Fully Automated Deep Learning Segmentation of RPE Loss and Ellipsoid Zone Integrity in Geographic Atrophy on Real-World OCT

arXiv:2606.31502 · cs.CV · Submitted 2026-06-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Fully Automated Deep Learning Segmentation of RPE Loss and Ellipsoid Zone Integrity in Geographic Atrophy on Real-World OCT".

Jane: Here's the gist of the paper. Fully automated high-precision segmentation of retinal atrophy and ellipsoid zone thickness in OCT provides a reliable tool for real-world geographic atrophy monitoring The proposed pipeline uses three specialized semantic segmentation models to delineate RPE loss, EZ boundaries, and Bruch’s membrane Results demonstrated high segmentation accuracy (Dice RPE loss: 0.88, Dice EZ loss: 0.87) and a strong reproducibility score (ICC > 0.98).

Tom: Pruning redundant reasoning tokens Geographic atrophy secondary to age-related macular degeneration requires precise monitoring of relevant structural biomarkers to assess disease stage, progression, and treatment response High-precision automated segmentation of RPE loss and EZ atrophy represents the advanced ”dry” stage of AMD, characterized by progressive loss of retinal pigment epithelium (RPE), overlying photoreceptors (PRs), and the underlying choriocapillaris The method has been validated on an independent real-world dataset How the training signal is built The models were developed on a diverse dataset of 298 SD-OCT volumes representing the full phenotypic spectrum of AMD (GA:222, intermediate AMD: 40, neovascular AMD: 17, healthy: 19) and validated on an independent external dataset (n=43) Data selection was curated for maximum phenotypic variability, including a broad range of lesion sizes and edge cases challenging prior automated segmentation iterations Ground truth was established by four experienced graders according to a clinically validated protocol, defining RPE loss as the complete absence of the EZ band or downward displacement of the outer nuclear layer (ONL) and outer plexiform layer (OPL) Where the gains come from The framework utilizes three semantic segmentation models: 1.

Jane: BM segmentation model: Implemented using a 2D CNN UNet++ featuring an EfficientNet-B5 encoder to predict BM positions per individual A-scan at subpixel resolution 2.

Tom: EZ layer segmentation model: Utilizes a 2D U-Net architecture with a DenseNet-201 encoder to segment the EZ across three distinct anatomical regions 3.

Paper summary: Tom: So we're looking at this paper today: "Fully Automated Deep Learning Segmentation of RPE Loss and Ellipsoid Zone Integrity in Geographic Atrophy on Real-World OCT." Basically, they've built a system to automatically find and measure retinal pigment epithelium loss, ellipsoid zone loss, and thinning in geographic atrophy using optical coherence tomography.

Jane: That sounds really precise. So what’s the main idea behind this approach? What are they trying to achieve with this automated segmentation?

Tom: The thesis is that geographic atrophy in age-related macular degeneration needs very accurate monitoring of these structural markers to figure out how bad the disease is, how it's progressing, and if a treatment is working. So instead of someone manually looking at every scan, they propose a fully automated pipeline using three different semantic segmentation models to draw the lines for RPE loss and EZ boundaries and Bruch’s membrane.

Lu: I'm interested in how they structured those three models because that's where the real depth is, you know, how you handle all those different types of data input.

Meng: From a deployment standpoint, it sounds like a lot of complex architecture for what we want to be something that actually runs reliably on a clinic's hardware. What kind of complexity are we talking about here?

Lalam: This paper focuses on using three specific models: the BM segmentation model which uses a two dimensional convolutional neural network called a UNet plus with an EfficientNet B five encoder to predict Bruch’s membrane positions per individual scan at subpixel resolution.

Tom: Okay, so one of those models is focused on the Bruch’s membrane, using that UNet plus structure with the EfficientNet B five encoder to pinpoint its location on each scan very finely.

Jane: And then they have a second model for segmenting the EZ layer itself, which uses a two dimensional U-Net architecture with a DenseNet twenty one encoder to segment the EZ across three different anatomical regions.

Lu: That DenseNet twenty one architecture is interesting because it suggests a strong focus on capturing features in dense data, and they use it to delineate the EZ layer across those distinct anatomical areas.

Meng: Three models sounds like high computational load, which makes me wonder about the inference time for real-time clinical use. How fast is this whole process supposed to run?

Paper summary: Tom: They’ve actually developed a third model for RPE loss segmentation, and that one uses a three dimensional to two dimensional Feature Projection Network called an FPN to transform full three dimensional OCT volumes into a two dimensional en face map.

Jane: So, they are tackling the RPE loss by projecting the whole three dimension volume down into a flat image where they can segment it. That sounds like a big step in handling that volumetric information.

Lalam: They trained these models on a diverse dataset of two hundred ninety eight SD OCT volumes representing the full spectrum of age related macular degeneration, including geographic atrophy, intermediate am amd, neovascular am amd, and healthy scans.

Tom: A really broad phenotypic range for training is important. So they validated this entire framework using an independent external dataset of forty three volumes to make sure it wasn't just overfitting to the initial training pool.

Jane: And how did they define what counts as RPE loss in their ground truth? That’s a crucial point because 'loss' can be defined in different ways depending on who you ask.

Tom: They defined RPE loss as the complete absence of the EZ band or a downward displacement, or subsidence, of the outer nuclear layer and outer plexiform layer.

Lu: I like that definition because it ties the structural loss directly to observable changes in those specific retinal layers. It’s concrete.

