Combination of optical coherence tomography-derived shape and texture features are associated with development of sub-foveal geographic atrophy in dry AMD

Abstract

Geographic atrophy (GA) is an advanced form of dry age-related macular degeneration (AMD) that leads to progressive and irreversible vision loss. Identifying patients with greatest risk of GA progression is important for targeted utilization of emerging therapies. This study aimed to comprehensively evaluate the role of shape-based fractal dimension features ($F_{\mathit{fd}}$) of sub-retinal pigment epithelium (sub-RPE) compartment and texture-based radiomics features ($F_t$) of Ellipsoid Zone (EZ)-RPE and sub-RPE compartments for risk stratification for subfoveal GA (sfGA) progression. This was a retrospective study of 137 dry AMD subjects with a 5-year follow-up. Based on sfGA status at year 5, eyes were categorized as Progressors and Non-progressors. A total of 15 shape-based $F_{\mathit{fd}}$ of sub-RPE surface and 494 $F_t$ from each of sub-RPE and EZ-RPE compartments were extracted from baseline spectral domain-optical coherence tomography scans. The top nine features were identified from $F_{\mathit{fd}}$ and $F_t$ feature pool separately using minimum Redundancy maximum Relevance feature selection and used to train a Random Forest (RF) classifier independently using three-fold cross validation on the training set ($S_t$, N = 90) to distinguish between sfGA Progressors and Non-progressors. Combined $F_{\mathit{fd}}$ and $F_t$ was also evaluated in predicting risk of sfGA progression. The RF classifier yielded AUC of 0.85, 0.79 and 0.89 on independent test set ($S_v$, N = 47) using $F_{\mathit{fd}}$, $F_t$, and their combination, respectively. Using combined $F_{\mathit{fd}}$ and $F_t$, the improvement in AUC was statistically significant on $S_v$ with p-values of 0.032 and 0.04 compared to using only $F_{\mathit{fd}}$ and only $F_t$, respectively. Combined $F_{\mathit{fd}}$ and $F_t$ appears to identify high-risk patients. Our results show that FD and texture features could be potentially used for predicting risk of sfGA progression and future therapeutic response.

Introduction

Age-related macular degeneration (AMD) is one of the leading causes of acute vision loss in developed countries¹. In its advanced form of dry AMD, the disease may convert to manifestations of geographic atrophy (GA), which is characterized by the progressive death of photoreceptor cells and retinal pigment epithelium (RPE), gradual loss of underlying choriocapillaris, and thickening of the Bruch’s membrane (BM)². Initially, GA may appear in the extrafoveal regions; however, as the disease progresses, it gradually affects the foveal center also. Apart from RPE degeneration, structural alterations of the sub-RPE (the space bound by the RPE and BM) surface is another hallmark of subfoveal GA (sfGA)³. Subretinal drusenoid deposits (SDDs) are also a notable finding of dry AMD that exacerbate local and global growth of GA⁴. The occurrence and progression of GA is influenced by several factors and mechanisms, making the treatment and prevention of GA extremely challenging^3,5. Given that no clinically-approved medications or therapeutics are currently available to prevent GA progression^6,7, the characterization of GA and identifying patients that are at an increased risk of converting to sfGA in dry AMD are of great importance for disease monitoring, treatment decision making, and clinical trial enrichment.

The dominant imaging modality used by clinicians for AMD and GA assessment is spectral domain-optical coherence tomography (SD-OCT) that allows three-dimensional (3D) depiction of the retina at micrometer resolution³. It provides excellent visualization of the enlargement and morphologic alterations within the atrophic area and its surrounding microenvironment^8,9. Additionally, the SD-OCT images allow characterization of size, location and progression of GA lesions as well as quantitative characterization of imaging features associated with future GA growth⁹. The role and association of SD-OCT biomarkers (such as hyper-reflective foci [HRF], SDD) with local and global growth of GA have been widely investigated in previous literature^10,11. Higher-order quantitative OCT biomarkers, such as EZ integrity and sub-RPE compartment features, have also been found to be associated with the development of sfGA¹². Most of the studies on GA progression using machine learning (ML) are predominantly based on lesion segmentation, detection and classification¹³, and mainly focus on lesion size^14,15, shape^16,17 and location¹⁷ to predict GA progression. However, there is still an unmet need for developing ML-based imaging biomarkers in predicting sfGA progression (especially for eyes that are at the greatest risk) that might be crucial for treatment development and providing therapeutic benefits. A multi-feature integrative approach may provide an important asset to risk assessment.

