Abstract
This study aims to propose a generative deep learning model (GDLM) based on a variational autoencoder that predicts macular optical coherence tomography (OCT) images following full-thickness macular hole (FTMH) surgery and evaluate its clinical accuracy. Preoperative and 6-month postoperative swept-source OCT data were collected from 150 patients with successfully closed FTMH using 6 × 6 mm2 macular volume scan datasets. Randomly selected and augmented 120,000 training and 5000 validation pairs of OCT images were used to train the GDLM. We assessed the accuracy and F1 score of concordance for neurosensory retinal areas, performed Bland–Altman analysis of foveolar height (FH) and mean foveal thickness (MFT), and predicted postoperative external limiting membrane (ELM) and ellipsoid zone (EZ) restoration accuracy between artificial intelligence (AI)-OCT and ground truth (GT)-OCT images. Accuracy and F1 scores were 94.7% and 0.891, respectively. Average FH (228.2 vs. 233.4 μm, P = 0.587) and MFT (271.4 vs. 273.3 μm, P = 0.819) were similar between AI- and GT-OCT images, within 30.0% differences of 95% limits of agreement. ELM and EZ recovery prediction accuracy was 88.0% and 92.0%, respectively. The proposed GDLM accurately predicted macular OCT images following FTMH surgery, aiding patient and surgeon understanding of postoperative macular features.
Similar content being viewed by others
Introduction
Full-thickness macular hole (FTMH) is a retinal condition of tractional disruption to the whole foveola, including the Müller cell cone and external limiting membrane (ELM)1,2. Idiopathic FTMH is caused by traction to the fovea during posterior hyaloid detachment. Vitrectomy, internal limiting membrane (ILM) peeling, and fluid-air exchange with tamponade can close the macular hole (MH)3. For larger holes, additional procedures for ILM can be performed to increase the closure rate4,5. Recently, the closure rate of idiopathic FTMH has reached as high as 96–98% after primary vitrectomy with ILM peeling6,7,8. Even if the first ILM flap technique fails to close large holes, the closure rate can reach 100% through reoperation9. With ongoing efforts of retinal surgeons for hole closure, most idiopathic FTMH cases can be closed. Therefore, the focus of FTMH surgery has shifted from predicting the hole closure to understanding how macular holes close morphologically and their association with visual outcomes10.
Visual acuity (VA) in patients with FTMH is primarily associated with the outer retinal conditions, particularly the ELM and ellipsoid zone (EZ) status. Following a successful closure through vitrectomy, the ELM and EZ show gradual recovery on optical coherence tomography (OCT), leading to improved VA11,12. Various factors extractable from preoperative OCT influence ELM and EZ recovery. Lee et al. revealed that axial length, minimum linear diameter, and preoperative intraretinal layer thickness are related to EZ recovery13. However, measuring these diverse factors in clinics manually and inputting them into a prediction model may be impractical due to the substantial time and effort.
Deep learning (DL) models have been proposed to identify risk factors or assess retinal disease prognoses from OCT images14. Several DL models have been proposed for the prognosis of FTMH surgery. These models utilize a combination of preoperative OCT and clinical data to predict the postoperative hole closure15, as well as preoperative OCT to estimate postoperative VA16. However, previous DL models for FTMH could not predict the detailed anatomical conditions of the macula, including the ELM and EZ restoration.
Generative deep learning models (GDLMs) may be suitable for constructing expected postoperative OCT images based on preoperative OCT images in eyes with FTMH. GDLMs are divided into the following categories: variational autoencoders (VAEs) and generative adversarial networks (GANs)17. VAEs are widely utilized for image-to-image conversion and super-resolution and have shown remarkable outcomes in producing high-quality images in computer vision and art17. Because GANs are known to have unstable training and are prone to spatial deformity and mode collapse phenomenon18, we selected VAE as the preferred GDLM for a reliable prediction of the postsurgical macular anatomy19.
We herein proposed a GDLM based on VAE to predict postoperative OCT images using preoperative OCT images following FTMH surgery. Further, we evaluated the accuracy and performance of the GDLM by comparing the predicted and actual postoperative OCT images, including concordance for neurosensory retinal areas and foveal anatomy, as well as ELM and EZ restoration.
