ProptoView: AI-based digital exophthalmometry using multi-view facial images in a multinational validation study

Abstract

Background

Accurate proptosis measurement is vital for managing thyroid eye disease (TED) and other orbital conditions. However, current approaches have certain limitations: the Hertel exophthalmometer is convenient but imprecise, while computed tomography (CT) is accurate but costly and exposes patients to radiation.

Methods

We developed ProptoView, an AI-based digital exophthalmometer, using 5676 images from 2516 eyes across 1258 visits of 763 TED patients with CT and Hertel measurements. For external validation, we used an additional 644 images from 648 eyes of 324 patients with TED and other orbital diseases, collected across three countries and five hospitals. Patients provided up to five images from four views. A three-stage deep learning approach, optimized with Adam and validated via five-fold cross-validation, helped develop three AI models: single-view, multi-view, and dynamic input.

Results

Compared with CT, the single-view model achieved an intraclass correlation coefficient (ICC) of 0.859, slightly lower than the Hertel exophthalmometer’s ICC of 0.888. The multi-view model achieved an ICC of 0.890, surpassing the Hertel exophthalmometer (0.871). The dynamic input model achieved the highest accuracy with an ICC of 0.901. Among two-view combinations, pairing the frontal view with another angle showed the highest agreement when paired with the upward gaze view (ICC = 0.855). In external validation, ProptoView showed robust concordance with the Hertel exophthalmometer (ICC = 0.845), comparable to its agreement in the development dataset. Additionally, ProptoView reduced misclassification at the 19-mm threshold (14.7% vs. 20.5% with the Hertel exophthalmometer).

Conclusion

ProptoView provides an accurate, non-contact, and cost-effective solution for proptosis measurement. Its flexibility and precision suggest significant potential for streamlining clinical workflows and enabling telemedicine applications.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-08401-w.

Background

Proptosis, defined as the abnormal protrusion of the eyeball, reflects a growing burden in global ophthalmic and endocrine disorders, with an impact extending beyond vision, affecting psychological well-being and social functioning [1, 2]. It is commonly observed in a variety of orbital and systemic diseases, such as thyroid eye disease (TED), orbital inflammation, orbital tumors, among others [3–5]. Although computed tomography (CT) is the gold standard to quantify the degree of proptosis, its use is constrained by cost, radiation, and limited accessibility [6]. Therefore, Hertel exophthalmometer remains the most commonly used tool for measuring proptosis, despite its manual measurement error and reported deviations up to 3 mm compared to CT [7], which may lead to misclassification of disease severity and treatment selection [8]. Furthermore, as proptosis is encountered not only by orbital specialists but also by general ophthalmologists and endocrinologists, the lack of familiarity with Hertel by non-orbital specialists may lead to diagnostic delays and disease progression. Therefore, there is an urgent need to develop an accurate and reproducible method of exophthalmometry for clinical practice.

The facial and orbital region provide a non-invasive window for observation. Recent advances in 3D scanning technology [9–11], such as smartphone-based exophthalmometers [12], have improved measurement accuracy compared to the Hertel exophthalmometer. However, these methods have limitations such as insufficient validation, dependence on manual processing, and imaging inconsistencies. In the face of these shortcomings, advances in artificial intelligence (AI) may pave the way towards reliable, user-friendly, and cost-effective solutions. As AI demonstrates growing capabilities in facial analysis, image recognition and anatomical landmark detection [13–17], it offers promising prospects for enhancing disease screening, diagnosis, treatment planning, and longitudinal monitoring.

To address these challenges, we developed ProptoView, an AI-powered digital exophthalmometer designed to deliver accurate, reproducible, and user-independent proptosis measurements. This study evaluated its performance against the Hertel exophthalmometer, using CT measurements as the reference standard. ProptoView was further validated across multiple centers and ethnically diverse populations, aiming to pave the way for broader clinical adoption and practice transformation.

Materials and methods

Study design

This multinational study was initiated and led by Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (SH9H, Shanghai, China). The study adhered to the Declaration of Helsinki and was approved by the Ethics Committee of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. SH9H-2022-T380-1), as well as the ethics committees of all participating centers. Written informed consent was obtained from all participants. The collaborating institutions included: Peking Union Medical College Hospital (PUMCH, Beijing, China), the Second Hospital of Dalian Medical University (SAHD, Liaoning, China), the Second Affiliated Hospital of Chengdu Medical College (SAHC, Sichuan, China), Khoo Teck Puat Hospital (KTPH, Singapore), and Mettapracharak (Wat Rai Khing) Hospital (WRKH, Nakhon Pathom, Thailand) (Fig. 1a).