As the atrophy ensues, the sub-RPE compartment undergoes several structural changes¹⁸. An arbitrary complex, irregular and chaotic deformation in the sub-RPE contour is observed for patients that are at the onset and highest risk for sfGA conversion (Progressors) in their respective SD-OCT scans. The SDDs in the vicinity of the RPE layer cause disruption of the EZ anatomy⁴. The retinal layers affected by GA are well visualized in the respective SD-OCT scans. Previous studies have extensively investigated the role of fractal dimension (FD) features in describing complex pathological changes in tumor growth¹⁹, irregularities in tumor shape^20–22, and microvascular changes of retinal vessel structure in different diabetic eye diseases^23–25. The irregular surfaces and structures associated with human organs (such as heart, lungs, brain) cannot be adequately described by traditional Euclidean distance measure^20–22. FD, on the other hand, allows for quantitative characterization of irregular shapes and structures at all scales and levels of magnification using the spatial self-similarity pattern concept. Thus, FD features can capture the shape and structural disorders along the contour of natural objects^26–28. In the present study we hypothesized that the significant differences between the shape characteristics of the sub-RPE compartment boundary between the Progressors and Non-progressors of sfGA can be accurately characterized by FD based shape descriptors. In addition, prior work from our group²⁹ has provided foundational data on the relevance of shape-based fractal features for characterizing sub-RPE contour deflection and their association with sfGA progression on the same dataset. Separately we have shown the role of varied texture-based radiomic descriptors from the fluid and retinal tissue compartments of OCT images in predicting therapeutic response in different diseases, such as diabetic retinopathy (DR) and diabetic macular edema (DME)³⁰.

In this study we sought to evaluate the role of SD-OCT derived shape-based fractal features of the sub-RPE surface and texture-based radiomic descriptors of the sub-RPE and EZ-RPE compartments in identifying dry AMD patients (N = 137) who are at the greatest risk of sfGA progression by securing 5 years of follow-up observation data. A comprehensive assessment of the shape and texture features was performed to evaluate their discriminability between the Progressors and the Non-progressors. Additionally, the combination of shape and texture features were also evaluated in their ability to discriminate between the two groups of patients.

Methods

Study description

This retrospective cohort study was approved by the Institutional Review Board (IRB) of the Cleveland Clinic and included 137 subjects with dry AMD with a follow-up of 5 years. All the methods were carried out in accordance with the Declaration of Helsinki³¹. As this was a retrospective and minimal risk analysis, the IRB of the Cleveland Clinic determined that informed consent was not required and waived this requirement for this study. A total of 216 subjects were initially included in the study, based on the main inclusion criteria of presence of intermediate or advanced dry AMD based on ICD-9 code with concurrent OCT scans obtained. After initial screening of the electronic medical record by one of the physicians involved in the study, patients with eyes with baseline advanced dry AMD with sfGA, history of neovascular AMD or anti-vascular endothelial growth factor (anti-VEGF) injections, other retinal disease, history of vitreoretinal surgery, and/or poor image quality were excluded from the study. sfGA was determined based on SD-OCT image review to confirm the absence of atrophy at the fovea. The most common reasons for exclusion were incomplete follow-up, concurrent neovascular AMD/previous anti-VEGF therapy. The final cohort consisted of 137 eyes (See Flowchart in Fig. 1). Clinical characteristics including demographic features were collected. The baseline characteristics of the patients are presented in Table 1.

Figure 1.

Flowchart showing inclusion–exclusion criteria for the study. SD-OCT Spectral Domain-Optical Coherence Tomography, AMD age-related macular degeneration, sfGA subfoveal Geographic Atrophy.

Table 1.

Baseline characteristics.

	Total group
Age
Mean	67 years
Range	56–90 years
Gender
Male	45 (32.8%)
Female	92 (67.2%)
Eye
OD	67 (48.9%)
OS	70 (51.1%)
Visual acuity
Mean	20/26
Range	20/20–20/125

SD-OCT was performed at four different time points (baseline, year 1, year 3, and year 5). At year 5, eyes were categorized into two subgroups: Progressors (N = 21) and Non-progressors (N = 116), based on the sfGA conversion status of the eyes. Progressor/Non-progressor classification was determined based on SD-OCT image review by an expert reader. sfGA was defined as RPE loss/atrophy with associated photoreceptor degeneration (EZ loss/outer nuclear layer (ONL) thinning) at the foveal center. The final cohort of 137 patients (Fig. 1) was randomly divided into a training set ($S_t$, containing 2/3rds of the dataset) that consisted of 90 patients (14 Progressors and 76 Non-progressors) and a test set ($S_v,$ containing 1/3rd of the dataset) that comprised of 47 patients (7 Progressors and 40 Non-progressors) with no overlap of patients between $S_t$ and $S_v$. The balance in the number of Progressors and Non-progressors in the training and test set were ensured³².

