Squint ASSIST: A Novel Application Program for Detecting Strabismus and Measuring Angle of Deviation

Abstract

Purpose

To develop and validate Squint ASSIST (Squint Automatic Strabismus Sizing using Image Sensing Technology), an application program for strabismus detection and angle measurement, using corneal light reflex deviation in eye photographs.

Methods

This prospective cross‑sectional clinic‑based study enrolled 136 patients and 1384 photographs. Two analytic methods were evaluated: Method 1 (inter-ocular comparison; comparing corneal light reflex positions between both eyes in one image) and Method 2 (intra‑ocular comparison; comparing the same eye in a straight vs deviated image). Photographs were processed with standardized processing images. The participants’ angles of deviation were measured using this program, which converts the light reflex deviation to misaligned angles in prism diopters. The main outcome was the sensitivity, specificity, and accuracy of the program for detecting strabismus. The agreement between the angles measured by the program and those measured by pediatric ophthalmologists using the Krimsky test and the Alternate Prism Cover Test (APCT) was analyzed using the Bland-Altman plot. Intra-observer and inter-observer reproducibility were also evaluated.

Results

Of those 136 patients with strabismus, eighty‑seven patients had esotropia, 43 had exotropia, and six had orthophoria after strabismic surgery. The mean age was 12.8 (1–69) years. Squint ASSIST detected strabismus with high performance. Method 1: sensitivity 96.39%, specificity 94.34%, accuracy 95.59%; Method 2: sensitivity 96.83%, specificity 90.48%, accuracy 94.29%. Pearson r versus clinician measures were 0.962 (Method 1) and 0.955 (Method 2) (P < 0.001). When compared the program with ophthalmologists, the Bland–Altman analysis showed that the mean biases were small (Method 1: −1.23 PD; Method 2: −2.00 PD), but the 95% limits of agreement were wide (−29 to +27 PD overall). Repeatability and inter-observer reproducibility was strong.

Conclusions

The Squint ASSIST offers a reliable method for detecting and measuring strabismic angles in clinical settings. The small mean bias and strong intra-observer and inter-observer consistency support substantial clinical utility for screening, standardized photographic documentation, and telemedicine follow‑up. However, due to its wide limits of agreement for absolute angle estimates, pediatric ophthalmologists remain the reference for situations that demand maximal precision such as pre‑surgical planning.

Introduction

The prevalence of strabismus is approximately 2% globally, varying across regions, age groups, and ethnicities.¹ Timely and appropriate treatment can prevent complications such as amblyopia, which can lead to long-term visual loss.² Surgical success rates vary depending on factors such as age at the time of surgery, visual acuity, and the type of strabismus.³ However, accurately measuring the deviation angle remains a challenge and is critical for effective treatment and outcomes.

Inter‑examiner variability in angle measurement using common clinical methods such as the Krimsky test and the Alternate Prism Cover Test (APCT) is well documented and can influence management decisions.⁴ Reported inter‑observer differences vary by technique, patient cooperation, and examiner experience. Clinically acceptable limits of agreement depend on context, small different 5–10 prism diopters (PD) can be clinically meaningful for follow‑up, whereas surgical planning typically requires more precise measurements. Digital, photograph‑based approaches have the potential to reduce subjective variability through standardized capture and automated landmark detection, but they must be validated against clinical reference standards across a range of deviation magnitudes to determine where agreement is acceptable for clinical use. While several studies have explored the use of modern technologies to screen strabismus patients,^2,5–8 enabling timely awareness and treatment, many focus on laboratory settings rather than clinical practice.^6–8 There is a need to validate photographic methods for measuring strabismus angles in real clinical environments.

Strabismus care is inherently longitudinal, with deviations fluctuating over time and varying by strabismus type (eg, esotropia vs exotropia, small- vs large-angle deviations, and postoperative alignment changes) and the angle estimates vary across clinicians. Systematic photographic documentation, analogous to other imaging techniques in ocular diseases such as optical coherence tomography (OCT) for macular diseases or retinal nerve fiber layers changes in glaucoma, can provide objective records of alignment in the preoperative and postoperative periods. A validated, software-assisted approach to quantify deviation from photographs offers standardized, reproducible measurements, mitigates inter- and intra-observer variability, and enable direct, visit-to-visit comparisons for remote follow-up. Such a framework could support telemedicine care, and broaden access, while preserving in-person pediatric ophthalmologist assessments for surgical decision-making.