Fig. 1.

Study design. (a) Images were collected using various equipment from three countries and six centers representing seven ethnic groups, forming both the development and external test datasets. (b) Development of ProptoView involved a novel three-stage approach. Three different AI models were trained with varying inputs: the single-view image-based AI utilized only frontal view images; the multi-view image-based AI employed images from all four views; and the dynamic input image-based AI used a frontal view combined with any available additional views. (c) Performance evaluation of ProptoView included intraclass correlation coefficient analysis, identification of the most contributory image views, assessment of classification accuracy, model visualization, and testing under various clinical conditions

Patients with TED from SH9H formed the development dataset and were recruited between April 2018 and May 2023. Between July 2022 and October 2023, retrospective data of TED patients were collected from PUMCH, SAHD, KTPH, and WRKH. KTPH and WRKH also included patients with TED, orbital inflammation, tumor, and fracture. Additionally, prospective data of TED patients from SAHC were collected between July 2023 and December 2023.

Participants

Study inclusion criteria were: (i) diagnosis of TED based on the Bartley criteria [18], including newly diagnosed and post-orbital decompression (with orbital rim repositioning); (ii) diagnosis of other orbital diseases, such as orbital inflammation, tumors, or fractures (not involving the orbital rim); (iii) availability of Hertel exophthalmometer measurements and/or CT scans; and (iv) unobstructed external ocular regions without significant occlusions by scars, birthmarks, tattoos, masks, heavy makeup, or hair. Exclusion criteria included uncooperative behavior and lesions affecting the orbital rim.

Data collection and labeling

For each patient, clinicians obtained multiple photographs across four standardized views under different capture conditions (distance, lighting, background) using a range of equipment (digital single-lens reflex [DSLR] cameras including Canon, Sony and Nikon cameras, and smartphones including Apple, Samsung and Huawei devices). All images were acquired during routine evaluations in clinics or wards. At SH9H, facial images were collected following the photography protocol [19]. Other centers gathered only external ocular views to uphold patient privacy. View 1 (frontal) was mandatory at all sites. Based on local feasibility, four additional images of three views were captured: a closed-eye frontal view (View 2), an upward gaze view (View 3), and left or right lateral 90° views (View 4). Each photo was labeled with the view number.

Measurement and classification of proptosis

Exophthalmometry was performed using the Hertel exophthalmometer (Oculus, Germany) at all sites. The Hertel exophthalmometer utilizes prism mirrors to measure the distance from the lateral orbital rim to the anterior surface of the cornea. CT scan was performed at SH9H using Mimics software (Materialize, Leuven, Belgium). Because CT imaging was not routinely performed for exophthalmometry elsewhere, only SH9H compared ProptoView outcomes with CT-based data. Two experienced ophthalmologists (> 8 years in practice) performed independent Hertel and CT measurements for proptosis; if their measurements differed by < 3 mm, the average was taken; if the difference was ≥ 3 mm, a senior orbital specialist (> 15 years in practice) reassessed. Proptosis severity of TED patients was classified at a 19 mm cutoff; in addition, we performed a dedicated sub‑analysis for measurements near this clinically important threshold by evaluating method performance between 17 and 21 mm (i.e., within 2 mm of the cutoff) [20]. We then examined how each method affected this classification (Fig. 1c).

Development of ProptoView

ProptoView included a novel three-stage pipeline for automated exophthalmometry using periocular features (Figs. 1b and 2). Specifically, the three stages were: (1) human facial landmark detection; (2) landmark-based periocular region extraction; and (3) deep learning-based multi-view proptosis estimation.

Fig. 2.

Architecture of the designed AI-based proptosis estimation pipeline. (a) The pipeline contains three stages, human facial landmark detection, landmark-based periocular region extraction, and deep learning-based multi-view proptosis estimation. (b) Detailed architecture of the designed ProptoView

In the first stage, we applied facial landmark detection to pinpoint key points on the face. We employed the PyTorch Face Landmark Detector (PFLD) [21] due to its robustness across images captured from four different views, including profile views. PFLD integrates a MobileFaceNet backbone [22] with a RetinaFace detector [23] for efficient and accurate landmark detection. Before processing, the facial images were resized to a width of 384 pixels while maintaining the original aspect ratio. PFLD then generated 68 facial landmarks (Fig. 2a), which we subsequently mapped back to the original image dimensions.

