In this study, using the pre-trained Inception-V3 with transfer learning, the proposed DL model showed the capability of automatically detecting shallow or deep AC directly from overview anterior segment photographs without slit-lamp illumination. A comparison of the diagnostic accuracy between ophthalmologists and the DL model revealed that ophthalmologists were less likely to detect shallow AC through the anterior segment appearance. To the best of our knowledge, this study is the first to report the classification ability of a DL model with high accuracy in shallow AC detection using anterior segment photographs.
ACD measurements have several important applications in ophthalmology. For example, shallow ACD has been previously identified as a risk factor for PACG. In the Singapore population, Lavanya et al. reported that AC depth of less than 2.80 mm was associated with a 42 times higher odds of angle-closure compared with AC depth of at least 3.00 mm (3). Shallow ACD is associated with ocular and general parameters. In Chinese adults, Xu et al. demonstrated shallow AC to be associated with age, female gender, hyperopia, small optic disk, short body stature, and chronic angle-closure glaucoma (22). Several clinical techniques have been proposed for AC depth measurements, such as IOLMaster, A-Scan ultrasound, and Pentacam (1, 5, 6, 7). However, these techniques are expensive and require trained nurses or technicians. In the current study, the DL model requires only photos and shows a higher accuracy (0.83 with 95% CI, 0.80–0.86) in screening shallow AC (AC depth < 2.4 mm) in the clinical hold-out dataset. Most medical DL systems adopt senior doctors’ grading as ground truth, but this grading system is time consuming and inherently subjective. The proposed DL model used the quantitative measurement of Pentacam as the gold standard to grade the dataset, and it made the results more objective and reliable. Such advantages make the DL technique an efficient means of screening the general population.
AC depth is a 3D biometric parameter associated with the anatomical structures of the anterior segment, such as the lens vault and the posterior corneal arc length (23). In clinical practice, ophthalmologists can qualitatively assess the AC using the pen torch method, the slit-lamp van Herick technique (24), or the Smith method (25). As knowledge and clinical experience vary among different individuals, human performance shows large variations in these techniques. Moreover, it is difficult to detect shallow AC directly from anterior segment photographs because of the limitation of 3D information. DL may address this by learning the critical features from a high-dimensional space (26, 27). For classification tasks, higher layers of the DL model amplify the aspects of the input that are important for the discrimination and suppression of irrelevant variations. Varadarajan (27) successfully used DL to make predictions using simple 2D images by fundus photography without sophisticated 3D imaging equipment in diabetic macular edema grading. In our study, we used t-SNE to create a 2D reduced representation of the 256-dimensional space extracted from the last fully connected layer of the DL model (Fig. 3). Our result shows that the DL model is able to automatically generate features that roughly detect shallow AC from anterior segment photographs using AC depth of less than 2.4 mm as a reference standard.
Grad CAM is an algorithm used to create heatmap images that indicate where the DL model is focused. Note that Grad CAM highlights the central AC, cornea, and surrounding iris area, which is also where ophthalmologists assess the AC during routine clinical practice (25). Grad CAM may also uncover the reasons that cause false predictions of the DL model (28). In the current study, the most common reason for misclassification is the images with coexisting photo conditions, especially those that were defocused during photography. Figures 4C and D show that the DL model highlights the eyelid area, which was focused when the photograph was taken. This issue can be solved using an advanced imaging technique, such as auto-focus (29). Another reason for misclassification is the images with coexisting eye conditions, such as high lacrimal meniscus (Figs. 4E–F). These nontraditional highlighted regions may offer some additional information for further investigation by eye care professionals.
This study has several limitations. First, the Pentacam camera uses a monochromatic slit-light source to produce only black and white images. Fortunately, the DL model can be adopted to train with color images with other imaging modalities. Second, the sample size of the training dataset is relatively small, and the model can only predict shallow or deep AC, not a specific value of AC depth. Third, all the subjects involved in the study were Chinese. Future studies with more subjects of multiple ethnicities and multiple imaging modalities, such as mobile phone eye photography, will be beneficial to provide more general predictions for clinical practice or community screening.
In conclusion, we proposed a DL model that can automatically detect a shallow AC based on anterior segment photographs. The results suggest that this DL model may be a potential tool for routine eye screening. Future efforts involving multiple ethnicities and multiple imaging modalities are warranted to investigate the application of this technology in the clinical and research setting or in community screening.