Scoliosis is a complex spinal deformity that occurs in all three planes, not just the coronal plane. Adolescent idiopathic scoliosis (AIS) can progress over time, especially during adolescence, and can cause musculoskeletal, lung, and psychological problems [1].

Measurement of the Cobb angle on standing posteroanterior full-length spine radiography remains the gold standard for diagnosing and monitoring changes in AIS [2,3] and is necessary for assessing the severity of scoliosis and quantifying the risk of progression [4].

In many cases, patients visit the hospital only after their scoliosis has progressed considerably to the extent that the curvature can be confirmed with the naked eye. However, no separate mandatory or regular early screening test for scoliosis currently exists for children/adolescents. After all, early diagnosis requires the adolescents to visit a hospital regularly and undergo diagnostic imaging, particularly radiography or computed tomography. However, radiological examinations expose individuals to radiation doses that can have detrimental effects, highlighting the need for alternative, reliable, noninvasive examinations [5,6].

Other studies have attempted to identify alternative, reliable, noninvasive examination methods. Navarro et al. [7] analyzed the Cobb angle, kyphosis, and lordosis using rasterstereography, whereas Turner-Smith et al. [8] attempted to determine the Cobb angle by scanning the back shape using an ISIS scanner (“Vicon”; Oxford Metrics Ltd., Oxford, UK). However, these methods require a measuring instrument and may produce variable classification outcomes depending on the observer’s skill level. Another study attempted to predict scoliosis by determining the patient’s appearance over the phone [9], but this approach also had its limitations given that it is not a self-examination method but instead requires an expert to call in and evaluate the other person’s condition. Therefore, a more accessible and useful classification method is certainly needed.

The current study aimed to (1) determine whether the use of visual body images instead of radiographs can successfully classify patients with scoliosis and controls based on the presence or absence of scoliosis with a Cobb angle of ≥10°, (2) classify scoliosis patients with a Cobbs angle of ≥20°, and (3) evaluate the intra- and interobserver reliabilities of the classification method.

From January 1, 2019 to December 31, 2022, 136 patients newly diagnosed with AIS and 124 healthy candidates from the Pusan National University Hospital, Busan, Korea were enrolled. This research was approved by the Clinical Research Ethics Committee of our institution (IRB approval no., 1909-001-082) and was retrospectively registered in a clinical trials registry (registration: NCT04867148). All parents of the subject provided written informed consent. AIS patients were recruited from individuals visiting the outpatient clinic.

Subjects who received any prior treatment for scoliosis were excluded. Those with a history of congenital deformities, neuromuscular disease, endocrine disease, skeletal dysplasia, connective tissue abnormalities, and mental retardation or diseases, as well as those using medications known to affect bone metabolism, unable to maintain an upright position without help, or who have a leg length discrepancy of >10 mm, were excluded.

Before the examination and measurements, all subjects and their parents provided informed consent. The demographic and disease characteristics of the participants were collected. At the initial visit, normal standing whole spine posteroanterior radiographs were obtained for each patient with AIS using a standard technique for measuring the Cobb angle.

Radiographs and photographs were obtained from the patients with AIS and healthy candidates. Images were obtained while positioning the patient against a background drawn horizontally on the wall at 10-cm intervals to facilitate observation of shoulder height and waist irregularities while looking directly behind. The photographer guided the subjects on how to avoid pelvic tilt and instructed them to look directly behind without clockwise or counterclockwise rotation based on the back of the head, shoulders, and pelvis while maintaining a comfortable posture.

This study classified body images into five types based on the visual confirmation of the positional relationship of the body. The standards for classification include (1) difference in height between the shoulders on both sides, (2) positional relationship of the scapulae, (3) concavity and convexity of the back, and (4) positional relationship between the flank lines on both sides when viewed from the rear.

Based on the position of the vertebral body, the thoracic vertebrae spanned from the upper end plate of T2 to the lower end plate of T11, whereas the thoracolumbar/lumbar vertebrae spanned from the lower end plate of T11 to the lower end plate of L5.

