1. Introduction

Glaucoma is one of the leading causes of irreversible vision loss worldwide [1], and can be classified into primary angle-closure glaucoma (PACG) and primary open angle glaucoma (POAG). The predominant glaucoma subtype is POAG [2], accounting for 65.2% of all glaucoma, but PACG is the most visually destructive form of the disease [3]. The most conspicuous feature of PACG is an anatomical angle closure caused by appositional contact between trabecular meshwork (TM) and the iris [4]. Angle closure can cause very high intraocular pressure and reduce aqueous humor flow from the posterior to anterior chambers, which are important risk factors for glaucomatous optic neuropathy [3]. In addition to the natural angle closure, implantable collamer lens (ICL) implantation for correcting myopia also carries the risk of acute angle closure, resulting in a remote increase in intraocular pressure and pupillary block [5]. Nevertheless, PACG can be treated by timely diagnosis and intervention to halt the angle closure process at the early stage to prevent irreversible blindness [6]. Treatments to open angle include laser peripheral iridotomy (LPI) and trabeculectomy, as well as cataract extraction with intraocular lens (IOL) implantation [7].

There are three main ways to assess the closure of the anterior chamber (AC) angle: gonioscopy, ultrasound biomicroscopy (UBM), and anterior segment optical coherence tomography (AS-OCT). Gonioscopy is a contact examination method and the current clinical gold-standard for evaluating the AC anatomy and angle closure, allowing for the assessment of the AC closure angle over 360 degrees [8]. However, gonioscopy is highly subjective with poor reproducibility, requiring extensive experiences of ophthalmologists, which limits its clinical use [9]. UBM is a non-invasive imaging tool and can provide high-resolution scanning images with an axial resolution of 25 µm and a lateral resolution of 50 µm, and a depth of penetration of approximately 4 mm [10] [11]. Anterior segment optical coherence tomography (AS-OCT) is an alternative in vivo imaging method for the assessment of angle closure, and can provide detailed two and three-dimensional cross-sectional images for quantitative measurement of AC angle parameters, including angle opening distance (AOD), trabecular iris space area (TISA), trabecular iris angle (TIA), and anterior chamber angle (ACA). Although AS-OCT and UBM both show similar diagnostic values and reproducibility [10], AS-OCT has undoubtedly faster acquisition speed and higher resolution (<15 µm) than UBM [12]. As shown in Fig. 1, iris root (IR) and scleral spur (SS) are two critical anatomical landmarks of AC, where IR refers to the vertex of AC angle recess, and SS represents the junction between the inner wall of TM and the sclera [12]. Accurate detection of IR and SS is crucial for quantifying AC angle parameters. Nevertheless, current commercial SS-OCT systems such as Heidelberg ANTERION [13] only provide semi-automatic measurements of AC angle parameters requiring the users to label IR and SS manually. However, manual identification of these landmarks is time-consuming and subjective [14]. Therefore, an automated algorithm for localization of IR and SS and for quantitative measurement of AC angle parameters can drastically alleviate the workload of ophthalmologists and reduce inter- and intra-observer variability.

 

Fig. 1. Illustration of AS-OCT imaging and anatomical structures of anterior chamber. Each AS-OCT volume consists of 6 refraction-corrected two-dimensional cross-sectional AS-OCT images acquired along 6 scan lines separated by 30 degrees apart. The shadow region * represents the area that is not refraction corrected. Red/green point: SS/IR.