In the second stage, we used key facial landmarks (e.g., cheek, mid-face, eyebrow, nasal wing) to define the periocular region for proptosis assessment (Fig. 2a). For the right eye, we set the horizontal boundary from the cheek landmark to the mid-face landmark, and the vertical boundary from the top of the eyebrow to the nasal wing. This was mirrored for the left eye. In profile images, only the visible eye region was extracted. This method preserved critical morphological cues while minimizing background noise.

In the third stage, we developed a deep learning framework for multi-view proptosis estimation (Fig. 2b). The framework used a four-branch architecture, with each branch corresponding to one imaging view. A primary branch (ResNet34 [24]) focused on the frontal image, while three auxiliary branches (ResNet18 [24]) processed additional views. This architecture allowed ProptoView to leverage information from multiple views to enhance its predictive capability [25–27]. To handle missing views, we concatenated available features along the channel axis so that other views could compensate. The combined features were fed into an Information Aggregation Head (convolution, batch normalization, and ReLU layers) to generate the final proptosis measurement.

Implementation details

We employed five-fold cross-validation, ensuring that images from the same patient remained in the same fold. In total, 5676 images (2516 eyes) from SH9H were used for model development. CT measurements at SH9H served as ground truth (gold standard) for training ProptoView. Not all eyes had complete multi-view images or matched CT/Hertel data, resulting in slightly different sample sizes for each model. To further assess the accuracy and generalization of ProptoView, we conducted external validation using five independent test datasets from three Chinese hospitals and two institutions in Singapore and Thailand.

All extracted periocular regions were resized to 224 × 224 pixels before being fed into the neural network model. Experiments were conducted using PyTorch on a single NVIDIA TITAN GPU, using the Adam optimizer [27] with default parameters, a batch size of 16, and a base learning rate of 0.0004. We applied image augmentations during training, including random cropping, random horizontal flipping, random blurring, Gaussian noise, and contrast adjustments. Training comprised 200 epochs, with the validation set evaluated three times per epoch based on the total number of batches, and early stopping mechanisms were used to prevent overfitting. We trained three AI models: (1) single-view image-based AI: used only frontal images (largest n); (2) multi-view image-based AI: used four standardized views (subsample with all 4 views available); (3) dynamic input image-based AI: accepted any available views (a flexible approach if views were missing).

Model robustness assessment

To evaluate ProptoView’s real‑world robustness, we simulated three scenarios: varying capture distance (by resizing images to 0.5× and 2× to represent different distances), angular rotation (5°–15°), and a missing lateral view (replaced with a black image). The dynamic input model was also compared against depth‑aware models (Zoedepth) [28] through a head‑to‑head comparison simulating a single frontal view (without/with depth). Performance was assessed using mean absolute error (MAE), root mean square error (RMSE) and intraclass correlation coefficient (ICC) across these conditions. Additionally, to address hardware‑induced variability, we performed pairwise Cohen’s d comparisons across hospital groups using different devices (smartphones vs. DSLR). Moreover, using the KTPH dataset as an example of Asian ethnic diversity, we compared ICC between Chinese and non‑Chinese patients (Malay, Indian, Thai, others).

Visualization of ProptoView

We used GradCAM + + [29, 30] (Fig. 1c) to visualize the features influencing ProptoView’s predictions. By highlighting both left and right eyes in several representative cases, we demonstrated ProptoView’s reliability under different conditions. We also generated t-distributed stochastic neighbor embedding (t-SNE) plots from the final layer to reduce high-dimensional features into two dimensions, aiming to confirm that ProptoView could effectively distinguish different proptosis levels and learn meaningful representations.

Statistical analysis

Statistical analyses were conducted using R (version 4.4.1). We used MAE, RMSE, and ICC to assess agreement among the three methods—Hertel exophthalmometer, CT, and ProptoView—comparing Hertel vs. ProptoView, CT vs. ProptoView, and CT vs. Hertel (Fig. 1c). Mean ICCs and 95% confidence intervals (CIs) were also reported, and permutation tests with Holm-adjusted P values handled multiple comparisons. A two-sided P < 0.05 was considered to be significant. In addition, Bland-Altman plots visualized measurement agreement, and accuracy was calculated to compare TED severity classification between Hertel and ProptoView.