Methods
Patient selection and ethics statements
We consecutively enrolled patients with FTMH who underwent pars plana vitrectomy, ILM peeling, and fluid-air exchange between January 2018 and December 2022. The selection of tamponade materials (air or SF6 18%), ILM staining dyes (0.025% brilliant blue G or 0.05% indocyanine green), and additional ILM procedures (inverted ILM flap or autologous ILM insertion) were left to the surgeon’s discretion. The inclusion criteria were FTMH cases identified through macular volume scans using a swept-source OCT device, the DRI OCT-1 Atlantis (Topcon Corp., Tokyo, Japan), before and 6 months after surgery. The eyes were scanned using a 6 × 6 mm2 scan protocol centered on the fovea by an experienced technician. The study protocol was approved by the institutional review boards of the Pusan National University Hospital (PNUH, approval no. 2304-016-126) and Pusan National University Yangsan Hospital (PNUYH, approval no. 05-2023-112). One hundred twenty-five patients who had met the inclusion criteria were excluded from the study based on the exclusion criteria summarized in Supplementary Table S1.
Data collection
Baseline parameters (age, sex, laterality, axial length, best corrected visual acuity [BCVA], central subfield macular thickness [CSMT], hole size, and FTMH stage based on the Gass classification20), intraoperative parameters (combined phacoemulsification, ILM peeled area in disc diameter, ILM manipulation technique, surgeon, ILM staining dye, and tamponade material), and 6-month postoperative parameters (BCVA, CSMT, ELM restoration, and EZ recovery) were assessed. The hole size was measured using the ImageNet 6 ver. 1.24 software (Topcon Corp., Tokyo, Japan) by determining the longest distance between the split ends of the ELM21. Successful ELM restoration was defined as a continuous ELM line that was clearly distinguishable between the outer nuclear layer and photoreceptor (PRL) in the fovea. Successful EZ recovery defines the continuous bright band between ELM and RPE in postoperative OCT.
Automatic registration and preparation of macular volume scan datasets
The macular volume OCT scan contains 256 continuous slices, each consisting of 992 vertical and 512 horizontal pixels of an 8-bit grayscale image, stored in a three-dimensional (3D) array using the code distributed by Graham (Graham, M. 2020. OCT-Converter. Version v0.5.0. https://github.com/marksgraham/OCT-Converter. Accessed 7 March 2023). It is evident that the retinal vasculature and retina itself can change after FTMH surgery22. Choroid may also vary post-vitrectomy in patients with vitreous traction23. Hence, we based our registration on the RPE surface, which was expected to undergo minimal changes and was not manipulated during surgery (Fig. 1).
For supervised learning the pairs of preoperative and postoperative slices were prepared according to the following sequence: An image with a resolution of 448 × 448 pixels was cropped, centered on the center of mass, and then resized by 50%. Only the 200 slices centered on the fovea were selected. Consequently, paired OCT images of 224 × 224 pixels (vertical 1164.8 μm and horizontal 5250.0 μm) were extracted from the volumetric OCT dataset for each patient. Augmentation techniques were then applied to the training set. This process increased the overall training data by six-fold (Fig. 2). All postoperative OCT images were designated as ground-truth OCT (GT-OCT) images. Since the foveal morphology and macular deformation vary according to the distance from the foveola22, different conditions were adopted for each image slice (See Supplementary Figure S2).
Structure of generative deep learning model
Conditional VAE was used for the GDLM and was designed to receive preoperative OCT image slices and condition vectors. This artificial intelligence (AI) model can generate postoperative OCT (AI-OCT) image slices and comprises four units: encoder, sampler, decoder, and loss function unit. In Fig. 3, these structures are explained in detail. The loss function unit calculated the difference between the AI-OCT images and the GT-OCT images and updated the weights of the GDLM to reduce this difference. In the late stage of the training process, perceptual loss functions were employed to generate fine AI-OCT images (See Supplementary Figure S3)24,25. The training process ended with 600 epochs and stopped if overfitting was detected. The epoch with the smallest loss was determined using the validation sets, and the weights of this epoch were loaded into the GDLM and used to evaluate the test set.
Verification for accuracy and validity of generative deep learning model
For quantitative assessment, a representative pair of preoperative and postoperative cross-sectional OCT slices was selected from each test volumetric OCT dataset using the following steps: (1) The RPE surface from both preoperative and postoperative volumetric OCT data were extracted in each case within the test set. (2) Similar to the training set, we aligned the preoperative 3D volumetric OCT data with the RPE surface of the postoperative volumetric OCT through rotation and translation for registration. (3) The 128th OCT slice from the registered preoperative volumetric OCT was fed into the GDLM. (4) The 128th slice from the postoperative volumetric OCT was chosen as the GT-OCT slice. (5) The GDLM's output was designated the postoperative predicted OCT (i.e., AI-OCT) slice.