Meng: But what about clinical variability? If you're using this in a real clinic setting, how does it handle different lesion sizes or scan densities? That’s where models often fail when they move from research to reality.

Tom: They did test for robustness by stratifying their analysis by lesion size and looked at the relevance of OCT b scan density on measurement accuracy across standard B scan patterns like thirty one four, two three nine, and twelve two microns.

Jane: So they checked if the performance holds up when the lesion is very small or when there are gaps in the B scans between slices. What did those tests show about that?

Tom: For small lesions with a size less than zero point five square millimeter, metrics like disc score were strongly influenced by that size, but surface distance metrics like HD nine five and ASSD proved more robust to those variations.

Lu: That’s a key finding for representation learning because it tells us which features are truly stable across different scales in the image.

Paper summary: Meng: When you look at the performance numbers, what are the actual scores they report for RPE loss and EZ loss area measurements? I need to know if this is actually hitting targets.

Tom: They achieved a mean disc score of zero point eight eight for RPE loss and zero point eight seven for EZ loss. And their accuracy of those area measurements correlated very highly with the manual ground truth, with a pearson r of nine nine nine for RPE loss and nine nine six for EZ loss.

Jane: A correlation that high is impressive, especially when you factor in that inter reader reliability test they did on the area measurements using icc two one values around zero point nine eight five.

Lu: That inter reader reliability test is a good check because it means the system isn't just accurate against one set of eyes, but consistently reproducible across different people.

Meng: But there was still a slight systematic bias mentioned in the Deming regression analysis; they saw a marginal over segmentation of RPE loss area by zero point three millimeters and an under segmentation of EZ loss area by negative zero point three five millimeters. That’s something we have to consider when deploying it.

Tom: It’s a small offset, but it shows that even this high-precision tool isn't perfectly neutral; the prediction might be slightly too big or slightly too small depending on what you're measuring.

Jane: So, looking at the overall implications for someone who just watches this show, what does this automated segmentation actually change about how we monitor geographic atrophy in the real world?

Tom: It means that with a fully automated system like this pipeline—the one they described in "Fully Automated Deep Learning Segmentation of RPE Loss and Ellipsoid Zone Integrity in Geographic Atrophy on Real-World OCT"—we can get high precision for tracking structural loss without needing constant manual oversight.

Lu: The implication is that we could potentially alleviate the need for burdensome microperimetry by providing a direct translation of structural data into functional insights for routine clinical practice.

Meng: From a production view, if you can automate this, you reduce the human error rate significantly in the initial assessment stage, which is where most of our pipeline failures happen.

Jane: So it moves the focus from manual interpretation to automated measurement, and they claim that optimized EZ thickness assessments could help clinicians get functional insights for routine practice.

Tom: Exactly. They’ve created a reliable tool for real world geographic atrophy monitoring by using this fully automated deep learning segmentation framework in OCT images >

Conclusion: Tom: So we've been looking at this paper, "Fully Automated Deep Learning Segmentation of RPE Loss and Ellipsoid Zone Integrity in Geographic Atrophy on Real-World OCT." Basically, they've built a system to automatically find and measure retinal pigment epithelium loss, ellipsoid zone loss, and thinning in geographic atrophy using optical coherence tomography.

Jane: That sounds really precise. So what’s the main idea behind this approach? What are they trying to achieve with this automated segmentation?

Tom: The thesis is that geographic atrophy in age-related macular degeneration needs very accurate monitoring of these structural markers to figure out how bad the disease is, how it's progressing, and if a treatment is working. So instead of someone manually looking at every scan, they propose a fully automated pipeline using three different semantic segmentation models to draw the lines for RPE loss and EZ boundaries and Bruch’s membrane.

Lu: I'm interested in how they structured those three models because that's where the real depth is, you know, how you handle all those different types of data input.

Meng: From a deployment standpoint, it sounds like a lot of complex architecture for what we want to be something that actually runs reliably on a clinic's hardware. What kind of complexity are we talking about here?

Lalam: This paper focuses on using three specific models: the BM segmentation model which uses a two dimensional convolutional neural network called a UNet plus with an EfficientNet B five encoder to predict Bruch’s membrane positions per individual scan at subpixel resolution.

Jane: Okay, so one of those models is focused on the Bruch’s membrane, using that UNet plus structure with the EfficientNet B five encoder to pinpoint its location on each scan very finely.

Tom: And then they have a second model for segmenting the EZ layer itself, which uses a two dimensional U-Net architecture with a DenseNet twenty one encoder to segment the EZ across three different anatomical regions.

Lu: That DenseNet twenty one architecture is interesting because it suggests a strong focus on capturing features in dense data, and they use it to delineate the EZ layer across those distinct anatomical areas.

Meng: Three models sounds like high computational load, which makes me wonder about the inference time for real-time clinical use. How fast is this whole process supposed to run?

Tom: They’ve actually developed a third model for RPE loss segmentation, and that one uses a three dimensional to two dimensional Feature Projection Network called an FPN to transform full three dimensional OCT volumes into a two dimensional en face map.

Wrap-up: Jane: So, they are tackling the RPE loss by projecting the whole three dimension volume down into a flat image where they can segment it. That sounds like a big step in handling that volumetric information.

Lalam: They trained these models on a diverse dataset of two hundred ninety eight SD OCT volumes representing the full spectrum of age related macular degeneration, including geographic atrophy, intermediate am amd, neovascular am amd, and healthy scans.

Tom: A really broad phenotypic range for training is important. So they validated this entire framework using an independent external dataset of forty three volumes to make sure it wasn't just overfitting to the initial training pool.

