Smartphone-Acquired Anterior Segment Images for Deep Learning Prediction of Anterior Chamber Depth: A Proof-of-Concept Study

Abstract

Purpose: To develop a deep learning (DL) algorithm for predicting anterior chamber depth (ACD) from smartphone-acquired anterior segment photographs.

Methods: For algorithm development, we included 4,157 eyes from 2,084 Chinese primary school students (aged 11-15 years) from Mojiang Myopia Progression Study (MMPS). All participants had with ACD measurement measured with Lenstar (LS 900) and anterior segment photographs acquired from a smartphone (iPhone Xs), which was mounted on slit lamp and under diffuses lighting. The anterior segment photographs were randomly selected by person into training (80%, no. of eyes = 3,326) and testing (20%, no. of eyes = 831) dataset. We excluded participants with intraocular surgery history or pronounced corneal haze. A convolutional neural network was developed to predict ACD based on these anterior segment photographs. To determine the accuracy of our algorithm, we measured the mean absolute error (MAE) and coefficient of determination (R ²) were evaluated. Bland Altman plot was used to illustrate the agreement between DL-predicted and measured ACD values.

Results: In the test set of 831 eyes, the mean measured ACD was 3.06 ± 0.25 mm, and the mean DL-predicted ACD was 3.10 ± 0.20 mm. The MAE was 0.16 ± 0.13 mm, and R ² was 0.40 between the predicted and measured ACD. The overall mean difference was -0.04 ± 0.20 mm, with 95% limits of agreement ranging between -0.43 and 0.34 mm. The generated saliency maps showed that the algorithm mainly utilized central corneal region (i.e., the site where ACD is clinically measured typically) in making its prediction, providing further plausibility to the algorithm's prediction.

Conclusions: We developed a DL algorithm to estimate ACD based on smartphone-acquired anterior segment photographs. Upon further validation, our algorithm may be further refined for use as a ACD screening tool in rural localities where means of assessing ocular biometry is not readily available. This is particularly important in China where the risk of primary angle closure disease is high and often undetected.

Keywords: anterior chamber depth; deep learning; glaucoma; primary angle-closure glaucoma; smartphone.

Figure 1

Smartphone mounted on slit lamp in use. Anterior segment photographs were captured on study eyes using a smartphone (iPhone Xs, Apple Inc, CA, USA) attached to a slit lamp. The smartphone was fixed on the eyepiece with an adapter (Celestron 81035, Celestron Acquisition LLC, CA, USA), making the camara lens in line with the eyepiece. We used the default mode of iPhone camara with a minimal magnification (1 X) to take photographs. A Bluetooth trigger for a one-tap image capture was fixed on the joystick making the procedure of taking photographs quickly and stably. Diffuse illumination of slit-lamp was used at 45-degree angle, with magnification set at 16X.

Figure 2

Scatterplot illustrating the relationship between deep learning-predicted and actual anterior chamber depth (ACD) measurements from Lenstar (n = 831, r = 0.63, P < 0.001).

Figure 3

Bland-Altman plots illustrating agreement between deep learning-predicted and actual anterior chamber depth (ACD) measurements from Lenstar (n = 831).

Figure 4

Averaged heatmap shows the regions of the anterior segment photograph that were most important for the deep learning algorithm predictions in test set. Hotter colors (reds) indicate higher activity while cooler colors (blues) represent lower activity. (A) an example of original anterior segment photograph (right eye) obtained by iPhone Xs; (B) the heatmap of the corresponding photograph; (C) an example of original anterior segment photograph (left eye); (D) the heatmap of the corresponding photograph; (E) averaged heat map crossed all images (n = 831).