Retinal detachment (RD) occurs when the neurosensory retina separates from the retinal pigment epithelium.[1] In general, the annual incidence of RD is 12 per 100 000.[2] Delayed treatment may lead to irreversible visual loss. The primary method to reattach the retina is surgical treatment.[3] The effect of retinal reattachment surgery is mainly evaluated by the patient's visual and anatomic outcomes.[4] Although RD surgery is well developed, up to 10% of all cases require additional interventions to ultimately repair recurrent detachments.[5] The overall functional prognosis is unfavorable in recurrent RDs so that the early detection of recurrent detachments is of great significance.[6]
Postoperative evaluation of RD surgery requires experienced ophthalmologists to examine the whole retina through a dilated pupil. Compared with the detection for RD, it is more difficult to detect recurrent RD. This is because RD repairment usually requires varied treatments, which may cause varied iatrogenic lesions, such as laser spots and retinal degeneration. Repeated RD may also cause some retinal lesion such as retinal hemorrhage and retinal hole. These lesions make it more difficult to detect recurrent RD. In short, manual detection is time-consuming and labor-intensive, especially in remote area with few ophthalmologists. Recently, the combination of artificial intelligence and fundus photographs has a wide application prospect in various ocular disease, such as diabetic retinopathy,[7] papilledema,[8] and glaucomatous optic neuropathy.[9] Studies have proved the feasibility of artificial intelligence in the diagnosis of RD.[10] At present, however, there is no studies on detection of recurrent RD using artificial intelligence combined with ultra-widefield (UWF) imaging.
In this study, we aimed to develop a deep learning (DL) system using UWF fundus images to automatically detect recurrent RD after retinal reattachment surgery.
Experimental Section Study Design and ParticipantsThis study was performed in Beijing Tongren Hospital (Beijing, China). In the first step, UWF images were retrospectively collected between 6 April 2020 and 11 March 2021 using a nonmydriatic camera Clarus 500 (Zeiss Clarus, Carl Zeiss, Germany). The visible scope of each fundus image was 133 degree, and two to eight optic fundus images centered around the optic disk of each eye were captured. The number of fundus images captured per eye depends on the location and the extent of the RD. The extent and location of RD can vary. Partial RD can be captured by camera with one or two fundus images. As a result, the fundus images from the same eye may be classified into different groups. Extensive RD needs more fundus images (>2) to show the extent of RD and location of retinal tear. In order to improve the generalization of the DL system, we captured several fundus images each time showing different parts of the fundus (Figure 1). These data were randomly divided into two independent datasets with fourfold cross-validation:[11–14] a development dataset and an internal validation dataset. At the same time, we adopted subject independent manner to carry out cross-validation, all the images of one patient will only be split into training or testing dataset. We collected the clinical data including the basic demographic information and operation procedure of the included patients. Various examination results, such as color Doppler imaging, optical coherence tomography (OCT), and ophthalmoscope, were also collected. Clinical data include demographic data, types of RD, surgical procedure, and perioperative management. Surgical procedures include pars plana vitrectomy (PPV) and scleral buckling (SB). For patients undergoing multiple operations, time after operation is calculated based on the date of the last operation.
Enlarge this image.Figure 1. Examples of ultra-widefield fundus images from the same eye. A) An extensive RD. Four fundus images were captured. Images show the nasal (A1), temporal (A2), superior (A3), and inferior (A4) retina, respectively; B) A partial RD. Two images show the temporal (B1) and nasal (B2) retina. The B1 was labeled recurrent RD and the B2 was labeled retinal reattachment; C) A localized retinal detachment occurs in the macular area. Two fundus images were captured. Both C1 and C2 show the same RD. However, C1 and C2 mainly show the nasal and temporal retina, respectively. The C2 not only shows the macular hole RD but also shows the chorioretinal atrophy and hyperpigmentation.
In the second step, to further evaluate the applicability of the DL system, patients who undergone retinal reattachment surgery from 1 June 2021 to 1 September 2021 in Beijing Tongren Hospital were prospectively collected as a prospective dataset. Written informed consent was obtained from each subject.
This study was conducted in accordance with the Declaration of Helsinki, and the Ethics Committee of Beijing Tongren Hospital approved the study. The sample images are shown in Figure 2.