Although previous photographic and automated approaches have demonstrated potential, important gaps remain in literature. Many earlier studies were conducted in controlled or laboratory settings or used small, narrowly selected cohorts, limiting their generalizability to routine clinical practice. Moreover, few studies have formally quantified intra- and inter-observer reliability across different photographers and sessions.

To address these limitations, this study aimed to validate the accuracy of Squint ASSIST (Squint Automatic Strabismus Sizing using Image Sensing Technology), an application program we developed to detect and measure the strabismus angle from patient photographs by quantifying corneal light-reflex displacement. Specifically, we compared the angles generated by the Squint ASSIST with those measured by a pediatric ophthalmologist. We will evaluate the agreement analysis of Squint ASSIST measurements from patient photographs and assess the reliability across users and sessions, including reproducibility of measurements by different photographers and repeatability of measurements by the same photographer over time.

Methods

This prospective cross-sectional study adhered to the Declaration of Helsinki and ICH Good Clinical Practice Guidelines, and was approved by the Center for Ethics in Human Research, Khon Kaen University (HE654005). All participants were informed about the purpose and procedures of the study prior to enrollment. Written informed consent was obtained from all adult participants and from parents or legal guardians of all minor participants. For children younger than 7 years, verbal assent was obtained from the child and written informed consent was obtained from the parent or legal guardian. The study included 136 participants with strabismus recruited from the KKU Eye Center, Faculty of Medicine, Khon Kaen University between 2021 and 2023. Inclusion criteria were patients with concomitant horizontal strabismus, including those who underwent surgical procedures for strabismus correction, aged one year or older, and with a refractive error within ±2.5 diopters. Exclusion criteria included patients with intermittent deviation, large angles of hypertropia or hypotropia that obstructed eye visibility under the eye lid, wearing glasses or visual aids, and nystagmus.

Data Collection and Procedures

Krimsky test or APCT was used according to each patient’s age and level of cooperation. Facial photographs with the flash on were taken by two photographers from a distance of 30 cm using a smartphone camera. The angle of deviation was evaluated using two methods: Method 1: comparing the corneal light reflex deviation between both eyes (Figure 1) and Method 2: comparing the corneal light reflex deviation within the same eye when it was straight and deviated (Figure 2). The second method was designed to reduce the confounding effect of the large-angle kappa.

Figure 1.

Method 1 (inter‑ocular comparison); corneal light‑reflex positions are compared between both eyes. Red circles indicate the limbus, red dots indicate the pupil centers, and yellow crosses indicate the corneal light reflexes.

Figure 2.

Method 2 (intra‑ocular comparison; comparing the same eye in a straight vs deviated image). Red circles indicate the limbus, red dots indicate the pupil centers, and yellow crosses indicate the corneal light reflexes.

Each photographer took two photographs per method for each patient to analyze intra-observer repeatability and inter-observer reproducibility in the analysis process (Figure 3). The camera used for capturing the photographs had a resolution greater than 960 × 640 pixels, with the flash consistently on in a well-lit room. If a smartphone has multiple camera lenses, focusing on any nearby lens of the same camera would yield consistent angles of deviation. After the photographs were collected, they were analyzed as a blind dataset, with no information on the deviated angles provided.

Figure 3.

Diagram showing workflow for photography of an individual patient.