For automated segmentation of RPE and BM, SD-OCT macular cube data was exported into a previously-described ML based multi-layer segmentation platform (OCTViewer, Cleveland Clinic)^33,34. The initial automated segmentation was followed by line-by-line review of each scan by two expert readers with manual correction of any segmentation errors, as required such as correction of proper line segmentation, identifying retinal fluids. Each OCT scan in the dataset contained 128 B-scans. Given the key interest in subfoveal progression, the central 2 mm × 6 mm (i.e., the central 43 slices (Slice 43–85) were considered for a radiomics-based assessment of the central macular area.

Spatial localization of retinal tissue compartments

An image ($I)$ was considered as a 3D spatial grid of voxels corresponding to the volume of OCT scans. Sub-volumes $I_r$ and $I_e$ were defined corresponding to the segmentation of the retinal tissue compartments between RPE-BM (sub-RPE) and EZ-RPE, respectively.

Shape-based fractal feature extraction

A total of 15 shape-based one-dimensional (1D)³⁵, two-dimensional (2D)^36,37, and 3D³⁸ fractal features ($F_{\mathit{fd}}$) were extracted in order to characterize the structural complexity assessment along the contour of the sub-RPE compartment from the respective OCT scans. The 1D fractal feature was extracted as a time series measurement from each individual slice from within the 3D volume using the central pixels-to-boundary distance. Box-counting (BC) method was used to measure 2D FD from the OCT images and boundary of the sub-RPE compartment masks for each patient³⁷. Statistics of mean, median, standard deviation, and skewness of each 2D FD feature was calculated for each of the 43 OCT slices and the mean of these values across all 43 slices was computed for each patient. Power spectral density analysis as described by Zhang³⁸ was used to compute the 3D fractal dimension of segmentation mask (describing the degree and direction of anisotropy). A more detailed description of $F_{\mathit{fd}}$ is provided in Section I of the Supplementary Material.

Texture-based radiomic feature extraction

A total of 494 3D texture-based radiomics features ($F_t$) were extracted from each of the $I_r$ and $I_e$ sub-volumes using the MATLAB platform (version 2022b; Mathworks, Natick, Mass). The $F_t$ included 65 Haralick features³⁹ (that quantified the spatial gray-level co-occurrence matrix [GLCM] within local neighborhoods around each pixel in an image and captures the structural heterogeneity within the region of interest), 152 Laws energy⁴⁰ (capturing presence of spots, edges, waves, and ripples), 225 Gabor wavelet⁴¹ (that considers oriented textures via changes in direction and scale to capture microarchitectures) and 52 CoLlage⁴² (capturing anisotropic tensor gradient differences) features. Statistics of median, standard deviation, skewness, and kurtosis were then calculated from the features within all regions of interest and a total of 1976 statistical features from each of the sub-RPE and EZ-RPE compartments were obtained.

To remove redundant/collinear features, the Pearson Correlation Coefficient (PCC) for all possible combinations of features were computed and the features with higher p-values, i.e., with PCC > 0.8 were pruned. This resulted in 949 and 1143 texture features from sub-RPE and EZ-RPE compartment, respectively. All remaining feature values were normalized (mean of 0 and standard deviation of 1). A detailed description of the texture-based radiomics features is provided in Supplementary Material Section II.

Statistical analysis

We developed three ML models in the present study. In the first experiment, to determine the subset of shape-based radiomic features from the sub-RPE contour that best discriminated between Progressors and Non-progressors on SD-OCT scans, the top nine features were selected from $F_{\mathit{fd}}$ using the minimum Redundancy maximum Relevance (mRmR)⁴³ feature selection method in a three-fold cross validation setting over 500 iterations. In each fold and run, these top nine features were used to train a Random Forest (RF) classifier on $S_t$ (N = 90). The RF model ($M_s$) was subsequently applied to predict the class level of $S_v$ (N = 47) and performance of $M_s$ on $S_v$ was evaluated using the area under the Receiver Operating Characteristics Curve (AUC) metric.

In experiment 2, to determine the texture-based radiomic features that favorably distinguished between Progressors and Non-progressors, the top nine features were selected from $F_t$ using mRmR feature selection method⁴³ and evaluated in conjunction with RF classifier ($M_t$) in a cross-validated approach on $S_t$. Performance of $M_t$ was subsequently validated on $S_v$ using the AUC metric.