Jane: And how did they define what counts as RPE loss in their ground truth? That’s a crucial point because 'loss' can be defined in different ways depending on who you ask.

Tom: They defined RPE loss as the complete absence of the EZ band or a downward displacement, or subsidence, of the outer nuclear layer and outer plexiform layer.

Lu: I like that definition because it ties the structural loss directly to observable changes in those specific retinal layers. It’s concrete.

Meng: But what about clinical variability? If you're using this in a real clinic setting, how does it handle different lesion sizes or scan densities? That’s where models often fail when they move from research to reality.

Tom: They did test for robustness by stratifying their analysis by lesion size and looked at the relevance of OCT b scan density on measurement accuracy across standard B scan patterns like thirty one four, two three nine, and twelve two microns.

Jane: So they checked if the performance holds up when the lesion is very small or when there are gaps in the B scans between slices. What did those tests show about that?

Tom: For small lesions with a size less than zero point five square millimeter, metrics like disc score were strongly influenced by that size, but surface distance metrics like HD nine five and ASSD proved more robust to those variations.

Lu: That’s a key finding for representation learning because it tells us which features are truly stable across different scales in the image.

Meng: When you look at the performance numbers, what are the actual scores they report for RPE loss and EZ loss area measurements? I need to know if this is actually hitting targets.

Wrap-up: Tom: They achieved a mean disc score of zero point eight eight for RPE loss and zero point eight seven for EZ loss. And their accuracy of those area measurements correlated very highly with the manual ground truth, with a pearson r of nine nine nine for RPE loss and nine nine six for EZ loss.

Jane: A correlation that high is impressive, especially when you factor in that inter reader reliability test they did on the area measurements using icc two one values around zero point nine eight five.

Lu: That inter reader reliability test is a good check because it means the system isn't just accurate against one set of eyes, but consistently reproducible across different people.

Meng: But there was still a slight systematic bias mentioned in the Deming regression analysis; they saw a marginal over segmentation of RPE loss area by zero point three millimeters and an under segmentation of EZ loss area by negative zero point three five millimeters. That’s something we have to consider when deploying it.

Tom: It’s a small offset, but it shows that even this high-precision tool isn't perfectly neutral; the prediction might be slightly too big or slightly too small depending on what you're measuring.

Jane: So, looking at the overall implications for someone who just watches this show, what does this automated segmentation actually change about how we monitor geographic atrophy in the real world?

Tom: It means that with a fully automated system like this pipeline—the one they described in "Fully Automated Deep Learning Segmentation of RPE Loss and Ellipsoid Zone Integrity in Geographic Atrophy on Real-World OCT"—we can get high precision for tracking structural loss without needing constant manual oversight.

Lu: The implication is that we could potentially alleviate the need for burdensome microperimetry by providing a direct translation of structural data into functional insights for routine clinical practice.

Meng: From a production view, if you can automate this, you reduce the human error rate significantly in the initial assessment stage, which is where most of our pipeline failures happen.

Jane: So it moves the focus from manual interpretation to automated measurement, and they claim that optimized EZ thickness assessments could help clinicians get functional insights for routine practice.

Tom: Exactly. They’ve created a reliable tool for real world geographic atrophy monitoring by using this fully automated deep learning segmentation framework in OCT images, but we have to keep an eye on those small systematic offsets they found in the area measurements.

Wolf-Dieter Vogl, Hlynur Skulason, Oliver Leingang, Ursula Schmidt-Erfurth, Amir Sadeghipour, Ariadne Whitby

RetInSight GmbH · Laboratory of Opthalmic Image Analysis (OPTIMA) · Medical University of Vienna

cs.CV

Submitted: 2026-06-30

Updated: 2026-11-16

Comments: 35 pages, 6 tables, 8 figures, contain 3 supplemental figures and 2 supplemental tables

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 83/100

The gist: Fully automated high-precision segmentation of retinal atrophy and ellipsoid zone thickness in OCT provides a reliable tool for real-world geographic atrophy monitoring The proposed pipeline uses…

Key concepts

RPE Loss
This refers to the progressive damage or complete absence of the retinal pigment epithelium, which is a layer of cells situated beneath the photoreceptors. In this study, it was defined as the complete lack of the ellipsoid zone band or any downward displacement of deeper retinal layers.
Ellipsoid Zone (EZ) Integrity
The ellipsoid zone is a critical structure composed of photoreceptor segments that connect to Bruch's membrane. Segmenting this zone helps assess the structural health and integrity of the photoreceptors, which is a key biomarker for Geographic Atrophy (GA).
Semantic Segmentation Models
These are advanced AI models (like UNet++ and DenseNet-201) trained to automatically classify every pixel in an OCT image into specific categories, such as RPE loss or EZ boundaries. This allows the system to delineate complex structures with high precision.
Geographic Atrophy (GA)
GA is a form of Age-related Macular Degeneration characterized by progressive, patchy loss of retinal pigment epithelium and overlying photoreceptors. The study focuses on using automated segmentation to accurately measure this structural loss for monitoring disease progression.

Terminology

Summary

Fully automated high-precision segmentation of retinal atrophy and ellipsoid zone thickness in OCT provides a reliable tool for real-world geographic atrophy monitoring The proposed pipeline uses three specialized semantic segmentation models to delineate RPE loss, EZ boundaries, and Bruch’s membrane Results demonstrated high segmentation accuracy (Dice RPE loss: 0. p. 0

Where the paper says it

This paper presents a fully automated, deep learning-based framework for the high-precision, pixel-wise segmentation of key biomarkers in optical coherence tomography (OCT) imaging: retinal pigment epithelium (RPE) loss, ellipsoid zone (EZ) loss, and EZ thinning. The proposed pipeline uses three specialized semantic segmentation models to delineate RPE loss, EZ boundaries (including interruptions), and Bruch’s membrane.