Enlarge this image.Figure 2. Examples of ultra-widefield fundus images. A) normal fundus image; B) recurrent retinal detachment; C) retinal reattachment after pars plana vitrectomy; D) retinal reattachment after scleral buckling; E) retinal reattachment after pars plana vitrectomy and scleral buckling.
Two experienced retinal specialists with at least 15 years of clinical experience screened all UWF images independently and classified them into two groups: recurrent RD after retinal reattachment surgery and retina reattachment after retinal reattachment surgery (Figure 2). The diagnosis of retina reattachment is based on the result of UWF images, color Doppler images, OCT, and indirect ophthalmoscope. Any disagreement between the two human graders was further diagnosed by another senior retinal specialist. The images were excluded if all three human graders gave different diagnoses.
Image Preprocessing and Quality ControlTo improve the DL analysis, we resized the images to a resolution of 512 × 512 pixels before developing the algorithm. In the quality control process, we assessed the image quality and filtered out unqualified images after mask removal. Images with poor quality were excluded based on several arguments, such as the readable region ratio, illumination, blurriness, and image contents. The pixel values of the selected images were applied to a linear mapping with a pixel value ranging from (0, 255) to (0, 1).
Algorithm DevelopmentWe applied a convolutional neural network to automatically detect recurrent RD. We first compared the performance of some architectures including ResNet-50, ResNet-101, Inception-V3, Inception-V4, and Inception-ResNet-V2.[15–17] ResNet-50 (Model1) and Inception-ResNet-V2 (Model 2), which achieve better performance, were finally chosen to complete the task. We adopted fourfold cross-validation to develop the models and selected the optimal one. To further test the performance of the DL models, we then prospectively validated the datasets. All models were developed with Tensorflow 1.10.0 and Keras 2.2.4 on the server with three NVIDIA 1080 GPUs (Graphical Processing Units) and were pretrained with imagenet dataset.[18] In addition, the training dataset is unbalanced so that we adopted class weight policy in the training process.
Comparison Between Human and DL SystemTwo senior ophthalmologists (with more than 15 years clinical experience), two junior ophthalmologists (with more than 5 years clinical experience), and two medical students (with more than 1-year clinical experience) were invited to independently diagnose the recurrent RD in the prospective validation dataset. The results of the DL system and definite diagnosis were not available to any human doctors. We compared the performance between these human ophthalmologists with the DL system.
Statistical AnalysesAll statistical analyses were performed using Python 3.7.3 (Wilmington, DE, USA) and MATLAB R2016a (https://www.mathworks.com/). We used the accuracy, sensitivity, specificity, and receiver operating characteristic curve to assess the performance of the DL model. The area under the curve (AUC) with 95% confidence interval (CI) was calculated.
ResultsAs shown in Figure 3, after deleting 32 poor-quality images, 554 qualified UWF images from 173 patients (mean age 39.2 ± 16.2 years, 66.5% male) were retrospectively gathered for the training, tuning, and internal validation datasets of the DL system (Table 1). Four hundred forty-four UWF images from 138 patients (mean age 39.1 ± 15.6 years, 70.3% male) were labeled as retinal reattachment after retinal reattachment surgery, while 106 UWF images from 35 patients (mean age 39.6 ± 18.6 years, 51.4% male) were labeled as recurrent RD. Rhegmatogenous retinal detachment (RRD) attributed to the most cases (156, 90.2%). Surgical procedures include PPV, SB, and PPV + SB, accounting for 27.2%, 57.8%, and 15%, respectively. The average time after operation (duration from the day of surgery to capturing fundus images) were 8.8 ± 10.7 months. Eighty-nine qualified UWF images from 23 patients (mean age 31.4 ± 12.3 years, 65.2% male) who have had retinal reattachment surgery were collected as the prospective validation datasets. In these UWF images, 35 UWF images from 7 patients (mean age 24.6 ± 10.6 years, 71.4% male) were labeled as recurrent detachment and 55 UWF images from 16 patients (mean age 34.4 ± 12.1 years, 62.5% male) were labeled as retinal reattachment. All patients from the prospective validation dataset had an operation because of RRD, among whom 6 (26.1%) underwent PPV, 13 (56.5%) underwent SB, and 4 (17.4%) underwent PPV + SB.