Finally in experiment 3, $F_{\mathit{fd}}$ and $F_t$ were combined and evaluated using RF classifier ($M_{\mathit{st}}$) in a similar way to identify the patients at highest risk of sfGA progression (Progressors) followed by its validation on $S_v$.

The computational pipeline for the approach employed is illustrated in Fig. 2.

Figure 2.

Overview of methodology. (a) SD-OCT scans were retrospectively collected. Segmentation of (b) sub-RPE and (c) EZ-RPE compartments using an ML based multi-layer segmentation platform OCTViewer. (d) 1D, 2D and 3D FDs were measured. The 1D fractal feature was extracted as a time series measurement from each individual slice from within the 3D volume. 2D and 3D fractal dimension measurements were then extracted using the Box-counting and Power Spectral Density Analysis methods, respectively. (e) Shape-based FD and texture-based radiomic feature extraction (f) mRmR feature selection was used to select the top nine features to train a RF classifier on ${\text{S}}_{\text{t}}$ (N = 47). (g) Classifier performance (in terms of AUC) validated on 47 patients in ${\text{S}}_{\text{v}}$. 2D scatter plot between the topmost two features show clear separation between the two groups of patients. SD-OCT Spectral Domain-Optical Coherence Tomography, RPE Retinal Pigment Epithelium, EZ Ellipsoid Zone, ML Machine Learning, 1D One dimensional, 2D Two Dimensional, 3D Three Dimensional, FD Fractal Dimension, mRmR minimum Redundancy maximum Relevance, RF Random Forest, AUC Area under Receiver Operating Characteristics Curve.

Results

Baseline characteristics

This study included a total of 137 patients with dry AMD, with a mean age of 76 ± 7 years. During the follow-up period of 5 years, 15.3% of the patients (N = 12) were found to have sfGA progression (Progressor), whereas the remaining 84.6% (N = 116) were identified with no progression (Non-progressor). The Progressor group included 57% (N = 12) females and 43% (N = 9) males, whereas 65% (N = 75) females and 35% (N = 41) males were in the Non-progressor group. The mean baseline visual acuity (VA) for Progressors and Non-progressors was 20/27 and 20/32 (p = 0.87), respectively.

Experiment 1: shape based fractal dimension features of sub-RPE compartment predicts risk of sfGA progression

In order to identify the most discriminating shape-based FD features from the sub-RPE contour that discriminate between Progressors and Non-progressors of sfGA, mRmR feature selection method was applied only to the features in $F_{\mathit{fd}}$. The most discriminating features within $F_{\mathit{fd}}$ that were consistently selected by $M_s$ over the course of three-fold cross-validation over 500 iterations are reported in Table 2.

Table 2.

The most discriminating shape-based fractal and texture-based radiomic features to distinguish between progressors and non-progressors of sfGA in Dry AMD.

Experiment	Feature	Statistics	Compartment	p-value
Experiment 1 (based on fractal features)	Entropy	Mean	Sub-RPE	< 0.001
	2D FD	Standard Deviation	Sub-RPE	< 0.001
	3D FD	Mean	Sub-RPE	< 0.001
	Entropy	Skewness	Sub-RPE	< 0.001
	Entropy	Kurtosis	Sub-RPE	0.004
	FD 2D Mask Min	Kurtosis	Sub-RPE	< 0.001
	FD 2D Mask Max	Median	Sub-RPE	0.001
	FD 2D Mask Min	Median	Sub-RPE	0.004
	FD 2D Mask Max	Mean	Sub-RPE	< 0.001
Experiment 2 (based on texture features)	Laws Energy S3L3E3	Kurtosis	Sub-RPE	0.0329
	Laws Energy L3L3E3	Median	Sub-RPE	0.04
	Laws Energy L3S3E3	Median	Sub-RPE	0.035
	Laws Energy S5L5L5	Kurtosis	Sub-RPE	0.035
	Laws Energy S5L5E5	Median	Sub-RPE	0.042
	Laws Energy L3E3L3	Kurtosis	EZ-RPE	< 0.001
	Laws Energy S5E5E5	Median	EZ-RPE	< 0.001
	Laws Energy W5E5E5	Skewness	EZ-RPE	< 0.001
	Laws Energy W5E5E5	Kurtosis	EZ-RPE	< 0.001

The ‘Mean entropy’ (p-value < 0.001) was identified as the topmost shape-based FD feature that best discriminated between the Progressors and Non-progressors. The discriminability of “Mean Entropy” is visually illustrated in Fig. 3 for two cases each of Progressors and Non-progressors. Higher values of entropy were found to be associated with the Progressors, which suggests that more architectural disorder and complexity in the sub-RPE compartment structure tends to be associated with patients that are at the early stages of sfGA progression (Progressors). This trend also appears to be reflected in the box and whisker plot, as illustrated in Fig. 3e.