According to the items listed earlier, images can be classified into five types. Type 1 is characterized by a unilateral curvature of the thoracic spine, indicating protrusion of the scapula one area (either the right or left side), with shoulder elevation on ipsilateral side. Type 2 is characterized by double curvatures of the thoracic spine, a difference in shoulder height on one side, and a protruding scapula on the opposite side. Type 3 is characterized by curvatures of the thoracic and lumbar spines, regardless of the size of the curvature, indicating a unilateral scapular protrusion and ipsilateral concave waist. Type 4 is characterized by two curvatures in the thoracic spine and one curvature in the lumbar spine, resulting in a raised shoulder, contralateral scapular protrusion, and a concave waist on the ipsilateral side of the scapular protrusion. Type 5 is characterized by only one lumbar spine curvature, resulting in a concave waist on one side. Examples and schematic images of the aforementioned types are shown in Figs. 1–5.

Visual observation was used to determine the presence of scoliosis >10°. To verify the classification method, we calculated its sensitivity, specificity, and accuracy. Accuracy was calculated by assuming that the cut value is positive with a Cobb angle of ≥20°. The classification of the patients’ visual images was determined by three researchers using the proposed method. Disagreements between the three researchers were resolved through discussions, after which the type was decided on unanimously. The classification was conducted by researcher 1 (the author of this paper), researcher 2 (an orthopedic spine surgeon who participated in the classification design and understood the paper), and researcher 3 (a professor specializing in spine surgery with over 10 years of experience in scoliosis surgery).

We calculated the intra- and interobserver reliabilities of body image classification using radiographs as the validation criterion. All naked eye photographs were classified twice, each on separate days, to analyze intra- and interobserver reliabilities. Intra- and interobserver reliabilities were measured by calculating the kappa coefficient within the 95% confidence interval. According to Svanholm et al. [10], a kappa coefficient ≥0.75, 0.5–0.75, and <0.5 indicates good, moderate, and poor validity, respectively. To determine the sample size, we used G*Power ver. 3.1.9.2 (Heinrich-Heine-Universität Düsseldorf, Düsseldorf, Germany; http://www.gpower.hhu.de/) with goodness of fit tests, a power of 95%, a probabilistic error of 5%, an effect size of 3, and 4 degrees of freedom, resulting in a sample size of 207 subjects, including the control group.

A total of 136 patients with scoliosis and 124 controls were investigated. No significant differences in gender, age, body mass index, Risser stage, and characteristics were observed between the investigated patients (Table 1).

The presence or absence of scoliosis was determined by three researchers based on the visual body images, defining patients with scoliosis as those with a Cobb angle of ≥10°. Our results showed a sensitivity of 98.1%, specificity of 98.9%, and accuracy of 98.4%.

As shown in Table 2, the three researchers categorized patients with scoliosis into types 1–5 based on visual body images, whereas normal subjects were categorized as type 0. The prevalence of types 1–5 was 13.5%, 11.7%, 41.4%, 23.0%, and 10.4%, with a sensitivity of 70.0%, 46.1%, 82.6%, 80.4%, and 73.9%; specificity of 96.7%, 99.3%, 93.2%, 97.2%, and 94.2%; and accuracy of 95.2%, 96.7%, 91.3%, 95.6%, and 93.3%, respectively.

Researchers 1, 2, and 3 tested the intra-observer reliability of the two tests at 24-hour intervals using kappa statistics, with their results showing a kappa statistic of 0.855, corresponding to good reliability. Interobserver reliability was measured three times using two of the three researchers in pairs, after which the mean value was calculated. The average of the paired values showed a kappa value of 0.751.

The current study aimed to determine whether our body image classification method could visually classify body images of AIS, evaluate its accuracy through sensitivity and specificity, and assess the reliability of the classification method.

First of all, our findings apparently show that our classification method is a good evaluation tool for screening patients with a Cobb angle of >10°, with a sensitivity, specificity, and accuracy of determining the presence of scoliosis exceeding 98%.