Download Full Size | PDF

Deep learning has been proven to be a powerful technique for automated medical image analysis [15–18]. Prior to this research, several methods based on deep learning were proposed to automatically detect SS in AS-OCT [14,19–23] and UBM images [24] with visible SS. Pham et al. [19] firstly generated a small rectangular focal region with SS as its center, and used a modified U-Net [16] to segment SS. Xu et al. [20] adopted a deep neural network based on ResNet-18 [25] with several fully connected layers to directly predict the coordinates of SS in AS-OCT images. Later, the first angle closure glaucoma evaluation challenge (AGE challenge) [14] providing a public AS-OCT dataset was introduced to facilitate the development of automated algorithms for angle closure assessment, including SS localization and angle closure classification. The top-ranked methods [14,21,22] proposed in the AGE challenge have built coarse-to-fine frameworks for heatmap regression of SS. Comparing with AS-OCT images, the method [24] for SS localization in UBM images showed a more significant prediction error due to the low signal-to-noise ratio of UBM images. These studies have demonstrated exemplary performance with a small mean absolute localization error (MAE). However, due to the limitation of AS-OCT, the location of SS may be invisible and could not be determined in approximately 30% of AS-OCT images [26]. Therefore, previously reported algorithms [19–23] that were only designed for SS localization could in fact generate errors for quantitative evaluation of AC angle parameters in AS-OCT images without detectable SS. For localization of IR, one previous method [27] extracted the position of IR from iris segmentation map and used IR as a salient landmark to locate the coarse AC angle region. However, the IR in the research [27] was not defined as the vertex of AC angle recess and could be wrongly recognized.

Although CNN-based methods [15–18] have shown good performance in some medical image analysis tasks, their intrinsic local operations based on convolutions inevitably neglect the global semantic information which limits their performance in certain applications requiring global context. Recently, the transformer [28] with self-attention mechanism that is widely used in natural language processing (NLP) was introduced into the computer vision domain [29], and has achieved great success in modeling global and long-range image features [30]. Hence, combining CNN with the transformer [28] can potentially improve the performance of SS and IR detection by exploiting both local and global contexts.

The purpose of this study was to fulfill the unmet clinical need by developing a new method to automatically differentiate whether SS is present, and further detect IR and SS on a full spectrum of possible cases, and accurately calculate AC angle parameters including AOD, TISA, TIA, and ACA. To this end, we proposed a hybrid CNN & transformer model to capture both local and global contextual information of IR and SS, and extensively validated the method by comparing the algorithm performance with two experienced analysts as well as with other state-of-the-art methods for AS-OCT and medical image analysis. In addition, we assessed the implications of AC angle parameters during IOL and ICL implantation by applying the proposed method to a clinical case of PACG patient undergoing IOL implantation, and a patient with high myopia undergoing ICL implantation. Finally, we discuss the significance and limitations of our method in a clinical context.

2. Methods

2.1 Datasets

AS-OCT images from 362 eyes of 203 patients who underwent AS-OCT imaging to diagnose cataract and high myopia were collected by a commercially available AS-OCT system (ANTERION, Heidelberg Engineering GmbH, Heidelberg, Germany) that uses swept-source OCT technology with a center wavelength of 1300nm, an axial resolution of less than 10 µm and a transversal resolution of 45 µm [31], at Shanxi Eye Hospital, Taiyuan, China. The mean age of patients was 58.1$\pm$21.6 years. AS-OCT images with severe artifacts such as eyelids that hampered attempts to identify IR and SS were excluded. The final dataset included for image analysis were 1774 refraction-corrected AS-OCT images from 203 patients (an average of 8.7$\pm$7.2 images/patient acquired with different radial scans) and 1531 refraction uncorrected AS-OCT images from a subset of 66 patients to increase the diversity of training. Images with visible IR and SS respectively accounted for approximately 99% and 67% of the dataset, which was consistent with former research [26]. An experienced ophthalmologist with 10 years of clinical experience labeled IR and SS using an open source software (Labelme, MIT, Cambridge, MA) [32]. The dataset was divided into a training set consisting of 2145 images, 205 eyes, and 116 patients to train the model, a validation set with 572 images, 75 eyes, and 41 patients to evaluate the hyperparameters, and a test set containing 588 images from 82 eyes and 46 patients to evaluate the performance of the algorithm. The training and validation data contained both refraction-corrected and uncorrected images, but the test set only contained refraction-corrected images. The percentage of images with visible SS in the training, validation, and test set was 68%, 72%, and 59%, respectively. To evaluate inter-observer variability, the test set was labeled again by another ophthalmologist with 2 years of experience blinded to the results from the first ophthalmologist and the algorithm. Moreover, since it was necessary to extract the boundaries of AC for the measurement of AC angle parameters, we constructed another dataset by randomly sampling 220 AS-OCT images from the entire dataset for developing and validating a separate algorithm for AC segmentation. The dataset was also divided into a training set (60%), a validation set (20%), and a test set (20%). This study was approved by the local IRB offices of the authors’ institutions. All study patients provided informed consent and patients’ information was de-identified before analysis.