Figure 3.

Illustration of the discriminability of the shape-based FD feature “Mean Entropy” for the sub-RPE compartment contour. Original OCT image for two cases of (a) Progressors and (b) Non-progressors. The segmentation of sub-RPE compartment was done by “OCTViewer”, an ML based multi-layer segmentation platform built in Cleveland Clinic. (c), (d) Zoomed-in Mean Entropy feature expression for (a), (b), respectively. Higher values of the shape-based fractal feature “Mean Entropy” is reflected by the higher feature expression (warmer color tones) in the sub-RPE contour for the Progressor. (e) Box and Whisker plot on the left corresponds to Mean Entropy feature values for the Progressors (N = 19) and that on the right corresponds to Mean Entropy feature values for the Non-progressors (N = 118). (f) t-SNE plot of the shape-based fractal features shows two distinct clusters of Progressors and Non-progressors. ML Machine Learning, OCT Optical Coherence Tomography, RPE Retinal Pigment Epithelium.

$M_s$ yielded an AUC of 0.85 (95% Confidence Interval [C.I] 0.62, 0.99) on $S_v$. The other performance metrices included accuracy, sensitivity and specificity with values of 85%, 82% and 83%, respectively. The AUC values yielded by $M_s$ on $S_t$ and $S_v$ are reported in Table 3.

Table 3.

AUC values yielded by different models on training ($S_t$) and test set ($S_v$).

Sl No.	Features	Model	${\mathit{S}}_{\mathit{t}}$ (95% confidence interval)	${\mathit{S}}_{\mathit{v}}$ (95% confidence interval)
1	Fractal	$M_s$	0.92 (0.7, 0.95)	0.85 (0.62,0.99)
2	Texture	$M_t$	0.85(0.69, 0.92)	0.79 (0.62, 0.84)
3	Combined fractal and texture	$M_{\mathit{st}}$	0.95(0.7, 0.9)	0.89 (0.77, 0.99)

Experiment 2: texture-based radiomic features of sub-RPE and EZ-RPE compartment predicts risk of sfGA progression

In the second experiment, the most discriminating texture-based radiomic features from the sub-RPE and EZ-RPE compartments that favorably distinguished between the Progressors and Non-progressors of sfGA, mRmR feature selection method was applied only to the features within $F_t$. The top nine texture features that were consistently selected by $M_t$ during training are presented in Table 2. Of these nine features, five belonged to the sub-RPE compartment and the remaining four features corresponded to the EZ-RPE compartment. The Kurtosis Laws S3L3E3 (p-value = 0.0329) feature from the sub-RPE compartment and Skewness Laws W5E5E5 (p-value = 3.0711e − 04) feature from the EZ-RPE compartment were identified as the two most discriminating texture features.

The feature map for the topmost texture feature “Laws S3L3E3” within the sub-RPE compartment is illustrated in Fig. 4 for two cases each of Progressor and Non-progressor. The Laws S3L3E3 feature captures the textural patterns of spots, levels and edges in the horizontal, vertical and diagonal directions, using the 3X3X3 convolutional kernel. As observed from Fig. 4, the Laws S3L3E3 feature is over-expressed for the Progressors and is reflective of the higher order of textural heterogeneity within the sub-RPE compartment for the Progressors. The box and whisker plot of the Kurtosis Laws S3L3E3 feature (p-value = 0.0329) presented in Fig. 4e reveals a significant statistical difference between the two patient groups.

Figure 4.

Illustration of the discriminability of the “Laws S3L3E3” texture feature within the sub-RPE compartment. Original OCT image for two cases of (a) Progressors and (b) Non-progressors. (c), (d) Zoomed-in Laws S3L3E3 feature expression on (a), (b), respectively. Higher feature expression for the Progressor represent higher order of textural discrepancy within sub-RPE compartment for the Progressor. (e) Box and Whisker plot on the left corresponds to kurtosis Laws S3L3E3 feature values for the Progressors (N = 19) and that on the right corresponds to kurtosis Laws S3L3E3 feature values for the Non-progressors (N = 118).

