January 2025
Volume 66, Issue 1
Open Access
Retina  |   January 2025
Deep Learning–Based SD-OCT Layer Segmentation Quantifies Outer Retina Changes in Patients With Biallelic RPE65 Mutations Undergoing Gene Therapy
Author Affiliations & Notes
  • German Pinedo-Diaz
    Center for Research and Advanced Studies, Cinvestav, Zapopan, Mexico
  • Birgit Lorenz
    Dept of Ophthalmology, University Hospital, Bonn, Germany
    Transmit Centre for Translational Ophthalmology, c/o Justus-Liebig-University Giessen, Giessen, Germany
  • Sandrine H. Künzel
    Dept of Ophthalmology, University Hospital, Bonn, Germany
  • Sarah Thiele
    Dept of Ophthalmology, University Hospital, Bonn, Germany
    Department of Ophthalmology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany
  • Susana Ortega-Cisneros
    Center for Research and Advanced Studies, Cinvestav, Zapopan, Mexico
  • Eduardo Bayro Corrochano
    Center for Research and Advanced Studies, Cinvestav, Zapopan, Mexico
  • Frank G. Holz
    Dept of Ophthalmology, University Hospital, Bonn, Germany
  • Alexander Effland
    Institute for Applied Mathematics, University of Bonn, Bonn, Germany
  • Correspondence: Alexander Effland, Institute for Applied Mathematics, University of Bonn, Endenicher Allee 60, Bonn 53115, Germany; [email protected]
  • Footnotes
     GPD and BL contributed equally.
Investigative Ophthalmology & Visual Science January 2025, Vol.66, 5. doi:https://doi.org/10.1167/iovs.66.1.5
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      German Pinedo-Diaz, Birgit Lorenz, Sandrine H. Künzel, Sarah Thiele, Susana Ortega-Cisneros, Eduardo Bayro Corrochano, Frank G. Holz, Alexander Effland; Deep Learning–Based SD-OCT Layer Segmentation Quantifies Outer Retina Changes in Patients With Biallelic RPE65 Mutations Undergoing Gene Therapy. Invest. Ophthalmol. Vis. Sci. 2025;66(1):5. https://doi.org/10.1167/iovs.66.1.5.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose: To quantify outer retina structural changes and define novel biomarkers of inherited retinal degeneration associated with biallelic mutations in RPE65 (RPE65-IRD) in patients before and after subretinal gene augmentation therapy with voretigene neparvovec (Luxturna).

Methods: Application of advanced deep learning for automated retinal layer segmentation, specifically tailored for RPE65-IRD. Quantification of five novel biomarkers for the ellipsoid zone (EZ): thickness, granularity, reflectivity, and intensity. Estimation of theEZarea in single and volume scans was performed with optimized segmentation boundaries. The control group was age similar and without significant refractive error. Spherical equivalent refraction and ocular length were evaluated in all patients.

Results: We observed significant differences in the structural analysis of EZ biomarkers in 22 patients with RPE65-IRD compared with 94 healthy controls. Relative EZ intensities were already reduced in pediatric eyes. Reductions of EZ local granularity and EZ thickness were only significant in adult eyes. Distances of the outer plexiform layer, external limiting membrane, and Bruch's membrane to EZ were reduced at all ages. EZ diameter and area were better preserved in pediatric eyes undergoing therapy with voretigene neparvovec and in patients with a milder phenotype.

Conclusions: Automated quantitative analysis of biomarkers within EZ visualizes distinct structural differences in the outer retina of patients including treatment-related effects. The automated approach using deep learning strategies allows big data analysis for distinct forms of inherited retinal degeneration. Limitations include a small dataset and potential effects on OCT scans from myopia at least −5 diopters, the latter considered nonsignificant for outer retinal layers.