2.2 Methodology overview

The overall workflow of our method consists of AC segmentation, IR and SS detection, and AC angle parameter measurement (Fig. 2). We adopted the classic U-Net-like [16] CNN to segment the AC, which consists of a feature encoder module based on Efficientnet-B4 [33] and a feature decoder module with four de-convolutional layers to recover the spatial resolution reduced by the feature encoder module. The anterior iris surface and posterior corneal surface extracted from the AC segmentation map were used as reference boundaries for subsequent AC parameter measurements. At the stage of IR and SS detection, the original AS-OCT image was split along the median of AC into left and right sub-images which were subsequently input into a hybrid CNN & transformer model. With the boundaries of AC and the coordinate information of IR and SS, the AC angle parameters (AOD, TISA, TIA, and ACA) were subsequently calculated.

 

Fig. 2. Flow diagram illustrating the overall methodology for AC segmentation, IR and SS detection, and AC angle parameter measurement. The CNN model took the full image as the input to segment AC. For IR and SS detection, two sub-images cropped from the full image along the median of AC were individually input to a hybrid CNN & transformer model. Red/green point: SS/IR.

Download Full Size | PDF

2.3 Vision transformer

Over the past decade, CNNs have been widely used and have shown remarkable performance in various medical image analysis tasks [34]. It is universally acknowledged that CNNs can effectively extract local features by virtue of local receptive fields of convolution operations. However, CNN-based methods exhibit inherent limitations to model long-range dependencies because they focus on local patterns instead of global context. Recently, the transformer [28] was introduced into the computer vision field [29] and demonstrated a powerful capability to capture the long-range relation and to model global context by using the self-attention mechanism [28]. We first review the concepts of transformer below.

An input image $x_{in}\in R^{3\times H\times W}$ with height H and width W can be divided into a sequence of flattened 2D image patches {$x^{i}_{p}\in R^{P^2\cdot 3}| i = 1,\ldots, N$}, where the size of each patch is P$\times$P and N = $\frac {H\times W}{P^2}$ is the number of the resulting image patches. To be consistent with the refraction-corrected AS-OCT image with a three-channel input $x_{in}\in R^{3\times H\times W}$, the gray AS-OCT image $x\in R^{1\times H\times W}$ without refraction correction was converted into $x_{in}\in R^{3\times H\times W}$ by duplicating the image along the channel dimension. The flattened image patches are mapped into a latent D-dimensional embedding space using a trainable patch embedding projection E $\in R^{(P^2\cdot 3)\times D}$, generating a sequence of patch embeddings. A learnable position embedding E$_{pos}\in R^{N\times D}$ is then added to the patch embeddings to retain positional information:

The attention mechanism maps a query and a set of key-value pairs to the output using scaled dot-product attention and multi-head self-attention (MSA). Queries, keys, and values are packed into matrices Q, K, and V. The function of scaled dot-product attention with a scaling factor of $\frac {1}{\sqrt {d_k}}$ can be represented as follows:

Multi-head self-attention linearly projects the queries, keys, and values h times to $d_k$, $d_k$, and $d_v$ dimensions:

2.4 Hybrid CNN & transformer model

Accurate detection and localization of SS require a combination of global contexts and local fine-grained features. Therefore, we designed a hybrid CNN & transformer model as shown in Fig. 3, where the feature encoder consists of a CNN-based branch and a transformer-based branch. The two branches take the same image as the input and respectively output a feature map encoding both fine-grained local context and long-range global information.