The feature map for the topmost texture feature “Laws W5E5E5” within the EZ-RPE compartment is shown in Fig. 5 for two cases each of Progressor and a Non-progressor. The textural patterns of waves in the horizontal and edges in both vertical and diagonal directions, using the 5X5X5 convolutional kernel is captured by the Laws W5E5E5 feature. As observed from Fig. 5, the Laws W5E5E5 feature is over-expressed for the Progressors. This is reflective of a higher degree of textural heterogeneity within the sub-RPE compartment for the Progressors. The box and whisker plot of the Kurtosis Laws W5E5E5 feature (p-value = 3.0711e − 04) illustrated in Fig. 5e also shows a statistically significant difference between the two patient groups.

Figure 5.

Illustration of the discriminability of the “Laws W5E5E5” texture feature within the EZ-RPE compartment. Original OCT image for two cases of (a) Progressors and a (b) Non-progressors. (c), (d) Zoomed-in Laws W5E5E5 feature expression on (a), (b), respectively. The feature value is highly expressed for the Progressor which represents higher order of heterogeneity within EZ-RPE compartment is associated with the Progressors. (e) Box and Whisker plot on the left corresponds to skewness Laws W5E5E5 feature values for the Progressors (N = 19) and that on the right corresponds to skewness Laws W5E5E5 feature values for the Non-progressors (N = 118).

An AUC of 0.79 (95% CI 0.62, 0.84) was obtained by $M_t$ on $S_v$ using these top nine features in a supervised setting. The accuracy, sensitivity and specificity values yielded by the RF model was 78%, 77% and 85%. The AUC values yielded by $M_t$ on $S_t$ and $S_v$ are reported in Table 3.

Experiment 3: OCT-derived combined fractal and texture features predict risk of sfGA progression

Finally, the shape-based fractal and texture-based radiomic features were combined and evaluated using $M_{\mathit{st}}$ to discriminate between Progressors and Non-progressors. The most discriminating nine features from combined $F_{\mathit{fd}}$ and $F_t$ were selected by mRmR feature selection method from $S_t$ and used to train $M_{\mathit{st}}$ over 500 iterations of three-fold cross-validation. The topmost nine features included five sub-RPE shape features, two sub-RPE Laws texture features and two EZ-RPE Laws texture features.

By employing these most discriminating nine features, $M_{\mathit{st}}$ yielded a statistically significantly improved AUC of 0.89 (95% CI 0.77, 0.99) on $S_v$ with p-values of 0.032 and 0.04 compared to using only shape features and using only texture features, respectively. The AUC values yielded by $M_{\mathit{st}}$ on $S_t$ and $S_v$ are reported in Table 3. The other performance metrics included ACC of 83%, sensitivity of 96% and specificity of 86%. The ROC curves generated by the models $M_s$, $M_t$ and $M_{\mathit{st}}$ in Experiment 1, Experiment 2 and Experiment 3, respectively are presented in Fig. 6.

Figure 6.

ROC Curves generated using (a) shape-based fractal features from sub-RPE contour alone (Experiment 1), (b) texture features within sub-RPE and EZ-RPE compartments (Experiment 2) alone and (c) combined fractal and texture features (Experiment 3). ROC, Receiver Operating Characteristics Curve; RPE, Retinal Pigment Epithelium, EZ, Ellipsoid Zone.

Discussion

Geographic atrophy (GA) is the clinical end point of dry age-related macular degeneration (AMD). Since there is no approved treatment for subfoveal GA (sfGA), predicting risk of progression to sfGA for dry AMD patients could potentially allow for a new treatment regimen. The focus of the state-of-the-art methods using OCT images is predominantly on GA lesion segmentation^18,44, and future spatial GA progression¹⁰. The Dice Similarity Score (DSC) was used for assessment of segmentation performance and the reported values lies within 0.81–0.87 for SD-OCT images. On the other hand, predictive models on characterization of the trend for GA progression are mostly based on lesion segmentation and detection, followed by classification of GA using fundus autofluorescence (FAF)⁴⁵ and color fundus images (CFI)⁴⁶. Pfau et al.⁴⁵ reported the prognostic relevance of different shape-based descriptors such as lesion area, perimeter, circularity, and caliper diameter for the rapid progression of GA secondary to AMD. Liefers et al.⁴⁶ investigated the role of structural biomarkers such as area, filled area, convex area, convex solidity, eccentricity, roundness, foveal involvement, perimeter, and circularity in their association with GA progression. Additionally, a number of studies have developed deep learning (DL) based models to predict GA growth rate prediction^47,48. While stand-alone DL based “black box” approaches might be acceptable for screening or diagnostic decision-guidance, DL model fails to produce optimal performance⁴⁹ as compared to radiomics-based hand-crafted features^30,50 for a complex task like treatment response prediction. It is unclear whether these opaque strategies might be clinically employed for quantifying and monitoring changes related to GA progression. Also, there still exists a knowledge gap regarding which of these features are most strongly associated with GA progression. In the present study, we evaluated the role of a set of new interpretable, domain inspired computer-extracted shape and texture features from the spectral domain-optical coherence tomography (SD-OCT) scans that could best discriminate between the Progressors and Non-progressors of sfGA, and we identified certain shape and texture-based radiomic features that were found to be associated with disease progression.