Results

A total of 1087 patients with 1582 visits from three countries, six hospitals, and seven ethnic groups were enrolled, providing 6320 images across four distinct views. The development dataset at SH9H comprised 763 patients across 1258 visits with 2516 eyes (5676 images). The external test dataset comprised 324 participants with 648 eyes (644 images). The external test dataset comprised patients from China, Singapore and Thailand, with various ethnic groups including Chinese, Thai, Malay, Indian, Bangladeshi, Filipino, and Burmese. The development dataset primarily consisted of TED cases, whereas the test dataset included 27 cases of orbital inflammation, 36 cases of orbital tumor and 10 cases of orbital fracture. In test datasets 1, 2, and 5, only frontal view images were available; while in test datasets 3 and 4, 276 and 173 images of various views were included, respectively. In the development dataset, the average proptosis measurement was 19.00 ± 2.34 mm by Hertel exophthalmometer and 19.90 ± 2.58 mm by CT, with a statistically significant difference between the two methods (P < 0.01). In the test dataset, the mean proptosis measured by Hertel exophthalmometer was 19.65 ± 2.83 mm. Detailed characteristics of the dataset can be found in Table 1.

Table 1.

Clinical characteristics and proptosis measurements in development and test datasets

	Development dataset (SH9H, China)	Test dataset 1 (PUMCH, China)	Test dataset 2 (SAHD, China)	Test dataset 3 (SAHC, China)	Test dataset 4 (KTPH, Singapore)	Test dataset 5 (WRKH, Thailand)
Number of patients	763	52	43	57	72	100
Number of patient visits	1258	52	43	57	72	100
Number of eyes	2516	104	86	114	144	200
Total number of images	5676	52	43	276	173	100
Age, mean (SD), y	45.77 ± 14.01	45.00 ± 14.64	43.09 ± 11.81	43.11 ± 15.01	53.24 ± 14.07	51.37 ± 14.11
Sex
Male	289 (37.88)	19 (36.54)	13 (30.23)	20 (35.09)	35 (48.61)	55 (55.00)
Female	474 (62.12)	33 (63.46)	30 (69.77)	37 (64.91)	37 (51.39)	45 (45.00)
Race
Chinese	763 (100)	52 (100)	43 (100)	57 (100)	52 (72.22)	0
Thai	0	0	0	0	0	99 (99.00)
Malay	0	0	0	0	11 (15.28)	0
Indian	0	0	0	0	4 (5.56)	0
Others (Bangladeshi, Filipino, Burmese)	0	0	0	0	5 (6.94)	1 (1.00)
Disease type
Thyroid eye disease	763 (100)	52 (100)	43 (100)	57 (100)	30 (41.67)	69 (69.00)
Orbital inflammation	0	0	0	0	18 (25.00)	9 (9.00)
Orbital tumor	0	0	0	0	17 (23.61)	19 (19.00)
Orbital fracture	0	0	0	0	7 (9.72)	3 (3.00)
Proptosis measurements
Hertel, mean (SD), mm	19.00 ± 2.34	19.22 ± 2.64	18.73 ± 2.42	19.79 ± 2.24	19.40 ± 2.80	20.51 ± 3.30
CT, mean (SD), mm	19.90 ± 2.58	/	/	/	/	/

* SH9H, Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine; PUMCH, Peking Union Medical College Hospital; SAHD, the Second Hospital of Dalian Medical University; SAHC, the Second Affiliated Hospital of Chengdu Medical College; KTPH, Khoo Teck Puat Hospital; WRKH, Mettapracharak (Wat Rai Khing) Hospital

During ProptoView development, 2087 eyes (each with four views) were used to train the multi-view model. To maximize data utilization, 2516 eyes with at least a frontal view were included for both the single-view and dynamic input models. Due to missing Hertel data, 2216 eyes were ultimately included for ICC analysis in the single-view and dynamic input models, and 1848 eyes for the multi-view model (Table 2).

Table 2.

Intraclass correlation coefficient (ICC) of proptosis measured by ProptoView compared with Hertel exophthalmometer and computed tomography (CT)

	Single-view image-based AI (n = 2216)	Multi-view image-based AI (n = 1848)	Dynamic input image-based AI (n = 2216)
ProptoView vs. CT	0.859 (95% CI 0.847–0.869)	0.890 (95% CI 0.881-0.900)	0.901 (95% CI 0.892–0.908)
Hertel vs. CT	0.888 (95% CI 0.879–0.897)	0.871 (95% CI 0.860–0.882)	0.888 (95% CI 0.879–0.897)
Predicted proptosis values	20.12 ± 2.17 mm	20.15  ± 2.12 mm	20.21 ± 2.20 mm

Single-view image-based AI model

The mean proptosis measured by the single-view image-based AI model was 20.12 ± 2.17 mm (Table 2), close to the CT mean of 19.90 ± 2.58 mm. MAE was 0.958, and RMSE was 1.204. Reaching an ICC of 0.859 (95% CI, 0.847–0.869, n = 2216) indicated a close accuracy and precision agreement between CT and single-view image-based AI model. In comparison, the ICC between CT scans and the Hertel exophthalmometer was higher at 0.888 (95% CI, 0.879–0.897, n = 2216).