The quality of the AI-OCT image slices was assessed using three metrics: (1) image quality score, (2) agreement of the generated neurosensory retinal area, and 3) the learned perceptual image patch similarity (LPIPS) score. The image quality score ranges from 0 to 10, with one point assigned for each of the 10 distinct layers—the nerve fiber, ganglion cell, inner plexiform, inner nuclear, outer plexiform, outer nuclear layer, ELM, EZ, RPE, and choroidal vasculature in the AI- and GT-OCT image slice. A score of 0 was assigned when each retinal layer was not distinguishable or absent. To independently measure the image quality score, H.J.K. separated the AI-OCT images and GT-OCT images and distributed them to S.H.P. in PNUYH and J.H. in PNUH, respectively. If the judgment of the image quality score for each slice was ambiguous, the corresponding author provided the final decision.
The accuracy, precision, recall, and F1 scores were calculated to determine the concordant area of the neurosensory retina by binarizing and comparing the AI- and GT-OCT image slices (Fig. 4A–G). Manual corrections were applied to address boundary errors, and agreement metrics were extracted from all test sets. LPIPS score is one image similarity indicator that employs pre-trained image classification networks to evaluate the similarity between two images25. The comparison between AI- and GT-OCT image slices included assessments of LPIPS scores. A value close to 0 indicates a higher perceptual similarity between the two images25.
We performed Bland–Altman analyses to assess the agreement of foveolar height (FH), mean foveal thickness (MFT), and mean nasal/temporal parafoveal thickness (MNPT/MTPT) measured in AI- and GT-OCT (Fig. 4H,I). ELM restoration and EZ recovery were analyzed by assessing the accuracy and recall metrics between the two slices. Considering that the surgical procedures of the inverted ILM flap and conventional ILM peeling have different impacts on surgical outcomes and macular morphology4, the performance of the GDLM was quantitatively tested for two surgical techniques in the test set.
Comparison of synthetic capabilities of various generative deep learning models
To our knowledge, there is yet to be any research or datasets utilizing GDLMs for synthesizing postoperative OCT images about FTMH. A few studies conduct OCT image synthesis to predict postoperative or post-therapeutic macular anatomy for other retinal diseases, such as wet age-related macular degeneration, retinal vein occlusion, and epiretinal membrane26,27,28. These studies employed GAN-based GDLMs, specifically the Pix2Pix and Pix2PixHD models. CycleGAN has been applied to address the variability of OCT images across different devices29. Among various VAE models, the nouveau VAE (NVAE) has showcased state-of-the-art performance on the MNIST dataset, evaluated by the bits per dimension metric30.
We compared the synthetic capabilities of our proposed model against various GDLMs. All GDLMs were trained up to 600 epochs using the same training dataset. We selected each model with the smallest loss, and synthetic capabilities were compared across 25 FTMH validation sets. The performance of each GDLM was evaluated with the LPIPS score, Fréchet Inception Distance (FID)31, and image quality score between postoperative AI-OCT and GT-OCT images.
Statistical analyses
To analyze the data, the BCVA was converted to logMAR scale. Pearson's chi-square or Fisher's exact test is adopted to determine dependencies between two categorical variables. Continuous variables were compared using the Kruskal–Wallis test to compare each parameter among the three sets. Differences in FH, MFT, MNPT, and MTPT between AI- and GT-OCT slices were investigated using the Wilcoxon signed-rank test. Comparisons between ILM peeling and inverted ILM flaps were performed using the Mann–Whitney U test. A P value of < 0.05 was considered statistically significant. In Bland–Altman analyses to assess the agreement of macular morphology, upper and lower bounds of 95% limits of agreement (LoA) as a percentage difference using the Python pyCompare library (ver. 1.5.4). Bland–Altman plots were generated for each parameter, and we established an acceptable threshold if the 95% limits of LoA were both within ± 30%32. Univariate and multivariate logistic regression analyses were conducted to identify the baseline and intraoperative parameters that affect ELM and EZ disruption. Furthermore, we developed logistic regression models with training sets comprising statistically significant parameters and assessed its accuracy using test sets using the Python scikit-learn library (ver. 1.0.2).
Ethics approval
Written informed consent was obtained from all participants. The study protocol and informed consent were approved by the Institutional Review Board of the PNUH and the PNUYH, and all research was conducted in accordance with the Declaration of Helsinki.