Enlarge this image.Figure 3. The workflow of developing deep learning (DL) systems for identifying recurrent retinal detachment (RD) based on ultra-widefield fundus images.
Table 1 Demographics and characteristics of the datasets
| Characteristic | Development dataset and internal validation dataset | Prospective validation dataset |
| Total number of qualified images | 554 | 89 |
| Number of individuals | 173 | 23 |
| Number of men [%] | 115 (66.5) | 15 (65.2) |
| Age [mean ± SD, y] | 39.2 ± 16.2 | 31.4 ± 12.3 |
| Types of RD | ||
|
Number of rhegmatogenous RD [%] Number of exudative RD [%] Number of tractional RD [%] |
156(90.2) 8 (4.6) 9 (5.2) |
23 (100) 0 (0) 0 (0) |
| Surgical procedure | ||
|
PPV [%] Scleral buckling [%] PPV + SB [%] |
47 (27.2) 100 (57.8) 26 (15.0) |
6 (26.1) 13 (56.5) 4 (17.4) |
| Time after operation [mean ± SD, month] | 8.8 ± 10.7 | 5.6 ± 5.3 |
| Retinal reattachmenta) | 448 (80.9) | 35 (39.3) |
| Recurrent RDa) | 106 (19.1) | 54 (60.7) |
a)Data are presented as the number of images (percentage in total number of qualified images, %). SD, standard deviation; RD, retinal detachment; PPV: pars plana vitrectomy; SB: scleral buckling.
The performance of DL models for detecting recurrent RD is shown in Table 2. In the internal cross-validation, the average accuracy of model 1 and model 2 were 91.2% and 90.6%, respectively. The mean AUCs of model 1 and model 2 were 0.935 and 0.944, respectively. The average sensitivity and specificity of model 1 were 84.0% and 92.9%, respectively, and the average sensitivity and specificity of model 2 were 89.0% and 87.8%, respectively. The highest AUCs of model 1 and model 2 in the internal cross-validation were 0.969 and 0.953, respectively (Figure 4). For the prospective validation dataset, the average accuracy, sensitivity, and specificity were 84.4%, 94.3%, 78.2%, respectively, for mode 1, and 85.6%, 83.6%, and 88.6%, respectively, for model 2. The AUCs of model 1 and model 2 in the prospective validation were 0.929 and 0.930, respectively (Figure 5).
Table 2 Performance of the algorithms in different validation datasets
| Dataseta) | Accuracy [95% CI] | Sensitivity [95% CI] | Specificity [95% CI] | AUC [95% CI] |
| Cross-validationb) | ||||
| Model 1 | 91.2% (85.5–96.8) | 84.0% (73.6–94.5) | 92.9% (88.3–97.4) | 93.5% (88.8–98.2) |
| Model 2 | 90.6% (81.8–99.5) | 89.0% (77.8–100) | 87.8% (74.2–100) | 94.4% (93.1–95.2) |
| Prospective validation | ||||
| Model 1 | 84.4% | 94.3% | 78.2% | 92.9% |
| Model 2 | 85.6% | 83.6% | 88.6% | 93.0% |
a)Model 1: ResNet-50; model 2: Inception-ResNet-V2.
b)Data are presented as the mean value of the AUC in fourfold internal cross-validation. CI, confidence interval.
Enlarge this image.Figure 4. Performance of models in the internal cross-validation. ResNet-50, model 1; Inception-ResNet-V2, model 2.
Enlarge this image.Figure 5. Performance of models in prospective validation dataset and comparison with human ophthalmologists. ResNet-50, model 1; Inception-ResNet-V2, model 2; Senior Ophthalmologist 1 is the red diamond; Senior Ophthalmologist 2 is the blue diamond; Junior Ophthalmologist 1 is the red dot; Junior Ophthalmologist 2 is the blue dot; Medical student 1 is the purple dot; Medical student 2 is the red triangle.
When comparing the performance between human ophthalmologists and the DL system, we found that DL system performed slightly inferior to the senior ophthalmologists. However, we also found that the DL system reached a similar and even better diagnostic performance than junior ophthalmologists, and the algorithms performed much better than medical students (Figure 5).