Multi-view image-based AI model

The average proptosis measured by the multi-view image-based AI model was 20.15 ± 2.12 mm (Table 2), with MAE and RMSE being 0.842 and 1.068, respectively. The model’s ICC with CT was 0.890 (95% CI, 0.881–0.900, n = 1848), while CT’s ICC with Hertel was 0.871 in this subset (95% CI, 0.860–0.882). According to Table 3, when combining frontal with other views, upward gaze (View 3) contributed most (ICC = 0.855, 95% CI, 0.842–0.867) over single-view (ICC = 0.837), followed by lateral (View 4, ICC = 0.852, 95% CI, 0.839–0.864) and closed-eye frontal (View 2, ICC = 0.849, 95% CI, 0.835–0.861).

Table 3.

ICC of proptosis measured by ProptoView compared with CT among different shooting view combinations (n = 1848)

	ICC	95% CI
View 1	0.837	0.822–0.850
View 1 + 2	0.849	0.835–0.861
View 1 + 3	0.855	0.842–0.867
View 1 + 4	0.852	0.839–0.864
View 1 + 2 + 3 + 4	0.890	0.881-0.900

Annotations: View 1 refers to frontal view, view 2 refers to closed-eye frontal view, view 3 refers to upward gaze view, and view 4 refers to left or right lateral 90° view

Dynamic input image-based AI model

The average proptosis measured by the dynamic input image-based AI model was 20.21 ± 2.20 mm (Table 2), with MAE and RMSE being 0.827 and 1.029, respectively. Table 2 showed the dynamic model—which accommodated any available views—achieved the highest ICC with CT at 0.901 (95% CI, 0.892–0.908, n = 2216), surpassing both Hertel with CT (ICC = 0.888) and the single-view AI with CT (ICC = 0.859). Figure 3 illustrated results for all AI models compared to CT. Figure 3a showed scatter plots, and Fig. 3b depicted prediction differences across 1.5 mm bins (14–28 mm). Notably, the dynamic model provided the most stable predictions, especially at the extremes. Figure 3c further visualized these comparisons via Bland-Altman plots, highlighting measurement differences across the three AI models.

Fig. 3.

Correlation between single-view image-based AI, multi-view image-based AI, dynamic input image-based AI and CT. (a) Scatter plots illustrating the correlation of proptosis across various AI models and CT. (b) The plots demonstrating the difference between average values of the predicted labels and ground-truth with each 1.5 mm length of each bin. (c) Bland-Altman plots comparing the differences in proptosis between three AI models and CT. Blue lines represent the means, red and green lines indicate 95% confidence intervals (CIs)

External test performance

All external test datasets (n = 648 total) were based on Hertel exophthalmometer. In the test datasets, MAE ranged from 1.017 to 1.452, and RMSE ranged from 1.334 to 1.758 (Supplementary Tables 1–1). As detailed in Table 4, the overall ICC between single-view image-based AI and Hertel exophthalmometer was 0.782 (95% CI, 0.749–0.812). For dynamic input image-based AI, the ICC reached 0.845 (95% CI, 0.820–0.867). Notably, in some external test datasets, the AI-Hertel agreement was equal or higher than that in the development dataset, suggesting robust generalizability. For instance, test dataset 4 (KTPH, n = 144) achieved ICC = 0.855(95% CI, 0.796–0.897) with the dynamic input model, surpassing the 0.792 (95% CI, 0.776–0.807) observed internally for AI with Hertel. In addition, we tested the performance of ProptoView on the 17–21 mm proptosis range (near the 19 mm threshold). As reported in Supplementary Tables 1–2, the mode maintained strong robustness. Moreover, we compared ICC between Chinese and non‑Chinese patients in the KTPH dataset. The ICC for Chinese patients was 0.872 (95% CI, 0.810–0.914), while for non‑Chinese patients it was 0.741 (95% CI, 0.502–0.875).

Table 4.