R PE65 codes for an isomerohydrolase in the RPE essential for retinol recycling.14 Biallelic mutations in RPE65 are often associated with a rare form of inherited retinal degeneration (RPE65-IRD) with progressive atrophy of the RPE and the photoreceptors together with severe vision impairment to blindness.5 Successful phase 1, 2, and 3 studies with subretinal gene augmentation therapy6,7 have resulted in the approval of voretigene neparvovec (VN [Luxturna]) for clinical use. Several postmarketing studies have been published, including our own consecutive series of 30 eyes from 19 patients.8 An ongoing challenge is achieving objective, high-resolution quantification of the natural progression of the disease and the therapeutic effect, especially in larger patient cohorts and within routine clinical settings. 
The ellipsoid zone (EZ), formerly known as the inner/outer segment of photoreceptors, refers to the second hyper-reflective band in optical coherence tomography (OCT) of the retina. An intact EZ has been reported to be associated with good visual function and visual field in a variety of retinal diseases, including IRDs.912 
However, assessments of the EZ have been typically qualitative in nature or limited to a single B-scan. Traditional segmentation methods often require manual intervention and have limitations in handling complex image structures with ambiguous boundaries, making manual segmentation challenging and leading to low segmentation repeatability. Manual segmentation is prone to interobserver and intraobserver variability, which can lead to inconsistencies and errors in the results.13 
In recent years, deep learning (DL)–based segmentation methods such as convolutional neural networks, have shown promising results in segmenting medical images in IRDs and other retinal conditions such as fundus autofluorescence and SD-OCT images accurately without requiring manual intervention.1416 Convolutional neural network–based segmentation methods have proven robust in handling the variability in image quality and can provide segmentation results in real time.17 
The presented work highlights the power of SD-OCT layer segmentation using an automated DL-based approach in the context of 22 patients with RPE65-IRD compared with 94 healthy controls. We address the challenge of high time consumption for manual annotation and validate the results for repeatability and reproducibility. Using a modified U-Net DL architecture18 for precise retinal layer segmentation, we extend the segmentation to all outer retinal layers, that is, the outer plexiform layer (OPL), EZ, external limiting membrane (ELM), and Bruch's membrane (BM) in SD-OCT images. Our approach allows us to define novel biomarkers that include thickness, granularity, reflectivity, and relative intensity of EZ. 
We further quantify EZ area in volume scans and EZ diameter in single scans using the segmentation boundaries. These objective measures can then be compared with the subjective results of perimetric visual field testing. 
This approach enables comparative studies between RPE65-IRD patients and healthy controls, as well as longitudinal analysis of EZ biomarkers before and after treatment, offering insights into disease progression and treatment outcomes. 
Methods
Dataset and Participants
The dataset used in this retrospective study comprises SD-OCT scans stored at the University Hospital Bonn (UKB) together with data on refraction and ocular length. This study adhered to the principles of the Declaration of Helsinki, received approval from the institutional ethics committee (AZ325/22), and the data were anonymized before analysis. 
The study uses two datasets with distinct purposes, both containing data from the same RPE65-IRD and healthy participants. Dataset 1 (D1), retrospectively annotated manually by an expert to delineate boundaries of different retinal layers, consists of 321 single scans from 116 participants, including 22 individuals with RPE65IRD and 94 healthy controls used for training DL-based models. In contrast, dataset 2 (D2) is used for testing and longitudinal analysis, comprising 262 single scans (158 RPE65-IRD and 94 healthy controls) with single scans and volume scans. D1 and D2 include data from 18 treated patients at baseline and after treatment with VN, from Lorenz et al.8 and 4 new untreated patients with a milder phenotype. We included both bilateral (26 eyes) and unilateral (11 eyes) treated patients in our study, and unilateral untreated (7 eyes), totaling 44 eyes (right and left) for D2. Unlike D1, files in D2 are not annotated for retinal layer delimitation. 
The participants in D1 and D2 had an average age of 22.87 ± 9.87 years, with the control group averaging 22.55 ± 7.20 years and the RPE65-IRD group averaging 23.52 ± 8.90 years. Overall, there were 68 males and 48 females, with 54 males and 40 females in the control group, and 22 patients, of which 14 were males and 8 were females, in the RPE65-IRD group. Ocular axial length measured with an IOL master (Zeiss, Oberkochen, Germany) was available in 18 treated patients. The mean ocular length was 23.42 ± 1.22 mm. Refraction was available in all patients. The mean spherical refraction (SER) was −1.11 ± 3.99. High myopia (at least −5 diopters SER) was present in 9 of these 44 eyes (20.4%). 
For further details on D2, readers are referred to the Supplementary Material
SD-OCT
SD-OCT scans were obtained using the Spectralis HRA-OCT2 Heidelberg Engineering (Heidelberg, Germany). A single horizontal line B-scan traversing the fovea was captured in both high-speed (768 pixels) and high-resolution (1024 pixels) settings. These scans were divided into 496 × 128 patches with a 64-pixel overlap, serving as batched input for our DL-based model. Following Lorenz et al.,19 SD-OCT 30° × 25° scans were acquired using a Spectralis HRA-OCT2 with follow-up mode (121 B-scans, ART 25). For cases of nystagmus or unsteady fixation, single SD-OCT line scans were registered. The B-scans were taken on different visit dates. 
Dataset Preprocessing
D1 was divided into training and validation sets by patient using an 80/20 random split ratio. This resulted in 93 participants (75 controls and 18 patients) in the training set and 23 participants (19 controls and 4 patients) in the validation set (a total of 116 participants). This approach ensured no overlap between sets, maintaining the generalization capability of the DL-based model and avoiding data leakage. 
B-scans were divided into 496 × 128 patches with a 64-pixel overlap between adjacent patches. This preprocessing step was necessary to homogenize the input size for our DL-based model. Each section served as a batched input, allowing the model to process localized features across the entire scan. Moreover, each file contains a B-scan centered at the fovea, annotated by a clinical expert. An example delimitation is shown in Fig. 1A and is used to generate a mask as ground truth. Table 1 provides a comprehensive summary of the specifications and distributions within the UKB dataset. 
Figure 1.
 
Automated B-scan layer segmentation methodology. (A) Annotations include distinct outer layers: the OPL zone, ELM zone, EZ, and BM. The class map mask is labeled as Background (0), EZ (1), OPL zone (2), ELM zone (3), and IZ + RPE + BM (4). (B) Overview of the modified four-level UNet architecture with 1.2 million parameters.
Figure 1.
 
Automated B-scan layer segmentation methodology. (A) Annotations include distinct outer layers: the OPL zone, ELM zone, EZ, and BM. The class map mask is labeled as Background (0), EZ (1), OPL zone (2), ELM zone (3), and IZ + RPE + BM (4). (B) Overview of the modified four-level UNet architecture with 1.2 million parameters.
Table 1.
 
UKB Dataset Description
Table 1.
 
UKB Dataset Description
DL Approach
For retinal layer segmentation, we use a DL-based approach using a modified U-Net architecture,18 which is known for its effectiveness in semantic segmentation tasks. The U-Net architecture consists of an encoder–decoder network with skip connections designed to retain both local and global context information. Further information regarding data and training is provided in the Supplementary Material
Our modified four-level U-Net architecture, as depicted in Figure 1B, includes the following components: four-level double convolutional blocks in the encoder phase, starting with 16 filters, a kernel size of 5 × 5, and incorporating batch normalization with leaky ReLU activation. The encoder uses max pooling for downsampling. In the decoder phase, there are four levels of upsampling blocks using transposed convolutions that mirror the structure of the encoder. The architecture concludes with a final convolutional layer using five 1 × 1 filters to produce the predicted segmentation mask. 
The input SD-OCT B-scan underwent partitioning into batches of 496× 128 pixel sections, resulting in an output segmentation map sized 8 × 496 ×128 ×5 per scan. 
The loss function used for segmentation accuracy combines weighted cross-entropy and dice loss:  
\begin{eqnarray} {{L}_{DCE}}\left( {y,\hat{p}} \right) = \left( {1 - {\rm{\lambda }}} \right){{L}_{Dice}}\left( {y,\hat{p}} \right) + {\rm{\lambda }}{{L}_{CE}}\left( {y,\hat{p}} \right),\quad \end{eqnarray}
(1)
 
where y and \(\hat{p}\) denote ground truth and predicted probabilities, respectively, with λ ∈ (0, 1) as a hyperparameter. 
Data augmentation techniques were applied to enhance model robustness and generalization. Intensity transforms were applied to SD-OCT B-scans, and geometric adjustments were used for both images and segmentation masks (details in the Supplementary Material). 
Evaluating Retinal Changes in RPE65-IRD
Our focus centers on crucial and novel biomarkers primarily designed for the EZ, encompassing thickness, granularity, reflectivity, intensity, and volume area. Performing DL-based segmentation, we isolate individual layers within a 6-mm diameter region of interest, according to Early Treatment Diabetic Retinopathy Study specifications. Subsequently, we compute intensity profiles and quantify these biomarkers for both RPE65-IRD patients and healthy controls. Table 2 shows the summary of definitions of these biomarkers. 
Table 2.
 