 

Fig. 3. Framework of the hybrid CNN & transformer model. The encoder module of the network contains a local branch based on Efficientnet-B4 and a global branch based on transformer. The same image is input into the local branch and global branch, respectively outputting a feature map with fine-grained local context and one with global information. The two feature maps are concatenated in channels, and then upsampled by three decoder blocks and a de-convolutional layer, generating two heatmaps with the same width and height to precisely localize IR and SS, respectively.

Download Full Size | PDF

The CNN-based branch is based on the widely used Efficientnet-B4 [33]. The transformer-based branch consists of L transformer layers. The feature maps generated by the CNN-based and the transformer-based branches are concatenated along channels. A 1$\times$1 convolution operation is used to fuse the combined feature maps. Following the U-Net-like architectures [15], the fused feature maps containing the high-level semantic information are upsampled to recover the spatial information of the target landmarks through several decoder blocks. In addition, skip connections with addition operations fusing the feature maps from the encoder and decoder blocks are further added to preserve more detailed information and remedy resolution loss caused by max-pooling and convolutional layers. After a de-convolution operation in the final layer, two heatmaps respectively containing the position information of IR and SS are generated, which are further filtered using a Gaussian kernel with a size of 15$\times$15, and the positions of the peak values are extracted as the precise coordinates of IR and SS. Finally, we exclude the predicted IR and SS with peak values <0.5, which indicates the IR and SS are invisible in the input OCT images.

2.5 AC angle parameter measurement

In this study, we only focused on the measurement of AOD, TISA, and TIA at 500 µm, as shown in Fig. 2, and the corresponding parameters at 750 µm can be similarly calculated. AOD at 500 µm (AOD 500) represents the length of the line perpendicular to the plane of TM from the TM-500 (a point on the TM at 500 µm from the SS) to the opposing iris [35]. TIA at 500 µm (TIA 500) is defined as the angle measured with the apex in the SS and the arms of the angle passing through TM-500 and a perpendicular point on the opposing iris [35]. TISA at 500 µm (TISA 500) is defined as the trapezoidal area enclosed by AOD 500, the inner cornea-scleral wall, iris surface, and a line drawn from the SS perpendicular to the plane of TM [12]. ACA refers to the angle between the iris tangential line and that of the posterior corneal surface with IR [35].

2.6 Experimental details

Our model was implemented on PyTorch and Ubuntu 16.04 with 4 Nvidia Titan XP GPUs and was trained for 100 epochs with the mini-batch size set to 4. The input images were resized to 512$\times$512. For training, we used the Adam optimizer with the mean square error (MSE) loss and we set the initial learning rate to $2\times 10^{-4}$. The learning rate was multiplied by 0.5 if the loss value was not changed over 2 epochs. The training was early stopped if the loss value was not changed over 5 epochs. The hyper-parameters of the transformer-based branch were consistent with ViT [29], where the size P$\times$P of image patches was set to 16$\times$16, D, $d_{model}$, $d_k$, $d_v$, h, and L were set to 768, 768, 64, 64, 12, and 12, respectively. We did not use any image pre-processing or data augmentation.

2.7 Algorithm evaluation and statistical analysis

Detected IR and SS were classified as true positives (TP), false positives (FP), and false negatives (FN) with the human annotations as the ground truth. The prediction was classified as TP if the predicted landmark location was within 80 µm (approximately 20% of the length of TM) distance both in the X and Y-axis with the ground truth, consistent with previous studies [36]. The detection accuracy of IR and SS was evaluated using the metrics of precision, sensitivity, and F1 score computed as follows:

We also used the mean absolute localization error (MAE) to evaluate the localization accuracy of IR and SS by calculating the distance between the prediction and the ground truth. Given the predicted coordinate (x, y) and the ground truth ($x_g$, $y_g$), the MAE can be calculated as:

Dice coefficient was used to compute the similarity between the AC segmentation map and the corresponding ground truth:

We compared our proposed method with the winner of the AGE challenge [14]. The winner of the AGE challenge adopted a coarse-to-fine localization framework which consisted of two CNNs with the same Efficientnet-B5 [33] based architecture to supervise the regression of the landmarks, achieving the best localization accuracy for SS in the AGE challenge. We also compared our method with HRNet [37], U-Net [16], and recently proposed TransUNet [30]. HRNet was a widely used CNN for heatmap regression of key points. The original U-Net and TransUNet for medical semantic segmentation were modified with heatmap regression with the MSE loss.

