not attempted, the one-month home scan was used instead. If both early scans failed, the first successful home ST scan was included provided that six-month radiological assessment confirmed that there was no change in curve magnitude > 5°, allowing an appropriate clinic–home comparison.

This hierarchical approach was prespecified to maximize usable data while preserving temporal alignment with available radiographs. Additionally, this maximized data availability related to overall usability.

Statistical analysis

Descriptive statistics were used to characterize patient demographic and clinical profile. Means and SDs were calculated for continuous variables. Frequencies and percentages were calculated for categorical variables.

Agreement statistics were calculated through intraclass correlation coefficients (ICCs) with a two-way random effects model with absolute agreement. Comparisons with the reference standard were made with both in-clinic and at-home ST scans (i.e. in-clinic agreement and at-home agreement). Comparisons for inter-rater reliability were made between initial clinic and home ST scans. An ICC less than 0.5 was considered poor, 0.5 to 0.75 moderate, 0.75 to 0.9 good, and > 0.9 excellent.²⁰ Bland-Altman plots were constructed to display in-clinic and at-home agreement and inter-rater reliability comparisons.

The accuracy of the ST scans to identify moderate sized curves ≥ 25° and predict change > 5° between two clinic visits six months apart, progression, and absolute change (improvement or progression), were determined through sensitivity, specificity, and positive and negative predictive values (PPV and NPV, respectively). PPV and NPV vary with prevalence (i.e. low prevalence lowers PPV and increases NPV), so positive and negative likelihood ratios (+ LR and -LR, respectively), were also calculated, which are independent of prevalence. Typically, + LR values from 1 to 2 offer limited diagnostic value, 2 to 5 a small increase in likelihood, 5 to 10 moderate, and > 10 strong evidence to rule in a condition and -LR < 0.1 provides strong evidence to rule it out.

The association between clinical presentation and the degree of error between in-clinic agreement was explored. Pearson correlations were used for continuous variables (BMI, major curve magnitude, angle of trunk rotation), t-tests for binary variables (sex, brace-use), and one-way analysis of variance (ANOVAs) for categorical variables (location of major curve apex). Descriptive statistics were used to summarize the app’s usability, measured through the number of successful and failed clinic and home ST scan attempts. Data analysis was performed with SPSS v.29.0 (IBM, USA). A p-value < 0.05 was considered significant.

Results

Sample characteristics

Most participants (n = 53, 84%) had a successful initial clinic ST scan, including three ST scans that initially failed but were successful following a software update. Three patients (5%) withdrew from the study prior to their six-month clinic visit, and four more withdrew from the study (6%) at their six-month visit. Reasons for withdrawal include discomfort with the scan procedure (n = 4), wanting to avoid attention on their scoliosis (n = 1), and a lack of interest in performing the scans (n = 1). One patient did not specify the reason for withdrawal. Four patients were lost to follow-up (6%). Of the remaining 52 patients who were scanned during their second clinic appointment, nine patients (17%) had failed ST scans. Four patients (6%) had failed ST scans at both clinic appointments (Figure 2).

Fig. 2.

Flowchart of participants and successful compared with failed clinic surface topography (ST)-scans. Flowchart reflects ST scan counts based on the most up to date software update at time of study.

Agreement between clinic ST scan and reference standard

There were 59 patients (94%) that had at least one successful in-clinic ST scan ( 53 from the initial visit, and six from the six-month visit) and were included in the agreement analyses with their corresponding radiological measurement. There was moderate agreement between in-clinic ST scan predictions and the reference standard (ICC 0.535, 95% CI 0.254 to 0.717; p < 0.001). Compared with the reference standard, the in-clinic ST scan predictions had a mean (SD) absolute difference (MAD) of 8.5° (6.4°) and 40.7% of ST scans (n = 24) were within 5°of the reference standard. On average, ST scans underestimated the curve magnitude by a mean of 5.2° (SD 9.4°) (Figure 3a).

Fig. 3.

Bland-Altman plots comparing illustrating inter-rater reliability by comparing: a) radiological measurements with in-clinic AI-predicted curve magnitudes, b) radiological measurements with at-home AI-predicted curve magnitudes, and c) in-clinic with at-home AI-predicted curve magnitudes. Compared with radiological measurements: in-clinic scan predictions were underestimated a mean of 5.19° (95% limits of agreement from -13.18 to 23.55) and at-home scan predictions underestimated curve magnitude by 4.41° (95% limits of agreement from -16.75 to 25.57). Compared with in-clinic predictions, scans performed at-home underestimated curve magnitude by an average of 1.22° (95% limits of agreement from -22.16 to 24.60).

Agreement between home ST scan and reference standard

Home ST scans were compared with radiological measurements in 32 participants (51%) who took a successful immediate (n = 13) or a one-month ST scan following their initial clinic visit (n = 15) or showed no curve progression and took at least one home ST scan between clinic visits (n = 4).