Image Processing Method

The primary goal was to adjust the images to minimize the effects of varying lighting and contrast, which could arise from taking photographs under different conditions or at different times. Variations in skin and eye color among patients could also lead to inaccuracies in landmark detection.⁹ The preprocessing pipeline utilizes the OpenCV library to ensure uniformity across samples. The image preparation process involved three key steps. The first step reduced unwanted details in the photographs using a low-pass filter known as a Gaussian filter.¹⁰ The next step was converting the image to grayscale. This conversion was achieved by changing the image format from RGB (Red-Green-Blue) to HSV (Hue-Saturation-Value) and extracting only the brightness information from the V-Value channel. This black and white conversion ensured that skin color and lighting were not influencing the detection of important facial landmarks. The final step adjusted the brightness and contrast of all images to a uniform standard. This was accomplished through an algorithm called Histogram Equalization,¹¹ which minimized variations in image tones and brightness by enhancing the contrast. This comprehensive approach to image preparation ensured that all photographs were standardized, allowing for accurate identification of facial landmarks and precise calculation of deviated angles.

Assessment of the Angle of Deviation

This was achieved through a Python program that identified and measured specific eye features within the image.¹² We implemented the MediaPipe library to locate facial features such as cornea. These features, crucial for calculating eye dimensions, were converted from pixels to millimeters to determine the eye angle in prism diopters. The program automatically detected key points around eye boundaries (upper-lower and left-right) using a self-built algorithm. This algorithm integrated a face-alignment technique specifically designed for the accurate identification of essential facial landmarks in images.¹³

Once the program pinpointed the position of the eyes, the next step was to identify critical landmarks for measuring the strabismus angle. This involved locating the pupil center and the corneal light reflex. The corneal light reflex is located from the pixel with the highest value (the brightest spot). In individuals without strabismus, the pupil center and corneal light reflex are typically aligned or close and are influenced by the individual angle Kappa. In contrast, these positions were noticeably different in individuals with strabismus (Figure 4). The angle of strabismus was calculated by measuring the distance between the pupil center and corneal light reflex. This distance was converted from millimeters to prism diopters (PD) using the Hirschberg ratio (HR). Based on prior literature, HR values ranging from 14 to 22 were systematically evaluated.^14–21 The HR that produced the lowest measurement error (from statistical calculation) in our test dataset was selected and subsequently applied to the clinical dataset to optimize accuracy. Detailed algorithmic methodology, including image preprocessing steps, landmark detection, corneal light reflex identification, and clinical unit conversion logic, is provided in the Supplementary Materials.

Figure 4.

Illustrates how the strabismus angle is measured by comparing the corneal light reflex (red line) to the pupil center (blue line) in images showing strabismus. The angle is calculated by subtracting the non-strabismic eye’s angle from the strabismic eye’s angle, using the distances between the corneal light reflex and the pupil center in both eyes.

Statistical Analysis

Using MedCalc version 20.015 (MedCalc Software, Ostend, Belgium), sample size was calculated with the following parameters; significance level (α): 0.05, statistical power (1-β): 0.90, expected mean of differences: 0, expected standard deviation of differences: 2, maximum allowed difference (limits of agreement): 5. This resulted in a final sample of 136 participants. R version 4.3.1, package blandr (R Core Team, Vienna, Austria) was used for statistical analysis. The sensitivity, specificity, and accuracy of the program were analyzed to determine the diagnostic performance of Squint ASSIST in the detection of strabismus. Pearson’s correlation coefficient was used to study the correlation between the deviated angles measured by the ophthalmologist and the program. Bland-Altman plot analysis was used to assess the agreement of the angle of strabismus using two different measurement methods made by the pediatric ophthalmologist and the program, and to assess the intra-observer repeatability and inter-observer reproducibility from strabismic photographs obtained from two different photographers (Photographer 1 with two images and Photographer 2 with two images). Angle-based subgroup analysis was performed by classifying absolute deviation values into small (≤24.9 PD), medium (25–40 PD), and large (>40 PD) groups, followed by assessment of categorical agreement between Squint ASSIST and pediatric ophthalmologists using weighted Cohen’s kappa. P-value < 0.05 was considered to be statistically significant.

Results

Test Set Validation

Two distinct test sets were used to validate the Squint ASSIST program. The first set comprised 76 photographs from 72 patients, and the second included 133 photographs from 30 patients. These sets were analyzed using Method 1 and Method 2 to estimate the Hirschberg ratio (HR) value used to convert the pixel distance of the light reflex deviation into the angle of deviation in the clinical data set. The analysis revealed that an HR of 16 for Method 1 and an HR of 14 for Method 2 yielded the lowest error values. Therefore, these HR values were applied to the clinical data set to increased accuracy.