DiscussionOver the past few years, artificial intelligence, especially DL system, has played a major role in the field of medicine, including image recognition,[19] auxiliary diagnosis,[20] and drug development.[21] The use of fundus photography combined with artificial intelligence is an appealing means to detect retinal diseases.[8,22–27] In this study, we developed a DL system to detect recurrent RD after retinal reattachment surgery based on the newly developed UWF images. In this system, we developed two models based on different architectures on the Resnet-50 and Inception-ResNet-V2, respectively. This DL system shows great performance in detecting recurrent RD, reaching accuracy, sensitivity, and specificity of 91.2%, 84.0%, and 92.9%, respectively, for model 1, and 90.6%, 89.0%, 87.8%, respectively, for model 2 in the detection of recurrent RD. This is the first report that uses true color fundus images captured by Zeiss Clarus 500 combined with artificial intelligence to detect RD. In the prospective validation, our DL system diagnosed whether the retina was reattached before human experts made final diagnosis. The DL system performed similarly, and even better than junior ophthalmologists, and performed slightly inferior to the senior ophthalmologists. This is because both senior ophthalmologists are retina specialist with a great wealth of clinical experience, and our sample size is relatively small. We believe the DL system will reach a similar and even better diagnostic performance than senior ophthalmologists by training in a larger dataset in the future.
In the recent studies, the authors developed a DL system for RD detecting based on fundus photographs with 45°–50° fields[28] or UWF images captured by OPTOS system.[10] In our study, we focused on the recurrent RD and obtained the UWF images using Zeiss Clarus 500, which is a newly designed scanning laser ophthalmoscope that can obtain 133-degree field of the retina in a single image without mydriasis. Zeiss Clarus 500 has advantages of partially confocal optics with high resolution of 7.3 μm, high image quality, and true color imaging, which are not available on the OPTOS. The partially confocal optics reduces artifacts due to eyelashes and eyelids in retinal images. In some cases, the retina could not be examined by the OPTOS due to the presence of artifacts caused by retinal hemorrhage, but it can be detected by the Zeiss Clarus 500.[29] OPTOS images is a pseudo-color imaging system with direct bearing on the identification of lesions with different colors, which may lead to lesion missing.[30] The image captured by Zeiss Clarus 500 formed by the combination of red, green, and blue provides a true color fundus image. Various of lesions in the postoperative retina can be represented more clearly using true color imaging.[31] Compared with the 14-micron resolution of OPTOS, Zeiss Clarus 500 has a widefield retinal camera with a 7-micron resolution that can detect smaller retinal lesions. Although the visible scope of each fundus images was 133°, we can capture fundus with a 200° retinal view by making patients look in different directions. We can detect the lesion in most areas of fundus, including the peripheral retina, through 4–8 images. Relying on the features of partially confocal optics and true color imaging, Zeiss Clarus 500 will have widely application in detecting retinal lesion. It should be noted that other retinal lesions, such as retinal degeneration and retinal break, can often be found in our UWF images (Figure 6). Our system can detect the RD from these lesions. Some common retinal lesions in our datasets are listed in Table 3.
Enlarge this image.Figure 6. Some common retinal lesions in our UWF images. A,B) chorioretinal atrophy and hyperpigmentation; C) hyperpigmentation; D) retinal break and laser spots; E,F) laser spots and proliferative vitreoretinopathy; G) retinal hemorrhage; H) laser spots.
Table 3 Retinal lesions in our datasets
| Retinal lesions | Development dataset and internal validation dataset | Prospective validation dataset |
| Total number of qualified images | 554 | 89 |
| Chorioretinal atrophya) | 92 (16.6) | 18 (20.2) |
| Proliferative vitreoretinopathya) | 102 (18.4) | 27 (30.3) |
| Laser spots a) | 123 (22.2) | 21 (23.6) |
|
Retinal breaka) Retinal hemorrhagea) Hyperpigmentationa) |
24 (4.3) 19 (3.4) 66 (11.9) |
7 (7.9) 4 (2.1) 9 (10.1) |
a)Data are presented as the number of images (percentage in total number of qualified images, [%]).