Bland-Altman analysis was used to compare the difference between automatic and manual AC angle parameter measurement.

2.8 Clinical application of AC angle parameter measurement in IOL and ICL implantation

Eyes with angle closure inherently have higher intraocular pressure caused by a narrow AC angle and a thick and anteriorly vaulted lens than normal eyes [38]. However, the most common treatments for PACG, iridectomy and iridotomy, do not always permanently control intraocular pressure because of released pigment granules and iris tissue fragments in the case of laser iridotomy [39]. It was reported that cataract extraction might have a beneficial clinical outcome [39] in long-term intraocular pressure control. Phacoemulsification and IOL implantation are widely used in cataract extraction [40]. Therefore, measuring the AC parameters before and after cataract extraction with IOL implantation allows for assessing its clinical significance for PACG patients. We demonstrated the application of our method towards this need by comparing the AC angle parameters over 360 degrees of a left eye in a 56-year-old male patient with PACG before and after phacoemulsification and IOL implantation.

ICL is a toric posterior chamber phakic intraocular lens to correct myopia, hyperopia, and astigmatism [41]. After ICL insertion, acute angle closure may be caused by elevated intraocular pressure secondary to anterior vaulting of the ICL [42]. Moreover, a previous research [43] reported that the intraocular pressure elevation following ICL implantation may fluctuate over time and reach a peak several days postoperatively. Although ICL implantation is currently considered as a safe and effective treatment for patients with high myopia, its long-term safety and stability warrant further investigation [44]. We aimed to assess the variation of intraocular pressure elevation by measuring the AC angle parameters in a right eye from a 30-year-old female patient undergoing ICL implantation to treat high myopia.

3. Results

3.1 Algorithm performance on IR and SS detection and localization

With 588 test AS-OCT images from 82 eyes and 46 patients, our proposed method achieved a precision of 0.941, a sensitivity of 0.914, a F1 score of 0.927, and a MAE of 37.1$\pm$25.3 µm for IR, and a precision of 0.805, a sensitivity of 0.847, a F1 score of 0.826, and a MAE of 41.4$\pm$29.4 µm for SS (Table 1). Comparing with the winner of AGE and other state-of-the-art methods in the field of medical image analysis, our method achieved superior performance in both detection and localization accuracy. The results of our proposed IR and SS detection model for representative cases are presented in Fig. 4, alongside a comparison with other methods. Our proposed method exhibited accurate detection of IR and SS, achieving a smaller MAE compared to other methods in the general case (Fig. 4(a)), as well as in challenging scenarios, such as partial blockage by the eyelid (Fig. 4(b)) and angle closure (Fig. 4(c)). Furthermore, the proposed method automatically excluded invisible SS in AS-OCT images (Fig. 4(d)).

 

Fig. 4. Results of IR and SS detection of our proposed algorithm and other methods in four typical cases. (a): Normal AC angle. (a) Partial blockage by eyelids. (c): Severe AC angle closure. (d): Invisible SS. Red/green points: SS/IR.

Download Full Size | PDF

Table 1. Performance evaluation of the proposed method

View Table

To validate the effectiveness of including both the local and global branches in the feature encoder for IR and SS detection, we conducted two ablation studies by eliminating each branch of the proposed model and compared the resulting performance with the full model. As shown in Table 1, ablating either of the branches would deteriorate all the classification and localization metrics, which proved the importance of including both the CNN and transformer modules for IR and SS detection and localization.

