GlauNet: Glaucoma Diagnosis for OCTA Imaging Using a New CNN Architecture

Glaucoma is a neurodegenerative disease that affects the optic nerve head and causes visual field defect. Current investigations focus on neural component which may overlook other important factors such as the vascular cause. The optical coherence tomography angiography (OCTA) imaging has been developed and provided quantitative parameters that showed good diagnostic accuracy to detect glaucoma. However, those parameters are based on image processing of observed clinical findings, therefore, some image information can be lost. Convolutional neural network has been successfully applied for automatic feature extraction and object classification. In this study, the glaucoma diagnosis network, namely GlauNet, has been proposed. GlauNet consists of two sections: the feature-extraction section and the classification section. The feature-extraction section has three convolutional layers. Each convolutional layer is followed by rectified linear unit and maximum pooling layer. The classification section contains five fully connected layers. GlauNet was trained with 258 glaucomatous and 439 non-glaucomatous eyes. The visualization of the feature-extraction section showed the highlight in the area of optic nerve head and retinal nerve fiber layer in the superotemporal and inferotemporal regions. It was then tested on 27 glaucomatous and 48 non-glaucomatous eyes. Its sensitivity and specificity were 88.9% with 89.6%, respectively. The area under receiver operating characteristic curve of GlauNet was 0.89. GlauNet was robust against the artifacts. Its sensitivity and specificity were still higher than 80% (82.4% and 80.3%, respectively) when tested on 88 poor-quality images.

by multiple factors. The prevalence of glaucoma in the world 23 was 3.54% in the year 2014 and there will be about 111.8 mil-24 lion glaucoma patients in the year by 2040. The majority of 25 patients will be in Asia and Africa [1]. Most of the patients 26 are still asymptomatic until the late stage of the disease, 27 The associate editor coordinating the review of this manuscript and approving it for publication was Prakasam Periasamy . as a result, the prevalence of undiagnosed patients were 28 between 53% and 88% around the world [2], [3], [4], [5], 29 [6]. In addition, challenge in glaucoma diagnosis is based on 30 anatomical variations among people. The current definition 31 of glaucoma is based on evaluation of neural structure of 32 the optic nerve head and surrounding retina and visual field, 33 while intraocular pressure is a major risk factor. 34 The mainstay glaucoma diagnostic tools in current practice 35 are optic disc examination, optical coherence tomography 36 (OCT) and visual field or perimetry. Optic nerve head assess-37 ment is the most commonly used method, because it is easy to 38 Reduction of blood vessels density in optic disc and Besides the above parameters, there are other vascular 94 parameters related to glaucoma [27]. For examples, disc flow 95 index, blood vessels tortuosity, vessel perimeter index, ves-96 sel complexity, branchpoint analysis, and flow analysis. The 97 relation of these parameters to the glaucoma comes from 98 clinical observation, thus, some image information can be 99 overlooked. Furthermore, the features that were described in 100 previous literature are individually extracted/segmented by 101 tailor-made image processing techniques which is time con-102 suming and may gave an incorrect value in noisy image [28]. 103 Convolutional neural network (CNN) is an effective image 104 classifier. It is trained to automatically capture the image 105 features important for its problem. CNN can be the more 106 effective glaucoma detector, since it can be trained to use the 107 non-clinical features that are correlated with vascular mech-108 anism in the pathogenesis of glaucoma. CNN uses a patch 109 convolved over the entire image. It extracts dominant features 110 in the hidden layers while preserves the correlation between 111 neighboring pixels [29]. After that, the output signal is trans-112 formed into a vector and classified. CNN has been applied to 113 many ophthalmologic image modalities. Several studies eval-114 uated the performance of glaucoma detection in the fundus 115 photograph using transfer learning [30], [31], [32], [33] on 116 various conventional deep learning models such as VGG16, 117 VGG19, Inception-v3, ResNet50, and GoogLeNet. The area 118 under receiver operating characteristic curves (AUC) were 119 around 0.90 or more [32], [33], [34]. The results showed 120 the accuracy equal to or more than ophthalmology trainees 121 [34], [35]. Some investigators evaluated the performance of 122 deep learning in OCT retinal nerve fiber layer thickness and 123 found the AUC of more than 0.90 [36], [37], [38]. In addi-124 tion, a combination of OCT quantitative value to the fundus 125 photograph was another approach to improve the accuracy of 126 glaucoma diagnosis [39].