Summary of Definitions of Biomarkers
Table 2.
 
Summary of Definitions of Biomarkers
Weighted Peak Distances
Relative EZ reflectivity represents a pivotal biomarker used in appraising photoreceptor health.2023 The relative EZ reflectivity measures the pixel distance between the peaks of the intensity profiles of EZ and OPL, offering valuable information on retinal optical characteristics and their implications for retinal diseases. Based on this approach, we introduce four novel biomarkers called weighted peak distances of the respective layers. These biomarkers measure distances between the ELM and OPL peaks, ELM and EZ peaks, BM boundary and EZ peaks, and ELM boundary and BM along the x axis. For evaluation, we first segment the retinal layers and extract peak values from each A-scan. Distances are then computed along the x axis for the B-scan centered on the fovea. Finally, the arithmetic mean and SD of these distances are calculated along the foveal B-scan. In summary, the four novel biomarkers (rescaled to micrometers) read as:  
\begin{eqnarray} \begin{array}{@{}l@{}} OPL\_E{{Z}_i}\left( {{\rm{\mu }}m} \right) = \left| {{{\mathcal{P}}_{max}}\left( {E{{Z}_i}} \right) - {{\mathcal{P}}_{max}}\left( {OP{{L}_i}} \right)} \right| \cdot {{\mathcal{S}}_y}\\ ELM\_E{{Z}_i}\left( {{\rm{\mu }}m} \right) = \left| {{{\mathcal{P}}_{max}}\left( {EL{{M}_i}} \right) - {{\mathcal{P}}_{max}}\left( {E{{Z}_i}} \right)} \right| \cdot {{\mathcal{S}}_y}\\ BM\_E{{Z}_i}\left( {{\rm{\mu }}m} \right) = \left| {B{{M}_i} - {{\mathcal{P}}_{max}}\left( {E{{Z}_i}} \right)} \right| \cdot {{\mathcal{S}}_y}\\ ELM\_B{{M}_i}\left( {{\rm{\mu }}m} \right) = \left| {{{\mathcal{P}}_{max}}\left( {EL{{M}_i}} \right) - B{{M}_i}} \right| \cdot {{\mathcal{S}}_y}\,, \end{array}\quad \end{eqnarray}
(2)
where \({{\mathcal{P}}_{max}}( {E{{Z}_i}} )\), \({{\mathcal{P}}_{max}}( {EL{{M}_i}} )\), and \({{\mathcal{P}}_{max}}( {OP{{L}_i}} )\) denote the maximum peaks of EZ, ELM, and OPL, respectively, and BMi is the location BM boundary on the x axis. \({{\mathcal{S}}_{{y}}}\) represents the axial resolution of the A-scan SD-OCT system (approximately 3.5 µm per pixel, according to the Spectralis operations manual). This process is implemented using Python, with PyTorch for segmentation, NumPy for peak value extraction and distance calculations, and iteration over the B-scan to compute the mean and SD. 
The arithmetic mean and SD of the weighted peak distances are calculated along the B-scan x axis, resulting in average values for \(OPL\_EZ\), \(ELM\_EZ\), \(BM\_EZ\), and \(ELM\_BM\). These biomarkers are displayed in Figure 2, segmented by the Early Treatment Diabetic Retinopathy Study grid. 
Figure 2.
 
Patient P04’s right eye underwent a macular B-scan with 1536 A-scans, and baseline BCVA was 0.5 logMAR (Lorenz et al.19). Figure 1 shows automated segmentation results along the Early Treatment Diabetic Retinopathy Study grid (6 mm diameter). Within this region, comprising 529 A-scans, individual layer segmentation was performed. For example, intensity profiles extracted from a 2-mm segment were used to derive biomarkers like EZ reflectivity, thickness, and relative intensity, with BM localization at the boundary.
Figure 2.
 
Patient P04’s right eye underwent a macular B-scan with 1536 A-scans, and baseline BCVA was 0.5 logMAR (Lorenz et al.19). Figure 1 shows automated segmentation results along the Early Treatment Diabetic Retinopathy Study grid (6 mm diameter). Within this region, comprising 529 A-scans, individual layer segmentation was performed. For example, intensity profiles extracted from a 2-mm segment were used to derive biomarkers like EZ reflectivity, thickness, and relative intensity, with BM localization at the boundary.
EZ Thickness
In RPE65-IRD, EZ thickness seems to be different from healthy controls, which might therefore be a reliable biomarker for disease progression. We quantify EZ thickness by considering the lower and upper boundaries of the segmented EZ layer, that is:  
\begin{eqnarray} E{{Z}_{Th}}\left( {{\rm{\mu }}m} \right) = \frac{1}{N}\mathop \sum \limits_{x = 0}^N \left| {{{\mathcal{L}}_u}\left( {E{{Z}_x}} \right) - {{\mathcal{L}}_\ell }\left( {E{{Z}_x}} \right)} \right| \cdot {{\mathcal{S}}_y},\quad \end{eqnarray}
(3)
where N refers to the number of pixels in the x- direction and the thickness as a difference in pixels is subsequently scaled to µm. \({{\mathcal{L}}_{{u}}}( {E{{Z}_x}} )\) and \({{\mathcal{L}}_{{l}}}( {E{{Z}_x}} )\) are the lower and upper limits of the EZ layer, respectively. Sy is the axial resolution from A-scan SD-OCT system (approximately 3.5 µm per pixel digital according to the operations manual). 
rEZI
In the past, most assessments of EZ have been qualitative, based on observations of its presence, absence, or the spatial extent of the disturbance.24 In our study on RPE65-IRD, we introduce a novel measure called rEZI. 
Changes in retinal diseases, such as retinitis pigmentosa and AMD, which are potentially associated with mitochondrial dysfunction, have been reported. To evaluate rEZI, we measure the ratio of EZ reflectivity to OPL at each retinal location25 yielding  
\begin{eqnarray} {\rm{rEZI}} = \frac{1}{N}\mathop \sum \limits_{i = 0}^N \frac{{\left| {\mathcal{P}\left( {E{{Z}_i}} \right) - \mathcal{P}\left( {OP{{L}_i}} \right)} \right|}}{{\mathcal{P}\left( {E{{Z}_i}} \right)}},\quad \end{eqnarray}
(4)
where N denotes the number of pixels on the x axis, and \(\mathcal{P}( {E{{Z}_i}} )\), \(\mathcal{P}( {OP{{L}_i}} )\) are the maximum peaks of the EZ and OPL, respectively. 
Total Variation Denoising and EZ Granularity
EZ local granularity or EZ local variation (EZLV) associated with RPE65-IRD denote the changes in the intensity of the EZ within a neighborhood, and EZ granularity or EZ total variations (EZTV) describe its average in B scan (x axis). This metric measures the difference of pixel grey value in a region given by:  
\begin{eqnarray} E{{Z}_{LV}} = \mathop \sum \limits_{x = 0}^N \mathop \sum \limits_{k = - \beta }^\beta \mathop \sum \limits_{y = 0}^M \left| {{{\mathcal{P}}_{x + k,y}} - {{\mathcal{P}}_{x,y}}} \right|\quad \end{eqnarray}
(5)
 