ICC analysis for ProptoView and Hertel exophthalmometer in development and external test datasets

	Single-view image-based AI	Dynamic input image-based AI
Development dataset (n = 2216)	0.761 (95% CI, 0.743–0.778)	0.792 (95% CI, 0.776–0.807)
Overall test dataset (n = 648)	0.782 (95% CI 0.749–0.812)	0.845 (95% CI 0.820–0.867)
Test dataset 1  (PUMCH, n = 104)	0.752 (95% CI 0.655–0.825)	0.844 (95% CI 0.778–0.892)
Test dataset 2  (SAHD, n = 86)	0.717 (95% CI 0.596–0.806)	0.799 (95% CI 0.707–0.864)
Test dataset 3  (SAHC, n = 114)	0.755 (95% CI 0.663–0.824)	0.767 (95% CI 0.679–0.833)
Test dataset 4  (KTPH, n = 144)	0.824 (95% CI 0.755–0.875)	0.855 (95% CI 0.796–0.897)
Test dataset 5  (WRKH, n = 200)	0.807 (95% CI 0.745–0.855)	0.854 (95% CI 0.805–0.891)

* SH9H, Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine, China; PUMCH, Peking Union Medical College Hospital, China; SAHD, the Second Hospital of Dalian Medical University, China; SAHC, the Second Affiliated Hospital of Chengdu Medical College, China; KTPH, Khoo Teck Puat Hospital, Singapore; WRKH, Mettapracharak (Wat Rai Khing) Hospital, Thailand

Robustness assessment

We evaluated the robustness of ProptoView to variations in image capture distance and angle, missing lateral-view images and across different devices. Under simulated capture distance changes (0.5× and 2× resizing), MAE varied by ≤ 0.01 mm (ranging from 0.827 mm to 0.837 mm) and RMSE by ≤ 0.02 mm (ranging from 1.029 mm to 1.042 mm) (Supplementary Table 3). Under angular rotations up to 15°, MAE remained ≤ 0.868 mm and ICC being 0.880 (95% CI, 0.870–0.889) (Supplementary Table 4). When the lateral view was missing (replaced with a black image), MAE increased marginally from 0.827 mm to 0.860 mm, with ICC remaining at 0.887 (Supplementary Table 5). Compared with depth‑aware models (ZoeDepth) [28], while depth cues improved single view performance by reducing MAE from 1.164 mm to 1.101 mm, ProptoView still achieved substantially lower error with an MAE of 0.827 mm (Supplementary Table 2). Besides, Pairwise Cohen’s d comparisons across hospital groups using different devices showed absolute Cohen’s d values generally below 0.3 (Supplementary Fig. 1).

Classification of proptosis

Using a 19-mm threshold, Hertel misclassified 20.5% of eyes (n = 454), predominantly by underestimating proptosis (18.7% false negatives). In contrast, ProptoView misclassified 14.7% of cases, including 4.0% false positives and 10.7% false negatives, thereby demonstrating superior accuracy. Hertel’s accuracy was 79.5%, whereas ProptoView achieved 85.3%. These improvements were clinically meaningful, as misclassification near the 19-mm threshold may lead to suboptimal management strategies.

Visualization of ProptoView

Figure 4a showed two examples of ProptoView’s attention maps across four different views (frontal, closed-eye, upward gaze, and lateral). In every view, the periocular region consistently stood out as the key area for both eyes, underscoring how the model pinpoints features crucial for estimating proptosis. Figure 4b highlighted two patient cases—one with mild proptosis and one with severe proptosis—to illustrate scenarios with missing views. When actual eye images were available, ProptoView’s heatmaps (i.e., CAM visualizations) concentrated on the eye and its surrounding region. However, for any branch receiving a completely blank image, the heatmaps shifted to random or non-informative areas (e.g., the edges of the blank image). This behaviour confirmed that ProptoView relied only on genuine image data for its predictions and did not fabricate signals from non-existent or irrelevant content.

Fig. 4.

Visualization of ProptoView. (a) GradCAM + + visualization of ProptoView using all four views as input. (b) GradCAM + + visualization of ProptoView using dynamic input, comparing cases of mild and severe proptosis. (c) GradCAM + + visualization of ProptoView using dynamic input, comparing single input to four input views. (d) T-distributed stochastic neighbor embedding (t-SNE) plot of our model at proptosis measurement of 19 mm

Figure 4c illustrated ProptoView’s robustness by visualizing one patient’s data from two clinical visits, three months apart. Despite changes in both proptosis measurements and the quantity of input images, the model consistently generated meaningful attention maps, demonstrating reliability over time and potential for continuous monitoring. Lastly, Fig. 4d showed t-SNE plots of extracted features prior to proptosis prediction, segregating the dataset into two groups (≤ 19 mm vs. >19 mm). This clear separation confirmed that ProptoView’s features effectively differentiated various degrees of proptosis, capturing critical information that flagged abnormal values.