Clinical Data Set

Our clinical data set consisted of 1384 photographs from 136 patients with strabismus: 87 with esotropia, 43 with exotropia, and 6 with orthophoria after strabismic surgery. The mean age of the patients was 12.8 years, ranging from 1 to 69 years.

Sensitivity, Specificity, and Accuracy of Squint ASSIST

When compared to a pediatric ophthalmologist, the Squint ASSIST had a high diagnostic performance in the detection of strabismus. Method 1 (inter-ocular comparison; comparing corneal light reflex positions between both eyes in one image) of the program correctly detected photographs of strabismus with sensitivity, specificity, and accuracy of 96.39%, 94.34%, and 95.59%, respectively. Method 2 (intra‑ocular comparison; comparing the same eye in a straight vs deviated image) also had sensitivity, specificity, and accuracy of 96.83%, 90.48%, and 94.29%, respectively. Moreover, the program had a high diagnostic performance for correctly detecting types of strabismus. Details of the sensitivity and specificity of the program for detecting esotropia and exotropia are shown in Table 1.

Table 1.

Diagnostic Performance of Squint ASSIST to Detect Types of Strabismus

Methods	Esotropia (87)	Exotropia (43)
Method 1
Sensitivity	96.4%	97.6%
Specificity	94.3%	96.9%
Method 2
Sensitivity	96.8%	93.8%
Specificity	90.5%	95.9%

Correlation Between Ophthalmologist and Squint ASSIST

There was a strong correlation between the angles calculated by the program and those measured by a pediatric ophthalmologist (Figure 5). Method 1 with a Hirschberg ratio of 16 has a high Pearson correlation coefficient (r = 0.962, P <0.001), whereas Method 2 with a Hirschberg ratio of 14 also has a high correlation (r = 0.955, P <0.001).

Figure 5.

Correlation between angles of deviation measured by the ophthalmologist and Squint ASSIST in Method 1 with Hirschberg ratio = 16 (A) and Method 2 with Hirschberg ratio = 14 (B).

Agreement Between Ophthalmologist and Squint ASSIST

In Method 1, the Squint ASSIST program calculated the angle of deviation by comparing the light reflexes of both eyes. A total of 544 photographs of 136 patients were processed and analyzed. The Bland-Altman plot revealed a bias of −1.23 PD, representing the mean difference between the angles of deviation measured by an ophthalmologist and the program. The 95% limits of agreement were calculated as 26.76 PD in the upper range and −29.23 PD in the lower range (Figure 6A).

Figure 6.

Bland–Altman analysis of agreement between angles of deviation measured by the ophthalmologist and Squint ASSIST in Method 1 with Hirschberg ratio = 16 (A) and Method 2 with Hirschberg ratio =14 (B).

In Method 2, Squint ASSIST calculated the deviation angle by comparing the position of the eye when it was straight and deviated. A total of 840 photographs of 105 patients were included in the analysis. The Bland-Altman plot indicated a mean difference of −2.00 PD, representing the deviation angles measured by an ophthalmologist compared to the program. The 95% limits of agreement were calculated as 26.14 PD in the upper range and −30.14 PD in the lower range (Figure 6B).

Angle-Base Subgroup Agreement

Overall categorical agreement between Squint ASSIST and pediatric ophthalmologists was 69.5% for Method 1 and 68.8% for Method 2, with substantial weighted agreement (Method 1: κ = 0.69; Method 2: κ = 0.70). Subgroup-specific agreement demonstrated the highest concordance in small-angle deviations, with correct classification rates of 95.5% for Method 1 and 85.0% for Method 2. Agreement remained high in the large-angle subgroup (Method 1: 74.4%; Method 2: 76.3%), while the lowest agreement was observed in the medium-angle subgroup (Method 1: 43.2%; Method 2: 43.3%). Misclassification occurred predominantly between adjacent categories (small vs medium and medium vs large), and gross misclassification between small and large deviations was uncommon.