There was poor agreement between home ST scans and the reference standard (ICC 0.402, 95% CI 0.067 to 0.657, p = 0.004). Home ST scans showed a MAD of 9.4° (SD 6.7°) and 37.5% of ST scans (n = 12) had a MAD ≤ 5°compared with the reference standard. Home ST scans underestimated the curve magnitude by a mean of 4.4° (SD 10.8°) compared with the reference standards (Figure 3b).

Inter-rater reliability between clinic and home ST scans

In total, 32 patients with a successful clinic ST scan completed at least one successful home ST scan and were included in the inter-rater reliability analysis. There was poor inter-rater reliability between clinic and home ST scans (ICC 0.168, 95% CI -0.194 to 0.487; p = 0.180). The MAD between clinic and home ST scans was 9.9° (SD 6.5°) and the mean (SD) difference was 1.2° (11.9°) (Figure 3c). Overall, 25% of initial home ST scans were within 5° of the clinic ST scan.

Accuracy of the ST scan to detect moderate scoliosis

The accuracy of ST scans to detect curve magnitudes ≥ 25° was 69.5%, with a sensitivity of 66.7% and specificity of 73.1% (Table II). The PPV was 75.9% and NPV 63.3% based on a prevalence of 55.9% (n = 33/59) having moderate scoliosis. Those flagged as having a moderate curve were slightly more likely to truly have one (+LR = 2.48), whereas those classified as not having a moderate curve were modestly less likely to have one (–LR = 0.46).

Table II.

Performance of the surface topography scan to detect a moderate size curve.

	Reference standard
ST scan prediction	25° to 45°	< 25°
25° to 45°	22	7
< 25°	11	19

ST, surface topography.

Accuracy of ST scan to detect curve progression and absolute change

A total of 37 patients (59%) had successful ST scans at both initial and follow-up appointments. The accuracy of the ST scans to detect curve progression > 5° was 70.3% with a sensitivity of 66.7% and specificity of 70.6% (Table III). The PPV was 16.7% and the NPV was 96% based on a prevalence of 8.1% (n = 3/37). Those identified as having progressed were slightly more likely to have truly progressed (+LR = 2.27), while those classified as not progressing were modestly less likely to have progressed (–LR = 0.47).

Table III.

Performance of the surface topography-scan to detect curve progression.

	Reference standard
ST scan prediction	> 5°	≤ 5°
> 5°	2	10
≤ 5°	1	24

ST, surface topography.

The accuracy of the ST scans to detect absolute change in curvature > 5° (progression or improvement) was 35.1% with a sensitivity of 62.5% and specificity of 34.5% (Table IV). The PPV was 20.8% and the NPV was 76.9% based on a prevalence of 21.6% (n = 8/37). With a + LR of 0.95 and -LR of 1.09 there was minimal diagnostic value in detecting absolute change > 5°.

Table IV.

Performance of the surface topography-scan to detect curve change (improvement or progression).

	Reference standard
	> 5°	≤ 5°
ST scan prediction
> 5°	5	19
≤ 5°	3	10

ST, surface topography.

Usability/home engagement

In total, 42 patients (67%) attempted at least one home ST scan between clinic appointments, averaging 1.9 attempts of six. Across all at-home ST scans, 29.6% failed on first attempt. Among failed first attempts (n = 34), 26% (n = nine) were retried until a successful ST scan was produced (Figure 4).

Fig. 4.

Successful and failed home surface topography (ST) scans on first attempt. Light grey shows successful ST scan attempt, while dark grey shows failed ST scan attempt.

Clinical factors associated with ST scan prediction error

Curve magnitude was associated with both absolute difference (p = 0.021) and difference (p < 0.001) between the reference standard and clinic ST scans. Specifically, smaller curves tended to be overestimated with less error, and larger curves tended to be underestimated with greater error. Curve apex, axial trunk rotation, prior brace treatment, sex, and BMI were not significantly associated with absolute difference or difference between the reference standard and ST scans.

Discussion

This is the first validation study of this new curve prediction technology, focused on use with AIS patients with curvatures ≤ 45°. We tested agreement of ST scan curve magnitude predictions with reference-standard radiological measurements and inter-rater reliability between scans performed in clinic compared with at home. Additionally, accuracy to detect moderate curvatures (> 25°) and progression (> 5°) was calculated, and overall usability and factors associated with prediction error were assessed. Our findings demonstrate that the ST scans had poor to moderate agreement with the reference standard (ICC 0.402 for scans done at home to 0.535 for scans done in clinic), and poor inter-rater reliability (ICC 0.168). Further, ST scans failed to accurately classify 30% of moderate sized curves (25° to 45°). Unanticipated usability barriers reduced participant completion rates and ultimately limited our sample size. Transparent reporting of these challenges highlights the importance of assessing user feasibility and scan reliability alongside algorithmic performance in validation studies.