88, Dice EZ loss: 0. p. 0
Where the paper says it

The comprehensive evaluation was further strengthened using additional datasets to assess repeatability, inter-reader reliability, the impact of B-scan density on measurement accuracy, and subgroup performance stratified by lesion size. Results demonstrated high segmentation accuracy (Dice RPE loss: 0.

87) and a strong reproducibility score (ICC > 0. p. 0
Where the paper says it

Total EZ thickness measurements exhibited a sub-pixel average deviation of 2. 15µm, and segmentation reliability was confirmed by a strong reproducibility score (ICC > 0.

98).

Pruning redundant reasoning tokens

Geographic atrophy secondary to age-related macular degeneration requires precise monitoring of relevant structural biomarkers to assess disease stage, progression, and treatment response High-precision automated segmentation of RPE loss and EZ atrophy represents the advanced ”dry” stage of AMD, characterized by progressive loss of retinal pigment epithelium (RPE), overlying photoreceptors (PRs), and the underlying choriocapillaris The method has been validated on an independent real-world dataset

How the training signal is built

The models were developed on a diverse dataset of 298 SD-OCT volumes representing the full phenotypic spectrum of AMD (GA:222, intermediate AMD: 40, neovascular AMD: 17, healthy: 19) and validated on an independent external dataset (n=43) Data selection was curated for maximum phenotypic variability, including a broad range of lesion sizes and edge cases challenging prior automated segmentation iterations Ground truth was established by four experienced graders according to a clinically validated protocol, defining RPE loss as the complete absence of the EZ band or downward displacement of the outer nuclear layer (ONL) and outer plexiform layer (OPL)

Where the gains come from

The framework utilizes three semantic segmentation models:

  1. BM segmentation model: Implemented using a 2D CNN UNet++ featuring an EfficientNet-B5 encoder to predict BM positions per individual A-scan at subpixel resolution

  2. EZ layer segmentation model: Utilizes a 2D U-Net architecture with a DenseNet-201 encoder to segment the EZ across three distinct anatomical regions

  3. RPE loss segmentation model: Employs a 3D-to-2D Feature Projection Network (FPN) to transform full 3D OCT volumes into a 2D en-face segmentation map

Performance and Validation Metrics

The overall model achieved a mean DSC of 0.88 for RPE loss and 0. p. 0

Where the paper says it

com Abstract Geographic atrophy (GA) secondary to age-related macular degeneration (AMD) requires precise monitoring of relevant structural biomarkers to assess disease stage, progression, and treatment response. This paper presents a fully automated, deep learning-based framework for the high-precision, pixel-wise segmentation of key biomarkers in optical coherence tomography (OCT) imaging: retinal pigment epithelium (RPE) loss, ellipsoid zone (EZ) loss, and EZ thinning.

87 for EZ loss Accuracy of RPE-loss and EZ-loss area measurements showed high correlation with manual ground-truth annotations, with Pearson’s r of 0. p. 5
Where the paper says it

To enhance segmentation robustness and ensure anatomical consistency, we integrated the model outputs using a heuristic constraint: since RPE loss in GA regularly occurs within regions of EZ loss39;40, RPE loss predictions were dropped in A-scans where the corresponding EZ model did not indicate a loss. 4 Evaluation and Statistical Analysis Accuracy of RPE-loss and EZ-loss area measurement The performance of RPE-loss and EZ-loss area measurements was evaluated against manual ground-truth (GT) annotations from the validation dataset.

999 for RPE loss and 0. p. 9
Where the paper says it

999 (95% CI: 0. 0) for RPE loss and 0.

996 for EZ loss Layer segmentation accuracy was confirmed by mean differences between manual and predicted positions for the IB-EZ, OB-OPR, and total EZ thickness being less than a single pixel given the pixel spacing Inter-reader reliability was assessed using ICC(2,1) values of 0. p. 9
Where the paper says it

35 mm2 (95% CI: -0. Accuracy of layer segmentation and EZ thickness measurement: As summarized in Table 3, the mean differences between manual and predicted positions for the IB-EZ, OB-OPR, and total EZ thickness were 3.

985 for RPE loss area measurements

Robustness to Clinical Variability

The evaluation included several rigorous tests to ensure generalizability, such as a stratified analysis by lesion size and an analysis of the relevance of OCT B-Scan density on measurement accuracy For small lesions (≤ 0.5mm2), performance metrics like DSC were found to be strongly influenced by lesion size, but surface distance metrics like HD95 and ASSD proved more robust to size variations The study confirmed that the algorithm is a robust tool for automated GA assessment in both research and clinical settings

Conclusion

We have elaborated an accurate, fully automated, and reliable method for segmenting key GA features - RPE loss, EZ loss, and EZ thickness - in OCT images The speed and reliability of a fully automated, high-precision algorithm eliminate the need for simultaneous manual oversight Optimized EZ thickness assessments could alleviate the need for burdensome microperimetry, providing clinicians with a direct translation of structural data into functional insights for routine practice

References

1 arXiv:2606.31502v1 [cs. p. 0

Where the paper says it

Geographic 1 arXiv:2606. 31502v1 [cs.