To evaluate inter-observer variability, we compared the agreements between the algorithm and the two analysts, and the results are shown in Fig. 5. The precision, sensitivity, and F1 score of [analyst A VS. analyst B, analyst A VS. algorithm, analyst B VS. algorithm] for detecting IR are [0.941, 0.949, 0.924], [0.946, 0.959, 0.918], and [0.943, 0.954, 0.921], respectively. The precision, sensitivity, and F1 score of [analyst A VS. analyst B, analyst A VS. algorithm, analyst B VS. algorithm] for detecting SS are [0.933, 0.832, 0.830], [0.839, 0.889, 0.894] and [0.883, 0.860, 0.861], respectively. The MAE of [analyst A VS. analyst B, analyst A VS. algorithm, analyst B VS. algorithm] of IR and SS are [35.9$\pm$19.8 µm, 35.2$\pm$18.8 µm, 36.8$\pm$19.5 µm] and [31.2$\pm$18.3 µm, 35.8$\pm$20.1 µm, 34.3$\pm$18.1 µm] respectively. These results demonstrated that the algorithm achieved similar agreements with human analysts for both detection and localization of IR and SS.

 

Fig. 5. Comparison between the algorithm and different analysts for the detection (Left) and localization (Right) of IR and SS.

Download Full Size | PDF

3.2 Evaluation of quantitative measurements

We calculated the four AC angle parameters using the coordinates of the detected IR and SS and the extracted AC boundaries. The AC segmentation model achieved a Dice coefficient of 0.991$\pm$0.002 on the test set. Figure 6 shows that the mean differences of AOD 500, TISA 500, TIA 500, and ACA between automatic and manual measurement are −0.005$\pm$0.393 mm, 0.000$\pm$0.199 mm$^2$, −0.1$\pm$9.3°, and −0.2$\pm$5.0°, respectively, and there are slightly worse agreements between the automatic and manual measurement for the large open angle of AC. This is because the large open angle amplified the prediction error of SS localization.

 

Fig. 6. Bland-Altman analysis between automatic and manual measurement of AC angle parameters including (a) AOD 500, (b) TISA 500, (c) TIA 500, and (d) ACA.

Download Full Size | PDF

3.3 Clinical application of AC angle parameter measurement in IOL and ICL implantation

We analyzed the AC angle parameters from an eye of a patient with PACG before and after cataract extraction (Fig. 7). Figure 7(a)-(d) show that IOL implantation resulted in increased AC angle parameters. The mean pre- and post-operative [AOD 500, TISA 500, TIA 500, ACA] are [0.392$\pm$0.071 mm, 0.136$\pm$0.038 mm$^2$, 37.5$\pm$5.1$^{\circ }$, 32.8$\pm$5.7$^{\circ }$] and [0.669$\pm$0.145 mm, 0.228$\pm$0.049 mm$^2$, 52.2$\pm$5.4$^{\circ }$, 47.9$\pm$7.1$^{\circ }$], where the post-operative parameters were increased by [70.2$\pm$16.1%, 72.5$\pm$30.3%, 40.0$\pm$9.7%, 47.2$\pm$13.5%] respectively compared to the pre-operative values (p<0.0001).

 

Fig. 7. AC angle parameter measurement before and after IOL (a-d) and ICL (e-h) implantation. The AC angle parameters over 360 degrees were interpolated by applying cubic-spline interpolation. Blue/orange points: measured AC angle parameters of pre- and post-operation. Green points: interpolated AC angle parameters. Blue and orange curves respectively represent fitted AC angle parameter curves of pre- and post-operation.

Download Full Size | PDF

By contrast, as shown in Fig. 7(e)-(h), the mean pre- and post-operative [AOD 500, TISA 500, TIA 500, ACA] for the patient undergoing ICL implantation to treat high myopia are [0.655$\pm$0.169 mm, 0.241$\pm$0.057 mm$^2$, 51.6$\pm$6.3$^{\circ }$, 42.6$\pm$5.8$^{\circ }$] and [0.325$\pm$0.128 mm, 0.101$\pm$0.056 mm$^2$, 31.9$\pm$10.8$^{\circ }$, 24.8$\pm$4.7$^{\circ }$], where the post-operative AC parameters were reduced by [50.4$\pm$17.9%, 58.2$\pm$23.5%, 38.5$\pm$19.0%, 41.9$\pm$5.2%] respectively compared to the pre-operative values (p<0.0001). This indicated that the ICL implantation caused acute angle closure, which may elevate intraocular pressure that is one of the main risk factors for PACG.