Intra-Observer Repeatability

The intra-observer repeatability for photographer 1 showed a mean difference of −0.43 prism diopters (PD) (95% limit of agreement −1.44 to 0.59) for Method 1 and 0.70 PD (95% limit of agreement 0.04 to 1.36) for Method 2. Photographer 2 demonstrated a mean difference of 0.38 PD (95% limit of agreement −0.37 to 1.13) in Method 1 and 0.11 PD (95% limit of agreement −0.52 to 0.73) in Method 2. The results are shown in Figure 7.

Figure 7.

Bland–Altman analysis of the intra-observer repeatability of photographer 1 in Method 1 (A) and Method 2 (B) and photographer 2 in Method 1 (C), and Method 2 (D).

Inter-Observer Reproducibility

The mean difference of inter-observer reproducibility between photographer 1 and photographer 2 were 0.01 PD (95% limit of agreement −0.98 to 1.00) in Method 1 and −0.02 PD (95% limit of agreement −0.65 to 0.61) in Method 2 as shown in Figure 8.

Figure 8.

Bland–Altman analysis of the inter-observer reproducibility between the two photographers in Method 1 pic 1 (A), Method 2 pic 1 (B), Method 1 pic 2 (C), and Method 2 pic 2 (D).

Discussion

This study addressed the clinical application of photographic methods using the corneal light reflex to evaluate strabismic angle deviation, especially beyond laboratory settings.^2,9 Previous studies have primarily employed light reflex deviation for strabismus angle assessment in controlled environments, leaving real-world clinical data limited. Some are developed for mainly at screening purpose or self-monitoring as home-based application.⁵ Our development of the Squint ASSIST program aims to integrate traditional light reflex deviation detection with an AI platform capable of identifying key anatomical landmarks, such as the limbus boundary and the center of the cornea. This capability, combined with the detection of deviated light reflexes, allows Squint ASSIST to calculate angle deviations, making it useful in clinical practice.

The Squint ASSIST served as a practical tool for measuring strabismic angles, demonstrating excellent intra-observer repeatability and inter‑observer reproducibility, thereby providing consistent results within and across clinicians and enhancing data validity across users and institutions. Compared with pediatric ophthalmologists’ assessments by two methods, it showed small mean differences of −1.23 PD and −2.00 PD. These near-zero biases, together with low inter-technician and inter-session variability, support its use for screening, standardized clinical documentation, and consistent longitudinal tracking across visits and photographers, including telemedicine-based assessments where relative changes and detection of misalignment are of primary importance.

Although Squint ASSIST demonstrated high sensitivity and specificity for strabismus detection, the unstratified Bland–Altman limits of agreement remained wide (approximately −29 to +27 PD), indicating that absolute angle estimates from single photographs may imprecise for individual patients. Accordingly, Squint ASSIST should not replace the standard strabismus angle examination when maximal precision is required for surgical planning. Nevertheless, its overall accuracy and strong repeatability make it a valuable adjunct for routine monitoring and serial photographic documentation.

Comparatively, Dericioglu et al⁷ reported a mean error of −0.68 ± 6.1 PD in their study on digital photography and light reflex measurements, possibly achieving smaller variance due to the use of precise individual corneal diameters. In contrast, our study used the average corneal diameter to accommodate a broader demographic, including very young children. However, modern computer programs have the potential to overcome these limitations. By including a reference object of known size in the photograph, the program can measure the corneal diameter, adjust for individual variations and provide more accurate and personalized measurements of the strabismic angle. This capability would enhance the precision of strabismus assessments and broaden the clinical utility of Squint ASSIST, supporting individualized treatment planning while retaining accessibility and ease of use.

Miao et al⁶ and Yeh et al⁸ used virtual reality systems, showing low mean deviations of 0.88 ± 5.77 PD with 95% limit of agreement −10.43 to +12.19 PD but with higher costs, compared to our simple software, which limits its widespread clinical applicability.