CV] 30 Jun 2026

2 Laboratory of Opthalmic Image Analysis (OPTIMA), Medical University of Vienna, Austria

3 RetInSight GmbH, Vienna, Austria

4 Wolf-Dieter Vogl1, Hlynur Skulason1, Oliver Leingang1, Ursula Schmidt-Erfurth2, Amir Sadeghipour1 and Ariadne Whitby1 2 Methods

5 Spectralis device (Heidelberg Engineering GmbH, Heidelberg, Germany)

6 FILLY Phase-II clinical trial(NCT02503332)

7 Macula Clinic at the Medical University of Vienna (MUV)

8 Vienna Imaging Biomarker Eye Study (VIBES) real-world registry

9 OAKS and DERBY Phase-III clinical trials(NCT03525613 and NCT03525600)

10 Declaration of Helsinki, with informed consent obtained from all participants. Scans were fovea-centered with patterns ranging from 49 × 512 to 97 × 1024 (B-scans × A-scans), covering an approximate 6 × 6mm (20°) field of view. p. 2

Where the paper says it

All studies adhered to the Declaration of Helsinki, with informed consent obtained from all participants. Scans were fovea-centered with patterns ranging from 49 × 512 to 97 × 1024 (B-scans × A-scans), covering an approximate 6 × 6mm (20°) field of view.

11 RPE loss was defined as the complete absence of the EZ band, downward displacement (subsidence) of the outer nuclear layer (ONL) and outer plexiform layer (OPL), and substantial irregularities or absence of the typically continuous hyperreflective RPE band

12 Fig. 1. Annotation example. p. 3

Where the paper says it

High-precision automated segmentation of RPE loss and EZ RPE loss EZ loss EZ layer OB-OPR layer BM layer Fig. Annotation example.

En-face view (left) and central B-scan (right) showing retinal pigment epithelium (RPE) loss (blue) and ellipsoid zone (EZ) loss (green) annotations, as well as the layer annotations of inner boundary of the ellipsoid zone (IB-EZ), and outer boundary of outer photoreceptors (OB-OPR). p. 3
Where the paper says it

Annotation example. En-face view (left) and central B-scan (right) showing retinal pigment epithelium (RPE) loss (blue) and ellipsoid zone (EZ) loss (green) annotations, as well as the layer annotations of inner boundary of the ellipsoid zone (IB-EZ) in light red, outer boundary of outer photoreceptors (OB-OPR) in dark red, and Bruch’s membrane (BM) in yellow.

In light red, Bruch’s membrane (BM) in yellow

13 Fig. 2. Data flow diagram describing the inputs and outputs of each model. p. 4

Where the paper says it

High-precision automated segmentation of RPE loss and EZ Model input Model output SD-OCT scan Segmentations, Biomarkers Bruch's membrane (BM) segmentation model EZ layer segmentation model RPE loss segmentation model BM segmentation Post processing RPE loss segmentation BM segmentation EZ layer segmentation Fig. Data flow diagram describing the inputs and outputs of each model.

14 BM segmentation was implemented using the SD-Layernet framework leveraging a topological engine to incorporate anatomical priors

15 EZ layer segmentation model The EZ was segmented using a 2D U-Net architecture featuring a DenseNet-201 encoder

16 RPE loss segmentation model RPE loss was segmented using a 3D-to-2D Feature Projection Network (FPN)

17 Post processing During post-processing, the probability maps from the five cross-validation models for RPE loss were averaged to form an ensemble prediction

18 Accuracy of RPE-loss and EZ-loss area measurement The performance of RPE-loss and EZ-loss area measurements was evaluated against manual ground-truth (GT) annotations from the validation dataset

19 Measurement accuracy was quantified using absolute deviation (AD) and absolute percentage deviation (PD): AD = Apred − Agt, PD = AD / Agt × 100, where Apred and Agt are predicted and GT loss areas, respectively

20 To assess the relationship between predicted and GT areas, Pearson’s R correlation was employed, while Bland-Altman analysis was used to identify potential proportional bias relative to lesion size

21 Additionally, Deming regression (DR) was performed to evaluate systematic bias, while accounting for potential measurement errors within the ground truth

22 Accuracy of layer segmentations and EZ thickness measurement: We evaluated the segmentation performance of the EZ layers using the external validation dataset For each A-scan within the B-scans, manual GT annotations were provided for the IB-EZ and OB-OPR These same positions were predicted by the automated segmentation model

23 Inter-reader reliability: To assess inter-reader reliability, two independent readers who were not involved in the initial labeling process annotated RPE and EZ loss on a subset of the validation dataset (annotation group AN2) These readers were trained by the expert readers that provided the GT (AN 1) according to the established protocol; however, all AN 2 annotations were performed without further supervision

24 Inter-reader reliability was assessed using a two-way random-effects model to estimate the intraclass correlation coefficients (ICC) for single-score absolute agreement(ICC(2,1))

25 Reproducibility: To evaluate reproducibility, we compared paired RPE and EZ loss areas between the screening and baseline visits of the reproducibility dataset Agreement was assessed using ICC(2,1), LoA (Bland-Altman plots), and mean deviation (DR)

26 B-scan density: Table 6 reports the MAE and MAPE for RPE loss, EZ loss, and their progression areas across standard B-scan patterns Corresponding B-scan spacings were 314, 239, 122, 61, 47, and 31 µm

27 GA lesion progression was slow during the 3-month interval, making the progression area small relative to overall lesion size (e.g., 0.32 vs. 5.38 mm2 for RPE). p. 8

Where the paper says it

Dataset (Source) Training External validation (Vibes) Reproducibility (OAKS & DERBY) Inter-reader reliability validation (Vibes) B-scan density highresolution scans Number of patients / eyes / OCTs 298 / 229 / 221 43 / 43 / 43 68 / 68 / 136 25 / 25 / 25 44 / 61 / 106 # annotated B-scans RPE loss: 14602 — EZ layer: 2910 2156 N/A 1189 N/A Age mean ± std N/A 78. 03 Gender female % 63 55 71 N/A 59 RPE loss area [mm2] 6.