Discussions

ProptoView is an AI‑based approach for digital exophthalmometer using external ocular images, offering two key advantages: a novel multi‑view fusion mechanism and the ability to accept dynamic input. It provides a user‑friendly, non‑contact, zero‑radiation, and highly accurate alternative to the Hertel exophthalmometer and CT scans. Compared to Hertel, ProptoView reduces measurement time, eliminates operator dependence, avoids physical contact, and enables standardized longitudinal monitoring. In our study, its multi-view and dynamic input image-based AI models outperform the Hertel exophthalmometer in agreement with CT and show robust predictive capabilities across diverse equipment and clinical scenarios. Our research advanced the field through four key innovations: (1) an AI-driven, automatic measurement method using external ocular images; (2) a multi-view image fusion approach to enhance measurement accuracy; (3) a dynamic input mode that flexibly utilizes all available images; and (4) validation across multiple countries and diseases to confirm generalizability. These contributions aim to overcome existing limitations, paving the way for the broader development of accessible and highly precise medical AI tools.

ProptoView’s innovation stems from its 68-point facial landmark detection, which accurately localizes the eye region in photographs captured from different angles. This process reduces background interference and improves robustness in detecting proptosis. Incorporating multiple viewpoints also helps to capture all relevant ocular features for accurate assessment of proptosis. Our single-view model did not exceed the accuracy of the Hertel exophthalmometer, likely because ocular features can vary significantly between frontal and lateral perspectives, and relying exclusively on the frontal view may cause ProptoView to miss critical details [31]. Our multi-view model leveraged images from different angles, integrating data from multiple perspectives. This reduced both false-negative and false-positive errors through cross-validation of critical features and was more accurate than the Hertel exophthalmometer [32]. Comparing various combinations of image views, frontal view with upward gaze delivered the most significant improvement in accuracy. The clearer visualization of key anatomical structures in upward gaze view may have made proptosis more apparent. Lateral views also contributed notably as proptosis manifests distinctly along the sagittal axis [33].

We have conducted validation studies across a spectrum of proptosis severity levels, multiple centers, and diverse diseases. In five external test datasets, ProptoView’s correlation with Hertel measurements was comparable to the development dataset. Still, the dynamic input approach showed superior consistency and reliability among all AI models. However, a concern is that high agreement with Hertel might reflect reproduction of its known imprecision. Given that internal CT data confirm ProptoView agrees better with CT than Hertel does, the external ProptoView–Hertel concordance reflects ProptoView’s stable performance across populations, not replication of Hertel’s limitations. Notably, ProptoView performed well in the Singapore and Thailand datasets, which included a broader range of orbital diseases and ethnic anatomies beyond those in the training data. Besides, cross-hospital differences likely arise from anatomical variations across Asian ethnic groups, inconsistent manual measurements, and image quality variations. These are common in external validations and underscore the need for standardized AI-based exophthalmometry. Overall, ProptoView demonstrates robustness and wide applicability across clinical settings and diverse populations.

Our robustness assessment further demonstrated that ProptoView maintains high accuracy under varying capture distances, angular rotations, and even missing lateral views. These findings confirm that the dynamic model does not rely on any single view or precise acquisition geometry. Although part of its advantage may stem from a larger training dataset, its marked improvement over the single‑view approach and its resilience to incomplete or imperfect inputs underscore its clinical reliability. Given that patient compliance, equipment constraints, and individual physician practice often limit the type and number of obtainable photographs [34], the dynamic AI model offers a practical solution that maintains high accuracy even with fewer or imperfect images. Compared with depth‑aware models (ZoeDepth) [28], while depth cues improved single view performance by reducing MAE, ProptoView still achieved lower error, demonstrating the importance of true multi-view depth information. Moreover, pairwise comparisons across hospital groups using different devices (smartphones vs. DSLR) revealed only small between‑group differences (Cohen’s d < 0.3, Supplementary Fig. 1), suggesting that ProptoView’s performance is robust to hardware‑induced variability as well.

In addition, our results show that ProptoView greatly improves classification accuracy at the critical 19 mm proptosis threshold, an essential benchmark for disease grading and treatment planning. Patients with TED who exhibit proptosis below 19 mm are generally considered mild and managed with local treatments and regular monitoring, whereas those exceeding 19 mm often require more aggressive interventions [1–3]. Using this cutoff, ProptoView increased overall accuracy from 79.5% to 85.3% and significantly enhanced sensitivity—lowering false negatives from 18.7% to 4.0%—making it more suitable to facilitate earlier detection and intervention.