However, selection bias could not be avoided when recruiting participants who visited the tertiary hospital as outpatients given that they had been diagnosed with scoliosis in other hospitals or were suspected of having scoliosis after radiographic examination and visited for therapeutic purposes. Selection bias may have caused a discrepancy between the distribution of curve types in the global population of AIS patients and that in our study group. These discrepancies may have decreased the classification accuracy of those with double thoracic curvatures and lumbar curvatures, who were relatively few in number. This limitation might hinder the applicability of our classification method in the general population, which needs to be addressed before our system can be considered accurate. Future studies may include randomized trials to determine its suitability as a screening tool.

The sensitivities for types 1, 2, and 5 were 70.0%, 46.1%, and 73.9%, respectively, which were relatively low compared to types 3 and 4, which were above 80%. Types 1, 2, and 5 accounted for around 10% of all classified AIS subjects, whereas types 3 and 4 accounted for 41.4% and 23.0% of all subjects, respectively. These finding could indicate that the number of subjects evaluated was too small for accurate classification, which could be attributed to poor sensitivity.

Patients with scoliosis who are likely to progress to a curvature of ≥20 degrees may require treatment, such as bracing or surgery [11]. To determine the accuracy of the classification according to type, we defined patients with a Cobb angle of ≥20° as the disease group. This approach can help to identify patients who require active treatment, such as bracing or surgery.

The exceptionally low sensitivity in type 2 may be due to the variations in the elevation of one side of the shoulder depending on the angle at which the picture was taken and the patient’s posture.

In addition, investigation of interobserver reliability revealed high reproducibility at 0.855. Such a high degree of reproducibility indicates that our classification method can easily be utilized by nonprofessionals. Furthermore, our classification method introduced is highly accurate and can be used for screening scoliosis.

The difficulty in appearance classification, which the researchers perceived, lies in the difficulty of recognizing the positional relationship of the indicators, particularly among obese patients, and the likelihood that the image taken during appearance classification may not be a true posteroanterior view. In the process of organizing photographs and collecting data, one needs to consider the possibility that the tilt of the pelvis may not be uniform while standing and that accurate evaluation may be limited among patients who have difficulty with standing radiography. Other studies, such as those by Buyukaslan et al. [12] and Liu et al. [13], have shown that children with scoliosis may have leg length discrepancy; hence, pelvic tilt is affected in children who cannot stand up without assistance or have a leg length discrepancy of >1.5 cm. It is said that it can form the position of the spine or the false scoliotic line. Therefore, establishing the posture is necessary when obtaining images during research or screening. Another difficult aspect of appearance classification is that the parameters used for visual body image classification are “visually observable” differences. To reduce the ambiguity associated with this subjective expression, we plan to conduct further research on the threshold that can be distinguished by the naked eye to guide other researchers.

A number of studies have previously attempted to analyze the characteristics of scoliosis patients through visual body images. Turner-Smith et al. [8] attempted to obtain the Cobb angle by scanning the back shape using an ISIS scanner; however, several disadvantages had been identified, namely the inconvenience of having the physician attach the sticker, the difficult scanning method, and the time-consuming reconstruction using the program. Navarro et al. [7] conducted a study to analyze the Cobb angle, kyphosis, lordosis, and so forth using rasterstereography; however, some disadvantages were apparent, particularly the long scan times and need for special equipment. Given that the current study used outline images, separate researcher intervention was required. However, considering that no tools or devices were needed, the scanning time was quite short, which is an advantage of our approach.

Regarding further research based on the mentioned advantages and disadvantages, our classification method itself can be helpful in categorizing patients by reflecting clinical features and in setting the direction of examination for hospitals and schools. However, further efforts and research are needed to predict the size of the curvature by analyzing visual body images.

Additionally, strategies for improving categorization through photographs need to be identified. Correcting the errors caused by pelvic tilt, leg length differences, and inaccurate photography can certainly improve the accuracy of our classification method.

The classification method presented herein demonstrated a high level of sensitivity (98.1%), specificity (98.9%), and accuracy (98.4%), as well as high reproducibility and ease of accessibility. Consequently, we believe that our method can be utilized for the purpose of scoliosis screening.