Combined with large B-scan spacings, this resulted in high percentage errors

28 The primary cause of these errors is that the physical lesion progression is smaller than the spacing between slices, rendering the changes unobservable

29 Data sets and baseline characteristics: For model training and internal validation, 298 OCT volumes from 229 eyes of 221 patients were used This cohort comprised 225 scans with GA, 40 with iAMD and drusen, 17 with nAMD, and 19 healthy controls The external test set consisted of 43 volumes from 43 eyes of 43 patients, all diagnosed with GA Additionally, 68 longitudinal OCT volume pairs of 68 eyes from 68 patients in the OAKS and DERBY trials were included for the reproducibility study Finally, for the B-scan density study, 61 baseline scans from 44 patients and 46 3-month follow-up scans from 31 patients were included

30 Detailed baseline characteristics for all datasets are summarized in Table 1.

31 Lesion size is reported as mean ± std — median ± IQR. p. 8

Where the paper says it

Population characteristics for the datasets used for training, external validation, as well as reproducibility, inter-reader reliability, and B-scan density analyses. Lesion size is reported as mean ± std — median ± IQR.

32 Accuracy of RPE-loss and EZ-loss area measurement: Performance metrics are summarized in Table 2, with their distributions visualized via boxplots in Fig. p. 7

Where the paper says it

Detailed baseline characteristics for all datasets are summarized in Table 1. Accuracy of RPE-loss and EZ-loss area measurement: Performance metrics are summarized in Table 2, with their distributions visualized via boxplots in Fig.

3 Overall, the model achieved a mean DSC of 0. p. 7
Where the paper says it

Detailed subset analyses of the stratified data are provided in the Supplemental Material (Table S1, Figure S1). Overall, the model achieved a mean DSC of 0.

88 (95% CI: 0.84 to 0.90) for RPE loss and 0. p. 0
Where the paper says it

com Abstract Geographic atrophy (GA) secondary to age-related macular degeneration (AMD) requires precise monitoring of relevant structural biomarkers to assess disease stage, progression, and treatment response. This paper presents a fully automated, deep learning-based framework for the high-precision, pixel-wise segmentation of key biomarkers in optical coherence tomography (OCT) imaging: retinal pigment epithelium (RPE) loss, ellipsoid zone (EZ) loss, and EZ thinning.

87 (95% CI: 0.83 to 0.89) for EZ loss Stratification by lesion size revealed the correlation between loss area and overall DSC; performance for RPE loss ranged from 0. p. 7
Where the paper says it

89) for EZ loss. Stratification by lesion size revealed the correlation between loss area and overall DSC; performance for RPE loss ranged from 0.

84 for small lesions to 0. p. 7
Where the paper says it

Stratification by lesion size revealed the correlation between loss area and overall DSC; performance for RPE loss ranged from 0. 84 for small lesions to 0.

96 for large lesions, while EZ loss DSC ranged from 0. p. 7
Where the paper says it

84 for small lesions to 0. 96 for large lesions, while EZ loss DSC ranged from 0.

76 to 0.95

33 Table 1: Population characteristics for the datasets used for training, external validation, as well as reproducibility, inter-reader reliability, and B-scan density analyses. Lesion size is reported as mean ± std — median ± IQR

34 Table 2: Comparison of the total RPE-loss and EZ-loss area measurements between manual ground truth annotation and automated segmentation. Numbers in brackets are the 95% confidence interval (CI)

35 Metric RPE loss EZ loss Pearson’s r 0.999 [0.998 to 1.000] 0.996 [0.993 to 0.998] DR Intercept [mm2] 0. p. 9

Where the paper says it

Metric RPE loss EZ loss Pearson’s r 0. 998] DR Intercept [mm2] 0.

14 [0.24 to 15] - - -

36 Mean [7764 ± 3389 — median ± IQR] — median ± IQR

37 Area difference [mm2] AD loss area [mm2] 0.251 ± 0.222 — 0.160 ± 0.268

38 PD loss area [%] 3.33 ± 2.97 — median ± IQR

41 Bland-Altman Plot (top) and Deming regression fit (bottom) from automated segmentation compared to manual ground truth annotation for RPE loss areas (left) and EZ loss areas (right)

42 Pearson's r = 0.996 [0.993 to 0.998]

43 Deming regression parameters and Bland-Altman analysis (Fig. 4) indicated a slight systematic bias: a marginal over-segmentation of the RPE loss area (mean difference: 0. p. 9

Where the paper says it

Deming regression parameters and Bland-Altman analysis (Fig. 4) indicated a slight systematic bias: a marginal over-segmentation of the RPE loss area (mean difference: 0.