We evaluated two complementary photographic measurement approaches because each addresses different sources of error inherent to corneal light-reflex techniques. Method 1 compares light reflex displacement between the two eyes (inter-ocular comparison) and is straightforward to perform when both eyes are visible and when the angle kappa is small or symmetric. In our dataset, Method 1 produced a small mean bias compared to pediatric ophthalmologists (−1.23 PD) and a high diagnostic performance (sensitivity 96.39%, specificity 94.34%). These findings support Method 1 as an efficient screening and documentation tool for routine clinical imaging where cooperative positioning and symmetric ocular anatomy are present.

Method 2 compared the corneal light reflex position of the same eye in a straight gaze image vs a deviated image (intra‑ocular comparison). This design explicitly reduces confounding from individual angle kappa by using each eye as its own reference, at the cost of requiring two well‑matched images (straight and deviated) of the same eye. In our study, Method 2 showed similarly high sensitivity (96.83%) but slightly lower specificity (90.48%), a mean bias of −2.00 PD against clinician measures. Therefore, Method 2 may be preferable when an asymmetric angle kappa is suspected.

The strength of this study was its pragmatic approach. We paired a simple corneal light reflex deviation measurement with a Hirschberg ratio empirically calibrated from our own, predominantly Asian dataset and optimized it for both measurement methods. This population-matched, lowest error HR likely improves accuracy over legacy values derived largely from Caucasian cohorts, helping to capture anatomical variations that have been under characterized in Asian eyes. Although individualized, patient specific HRs may further refine precision, their added requirements for calibration, time, and training make them impractical for routine clinical workflows.

Despite its strengths, the Squint ASSIST primarily accommodates manifest strabismus, with diminished accuracy in cases with significant phoria components or large-angle deviations where detecting ocular landmarks is challenging. Squint ASSIST faces detection errors in large-angle strabismus, especially when the light reflex falls near or on the sclera, because off-axis gaze renders the corneal contour elliptical rather than circular and shifts the corneal light reflex toward the limbus or sclera (Figure 9). However, angle-based subgroup analysis demonstrated substantial agreement between Squint ASSIST and pediatric ophthalmologists (overall agreement 68.8–69.5%, weighted κ ≈ 0.69–0.70). The highest concordance occurred in small-angle deviations, with high agreement also observed in large-angle deviations and lower agreement in medium-angle deviations (likely reflects a boundary effect inherent to categorical classification). Misclassification was largely confined to adjacent categories, and gross misclassification between small and large angles was uncommon. These findings indicate that, while optical and geometric constraints may introduce increasing measurement variability at larger deviations, Squint ASSIST maintains clinically meaningful consistency across deviation severities.

Figure 9.

Example in a large-angle deviation: the algorithm reduced the limbal circle (red circle) to a smaller size and placed the corneal light reflex on the sclera (yellow cross). These errors hinder accurate localization; manual review may be needed to adjust landmarks.

Although the mean differences in our study were small, 95% limits of agreement were quite wide. We propose that they may arise from projecting three-dimensional ocular rotations, clinically as angular deviation measured by ophthalmologist, onto two‑dimensional images that capture linear reflex displacement of the corneal light reflex. Real‑world imaging factors such as camera-to-eye distance, lens characteristics, and image scaling or resolution can further broaden variability. The algorithm also assumes a spherical cornea. With off‑axis gaze the apparent corneal outline becomes more elliptical, as mention above, which makes automated localization of the corneal center more difficult and can reduce precision at larger deviations.

In conclusion, Squint ASSIST is a practical, repeatable tool for strabismus screening, standardized photographic documentation, and telemedicine follow-up. While wide limits of agreement preclude its use as a substitute for clinical measurements in surgical planning, its strong diagnostic accuracy, repeatability, and subgroup agreement establish Squint ASSIST as a valuable adjunct in real-world strabismus care. Continued refinement of image modeling and algorithmic robustness, particularly at larger deviations, will further expand its clinical utility. Future longitudinal studies are needed to evaluate the performance of Squint ASSIST across visits separated by clinically meaningful time intervals, including postoperative follow-up and disease progression.

Funding Statement

This study was supported by Research Fund, Faculty of Medicine, Khon Kaen University (RU66002).

Data Sharing Statement

All relevant data that support the findings of this study are available from the corresponding author on reasonable request.

Disclosure

The authors report no conflicts of interest in this work.