\begin{eqnarray} E{{Z}_{TV}} = \frac{{E{{Z}_{LV}}\left( {2{\rm{\beta }}} \right)}}{N},\quad \end{eqnarray}
(6)
where N is the total of pixels in the x axis, P is the pixel gray value, and β  ∈  {3, 5, 7} in the x axis is the neighborhood size. Owing to the strong dependency on noise, we additionally use a total variation-based denoising26 with smoothing parameter α in a preprocessing step. For further details, we refer the reader to the Supplementary Material. As shown in Figure 3, the local variation along B-scan (x axis), has been smoothed owing to TV denoising. To assess EZ granularity (EZTV) in this study, the local variation window β = 3 and the denoising parameter α = 0.05 are established empirically, which properly removes speckle noise typically found in tomography images while promoting data consistency. 
Figure 3.
 
Central macular SD-OCT B-scan comparison of EZ granularity in (A) healthy control and (B) RPE65IRD shown in Figure 5 patient P0419 using TV denoising with α  ∈  {0.5,  0.05,  0.005} and β  ∈  {3,  5,  7} as explained in equations (5) and (7). The TV in (A) has an oscillatory behavior, with a 0.75 mean value interpreted as better EZ granularity. In contrast, (B) shows a constant-like behavior with 0.47 EZ granularity value. Therefore, EZ granularity between healthy and RPE65-IRD is significant.
Figure 3.
 
Central macular SD-OCT B-scan comparison of EZ granularity in (A) healthy control and (B) RPE65IRD shown in Figure 5 patient P0419 using TV denoising with α  ∈  {0.5,  0.05,  0.005} and β  ∈  {3,  5,  7} as explained in equations (5) and (7). The TV in (A) has an oscillatory behavior, with a 0.75 mean value interpreted as better EZ granularity. In contrast, (B) shows a constant-like behavior with 0.47 EZ granularity value. Therefore, EZ granularity between healthy and RPE65-IRD is significant.
Statistical Analysis
Biomarker comparisons were made between age groups (pediatric and adult) and healthy controls. We used the Shapiro–Wilk test to check for normality. The parametric two-sided t test was applied for normally distributed data (P > 0.05), and the nonparametric Wilcoxon rank-sum test was used for non-normally distributed data (P < 0.05). 
Furthermore, we examined the correlation between the rEZI biomarker and others using Pearson's correlation coefficient. This analysis quantified the strength and direction of linear relationships between variables, with a correlation coefficient r > 0.3 indicating moderate to strong associations. 
All statistical analyses were performed using Python 3.9 with the SciPy library for the Shapiro–Wilk test and the t test, and the statsmodels library for the Wilcoxon rank-sum test. Pearson's correlation was calculated using the Pandas library. The Python scripts used to perform the analyses, including image processing and DL, described herein are freely and openly available at https://github.com/alonso59/ukb_rpe65_ird/
Results
Characteristics of Retinal Changes Before Treatment
Our comparative study of biomarkers extracted from SD-OCT scans at baseline diagnosed with RPE65-IRD and a healthy control cohort is summarized in Table 3. Data are presented as mean ± SD across three distinct comparisons: (a) comparing the control group with the pediatric patient group, (b) comparing the control group with the adult patient group, and (c) comparing the pediatric patient group with the adult patient group. 
Table 3.
 
Biomarkers Extracted From D2, Including RPE65-IRD Patients, Before Treatment and From a Healthy Control Group
Table 3.
 