3 mm2 (95% CI: 0. p. 8
Where the paper says it

Dataset (Source) Training External validation (Vibes) Reproducibility (OAKS & DERBY) Inter-reader reliability validation (Vibes) B-scan density highresolution scans Number of patients / eyes / OCTs 298 / 229 / 221 43 / 43 / 43 68 / 68 / 136 25 / 25 / 25 44 / 61 / 106 # annotated B-scans RPE loss: 14602 — EZ layer: 2910 2156 N/A 1189 N/A Age mean ± std N/A 78. 03 Gender female % 63 55 71 N/A 59 RPE loss area [mm2] 6.

17 to 0.43)) and a marginal under-segmentation of the EZ loss area (-0. p. 9
Where the paper says it

Deming regression parameters and Bland-Altman analysis (Fig. 4) indicated a slight systematic bias: a marginal over-segmentation of the RPE loss area (mean difference: 0.

35 mm2 (95% CI: -0. p. 8
Where the paper says it

Dataset (Source) Training External validation (Vibes) Reproducibility (OAKS & DERBY) Inter-reader reliability validation (Vibes) B-scan density highresolution scans Number of patients / eyes / OCTs 298 / 229 / 221 43 / 43 / 43 68 / 68 / 136 25 / 25 / 25 44 / 61 / 106 # annotated B-scans RPE loss: 14602 — EZ layer: 2910 2156 N/A 1189 N/A Age mean ± std N/A 78. 03 Gender female % 63 55 71 N/A 59 RPE loss area [mm2] 6.

64 to-0.068))

44 Accuracy of layer segmentations and EZ thickness measurement: As summarized in Table 3, the mean differences between manual and predicted positions for the IB-EZ, OB-OPR, and total EZ thickness were 3.33 µm, 5.17 µm, and 2. p. 7

Where the paper says it

12) libraries. Relevance of B-scan density on measurements: Spectralis OCT systems allow for variable B-scan spacing ranging from 11 to 300 µm, corresponding to 193 down to 19 B-scans in a 6 mm volumetric scan (standard defaults: 49 for high-speed; 97 for high-resolution).

15 µm respectively Given a pixel spacing of 3. p. 9
Where the paper says it

15 µm, respectively. Given a pixel spacing of 3.

9 µm, the average deviations for both IB-EZ position and EZ thickness are less than a single pixel, demonstrating high accuracy in an OCT setting

45 Table S 1: Comparison of the total RPE-loss (top) and EZ-loss (bottom) area measurements between manual ground truth annotation and automated segmentation; stratified by RPE loss area and EZ loss area

46 Stratification RPE loss area ≤ 0.5 [mm2] ≤ 2. p. 8

Where the paper says it

Dataset (Source) Training External validation (Vibes) Reproducibility (OAKS & DERBY) Inter-reader reliability validation (Vibes) B-scan density highresolution scans Number of patients / eyes / OCTs 298 / 229 / 221 43 / 43 / 43 68 / 68 / 136 25 / 25 / 25 44 / 61 / 106 # annotated B-scans RPE loss: 14602 — EZ layer: 2910 2156 N/A 1189 N/A Age mean ± std N/A 78. 03 Gender female % 63 55 71 N/A 59 RPE loss area [mm2] 6.

5 [mm2] ≤ 10 [mm2] > 10 [mm2]

47 Metric mean ± std — median ± IQR AD loss area [mm2] 0.06 ± 0.05 — median ± IQR

48 PD loss area [%] 23.31 ± 12.22 — median ± IQR

49 Area Prediction [mm2] 0.29 ± 0.23 — median ± IQR

50 Area GT [mm2] 0.25 ± 0.19 — median ± IQR

51 DSC 0.87 ± 0.18 — median ± IQR

52 HD95 [mm] 0.37 ± 0.19 — median ± IQR

53 ASSD [mm] 0.06 ± 0.12 — median ± IQR

54 Sensitivity 0.72 ± 0.14 — median ± IQR

55 Specificity 1.00 ± 0.99 — median ± IQR

56 Precision 13

Supplemental Content: Supplemental Figures

57 Figure S 1. Segmentation performance stratified by lesion size comparing manual and automated segmentations of RPE loss (top) and EZ loss (bottom). p. 8

Where the paper says it

Distribution of segmentation evaluation metrics comparing manual and automated segmentations of RPE loss (blue) and EZ loss (green) in the external validation dataset. Reported metrics are the absolute difference between measured and annotated lesion size, Dice similarity coefficient (DSC) for segmentation overlap, Hausdorff distance 95th percentile (HD95) 95% and average symmetric surface distance (ASSD) for lesion surface accuracy.

Reported metrics are the absolute difference between measured and annotated lesion size, DSC for segmentation overlap, Hausdorff distance 95th percentile (HD95) and average symmetric surface distance (ASSD) for lesion surface accuracy

58 Figure S 2. Deming regression fits for RPE loss area(left) and EZ loss area (right) inter-reader reliability comparing reader groups AN1 vs AN2 (top), AN1 vs automated segmentation (center) and AN2 vs automated segmentation (bottom)

59 Figure S 3. Effect of B-scan density on RPE loss and EZ loss measurement. p. 7

Where the paper says it

Relevance of B-scan density on measurements: Spectralis OCT systems allow for variable B-scan spacing ranging from 11 to 300 µm, corresponding to 193 down to 19 B-scans in a 6 mm volumetric scan (standard defaults: 49 for high-speed; 97 for high-resolution). To determine the effect of B-scan density on the quantification accuracy of EZ and RPE loss, we first applied our segmentation algorithm to high-density scans (193 B-scans) to establish ground-truth reference area measurements.