127
A number of deep learning methods have been applied to 128 an OCTA image for diagnosing retinal diseases [40], [41], 129 [42]. Nagasato et al. [40] applied VGG16 and supported 130 vector machine to detect non-perfusion area in retinal vein 131 occlusion. 322 OCTA images of normal and retinal vein 132 occlusion were augmented and used. The AUC of VGG16 133 was 0.98 which was better than supported vector machine and 134 ophthalmologists. Le et al. [41] also used VGG16 to detect 135 diabetic retinopathy. The VGG16 was trained with a very 136 small dataset (32 controls and 99 diabetic eyes). The best 137 accuracy (AUC of 0.97-0.98) was achieved when the final 138 9 layers were retrained. Another study on diabetic retinopathy 139 was done by Heisler, et al. [42]. They use ensemble method 140 on multiple layers of foveal avascular zone area in 380 eyes 141 and found that VGG19 provided the best performance with 142 an AUC of 0.90-0.92.

143
The application of deep learning for glaucoma detection in 144 an OCTA image is still limited. Recently, Bowd, et al. [43] 145 compared the performance of VGG16 and gradient boosting 146 classifier model in 4.5 × 4.5 mm 2 of radial peripapillary 147 capillary vessel density layer of optic nerve head image. 148 The weight in the first 4 convolution blocks were frozen 149 also implemented other deep learning models and found to be 153 better than gradient boosting classifier.

154
In this paper, we applies CNN to diagnose glaucoma from 155 OCTA images. We hypothesized that the prominent of reti- modalities. We found only one public OCTA database [44]. 163 However, optic nerve head images are not routinely included.

164
In this paper, we proposes GlauNet to capture the glaucoma 165 features from limited data. The architecture of GlauNet was 166 based on our previous researches [45], [46] which shows that   To set the ground truth, three glaucoma specialists (SC, 194 KR, NU) independently graded optic disc photographs in 195 conjunction with OCT and/or visual field results and made 196 a diagnosis to be normal, glaucoma suspect or glaucoma. 197 The criteria for glaucoma (GL) diagnosis were adapted from 198 Li, et al. [32]. Normal visual field was determined by glau-199 coma hemifield test within normal limits and and pattern stan-200 dard deviation more than 5%. The final diagnosis depended 201 upon majority vote. If the majority vote is not reached, 202 a senior expert (AM) independently reviewed the investiga-203 tion results and made a diagnosis. The normal and suspected 204 eyes were grouped into non-glaucoma (NG) group. The data 205 were divided into training and test datasets. The demograph-206 ics of the training and test datasets are shown in Table 1.   Its input image, I input , is an 307 × 307 × 3 OCTA images as 234 shown in Fig. 3(a).
where z t is the pre-activation output of layer t; h t is the output 241 of layer t; * is the discrete convolution operator; W t is the 242 learnable n × n parameters of layer t and can be considered  ReLU

247
The maximum pooling layers (MP) seeks the strongest 248 response inside the s × s windows and can be formulated 249 as follows.
where h t xy is the output of layer t at (x, y) and Z t−1 (x+i)(y+j) is the 252 output of ReLU unit in the previous layer.
The classification section consists of five fully connected 278 layers (FC). All outputs in the previous layer are linearly 279 convoluted as follows.  Low-quality images were excluded from the training of 304 GlauNet as well as for the evaluation of the best architec-305 ture and hyperparameters. However, a subset of low-quality 306 images that contained optic nerve head and macula was used 307 to test for the robustness against artifacts.  The accuracy of the GlauNet with 5-fold cross validation for 365 the training set varied between 85.30% and 90.11%. For the 366 test set, the accuracy ranged from 79.86% to 87.05%. One 367 example of the accuracy and loss graph was shown in Fig. 4. 368 In our experiment, only GlauNet was capable of differentiat-369 ing glaucomatous and normal eyes. VGG16, ResNet50 and 370 EfficientNetV2 failed to learn and classified images either as 371 all normal or all glaucomatous eyes.

372
The visualization of the feature extraction in GlauNet was 373 shown in Fig. 3. The most highlighted area was the optic disc 374 and superotemporal and inferotemporal area. Yellow arrow in 375 Fig. 3(a:left) show the area with a decrease in vessel density. 376 The macula was highlighted only in the first convolutional 377 layer and became less prominent afterwards. In glaucomatous 378 eye, the loss of retinal vessel density was highlighted in the 379 last convolution layer, as shown inside the yellow rectangle 380 in Fig. 3(d:left). This finding was correlated with clinical 381 observation that found reduction of retinal vessel density 382 in the affected area (between arrows in Fig. 3(a:left)). This 383 finding was not presented in nonglaucomatous eye with the 384 same filter, as shown in Fig. 3(d:right).