Biomarkers Extracted From D2, Including RPE65-IRD Patients, Before Treatment and From a Healthy Control Group
Each comparison evaluates biomarkers such as rEZI, EZTV, EZTh, and weighted peak distances, providing the mean, SD, and P values for nasal and temporal regions at 0.5 mm from the foveal center. Notable differences are observed between groups for various biomarkers. For instance, significant discrepancies in rEZI are noted between the control and pediatric groups (nasal, P < 0.0013; temporal, P < 0.0008), similar to findings between the control and adult groups. 
However, when comparing pediatric and adult groups, no significant difference in rEZI is found (nasal, P < 0.1007; temporal, P < 0.3110). These results suggest that, although significant differences in rEZI are observed between healthy controls and both pediatric and adult RPE65-IRD patients, there is no significant difference in rEZI between pediatric and adult patients. This finding implies that rEZI values may stabilize or change differently with age within the RPE65-IRD cohort. 
The Pearson correlation analysis reveals several key relationships. In the first set, rEZI shows a moderate positive correlation with EZTV (r = 0.414). It is also moderately correlated with \(ELM\_BM\) (r = 0.329), while displaying a weak negative correlation with EZTh (r = −0.085). In the second set, EZTV demonstrates moderate positive correlations with OPLEZ (r = 0.333), and a weak positive correlation with EZTh (r = 0.107). Notably, \(ELM\_EZ\) shows almost no correlation with EZTV (r = −0.006). These results highlight the complex relationships among biomarkers, rEZI, and EZTV in RPE65-IRD. For a complete analysis of all correlations, we refer the reader to the Supplementary Material
Longitudinal Analysis After Treatment With VN
As the patients are followed for at least 5 years after VN therapy according to the requirements of the PERCEIVE registry imposed by the European Medicines Agency EMA (EUPAS31153, https://catalogues.ema.europa.eu/node/3021/data-management). More data can be collected at follow-up. As a general observation, the biomarkers showed fewer deviations in younger patients than in older patients compared with healthy controls. Exceptions were related to differences in phenotype severity, indicating some genotype–phenotype correlation of RPE65-IRDs. Unilaterally treated patients tended to have lower values both in their treated and untreated eyes. This worse performance was usually the reason for treating only one eye. 
Longitudinal analysis (Fig. 4) of biomarkers after treatment with VN revealed significant trends in pediatric and adult patients. Pediatric patients P05 and P04 showed consistent increases in outer retinal layer thicknesses (OPL_EZ, ELM_EZ) and EZ area over 24 months. P05 improved steadily in both OPL EZ and ELM_EZ, indicating a positive treatment response. Conversely, adult patients such as P16 and P02 exhibited varying responses, with differences in EZ thickness (BM_EZ) and variability (EZTV) between eyes. P16 showed moderate improvements in EZTh, whereas P02 consistently improved in both EZTh and EZTV. Untreated patients (UP01, UP03) demonstrated stable biomarker levels. Individual patient results are provided in the Supplementary Material
Figure 4.
 
A comprehensive longitudinal analysis spanning −3 to 24 months after treatment of RPE65-IRD. The data are based on the results reported in.19 The numbers refer to the number of eyes per visit.
Figure 4.
 
A comprehensive longitudinal analysis spanning −3 to 24 months after treatment of RPE65-IRD. The data are based on the results reported in.19 The numbers refer to the number of eyes per visit.
Figure 5.
 
Automated segmentation was used to obtain the EZ area in volume scans focused on region of interest Early Treatment Diabetic Retinopathy Study at 6 mm. For instance, in the right eye of Patient P04, analysis involved 60 B-scans with 1536 A-scans each. At baseline, the patient exhibited a best-corrected visual acuity of 0.5 logMAR, with an EZ area measuring 15.15 mm2 (refer to Lorenz et al.19). Refer to Table S2 for comprehensive numerical results regarding EZ area and EZ diameter for each RPE65-IRD patient.
Figure 5.
 
Automated segmentation was used to obtain the EZ area in volume scans focused on region of interest Early Treatment Diabetic Retinopathy Study at 6 mm. For instance, in the right eye of Patient P04, analysis involved 60 B-scans with 1536 A-scans each. At baseline, the patient exhibited a best-corrected visual acuity of 0.5 logMAR, with an EZ area measuring 15.15 mm2 (refer to Lorenz et al.19). Refer to Table S2 for comprehensive numerical results regarding EZ area and EZ diameter for each RPE65-IRD patient.
EZ Area and EZ Width
In our quantitative analysis, we emphasize the significance of the EZ area as a crucial biomarker derived from SD-OCT scans.11 The process involves precisely measuring its width for each B-scan. Our study, comprising 96 SD-OCT scans (22 from healthy controls and 74 from patients with RPE65-IRD), underscores the pivotal role of EZ area size within the region of interest. These measurements serve as quantitative indicators of retinal structural changes and can be correlated with visual field measurements (data not shown). It is important to note that the EZ area limits are calculated exclusively where the EZ exists. Subsequently, we use linear interpolation to fill gaps between B-Scans. Finally, pixels within the EZ limits are tallied and scaled to square millimeters using the digital pixel size (approximately 121 µm2 according to the Spectralis manual). Figure 5 provides an example of the EZ area and its width on the foveal center B-scan. 
As a result, significant differences were observed: the control group exhibited an EZ area of 34.476 ± 0.66 mm2, whereas the pediatric RPE65-IRD group showed 27.119 ± 14.30 mm2 and the adult RPE65-IRD group had an EZ area of 8.110 ± 8.88 mm2. The P value for the comparison between the control group and the pediatric RPE65-IRD group was 0.0170, whereas, for both the adult and pediatric RPE65-IRD groups against the control group, the P value was less than 0.0001. 
Figure 6 illustrates age-EZ width correlation, grouped by pediatric (<20 years) and adult categories, and further subdivided into treated (P) and untreated (UP) groups. Each data point represents left and right eyes, facilitating comparisons across age and treatment. Analysis indicates that untreated patients mostly exhibit EZ widths between 80% and 98% of the region of interest, suggesting larger widths reflecting less severe or less progressed phenotypes, which was a factor for deferring treatment in such patients. Moreover, most adult patients show lower EZ widths compared with pediatric ones, except for patients P09 and P11, whose widths closely align with 0% to 20%. Of note, P09 and P11 had severe nystagmus hampering reproducible SD-OCT scans, and impeding both morphological and functional characterization of the disease stage. 
Figure 6.
 
Relationship between age and EZ width expressed as a percentage (%) of the region of interest. Pediatric patients (<20 years) and adult patients are divided into treated (P) and untreated (UP) groups, with data marked for each eye (left and right). All measures of the treated eyes are taken at baseline. A trend can be observed of larger EZ width at younger age. Older patients with still relatively large EZ width reflect phenotypic differences, with some patients presenting milder disease. This knowledge is important for guiding treatment decisions.
Figure 6.
 