Results
Patients’ demographics
In total, 150 eyes with successfully closed FTMH met the inclusion criteria. Five surgeons performed the surgery on patients with an average age of 65.2 ± 9.0 years. The preoperative BCVA and CSMT were 0.773 ± 0.366 (20/119) and 320.5 ± 84.0 μm, respectively. Out of the total number of eyes, 47 (31.3%) were categorized as stage 2, 49 (32.7%) as stage 3, and 54 (36.0%) as stage 4 FTMHs, with a mean hole size of 448.0 ± 215.1 μm. Combined cataract surgery was conducted in 114 (76.0%) eyes. The inverted ILM flap technique was performed in 78 (52.0%) eyes, while ILM peeling in 72 eyes. The inverted ILM flap technique was performed in cases of larger hole size (538.8 ± 210.3 μm), compared to ILM peeling technique (349.7 ± 185.0 μm, P < 0.001). The postoperative BCVA and CSMT were significantly improved (0.393 ± 0.354 [20/49] and 259.6 ± 49.6 μm, respectively; P = 0.001 and P = 0.046, respectively). ELM restoration was achieved in 110 eyes (73.3%), and successful EZ recovery was observed in 101 eyes (67.3%). Eyes with complete restoration of ELM exhibited better BCVA than those without it (0.286 ± 0.267 vs. 0.688 ± 0.399, P < 0.001), and demonstrated greater BCVA improvement (0.457 ± 0.343 vs. 0.168 ± 0.222, P < 0.001). Additionally, eyes with EZ recovery exhibited better BCVA (0.289 ± 0.278 vs. 0.508 ± 0.394, P < 0.001) than those with failure of EZ recovery. All cases were randomly allocated to 100 training, 25 validations, and 25 test sets; none of the factors differed significantly across the three sets (Table 1).
Training GDLM and artificial intelligence OCT images
Of the 120,000 OCT image slice pairs in the training set, 25,800 pairs were set to condition vector 0; 25,200 pairs to condition vector 1; 25,800 pairs to condition vector 2; and 43,200 pairs to condition vector 3. As the number of epochs increased, the conditional VAE gradually reduced the loss, and the AI-OCT image slices resembled the GT-OCT image slices (See Supplementary Figure S3, S4, and Movie S5).
Comparison of various GDLMs with validation set
Proposed conditional VAE adopting LPIPS loss showed the lowest LPIPS and the highest image quality scores compared with GAN-based GDLMs (See Supplementary Table S6). Supplementary Figure S7 depicted representative cases of GT- and AI-OCT images predicted by various GDLMs. The Pix2Pix model achieved the lowest FID score, but spatial deformity was observed in specific cases. CycleGAN was excluded because synthesized OCT images are almost identical to preoperative OCT images.
Comparison of accuracy and validity between AI- and GT-OCT image slices
AI-OCT image slices from test sets had a mean image quality score of 9.80 ± 0.50, clearly distinguishing between the retinal layers and surrounding structures, and were no different from the image quality scores of GT-OCT image slices (9.96 ± 0.20). The accuracy of the agreement between AI- and GT-OCT image slices for retinal regions was 94.7 ± 2.0%, with a precision of 89.0 ± 6.6%, recall of 89.5 ± 4.9%, and F1 score of 0.891 ± 0.042. The mean LPIPS score was 0.135 ± 0.033. Eyes with ILM peeling (n = 14, 56.0%) exhibited higher accuracy (95.5% vs. 93.6%, P = 0.025), precision (91.4% vs. 85.9%, P = 0.038), and F1 scores (0.909 vs. 0.868, P = 0.018) compared to those with inverted ILM flaps (n = 11, 44.0%) (Table 2).
AI- and GT-OCT image slices revealed no statistically significant differences in the averages of FH (228.2 ± 51.8 vs. 233.4 ± 70.0 μm), MFT (271.4 ± 35.5 vs. 273.3 ± 55.7 μm), MNPT (316.6 ± 35.2 vs. 311.2 ± 34.3 μm), and MTPT (314.1 ± 32.7 vs. 309.7 ± 36.0 μm). No statistically significant difference was observed in retinal thickness between the ILM inverted flap and peeling groups (Table 2). The 95% LoA for all morphological parameters in both the test sets and the ILM peeling group remained within the range predefined by the cut-off value. However, in the case of the ILM inverted flap group, the upper 95% LoA for MFT exceeded 30% (See Supplementary Figure S8).