Developing smartphone applications derived from ProptoView could greatly expand the practical applications of exophthalmometry. For orbital disease specialists, accurate measurement of proptosis facilitates screening, diagnosis, treatment planning, and post-treatment follow-up. Meanwhile, its accessibility to non-orbital specialists, such as general ophthalmologists and endocrinologists, enhances assessments in non-specialty clinical settings. Beyond clinical settings, AI-driven smartphone applications can further enable at-home screening and longitudinal follow-ups, which could reduce the burden on clinicians delivering in-office care. However, images captured using smartphones could vary substantially from the standardized images used in model training, posing challenges related to generalizability and robustness [35]. In this study, three of the external test datasets included images from mainstream smartphone brands (Apple, Samsung, and Huawei), and the model demonstrated satisfactory performance across all devices. These findings not only underscore the feasibility of a non-contact, user-friendly, and cost-effective tool for broad public health applications, but also offer a foundation for future clinical guidelines to incorporate AI-based proptosis assessment.

There are some limitations in this study. First, the external test datasets only had Hertel exophthalmometer and not CT measurements of propotosis. Although CT scans provide a precise gold standard, they are not routinely used for exophthalmometry; thus, some centers understandably lack such data. This shortfall highlights the need for simple, automated alternatives that our model provides. Second, while our external validation included data from three Asian countries, the study is exclusively focused on Asian populations, and broader racial and ethnic diversity would be necessary to establish global generalizability across a wider range of facial morphologies. Finally, our findings are limited to the algorithmic level; future work should aim to create a user-friendly smartphone application to boost accessibility for broader clinical use.

Conclusion

We introduced ProptoView, a novel AI-based exophthalmometer leveraging multi-view external ocular images to bridge the gap between operator-dependent manual methods and resource-intensive CT scans. ProptoView not only outperforms the traditional Hertel exophthalmometer in accuracy but also exhibits robust performance across diverse imaging equipment, clinical scenarios, and patient populations. Our multicenter validation highlights its potential to enhance clinical practice and improve outcomes for patients with TED and other orbital diseases. By delivering a user-friendly, non-contact, zero-radiation and highly accurate exophthalmometry, ProptoView is well positioned for broad adoption in various healthcare settings and guiding future clinical practice.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AI

Artificial intelligence

CT

Computed tomography

DSLR

Digital single-lens reflex

ICC

Intraclass correlation coefficient

MAE

Mean absolute error

PFLD

PyTorch face landmark detector

RMSE

Root mean square error

TED

Thyroid eye disease

t-SNE

t-distributed stochastic neighbor embedding

Author contributions

H.Z., K.D. and C.L. contributed to conception and design of the study. C.L., P.S.P., J.D., S.W., X.L., W.L., D.W. and S.Z. contributed to data acquisition and interpretation. S.S. and K.D. built the deep learning model. J.F., C.L., X.L. and W.Z. performed the data analysis. C.L., S.S. and J.F. drafted the manuscript. C.C.Y., S.S., J.S., X.D., X.S., M.Z. and B.X. jointly supervised the project and provided critical revisions to the manuscript. All authors had access to the study data, critically read and reviewed the manuscript, and approved the final version for publication.

Funding

This work was supported by the National Natural Science Foundation of China (82388101, 82271122), National Key R&D Program of China (2024YFB4710200, 2024YFB4710205), Science and Technology Commission of Shanghai Municipality (20DZ2270800), Shanghai Key Clinical Specialty, Shanghai Eye Disease Research Center (2022ZZ01003), Shanghai Municipal Commission of Health and Family Planning Project (2022XD006), Shanghai Jiao Tong University 2030 Initiative (WH510272301), Shanghai Three-Year Plan for the Inheritance and Innovative Development of Traditional Chinese Medicine (2-5-1), Hainan Province Science and Technology Special Fund (ZDYF2024LCLH004), Research Development Fund of Xi’an Jiaotong-Liverpool University (RDF-24-01-110). This work was also supported by the AI for Digital Health Laboratory, School of Artificial Intelligence and Advanced Computing, Xi’an Jiaotong-Liverpool University.

Data availability

The de-identified data supporting our findings are available from the corresponding author upon reasonable request. The code employed for the analysis in this paper is available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study adhered to the Declaration of Helsinki and was approved by the Ethics Committee of Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine (No. SH9H-2022-T380-1), as well as the ethics committees of all participating centers. Written informed consent was obtained from all participants.

Consent for publication

Written informed consent was obtained from all participants, and all procedures were conducted in compliance with applicable guidelines and regulations.

Competing interests

The authors declare no conflicts of interest.