The retinal pigment epithelium (RPE) and Bruch’s membrane (BM) undergoes several structural alternations and deformations over the course of sfGA progression. Multiple studies have investigated the role of the sub-RPE (RPE-BM compartment) thickness in their association with the likelihood of sfGA progression where the sub-RPE thickness has been found to be significantly greater for the Progressor group¹². Distinct differences in RPE morphology and thickness were also found to be implicated with sfGA progression⁴⁹. However, to date, no quantitative studies have been reported on quantifying the structural complexity of the sub-RPE contour for dry AMD patients; more specifically, for the patients who are at the greatest risk of sfGA progression (Progressors). On the other hand, the subretinal drusenoid deposits (SDDs) may also have impact on the texture within the sub-RPE and ellipsoid zone-RPE (EZ-RPE) compartment among the patient groups as observed in their respective OCT scans. In the present study, we hypothesized that the utility of both shape and texture-based features and their successful fusion might play a vital role in predicting their association with sfGA progression and identifying high-risk patients.

In the present study we did three experiments to evaluate our hypothesis with respect to the shape and texture-based radiomic features that are most implicated in predicting likelihood of sfGA progression. In the first experiment, we evaluated the role of shape-based fractal dimension (FD) features of sub-RPE contours in association with sfGA progression. The use of FD to describe the complex geometrical properties of tissue, tumors²⁷, vessel architecture of different diabetic eye diseases^23–25 have all been previously interrogated. In the present study, the concept of fractal analysis was used to quantitatively describe the structural patterns of the sub-RPE contour. The sub-RPE contour was observed to exhibit a certain degree of randomness associated with their shape, the shape being more complex and irregular for the Progressors of GA. We identified the shape-based FD feature ‘Mean Entropy’ that strongly discriminated between the Progressors and Non-progressors, with higher values of mean entropy being associated with the Progressors. Entropy being a measure of coherence of architecture, higher entropy reflets higher degree of architectural and structural complexity as well as shape disorder in the sub-RPE contour for the Progressors, therefore signifying greater risk of sfGA progression, whereas the sub-RPE contour was found to be more structurally organized for the Non-progressors. This may be related to the structural alteration (elevation or attenuation) of the sub-RPE compartment on SD-OCT scans of the sfGA Progressors because of irregular thickening (due to accumulation of subretinal drusenoid deposit between RPE and BM) and thinning (due to RPE cell death) of the RPE bands in the atrophic area. This also corroborates with the previous finding where the RPE morphology was found to be deteriorating gradually towards the atrophic area on account of disease progression^44,51,52.

The goal of the second experiment was to identify the textural discrepancies within the sub-RPE and the photoreceptor outer segment [Ellipsoid Zone (EZ)-RPE] compartments. Laws texture features pertaining to the sub-RPE and EZ-RPE compartment were found to be mostly associated with sfGA progression with the feature values highly expressed for the Progressors. The sub-RPE and EZ-RPE compartments tended to have a more chaotic micro-architecture, and substantially more textural heterogeneity that might have been captured by the Laws texture features. The Laws features captures spots, waves, and ripple-like patterns that are most likely being reflected in the form of heterogeneity within the texture of OCT compartments between the patient groups. This could be linked with the textural discrepancy within the sub-RPE and EZ-RPE compartments due to internal drusenoid deposits that are correlated with local and global progression of GA in SD-OCT images^3,12.

In the third experiment, we sought to investigate whether a combination of shape-based fractal (of the sub-RPE contour) and texture-based radiomic features (within the EZ-RPE and sub-RPE compartments) improved the predictive performance of the classifier. In our experiment, the fusion of shape and fractal features yielded a statistically significant improvement (Delong test⁴⁵, p-value = 0.032) in the Random Forest (RF) classifier performance (with an Area under the Receiver Operating Characteristics Curve (AUC) of 89% on the independent test set) in predicting sfGA risk progression compared to the use of only shape features and only texture features (p = 0.032 and 0.04, respectively). These findings further appear to validate the role of these features in predicting risk of sfGA progression. These findings also suggest that the combination of different shape and texture features requires further evaluation as potential predictive biomarkers for predicting risk of sfGA progression.