The logistic regression analysis revealed that a larger hole size was the sole factor that increased the risk of both ELM and EZ disruption among the baseline and intraoperative parameters. The ILM peeling group demonstrated a higher probability of EZ recovery (See Supplementary Table S9). These logistic regression models predicted ELM and EZ recovery accuracy rates at 84.0% and 76.0%, respectively.
The accuracy of ELM restoration using GDLM was 88.0%, with recall rates of 94.4% for successful restoration and 71.4% for cases of restoration failure. For the EZ recovery, the GDLM achieved an accuracy of 92.0%, with recall rates of 100.0% for successful recovery and 80.0% for failures. The accuracy of GDLM model was superior to statistical analysis in prediction for postoperative foveal microstructure. The confusion matrices for predicting the ELM and EZ restoration are summarized in Supplementary Figure S10. Fig. 5 and Supplementary Figure S11 show representative cases of AI- and GT-OCT images.
Discussion
This study introduces the GDLM for predicting postoperative macular structures based on preoperative OCT images using a modified conditional VAE architecture. The AI-OCT images effectively depicted the distinct retinal layers as well as RPE and choroid. The high accuracy and F1 score indicated strong agreement with the neurosensory retinal region between the AI- and GT-OCT images. The retinal thickness did not differ between AI- and GT-OCT images. The accuracy of restoring the ELM and EZ was 85% or higher. These findings indicated that the proposed GDLM could generate high-quality AI-OCT images, similar to the actual postoperative OCT images. The average FH and MFT of AI-OCT images were comparable to those of GT-OCT images with a difference of not more than 5.2 μm, which is within a one-pixel level of mean differences. Bland–Altman analyses of the four thickness profiles showed that the upper and lower bounds of the 95% LoA did not exceed 30%. This indicates that the predicted thickness profiles can be considered acceptable as a new technique32.
Wakabayashi et al. identified that the ELM status at 3 months was related to the BCVA 1 year after surgery12. Further, ELM recovery may be a prerequisite for the EZ recovery13. However, presenting the postoperative ELM status as a simple probability value may lack persuasiveness for patients and physicians. Postoperative AI-OCT images generated by the GDLM could provide an intuitive prediction of ELM restoration by determining the presence of a continuous bright line corresponding to the ELM in a 2D OCT image. The GDLM exhibited a significant resemblance to the GT-OCT images in eyes with ILM peeling, accurately reproducing retinal structures compared to those with the inverted ILM flap. Following MH surgery, the inverted flap group experiences more glial proliferation around the fovea than the ILM peeling group33. Consequently, such foveal proliferation results in diversity for the ILM inverted flap group12,34. This diversity can potentially hinder the accurate prediction of macular structures by the GDLM in the ILM inverted flap group. Nevertheless, the proposed model effectively predicted not only other retinal thickness profiles but also ELM and EZ restoration in eyes experienced with ILM inverted flaps. The recall rates for ELM and EZ restoration failure were at 71.4% and 85.6%, respectively. This observation can be elucidated as follows: ELM and EZ disruption present as diverse signal intensities in postoperative OCT images, potentially introducing inaccuracies into the generative model. For instance, PRL defects may exhibit low signal intensity alongside outer foveal defects or moderate hyperreflective lesions resulting from glial cell proliferation, even in the ILM peeling group35. In addition, the diminished recall rate might be attributed to the limited number of ELM and EZ disruption cases. Additional training sets of ELM and EZ disruption cases may enhance the recall rate.
We employed a conditional VAE that was trained on foveal OCT images as well as parafoveal OCT images corresponding to different conditions. The OCT training set for conditions other than condition 0, located away from the fovea, was more than three times larger than that for condition 0. The parafoveal OCT images revealed detailed retinal layers and surrounding structures. This approach enabled the GDLM to effectively represent specific retinal layers and predict various anatomical conditions for both the foveal and parafoveal regions from limited training sets. The realistic generation of parafoveal OCT images using the proposed GDLM is depicted in Supplementary Figure S11.