386
The test set was augmented in the same manner as the training 387 set in order to evaluate the robustness against the change in 388 camera setting. The performance of GlauNet for the test set 389 and the augmented test set was shown in Table 2. There was 390 only a slight difference between the performance in these 391 two test sets, so it can be concluded that GlauNet was robust 392 against the change in camera setting. In the test set, the 393 classification by other three CNNs were poor. All eyes were 394 classified as glaucoma.

395
The predictive values in Table 2 can be used to explain 396 the probability of getting the correct diagnosis. GlauNet had 397 93.5% negative predictive value. This meant that if an image 398 was classified as non-glaucomatous, the probability of the 399 image being non-glaucomatous would be more than 90%. 400 The likelihood ratio of positive test (LR+) of 8.06 implied 401 that the probability of the image being glaucomatous was 402 moderately increased when the input was classified as glau-403 coma. Furthermore, the likelihood ratio of negative test 404 (LR -) of 0.12 meant that the probability of the image being 405 non-glaucomatous was moderately decreased when the input 406 was classified as glaucoma.  GlauNet was also tested for the robustness against OCTA 412 artifacts. Poor-quality images that still contained optic nerve 413 head and macula were used as the testing images. Out of 414  167 poor quality images, 88 images met our criterion. There 415 were 17 GL and 71 NG eyes. The result was shown in Table 2.

416
Though there was some drop in performance, but the sensitiv-417 ity, specificity and accuracy were still more than 80%. Exam-  set and 75 eyes in the test set. Only our proposed GlauNet was 430 capable of differentiating between glaucomatous and non-431 glaucomatous eyes. The moderate high likelihood ratio value 432 of GlauNet indicated that it was applicable to large dataset 433 (general population).

434
VGG16 and ResNet50 with transfer learning have been 435 applied to OCTA images to detect retinal vein occlusion [40], 436 diabetic retinopathy [41], [42] and glaucoma [43]. They pro-437 vided excellent results. However, pre-trained networks did 438 not work well with our data. The primary reason would be 439 a limited dataset. To deal with this issue, previous reports 440 modified the training steps to fit their data [40], [41], [42], 441 [43]. Another possible explanation was the distinct difference 442 between normal eyes and eyes with diabetic retinopathy or 443 retinal vein occlusion. The first character was the broadening 444 of white area with distinct border, such as retinal vessel 445 thickening and microaneurysm in diabetic retinopathy. The 446 second character was the black area inside the white retina 447 area. The black area was the result of the capillary loss in 448 diabetic retinopathy and non-perfusion area in retinal vein 449 occlusion [40], [42]. Both characters are prominent and easy 450 to identify, whereas the change of retinal vessels in a glau-451 coma eye is gradual and sometimes has ill-defined border.

452
Recently, Bowd, et, al. [43] applied modified VGG16 to 453 the peripapillary OCTA images and found the areas under 454 precision-recall curves of 0.97. In this study, we used wider 455 field of OCTA image that included retinal and capillary ves-456 sels outside the peripapillary area because the correlation 457 between foveal avascular zone and glaucoma was reported 458 in [22], [23], [24]. In addition, because the prevalence of 459 poor image quality was high [47], [48], [49], wider field of 460 view that includes more clinical important area may provide 461 more data for CNN to learn and predict in less severe case 462 of poor quality image. We evaluated the generalization of 463 VOLUME 10, 2022  Table 2.  [32]. In this study, 477 we did not exclude other eye diseases or patients who under-478 went intraocular surgery to make the model more generalized.

479
However, to be included in the study, the patients need to 480 have OCTA, and optic disc photograph within 6 months. GlauNet was proposed for glaucoma diagnosis based on 501 an OCTA image. It was trainable with small dataset and 502 robust against different camera projection setting as well as 503 image artifacts. The robustness can be further improved if 504 actual data with more variation are used for training. Varia-505 tion comes from both patients (ethnicities, severity, etc.) and 506 OCTA instruments (manufacturer, setting, etc.). Currently, 507 the input of GlauNet is retinal vessels in all layers of the 508 OCTA image. However, some layers may be more correlated 509 to glaucoma than the others. The correlation between the 510 specific layer and glaucoma should be further investigated for 511 both the understanding of the disease and better classification.  in Ophthalmology. In addition, she is the current Deputy Secretary General 813 of Asia Pacific Glaucoma Society. She also carries a lot of experiences in 814 many fields. She has been doing, mentoring and participating in glaucoma 815 diagnosis and therapeutic researches both locally and internationally. Under 816 her advisement, one of the advisees was granted Prince Mahidol Youth 817 Award, the renowned prestigious fund for medical student project. She 818 has been entrusted to lead the organizing team of numerous national and 819 international scientific meetings in which acclaimed at high success and 820 being invited to speak and chair sessions in various national and international 821 scientific conferences.