The app showed moderate ability to predict the degree of spinal curvature compared with the reference standard, when used by a trained researcher in-clinic. In this setting, ST scans tended to underestimate the curvature, on average 5.2°, but with significant variability in error (range -17° to 25°), and a MAD of 8.5°. Although the sensitivity (66.7%) and specificity (70.6%) for detecting > 5° progression suggest moderate discriminative ability, the PPV was low (16.7%), reflecting the small number of true progressors in the six-month study period (n = three). As a result, most ST scan progression alerts were false positives (n = ten). Had the ST scans been used to triage patients based on predicted curve progression, ten false positives would have been incorrectly prioritized for earlier follow-up, increasing cost burden.²¹ Although the NPV was high (96.0%), indicating that a negative ST scan result was reassuring for true radiological stability, and other AI-ST methods analyzing curve progression report similar accuracy,²² such misclassifications would increase clinic volume and reduce efficiency by diverting resources to patients not requiring urgent follow-up. Further, errors in predicted progression may also create patient anxiety while awaiting radiological confirmation. While early identification of curve progression is valuable in a population where timely intervention can significantly alter the course of treatment and prevent surgery, these findings demonstrate the need for further optimization and safe-guards before home ST scans can reliably be integrated in clinical decision-making pathways.

The app had lower usability at-home as shown by the higher failed rate compared with in-clinic (30% vs 16%). User-engagement with the app was also limited, with participants completing fewer than two monthly scans on average. Trends in home scans suggest participant fatigue with the highest rate of scan attempts at months zero to two followed by a decline over the remainder of the study period. This low adherence may reflect the declining novelty of the app, limited perceived benefit, or potential difficulties using the app. Further, poor inter-rater reliability with clinic scans may stem from variability in patient positioning, lighting, and camera angles as other AI-ST tools have required standardized equipment with fixed camera and patient positioning to achieve greater reliability.^18,23,24 Participants’ feedback cited challenges in aligning and steadying the camera and finding a large enough space to perform the scan. Additional patient education and exploration of factors limiting at-home engagement and scan quality is warranted as the low accuracy and poor inter-rater reliability currently bring into question the suitability for use by patients.

AI and machine learning have increasingly been integrated into AIS-related ST to enhance its diagnostic utility. One study by Minotti et al¹⁷ applied machine learning to rastersterography in AIS patients and found a MAD of 6.1° between radiological measurement and their predictions. The programme accurately classified 59% of curves according to severity type (non-scoliotic, mild, and moderate). Rothstock et al²⁴ used machine learning to classify curves according to mild, moderate, and severe magnitudes through 3D depth sensor ST analysis and found and achieved an accuracy of 90%. Finally, Kokabu et al²³ developed a deep learning algorithm to predict curve magnitude from a 3D depth sensor imaging system with a MAD of 4.4° to 4.7°.

Compared with these systems, the app used in our study demonstrated higher error margins and lower classification accuracy. However, this format offers a distinct advantage of accessibility. Unlike many AI-based health systems which require specialized equipment and personnel, this app leverages standard smartphone features, making it easily deployable to a variety of clinical settings. Thus, the app may be a compelling option for initial scoliosis screening, particularly in resource-limited environments or primary care settings, where access to radiography and diagnostic tools may be constrained. Addressing the unanticipated usability challenges will be essential for successful large-scale implementation.

This study has several limitations. First, the number of successful home ST scans (51%) and paired in-clinic ST scans (59%) was lower than anticipated. The study team underestimated the challenges encountered when completing scans, and the extent of incomplete or failed scans, despite standardized instruction and reminders. While these low completion rates limit the precision of agreement, reliability, and accuracy estimates, they also represent an important and unanticipated finding by highlighting important barriers to feasibility, including user fatigue, environmental constraints (space, lighting, camera stability), and variable engagement with monthly scanning. These findings suggest that usability and adherence, not only algorithmic performance, may be a major constraint on real-world deployment of mobile ST-scanning and require targeted refinement before considering broader clinical integration. Additionally, the short follow-up interval yielded few true progressors, further restricting the ability to evaluate detection of curve progression. Software updates during the study required reprocessing of earlier scans, which may not reflect real-world usage. Finally, the sample size was one of convenience, determined by feasibility rather than formal power calculations; thus, the results should be interpreted as early validation requiring further investigation in larger, longitudinal cohorts.

In conclusion, conceptually, the current application demonstrates potential as an accessible, radiation-free screening tool when used by trained personnel. However, clinicians should exercise caution if looking to replace radiological and/or routinely scheduled examinations. Continued refinement of the algorithm and user-interface may enhance its clinical value. Overall usability, including patient-related barriers to uptake, will remain a key consideration.

Take home message

- This AI-powered mobile phone application that uses remodelled surface topography to predict spinal curvature demonstrates potential as an accessible tool for adolescent idiopathic scoliosis screening.

- While the application shows moderate reliability and accuracy when used by trained personnel, patients had difficulties with the app as shown by poor reliability and higher rates of failed scans.

Data Availability

The data that support the findings for this study are available to other researchers from the corresponding author upon reasonable request.