Relationship between age and EZ width expressed as a percentage (%) of the region of interest. Pediatric patients (<20 years) and adult patients are divided into treated (P) and untreated (UP) groups, with data marked for each eye (left and right). All measures of the treated eyes are taken at baseline. A trend can be observed of larger EZ width at younger age. Older patients with still relatively large EZ width reflect phenotypic differences, with some patients presenting milder disease. This knowledge is important for guiding treatment decisions.
Discussion
In this work, we have proposed for the first time several novel concepts to quantitatively assess differences between RPE65-IRD patients and healthy controls in the outer retina imaged with SD-OCT with a focus on the EZ. Most of the biomarkers we introduced exhibited statistically significant differences. Despite the limited number of scans, the methods we used resulted in quite robust results showing the validity of our approach of DL to allow time-efficient quantitative biomarker analysis. Interestingly, the most significant differences were observed in the four weighted peak distances. In the future, these novel biomarkers could be correlated with visual function including best-corrected visual acuity and mesopic retinal sensitivity tested with microperimetry, as well as multifocal ERG measurements. The biomarkers quantifying the thickness of the EZ or its total variation were inferior indicators of RPE65-IRD in the pediatric group, but became significant in adults. Most of the novel biomarkers were changed significantly already early on, indicating that significant changes in the morphology of the outer retina may be present at birth or shortly afterward, in particular of the inner and outer segments and the outer nuclear layer. This questions up to which degree any therapy can reverse or prevent these changes, even if applied very early. An abnormal structure of EZ without further quantification has been observed in young children with SD-OCT. The data are in line with histology data of the retina of Briard dogs with a naturally occurring 4-bp deletion in RPE65.27 In any case, these novel biomarkers can be an additional way to monitor the in vivo treatment effects in patients and eventually also in animal studies as a noninvasive method. The longitudinal data we show have the problem of small numbers at the various time points after treatment impeding meaningful statistical analysis. This reflects the incomplete data collection of retrospective studies without a strict protocol such as the PERCEIVE register (EUPAS31153, https://catalogues.ema.europa.eu/node/3021/data-management), and variable compliance and logistics problems of the patients and their families to come to the recommended follow-up Examinations. Some of the patients live far away from the treatment site at the Department of Ophthalmology in Bonn. We confirm that EZ diameter and EZ area are valuable objective biomarkers of centripetal disease progression typical of rod–cone IRDs. Our novel approach allows quantification of both biomarkers within an acceptable time frame, thus enabling correlation with visual field measurements and quantification of the natural history of the disease in a large cohort of patients, as well as the definition of optimal timing of available and future treatments and treatment monitoring. 
A strength of our study is the precise segmentation of the outer retinal layers guaranteed by the DL-based algorithm with its relatively high dice similarity coefficient and sensitivity throughout the entire cohort. This is a prerequisite for the exact quantification of the nine biomarkers we identified as meaningful parameters to describe with high resolution the outer retina in RPE65-IRD. 
A limitation of our study is the relatively small data set of patients with RPE65-IRD and the available SD-OCT scans. In addition, 9 of the 44 eyes have high myopia (at least −5 diopter SER). However, according to the literature, the effect of high myopia on the outer retina is less significant than on the inner retina, and published data show a wide range of layer thicknesses for a wide range of refraction and ocular lengths.28 These uncertainties, however, still allow monitoring of individual disease courses with and without therapy. To further validate our algorithm, including more patients with RPE65-IRD is desirable. Further, to analyze how specific the novel biomarkers on EZ internal structure are, testing patients with other forms of IRDs will be interesting. 
Conclusions
Our study presents an automated DL-based approach for the precise evaluation of retinal changes in RPE65-associated inherited retinal dystrophy (RPE65-IRD) including treatment-related effects using OCT (SD-OCT) data. This method effectively overcomes the laborious and specialized nature of manual SD-OCT layer segmentation. We conducted a thorough quantitative analysis of nine biomarkers within the EZ, revealing significant differences between RPE65-IRD patients and healthy controls, notably in intensity profiles of EZ with the other layers of the outer retina and the volumetric area of EZ. Owing to the limitations of our study, a major use of our algorithms is the monitoring of individual disease courses. 
Acknowledgments
The authors thank Philipp Herrmann, MD, PhD, Department of Ophthalmology, University Bonn, Bonn, Germany, provided excellent patient care; Markus N. Preising, PhD, Department of Ophthalmology, Justus-Liebig University Giessen, Giessen, Germany, established a large database of IRD patients in Giessen. The research nurses Naike Cavriani, Laura Wirtz, and Kristina Küpper, Department of Ophthalmology, University Bonn, Bonn, Germany, performed high-quality SD-OCT imaging. Part of the OCT scans was collected within the PERCEIVE Registry (EUPAS31153, https://catalogues.ema.europa.eu/node/3021/data-management) funded by Novartis Pharma Suisse GmbH, Switzerland. 
Supported by the German Research Foundation under Germany's Excellence Strategy - EXC-2047/1 - 390685813 and - EXC2151 - 390873048. This work was supported by the Open Access Publication Fund of the University of Bonn. 