Studies have been conducted on the use of discriminative DL models in predicting VA or surgical outcomes in FTMH using preoperative cross-sectional OCT images. Obata et al. found the DL model to be more accurate in predicting VA than multivariate linear regression with baseline factors16. In contrast, Lachance et al. observed no significant improvement in visual prediction by adding clinical findings to a CNN-based DL model36. The use of preoperative cross-sectional OCT images in DL models influencing postoperative VA can lead to varied results. Predicting postoperative macular structures could be key to forecasting postoperative VA11,12. Xiao et al. introduced a DL model for predicting postoperative hole closure at one month using preoperative OCT images and clinical data15. This model achieved an accuracy of over 80% in predicting MH closure. However, their successful closure rate was less than 70%, significantly lower than the primary hole closure rate of over 95% reported in recent studies7,8. The prerequisite for using our GDLM model is knowledge of MH closure, which is impossible to determine before surgery. Nonetheless, since cases where idiopathic FTMH fails to close after the primary surgery are rare, these limitations will only affect a small number of patients.
In GAN-based GDLMs, spatial deformity (Supplementary Figure S7J and S7K) occurs in small training sets or spatial misalignment of image pairs18. Achieving exact structural and spatial alignment in image-to-image transformation with conditional GANs is crucial, often requiring further registration during preprocessing to ensure high-quality images37,38. However, the newly remodeled and migrated neural structure around the fovea makes accurate structural alignment impossible following FTMH surgery. Therefore, we designed a new VAE-based GDLM to predict postoperative OCT images in FTMH patients, as GAN-based GDLMs encounter these drawbacks. Considering the drawbacks of GANs, such as instability in training, mode collapse, and spatial deformity when generating images18, the stability of VAEs is more suitable for medical applications.
This study had other several limitations, including its small sample size and retrospective nature. The GDLM exclusively employed preoperative OCT data during the training phase, and did not utilize clinical factors which can affect the anatomical status of the macula after FTMH surgery. Chronicity, stage of MH, and preoperative VA are major factors affecting postoperative anatomical and functional status39,40. Incorporating these factors as supplementary condition vectors or integrating them with DL models holds the potential to enhance their performance. The proposed GDLM could not predict postoperative visual prognosis. Future research is required to develop an advanced DL model that can infer visual prognosis from predicted OCT images.
Despite these limitations, the proposed GDLM can serve as an explainable AI technique by providing predicted OCT images and anatomical profiles surrounding the macula rather than relying on simplistic measurements, such as VA or hole closure, as utilized in conventional CNN-based DL model. Further, the introduced GDLM can be applied to various fields where the anatomical status of the macula after intervention needs to be predicted through OCT images, not limited to FTMH patients but also in all retinal conditions requiring the involvement of retinal specialists. Therefore, this GDLM will serve as a significant initial step in post-intervention OCT prediction. This study is anticipated to provide valuable clinical assistance to patients and ophthalmologists in assessing postoperative prognosis. Cases that predict ELM or EZ defects in AI-OCT images may result in a poor visual prognosis. In such cases, strategies to enhance their integrity can be considered before or during surgery, such as advising early surgery to the patient or considering ILM inverted flap techniques instead of ILM peeling6,41,42.
In conclusion, the proposed GDLM demonstrated the ability to generate realistic and accurate predictions of postoperative OCT images. It successfully captured detailed retinal structures with a high degree of regional agreement and has the potential to provide valuable clinical insights by forecasting the restoration of ELM and EZ conditions closely related to postoperative prognosis. Sharing these predictive OCT images with patients scheduled for vitrectomy allows us to directly inform them about the surgical benefits and goals.
Data availability
The datasets for this study are protected patient information. Some data may be available for research purposes from the corresponding author upon reasonable request.
References
Gass, J. D. Müller cell cone, an overlooked part of the anatomy of the Fovea centralis. Arch. Ophthalmol. 117, 821–823 (1999).
Bringmann, A. et al. Different modes of full-thickness macular hole formation. Exp. Eye. Res. 202, 108393. https://doi.org/10.1016/j.exer.2020.108393 (2021).
Ando, F., Sasano, K., Ohba, N., Hirose, H. & Yasui, O. Anatomic and visual outcomes after indocyanine green-assisted peeling of the retinal internal limiting membrane in idiopathic macular hole surgery. Am. J. Ophthalmol. 137, 609–614 (2004).
Michalewska, Z., Michalewski, J., Adelman, R. A. & Nawrocki, J. Inverted internal limiting membrane flap technique for large macular holes. Ophthalmology. 117, 2018–2025 (2010).
Morizane, Y. et al. Autologous transplantation of the internal limiting membrane for refractory macular holes. Am. J. Ophthalmol. 157, 861-869.e1 (2014).