4. Discussion

In this paper, we proposed a novel hybrid CNN & transformer model for automated detection and localization of IR and SS in AS-OCT images. The model incorporates the knowledge of fine-grained local contexts and long-range global dependencies into its feature encoder and has been proven to be effective and robust by testing a full spectrum of possible cases in the clinic (Table 1). The proposed method is advantageous over existing methods [14,16,30,37] both in terms of performance, and by taking into consideration of the visibility of SS in AS-OCT images for the first time. Previous studies [20] [24] for SS detection reported that the CNN model trained on the reference labels provided by one human analyst would not generalize to other analysts. However, we demonstrated that our proposed method achieved similar performance compared with different analysts (Fig. 5).

The automatic detection and localization of IR and SS are preliminary steps for AC angle parameter measurement which are essential for assessing acute angle closure caused by ICL implantation. Implantation of the ICL narrows the angle width and depth [45] and high intraocular pressure can be developed postoperatively after ICL implantation [43] in some eyes, which are important risk factors for PACG. One previous clinical study [46] demonstrated that AC angle parameters measured by AS-OCT have a strong correlation with intraocular pressure. Automatic detection and localization of IR and SS can help physicians quickly evaluate the difference of AC angle parameters between pre- and post-operation, providing the clinical reference for managing the acute PACG, as demonstrated in this study shown in Fig. 7(e)-(h).

Contrary to ICL implantation, phacoemulsification and IOL implantation used in cataract treatment can effectively eliminate the pupillary block and well control the intraocular pressure in PACG patients [47], as was also demonstrated by the increase of AC angle parameters between pre- and post-operation in Fig. 7(a)-(d) of this study. The post-operative intraocular pressure peak is a significant problem that needs to be detected early and treated vigorously [47]. Therefore, automatic AC angle parameter measurement facilitates the monitoring of the progression of angle narrowing and angle closure over time, which can also help find the intraocular pressure peak.

The automatic detection and localization of IR and SS have great significance for standardizing the OCT-based evaluation of angle closure in patients with PACG. The determination of narrow angle is based on the space between the TM and the iris, which is interpreted qualitatively in AS-OCT images in current clinical practice [12] [35] [48]. Nevertheless, there is no consensus on which parameters or cut-off values to use to identify a narrow angle [49]. The previous clinical research [50] determined the thresholds of AOD and TISA for distinguishing open vs. narrow angle eyes with gonioscopy examination as the reference standard, and found that AOD had a better discriminative ability for detecting narrow angles than TISA. Automatic and quantitative measurement of AC angle parameters allows for developing and validating OCT-based diagnostic criteria for narrow angles in large clinical studies.

The automatic measurement of AC parameters is also crucial for determining the anatomical risk factors contributing to the pathogenesis of PACG. In addition to the static anatomic factors, such as shallow anterior chamber depth, greater lens thickness, and larger lens vault, a recent study [51] demonstrated that dynamic changes in AC angle parameters with light-to-dark transition are correlated with the disease progression of PACG. However, the clinical application of this finding is currently limited by the manual analysis of AC angle parameters in AS-OCT images. Automatic analysis of AC angle parameters could contribute to monitoring the dynamic changes of AC angle anatomy and identifying the rick factors associated with the progression of angle narrowing in patients with PACG.

This study has limitations. AS-OCT images in this research were collected from one AS-OCT imaging system (ANTERION). Transfer learning [52] can be implemented for generalizing our method to AS-OCT images from other commercial systems with different imaging configurations and image quality. Therefore, future multicenter study is required to improve the performance and to test the generalization ability of the proposed deep learning method.

5. Conclusion

In summary, we proposed a novel hybrid CNN & transformer model to automatically detect and localize IR and SS, and to automatically compute the AC angle parameters in AS-OCT images. The results demonstrated that our method achieved significantly better performance than state-of-the-art methods for detection and localization of IR and SS, and is highly accurate for AC angle parameter measurement, and has a great potential to be used in clinic for the screening, prevention, and treatment of PACG.