Disclosure: G. Pinedo-Diaz, None; B. Lorenz, Novartis Pharma (F, C, R), Janssen/Johnson & Johnson (C, R), Pfizer (R); S.H. Künzel, Novartis Pharma (C), Chiesi GmbH (R), Apellis (R); S. Thiele, Heidelberg Engineering (F, R), Novartis Pharma GmbH (F, R), Allergan (R), Bayer AG (R), Optos (F), CenterVue (F), Zeiss (F); S. Ortega-Cisneros, None; E. Bayro Corrochano, None; F.G. Holz, Novartis Pharma (F), Heidelberg Engineering (C, R), Optos (R), Zeiss (C, R), iCare/Centervue (R), 4D Molecular Therapeutics (C), Euretina (S), German Ophthalmological Society (S), Macula Society (S); A. Effland, relios.vision GmbH (O, P) 
References
Gu SM, Thompson DA, Srikumari CRS, et al. Mutations in RPE65 cause autosomal recessive childhood–onset severe retinal dystrophy. Nat Genet. 1997; 17(2): 194–197. [CrossRef] [PubMed]
Paunescu K, Wabbels B, Preising MN, Lorenz B. Longitudinal and cross-sectional study of patients with early-onset severe retinal dystrophy associated with RPE65 mutations. Invest Ophthalmol Vis Sci. 2000; 243(5): 417–426.
Chung DC, Bertelsen M, Lorenz B, et al. The natural history of inherited retinal dystrophy due to biallelic mutations in the RPE65 gene. Am J Ophthalmol. 2019; 199: 58–70, doi:10.1016/j.ajo.2018.09.024. [CrossRef] [PubMed]
Kiser PD. Retinal pigment epithelium 65 kDa protein (RPE65): an update. Prog Retin Eye Res. 2022; 88: 101013, doi:10.1016/j.preteyeres.2021.101013. [CrossRef] [PubMed]
Daiger SP, Sullivan LS, Cadena EL, Bowne SJ. History of finding genes and mutations causing inherited retinal diseases. Cold Spring Harb Perspect Med. 2024; 14: a041287. [PubMed]
Bennett J, Wellman J, Marshall KA, et al. Safety and durability of effect of contralateral-eye administration of AAV2 gene therapy in patients with childhood-onset blindness caused by RPE65 mutations: a follow-on phase 1 trial. Lancet. 2016; 388(10045): 661–672. [CrossRef] [PubMed]
Maguire AM, Russell S, Wellman JA, et al. Efficacy, safety, and durability of voretigene neparvovecrzyl in RPE65 mutation–associated inherited retinal dystrophy. Lancet. 2017; 126(9): 1273–1285.
Lorenz B, Tavares J, Born VDLI, et al. Current management of patients with RPE65 mutation associated inherited retinal degenerations in Europe: results of a 2-year follow-up multinational survey. Ophthalmic Res. 2023; 66: 727–748. [PubMed]
Cai CX, Locke KG, Ramachandran R, Birch DG, Hood DC. A comparison of progressive loss of the ellipsoid zone (EZ) band in autosomal dominant and X-linked retinitis pigmentosa. Invest Ophthalmol Vis Sci. 2014; 55(11): 7417. [CrossRef] [PubMed]
Birch DG, Locke KG, Felius J, et al. Rates of Decline in regions of the visual field defined by frequency-domain optical coherence tomography in patients with RPGR-mediated X-linked retinitis pigmentosa. Ophthalmology. 2015; 122(4): 833–839. [CrossRef] [PubMed]
Kumaran N, Georgiou M, Bainbridge JWB, et al. Retinal structure in RPE65-associated retinal dystrophy. Invest Ophthalmol Vis Sci. 2020; 61(4): 47. [CrossRef] [PubMed]
Hood DC, Ramachandran R, Holopigian K, Lazow M, Birch DG, Greenstein VC. Method for deriving visual field boundaries from OCT scans of patients with retinitis pigmentosa. Biomed Opt Express. 2011; 2(5): 1106. [CrossRef] [PubMed]
Kiefer LS, Fabian J, Lorbeer R, et al. Inter- and intra-observer variability of an anatomical landmark based, manual segmentation method by MRI for the assessment of skeletal muscle fat content and area in subjects from the general population. Br J Radiol. 2018; 91: 1089:20180019, doi:10.1259/bjr.20180019.
Lee CS, Tyring AJ, Deruyter NP, Wu Y, Rokem A, Lee AY. Deep-learning based, automated segmentation of macular edema in optical coherence tomography. Biomed Opt Express. 2017; 8(7): 3440–3448. [CrossRef] [PubMed]
Ehnes A, Wenner Y, Friedburg C, et al. Optical coherence tomography (OCT) device independent intraretinal layer segmentation. Transl Vis Sci Technol. 2014; 3(1): 1. [CrossRef] [PubMed]
Woof W, de Guimarães TAC, Al-Khuzaei S, et al. Quantification of fundus autofluorescence features in a molecularly characterized cohort of more than 3500 inherited retinal disease patients from the United Kingdom. medRxiv [Preprint]. 2024, doi:10.1101/2024.03.24.24304809.
Gopinath K, Rangrej SB, Sivaswamy J. A deep learning framework for segmentation of retinal layers from OCT images. Nanjing, China: 2017 Asian Conference on Pattern Recognition; 2017, pp. 888–893.
Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. Medical Image Computing and Computer Assisted Intervention. Cham: Springer Nature Switzerland; 2015; abs/1505.04597: 234–241.
Lorenz B, Künzel SH, Preising MN, et al. Single center experience with voretigene neparvovec gene augmentation therapy in RPE65 mutation–associated inherited retinal degeneration in a clinical setting. Ophthalmology. 2024; 131(2): 161–178. [CrossRef] [PubMed]
Wu Z, Ayton LN, Guymer RH, Luu CD. Second reflective band intensity in age-related macular degeneration. Ophthalmology. 2013; 120(6): 1307–1308.e1. [CrossRef] [PubMed]
Thiele S, Isselmann B, Pfau M, et al. Validation of an automated quantification of relative ellipsoid zone reflectivity on spectral domain-optical coherence tomography images. Transl Vis Sci Technol. 2020; 9(11): 17. [CrossRef] [PubMed]
Saßmannshausen M, Thiele S, Behning C, et al. Intersession repeatability of structural biomarkers in early and intermediate age-related macular degeneration: a MACUSTAR study report. Transl Vis Sci Technol. 2022; 11(3): 27. [CrossRef] [PubMed]
Thiele S, Wu Z, Isselmann B, Pfau M, Guymer RH, Luu CD. Natural history of the relative ellipsoid zone reflectivity in age-related macular degeneration. Ophthalmol Retina. 2022; 6(12): 1165–1172. [CrossRef] [PubMed]
Gin TJ, Wu Z, Chew SKH, Guymer RH, Luu CD. Quantitative analysis of the ellipsoid zone intensity in phenotypic variations of intermediate age-related macular degeneration. Invest Ophthalmol Vis Sci. 2017; 58(4): 2079. [CrossRef] [PubMed]
Lee KE, Heitkotter H, Carroll J. Challenges associated with ellipsoid zone intensity measurements using optical coherence tomography. Transl Vis Sci Technol. 2021; 10(12): 27. [CrossRef] [PubMed]
Chambolle A. An algorithm for total variation minimization and applications. J Math Imaging Vis. 2004; 20: 89–97. [CrossRef]
Annear MJ, Mowat FM, Occelli LM, et al. A comprehensive study of the retinal phenotype of Rpe65-deficient dogs. Cells 2021; 10(1): 115. [CrossRef] [PubMed]
Palazon-Cabanes A, Palazon-Cabanes B, Rubio-Velazquez E, Lopez-Bernal MD, Garcia-Medina JJ, Villegas-Perez MP. Normative database for all retinal layer thicknesses using SD-OCT posterior pole algorithm and the effects of age, gender and axial length. J Clin Med. 2020; 9(10): 3317. [CrossRef] [PubMed]
Figure 1.
 