Steel, D. H. et al. Factors affecting anatomical and visual outcome after macular hole surgery: findings from a large prospective UK cohort. Eye (London) 35, 316–325 (2020).
Bae, K. et al. Extent of internal limiting membrane peeling and its impact on macular hole surgery outcomes: A randomized trial. Am. J. Ophthalmol. 169, 179–188 (2016).
Almeida, D. R., Wong, J., Belliveau, M., Rayat, J. & Gale, J. Anatomical and visual outcomes of macular hole surgery with short-duration 3-day face-down positioning. Retina 32, 506–510 (2012).
Michalewska, Z. & Nawrocki, J. Repeat surgery in failed primary vitrectomy for macular holes operated with the inverted ILM flap technique. Ophthalmic Surg. Lasers Imaging Retina 49, 611–618 (2018).
Miura, G., Mizunoya, S., Arai, M., Hayashi, M. & Yamamoto, S. Early postoperative macular morphology and functional outcomes after successful macular hole surgery. Retina 27, 165–168 (2007).
Kang, S. W., Lim, J. W., Chung, S. E. & Yi, C.-H. Outer foveolar defect after surgery for idiopathic macular hole. Am. J. Ophthalmol. 150, 551–557 (2010).
Wakabayashi, T., Fujiwara, M., Sakaguchi, H., Kusaka, S. & Oshima, Y. Foveal microstructure and visual acuity in surgically closed macular holes: spectral-domain optical coherence tomographic analysis. Ophthalmology 117, 1815–1824 (2010).
Lee, M.-W., Kim, T.-Y., Song, Y.-Y., Baek, S.-K. & Lee, Y.-H. Changes in each retinal layer and ellipsoid zone recovery after full-thickness macular hole surgery. Sci. Rep. 11, 11351. https://doi.org/10.1038/s41598-021-90955-4 (2021).
Moraru, A., Costin, D., Moraru, R. & Branisteanu, D. Artificial intelligence and deep learning in ophthalmology—present and future (review). Exp. Ther. Med. 20, 3469–3473 (2020).
Xiao, Y. et al. Development and validation of a deep learning system to classify aetiology and predict anatomical outcomes of macular hole. Br. J. Ophthalmol. 107, 109–115 (2021).
Obata, S. et al. Prediction of postoperative visual acuity after vitrectomy for macular hole using deep learning–based artificial intelligence. Graefes. Arch. Clin. Exp. Ophthalmol. 260, 1113–1123 (2021).
Bond-Taylor, S., Leach, A., Long, Y. & Willcocks, C. G. Deep generative modelling: A comparative review of VAEs, GANs, normalizing flows, energy-based and autoregressive models. IEEE. Trans. Pattern. Anal. Mach. Intell. 44, 7327–7347 (2022).
You, A., Kim, J. K., Ryu, I. H. & Yoo, T. K. Application of generative adversarial networks (GAN) for ophthalmology image domains: a survey. Eye Vis. (Lond.) 9, 6. https://doi.org/10.1186/s40662-022-00277-3 (2022).
Lee, J.-Y. & Choi, S.-I. Improvement of learning stability of generative adversarial network using variational learning. Appl. Sci. 10, 4528. https://doi.org/10.3390/app10134528 (2020).
Gass, J. D. Idiopathic senile macular hole. Its early stages and pathogenesis. Arch. Ophthalmol. 106, 629–639 (1988).
Shin, J. Y., Chu, Y. K., Hong, Y. T., Kwon, O. W. & Byeon, S. H. Determination of macular hole size in relation to individual variabilities of fovea morphology. Eye (London) 29, 1051–1059 (2015).
Park, S. H. et al. Square grid deformation analysis of the macula and postoperative metamorphopsia after macular hole surgery. Retina 41, 931–939 (2020).
Kang, E. C., Lee, K. H. & Koh, H. J. Changes in choroidal thickness after vitrectomy for epiretinal membrane combined with vitreomacular traction. Acta Ophthalmol. 95, e393–e398 (2016).
Charrier, C., Knoblauch, K., Maloney, L. T., Bovik, A. C. & Moorthy, A. K. Optimizing Multiscale SSIM for Compression via MLDS. IEEE. Trans. Image. Process. 21, 4682–4694 (2012).
Zhang, R., Isola, P., Efros, A.A., Shechtman, E. & Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. arXiv. Preprint at arXiv:1801.03924 (2018).