With the advancement in the surgical techniques of RD repairment in the last few decades, recent studies have reported the retinal reattachment rate of more than 80–90% after surgery.[5] However, the increasing number of patients has brought a greater number of recurrences. Early detection of recurrent RD is crucial. Proliferative vitreoretinopathy (PVR) may be secondary to the recurrent RD in a short time, which may cause irreversible visual impairment.[32] Early recurrent RD is defined as detachments occurring within the first 6 weeks postoperatively, while late recurrent RD is defined as detachments occurring 6 or more weeks postoperatively.[33] The severity of PVR may be determined by the duration of recurrent RD, and early reoperation for recurrent RD can successfully reattach the retina and improve visual acuity in most cases.[32,34] As a result, the follow-up visits for patients after surgical repair are of great significance. Postoperative evaluation of RD surgery requires experienced ophthalmologists to examine the whole retina through a dilated fundus after mydriasis. However, the patients who have surgery in our hospital come from all over the country including many rural areas lacking of ophthalmologist. As a new type of UWF fundus imaging system, the market share of Zeiss Clarus 500 in China is relatively small. However, with health spending growing in China, it is growing rapidly in recent years. Nevertheless, the number of experienced ophthalmologists in remote area cannot be increased in a short time by increasing health spending. With the popularization of Zeiss Clarus 500 in the future, our DL system has the potential to bring great changes to the follow-up of postoperative patients in remote area. RRD is the most common RD, and the incidence of RRD is increasing in recent years.[35,36] It is estimated that there are 9000 to 10 000 new cases of RRD in China each year.[37] In this study, RRD attributed to the most cases, which is in accord with epidemiologic characteristics of RRD. As 60% to 70% of the RRD patients live in small towns or remote farming areas with limited access to qualified retina-vitreous surgeons, care for these patients presents a great challenge in China.[37] Our system will significantly improve the RRD patients’ ability to be followed up with and accessibility to medical resources. But it may also hinder the generalization performance of the models. Next, we plan to verify the performance of models in more balanced dataset.
Several limitations of this study should be noted. All included patients were collected from the same ophthalmic center, which did not cover patients from multiple ethnic and regional populations. In addition, because the Zeiss Clarus 500 was just on the market for several years as a new technology and the sample size was relatively small in this study, prospective studies with a larger sample size and comparing varying ethnic groups are required to further evaluate the performance of the models. Finally, it should be emphasized that our system only detects RD, and we will upgrade the system so that it can detect varied fundus lesion in the future.
ConclusionIn conclusion, the DL-based system performed well for detecting recurrent RD. Further studies are currently ongoing, which will integrate our DL system into the UWF camera and evaluate its real-word performance.
AcknowledgementsThis study was supported by the Capital Health Research and Development of Special (2020-1-2052), Science & Technology Project of Beijing Municipal Science & Technology Commission (Z201100005520045, Z181100001818003).
Conflict of InterestThe authors declare no conflict of interest.
Author ContributionsW-D.Z. and L.D. contributed equally to this work. Concept and design: W-D.Z., L.D., W-B.W. Acquisition, analysis, or interpretation of data: W-D.Z., L.D., K.Z., Q.Y., Y-M.L., L-J.F., X-H.S., C.Z., R-H.Z., H-Y.L., H-T.W., and W-B.W. Critical revision of the manuscript for important intellectual content: All authors.
Data Availability StatementThe data that support the findings of this study are available from the corresponding author upon reasonable request.
© 2022. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; Wang, Qian 1 ; Shao, Lei 1 ; Yang, Qiong 1 ; Yue-Ming, Liu 1 ; Li-Jian, Fang 3 ; Xu-Han, Shi 1 ; Zhang, Chuan 1 ; Rui-Heng, Zhang 1 ; He-Yan, Li 1 ; Hao-Tian, Wu 1
; Wen-Bin, Wei 1 1 Beijing Tongren Eye Centre, Beijing Key Laboratory of Intraocular Tumour Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Tongren Hospital, Capital Medical University, Beijing, China
2 InferVision Healthcare Science and Technology Limited Company, Shanghai, China
3 Beijing Tongren Eye Centre, Beijing Key Laboratory of Intraocular Tumour Diagnosis and Treatment, Beijing Ophthalmology & Visual Sciences Key Lab, Medical Artificial Intelligence Research and Verification Key Laboratory of the Ministry of Industry and Information Technology, Beijing Tongren Hospital, Capital Medical University, Beijing, China; Department of Ophthalmology, Beijing Liangxiang Hospital, Capital Medical University, Beijing, China