Boxplot showing distribution of measurement errors of RPE loss area, RPE loss growth area growth, EZ loss areas, and EZ loss growth area (top to bottom), comparing synthetically downsampled segmentations in the range of 19 to 128 B-Scans to ground-truth segmentation with 193 B-scans

60 Table S 1: Comparison of the total RPE-loss (top) and EZ-loss (bottom) area measurements between manual ground truth annotation and automated segmentation; stratified by RPE loss area and EZ loss area

61 Stratification EZ loss area ≤ 2.5 [mm2] ≤ 7. p. 8

Where the paper says it

Dataset (Source) Training External validation (Vibes) Reproducibility (OAKS & DERBY) Inter-reader reliability validation (Vibes) B-scan density highresolution scans Number of patients / eyes / OCTs 298 / 229 / 221 43 / 43 / 43 68 / 68 / 136 25 / 25 / 25 44 / 61 / 106 # annotated B-scans RPE loss: 14602 — EZ layer: 2910 2156 N/A 1189 N/A Age mean ± std N/A 78. 03 Gender female % 63 55 71 N/A 59 RPE loss area [mm2] 6.

5 [mm2] ≤ 15 [mm2] > 15 [mm2]

62 Metric mean ± std — median ± IQR AD loss area [mm2] 0.11 ± 0.09 — median ± IQR

63 PD loss area [%] 10.69 ± 7.68 — median ± IQR

64 Area Prediction [mm2] 1.29 ± 0.75 — median ± IQR

65 Area Groundtruth [mm2] 1.40 ± 0.74 — median ± IQR

66 DSC 0.87 ± 0.18 — median ± IQR

67 HD95 [mm] 0.37 ± 0.19 — median ± IQR

68 ASSD [mm] 0.06 ± 0.12 — median ± IQR

69 Sensitivity 0.72 ± 0.14 — median ± IQR

70 Specificity 1

Supplemental Content: Supplemental Tables

71 Table S 2: EZ loss / RPE loss area ratio split into 4 quartiles. Areas are based on segmentation measurements of all Spectralis scans of the OAKS & DERBY study eyes

72 Q1 < 1.37; Q2 ≥ 1.37 to < 1.67; Q3 ≥ 1.67 to < 2.17; Q4 ≥ 2.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Automated Geometric Feature Quantification for GA Monitoring

The improved system can perform high-precision, pixel-wise segmentation of key biomarkers in optical coherence tomography (OCT) imaging: retinal pigment epithelium (RPE) loss, ellipsoid zone (EZ) loss, and EZ thinning. This enables the accurate and consistent quantification of outer photoreceptor degeneration and RPE loss in both clinical trials and routine real-world ophthalmic care.

  1. Robust Generalizability Across Diverse Phenotypes

The system can achieve high generalizability by being trained on a diverse dataset spanning the full spectrum of AMD: from early-onset to latestage GA, as well as intermediate AMD (iAMD), neovascular AMD (nAMD), and healthy controls. This prevents reliance on clean data and improves performance when applied to diverse populations.

  1. Enhanced Measurement Accuracy via Multi-Model Integration

The system utilizes an ensemble of three specialized semantic segmentation models—for Bruch’s membrane, EZ layer boundaries, and RPE loss—whose outputs are integrated through post-processing using a heuristic constraint: since RPE loss in GA regularly occurs within regions of EZ loss. This integration ensures high accuracy in both area measurement and layer position.

  1. Objective Assessment of Inter-Reader Reliability

The system incorporates an assessment of inter-reader reliability using intraclass correlation coefficients (ICC), which shows that the algorithm performs on par with human annotators while offering the advantage of perfectly consistent results. This provides a reliable measure for assessing the performance consistency across different clinical annotator groups.

  1. Density-Dependent Error Correction for Longitudinal Monitoring

The system can be adapted to account for B-scan density effects by simulating lower densities and assessing measurement variance against ground truth. This allows it to accurately monitor slow GA progression, as it addresses the issue that the physical lesion progression is smaller than the spacing between slices.

Abstract

Purpose: To develop and validate a fully automated deep learning method for segmenting retinal pigment epithelium (RPE) loss, ellipsoid zone (EZ) loss, and EZ thickness in geographic atrophy (GA) on optical coherence tomography (OCT). Methods: Three convolutional neural networks segment Bruch's membrane, the EZ boundaries, and RPE loss in Spectralis OCT volumes; EZ loss and thickness are derived from the EZ boundaries. The networks were developed on 298 volumes (GA, intermediate and neovascular age-related macular degeneration, healthy) and evaluated against double-graded expert annotations in an independent real-world set of 43 eyes with GA. Further analyses assessed inter-reader reliability (25 eyes), test-retest repeatability between screening and baseline visits of two phase 3 trials (68 eyes), and the effect of B-scan density (61 eyes). Results: Mean Dice similarity coefficients were 0.88 (95% confidence interval [CI], 0.84-0.90) for RPE loss and 0.87 (95% CI, 0.83-0.89) for EZ loss; the mean absolute EZ thickness error was 2.08 μ m. Agreement with an independent reader group was similar to inter-reader agreement. Test-retest intraclass correlation coefficients were 0.995 (RPE loss) and 0.988 (EZ loss). Lower B-scan densities had little effect on lesion area but markedly increased the error in measuring 3-month change. Conclusions: On real-world OCT, the fully automated method was accurate, with agreement comparable to human readers, and reliable, with high test-retest repeatability. Translational Relevance: Fully automated, reproducible quantification of RPE and photoreceptor loss on routine OCT could support standardized structural endpoints in GA clinical trials and objective assessment of GA in routine ophthalmic care.

Related papers