Liu, Y. et al. Prediction of OCT images of short-term response to anti-VEGF treatment for neovascular age-related macular degeneration using generative adversarial network. Br. J. Ophthalmol. 104, 1735–1740. https://doi.org/10.1136/bjophthalmol-2019-315338 (2020).
Xu, F. et al. Predicting OCT images of short-term response to anti-VEGF treatment for retinal vein occlusion using generative adversarial network. Front. Bioeng. Biotechnol. 10, 914964. https://doi.org/10.3389/fbioe.2022.914964 (2022).
Kim, J. & Chin, H. S. Deep learning-based prediction of the retinal structural alterations after epiretinal membrane surgery. Sci. Rep. 13, 19275. https://doi.org/10.1038/s41598-023-46063-6 (2023).
Romo-Bucheli, D. et al. Reducing image variability across OCT devices with unsupervised unpaired learning for improved segmentation of retina. Biomed. Opt. Express. 11, 346–363. https://doi.org/10.1364/BOE.379978 (2019).
Vahdat, A. & Kautz, J. NVAE: A Deep Hierarchical Variational Autoencoder. Preprint at arXiv:2007.03898 (2021).
Kynkäänniemi, T., Karras, T., Aittala, M., Aila, T. & Lehtinen, J. The Role of ImageNet Classes in Fréchet Inception Distance. arXiv preprint arXiv:2203.06026 (2023).
Critchley, L. A. & Critchley, J. A. A meta-analysis of studies using bias and precision statistics to compare cardiac output measurement techniques. J. Clin. Monit. Comput. 15, 85–91 (1999).
Liu, L. et al. Microstructural and microperimetric comparison of internal limiting membrane peeling and insertion in large idiopathic macular hole. BMC Ophthalmol. 23, 274. https://doi.org/10.1186/s12886-023-03006-z (2023).
Caprani, S. M. et al. Macular hole surgery: The healing process of outer retinal layers to visual acuity recovery. Eur J Ophthalmol. 27, 235–239. https://doi.org/10.5301/ejo.5000905 (2017).
Ko, T. H. et al. Ultrahigh-resolution optical coherence tomography of surgically closed macular holes. Arch. Ophthalmol. 124, 827–836 (2006).
Lachance, A. et al. Predicting visual improvement after macular hole surgery: a combined model using deep learning and clinical features. Transl. Vis. Sci. Technol. 11, 6. https://doi.org/10.1167/tvst.11.4.6 (2022).
Shin, Y., Yang, J. & Lee, Y. H. Deep generative adversarial networks: Applications in musculoskeletal imaging. Radiol. Artif. Intell. 3, e200157 (2021).
Yoo, T. K. et al. Deep learning can generate traditional retinal fundus photographs using ultra-widefield images via generative adversarial networks. Comput. Methods Programs Biomed. 197, 105761 (2020).
Stec, L. A. et al. Vitrectomy for chronic macular holes. Retina. 24, 341–347 (2004).
Amram, A. L., Mandviwala, M. M., Ou, W. C., Wykoff, C. C. & Shah, A. R. Predictors of visual acuity outcomes following vitrectomy for idiopathic macular hole. Ophthalmic. Surg. Lasers. Imaging. Retina. 49, 566–570 (2018).
Ramtohul, P., Parrat, E., Denis, D. & Lorenzi, U. Inverted internal limiting membrane flap technique versus complete internal limiting membrane peeling in large macular hole surgery: A comparative study. BMC. Ophthalmol. 20, 11. https://doi.org/10.1186/s12886-019-1294-8 (2020).
Hu, X.-T., Pan, Q.-T., Zheng, J.-W. & Zhang, Z.-D. Foveal microstructure and visual outcomes of myopic macular hole surgery with or without the inverted internal limiting membrane flap technique. Br. J. Ophthalmol. 103, 1495–1502 (2018).
Author information
Authors and Affiliations
Contributions
Design and execution of the study: H.J.K. and I.B. Performed data collection and processing: J.H, S.Y.K., and S.H.P. Initiated the data analysis and interpretation of data: H.J.K., S.W.P., and I.B. Prepared, reviewed, and approved the manuscript: H.J.K., J.H., S.H.P., S.W.P., and I.B.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Supplementary Movie S5.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Kwon, H.J., Heo, J., Park, S.H. et al. Accuracy of generative deep learning model for macular anatomy prediction from optical coherence tomography images in macular hole surgery. Sci Rep 14, 6913 (2024). https://doi.org/10.1038/s41598-024-57562-5
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598-024-57562-5
Keywords
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.