Automated B-scan layer segmentation methodology. (A) Annotations include distinct outer layers: the OPL zone, ELM zone, EZ, and BM. The class map mask is labeled as Background (0), EZ (1), OPL zone (2), ELM zone (3), and IZ + RPE + BM (4). (B) Overview of the modified four-level UNet architecture with 1.2 million parameters.
Figure 1.
 
Automated B-scan layer segmentation methodology. (A) Annotations include distinct outer layers: the OPL zone, ELM zone, EZ, and BM. The class map mask is labeled as Background (0), EZ (1), OPL zone (2), ELM zone (3), and IZ + RPE + BM (4). (B) Overview of the modified four-level UNet architecture with 1.2 million parameters.
Figure 2.
 
Patient P04’s right eye underwent a macular B-scan with 1536 A-scans, and baseline BCVA was 0.5 logMAR (Lorenz et al.19). Figure 1 shows automated segmentation results along the Early Treatment Diabetic Retinopathy Study grid (6 mm diameter). Within this region, comprising 529 A-scans, individual layer segmentation was performed. For example, intensity profiles extracted from a 2-mm segment were used to derive biomarkers like EZ reflectivity, thickness, and relative intensity, with BM localization at the boundary.
Figure 2.
 
Patient P04’s right eye underwent a macular B-scan with 1536 A-scans, and baseline BCVA was 0.5 logMAR (Lorenz et al.19). Figure 1 shows automated segmentation results along the Early Treatment Diabetic Retinopathy Study grid (6 mm diameter). Within this region, comprising 529 A-scans, individual layer segmentation was performed. For example, intensity profiles extracted from a 2-mm segment were used to derive biomarkers like EZ reflectivity, thickness, and relative intensity, with BM localization at the boundary.
Figure 3.
 
Central macular SD-OCT B-scan comparison of EZ granularity in (A) healthy control and (B) RPE65IRD shown in Figure 5 patient P0419 using TV denoising with α  ∈  {0.5,  0.05,  0.005} and β  ∈  {3,  5,  7} as explained in equations (5) and (7). The TV in (A) has an oscillatory behavior, with a 0.75 mean value interpreted as better EZ granularity. In contrast, (B) shows a constant-like behavior with 0.47 EZ granularity value. Therefore, EZ granularity between healthy and RPE65-IRD is significant.
Figure 3.
 
Central macular SD-OCT B-scan comparison of EZ granularity in (A) healthy control and (B) RPE65IRD shown in Figure 5 patient P0419 using TV denoising with α  ∈  {0.5,  0.05,  0.005} and β  ∈  {3,  5,  7} as explained in equations (5) and (7). The TV in (A) has an oscillatory behavior, with a 0.75 mean value interpreted as better EZ granularity. In contrast, (B) shows a constant-like behavior with 0.47 EZ granularity value. Therefore, EZ granularity between healthy and RPE65-IRD is significant.
Figure 4.
 
A comprehensive longitudinal analysis spanning −3 to 24 months after treatment of RPE65-IRD. The data are based on the results reported in.19 The numbers refer to the number of eyes per visit.
Figure 4.
 
A comprehensive longitudinal analysis spanning −3 to 24 months after treatment of RPE65-IRD. The data are based on the results reported in.19 The numbers refer to the number of eyes per visit.
Figure 5.
 
Automated segmentation was used to obtain the EZ area in volume scans focused on region of interest Early Treatment Diabetic Retinopathy Study at 6 mm. For instance, in the right eye of Patient P04, analysis involved 60 B-scans with 1536 A-scans each. At baseline, the patient exhibited a best-corrected visual acuity of 0.5 logMAR, with an EZ area measuring 15.15 mm2 (refer to Lorenz et al.19). Refer to Table S2 for comprehensive numerical results regarding EZ area and EZ diameter for each RPE65-IRD patient.
Figure 5.
 
Automated segmentation was used to obtain the EZ area in volume scans focused on region of interest Early Treatment Diabetic Retinopathy Study at 6 mm. For instance, in the right eye of Patient P04, analysis involved 60 B-scans with 1536 A-scans each. At baseline, the patient exhibited a best-corrected visual acuity of 0.5 logMAR, with an EZ area measuring 15.15 mm2 (refer to Lorenz et al.19). Refer to Table S2 for comprehensive numerical results regarding EZ area and EZ diameter for each RPE65-IRD patient.
Figure 6.
 
Relationship between age and EZ width expressed as a percentage (%) of the region of interest. Pediatric patients (<20 years) and adult patients are divided into treated (P) and untreated (UP) groups, with data marked for each eye (left and right). All measures of the treated eyes are taken at baseline. A trend can be observed of larger EZ width at younger age. Older patients with still relatively large EZ width reflect phenotypic differences, with some patients presenting milder disease. This knowledge is important for guiding treatment decisions.
Figure 6.
 
Relationship between age and EZ width expressed as a percentage (%) of the region of interest. Pediatric patients (<20 years) and adult patients are divided into treated (P) and untreated (UP) groups, with data marked for each eye (left and right). All measures of the treated eyes are taken at baseline. A trend can be observed of larger EZ width at younger age. Older patients with still relatively large EZ width reflect phenotypic differences, with some patients presenting milder disease. This knowledge is important for guiding treatment decisions.
Table 1.
 
UKB Dataset Description
Table 1.
 
UKB Dataset Description
Table 2.
 
Summary of Definitions of Biomarkers
Table 2.
 
Summary of Definitions of Biomarkers
Table 3.
 
Biomarkers Extracted From D2, Including RPE65-IRD Patients, Before Treatment and From a Healthy Control Group
Table 3.
 
Biomarkers Extracted From D2, Including RPE65-IRD Patients, Before Treatment and From a Healthy Control Group
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×