For the last few decades, artificial intelligence (AI) has been rapidly evolving the ophthalmologic research domain and potentially implemented for diagnosis, surveillance and individualized therapy for multiple ocular diseases¹³. Given that till date there is no approved therapy to suppress sfGA progression^6,7, identifying patients who are at the onset of developing and converting towards sfGA would be highly beneficial for clinical decision making. In this regard, machine learning (ML) based predictive models could play vital role in identifying high-risk patients using different SD-OCT biomarkers. In particular, this study identifies certain shape and texture based radiomics features and presents their importance as potential biomarkers for predicting risk of sfGA progression and an important opportunity for clinical trial enrichment to identify eyes that are at highest risk of converting towards sfGA.

The present study had some limitations that must be acknowledged. The sample size of the study is relatively small (N = 137). The stability and consistency of the features need to be validated on a larger dataset. Since the OCT scans are not available for the follow-up interval of 2 years, the exact time of sfGA conversion is difficult to estimate. A majority number of the cases (66%) in the study converted to sfGA after 3 years, showing that the baseline case selection is appropriate in this respect. Also, a detailed analysis of the sensitivity of the segmentation method effects on derived features needs to be addressed. Further, the nature of the analysis is retrospective rather than prospective. Finally, this is a single-institution study that utilized a single type of OCT scanner. An obvious question rises here is the generalizability of the classifier on multi-institution data. The reproducibility and reliability of the proposed radiomic model needs further validation on independent cohorts of patients, including with other SD-OCT scanners to explicitly establish its role in predicting risk progression to sfGA.

Conclusion

In this preliminary study, we investigated the role of computer-extracted OCT-derived shape and texture-based radiomic features in discriminating the Progressors and Non-progressors with likelihood of sfGA conversion in dry AMD. The major findings of the study are the identification of certain shape and texture features that favorably distinguished between the two patient groups. With further validation from multi-institutional datasets, these features could be potentially used to build a clinical decision support tool for automated quantitative assessment of sfGA progression in dry AMD patients.

Abbreviations

sfGA

Subfoveal geographic atrophy

SD-OCT

Spectral domain-optical coherence tomography

AMD

Age-related macular degeneration

FD

Fractal dimension

RPE

Retinal pigment epithelium

Author contributions

S.S.K. J.P.E. and A.M. designed the study. H.C., J.A., S.K.S., J.P.E provided all image segmentation. S.S.K., H.C., A.M and J.P.E. analyzed and interpreted the data and the appropriateness of all statistical analysis. S.S.K. wrote the first draft of the manuscript, with all authors reviewing, editing, and approving the manuscript.

Data availability

Data are available upon reasonable request. Data request can be directed to JPE. Access to data sets from the Cole Eye Institute, Cleveland Clinic Foundation (used with permission for this study) should be requested directly from these institutions via their data access request forms. Subject to the institutional review boards’ ethical approval, unidentified data would be made available as a test subset.

Code availability

The MATLAB codes for feature extraction is available in https://github.com/SudeshnaSilKar/Feature-Extraction-MATLAB. This code is being made available only for research and non-commercial use.

Competing interests

Dr. Srivastava has research support from Regeneron, Allergan, and Gilead; is a consultant for Bausch and Lomb, Novartis, and Regeneron. Dr. Madabhushi has research Funding: Astrazeneca, Bristol Myers-Squibb, Boehringer-Ingelheim, Eli-Lilly. Equity: Picture Health Inc, Inspirata Inc., Elucid Bioimaging, Consultant: Aiforia, Caris, Roche, Biohme, Castle Biosciences, Simbiosys. Dr. Ehlers has research support from the following: Aerpio, Alcon, Thrombogenics/Oxurion, Regeneron, Genentech, Novartis, Allergan, Apellis, Iveric, Stealth, Perceive Biotherapeutics, Roche, Alexion; is a consultant for the following: Aerpio, Adverum, Alcon, Allegro, Allergan, Genentech/Roche, Stealth, Novartis, Thrombogenics/Oxurion, Leica, Zeiss, Regeneron, Santen, Perceive Biotherapeutics, Boerhinger-Ingelheim, Exegenesis, RegenxBIO, Roche, Alexion; and holds a patent with Leica. Dr. Sil Kar, Dr. Cetin and Dr. Abraham have no financial disclosures to report.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-024-68259-0.