Texture-Based Radiomic SD-OCT Features Associated With Response to Anti-VEGF Therapy in a Phase III Neovascular AMD Clinical Trial

Abstract

Purpose: The goal of this study was to evaluate the role of texture-based baseline radiomic features (F_r) and dynamic radiomics alterations (delta, F_Δr) within multiple targeted compartments on optical coherence tomography (OCT) scans to predict response to anti–vascular endothelial growth factor (VEGF) therapy in neovascular age-related macular degeneration (nAMD).

Methods: HAWK is a phase 3 clinical trial data set of active nAMD patients (N = 1082) comparing brolucizumab and aflibercept. This analysis included patients receiving 6 mg brolucizumab or 2 mg aflibercept and categorized as complete responders (n = 280) and incomplete responders (n = 239) based on whether or not the eyes achieved/maintained fluid resolution on OCT. A total of 481 F_r were extracted from each of the fluid, subretinal hyperreflective material (SHRM), retinal tissue, and sub–retinal pigment epithelium (RPE) compartments. Most discriminating eight baseline features, selected by the minimum redundancy, maximum relevance feature selection, were evaluated using a quadratic discriminant analysis (QDA) classifier on the training set (S_tr, n = 363) to differentiate between the two patient groups. Classifier performance was subsequently validated on independent test set (S_t, n = 156).

Results: In total, 519 participants were included in this analysis from the HAWK phase 3 study. There were 280 complete responders and 219 incomplete responders. Compartmental analysis of radiomics featured identified the sub-RPE and SHRM compartments as the most distinguishing between the two response groups. The QDA classifier yielded areas under the curve of 0.78, 0.79, and 0.84, respectively, using F_r, F_Δr, and combined F_r, F_Δr, and F_c on S_t.

Conclusions: Utilizing compartmental static and dynamic radiomics features, unique differences were identified between eyes that respond differently to anti-VEGF therapy in a large phase 3 trial that may provide important predictive value.

Translational Relevance: Imaging biomarkers, such as radiomics features identified in this analysis, for predicting treatment response are needed to enhanced precision medicine in the management of nAMD.

Introduction

Age-related macular degeneration (AMD) is one of the major causes of visual impairments in developed countries, especially for people over the age of 50 years.¹ A small percentage (<10%) of AMD cases are converted to its most advanced form of neovascular AMD (nAMD), which can lead to progressive and central vision loss.² This vision loss is characterized by the growth of abnormal choroidal blood vessels known as choroidal neovascularization (CNV).³ These abnormal blood vessels primarily break through Bruch's membrane (BM) and gradually profuse into the sub–retinal pigment epithelium (sub-RPE) space (i.e., the space bounded by the RPE and BM).¹ Exudation of these immature vessels might result in fluid accumulation in different retinal tissue compartments and subretinal hyperreflective material (SHRM) deposition in the subretinal space.³ Anti–vascular endothelial growth factor (VEGF) therapy through the inhibition of intravitreal VEGF agents has clinically been proven to block the pathophysiology of nAMD and restore retinal morphology.⁴ Recent advances in optical coherence tomography (OCT) technology offer a noninvasive and non-contact-based approach of retinal cross-section imaging in three dimensions (3D) at micrometer resolution.⁵ It also provides important information about the morphologic and structural alterations within different retinal tissue compartments that might be associated with the therapeutic response for nAMD. 

Multiple studies in the literature have assessed the role of different clinical features associated with visual acuity (VA) outcomes for patients with nAMD. Larger CNV area, elevation in RPE, presence of intraretinal fluid (IRF), and presence of geographic atrophy on OCT scans were identified as predictive biomarkers associated with worse VA or less VA gain at the first year of anti-VEGF therapy.⁶ Morphologic alteration within intraretinal cysts, subretinal fluid (SRF), and pigment epithelial detachments were found to be strongly affecting visual gain in patients with nAMD.⁷ Presence of SRF, smaller total CNV leakage area, and area of occult CNV at baseline were found to predict better/greater VA gain at the first year.^8,9 Additionally, the effectiveness of anti-VEGF therapy in ellipsoid zone integrity, SHRM, and the sub-RPE compartment on posttreatment OCT scans were assessed by Ehlers et al.¹⁰ However, limited studies have reported on the association between disease progression or the effectiveness of treatment with textural and quantitative morphologic alterations within different retinal tissue compartments, including the fluid, SHRM, and the sub-RPE compartment. 

Computationally derived imaging attributes or radiomics have shown promise in extracting and identifying biologically relevant features from clinical images.^11–14 Radiomics-based image interpretation is a promising area in ophthalmology for biomarker discovery, predicting therapeutic response, and treatment decision-making.^13,14 In a recent work¹⁵ from our group, we assessed texture-based baseline radiomic features and delta texture features (alteration of texture between baseline and month 3) within different fluid, SHRM, and sub-RPE compartments of spectral domain OCT scans of a prospective, randomized, phase 2 trial (OSPREY) study¹⁶ and identified that measures of fluid heterogeneity and delta SHRM texture features were associated with treatment response in patients with nAMD treated with brolucizumab 6 mg and aflibercept 2 mg. Although the analysis on the OSPREY study provided important proof-of-concept data, the sample size was still small (N = 81) and additional validation was clearly needed to further establish the potential predictive value of these biomarkers. In the present study, we hypothesized that a machine learning (ML) model trained and validated on a larger phase 3 clinical trial data set might identify more robust and consistent features associated with CNV and other underlying disease manifestations related to nAMD that might provide relevant information regarding the potential benefit of the anti-VEGF agents in patients with nAMD. 

To this aim, we evaluated the role of baseline and delta texture features in distinguishing between complete responders and incomplete responders of anti-VEGF therapy on the HAWK data set (N = 1082).¹⁷ It is a prospective, randomized phase 3 clinical trial data set, primarily designed to compare efficacy and safety of brolucizumab with aflibercept to treat patients with nAMD. This is a first of its kind radiomic-based assessment in a phase 3 clinical trial. Different texture-based radiomic features (F_r) from the fluid, SHRM, and sub-RPE compartments were extracted from the baseline and posttreatment (month 3) OCT scans of the HAWK study. The delta texture features (F_Δr) were obtained by computing absolute changes in the texture-based features between baseline and posttreatment OCT scans. The topmost discriminating features within F_r and F_Δr were identified using feature selection and evaluated independently using different ML classifiers on the training set (S_tr, n = 363) across 500 iterations of threefold cross-validation. Once trained, the classifier was locked and classifier performance was subsequently validated on an independent test set (S_t, n = 156). Additionally, a combination of F_r, F_Δr, and clinical parameters (F_c) was also used to train ML models on S_tr followed by its validation on S_t. 

Methods

Study Description and Patient Selection

HAWK is a phase 3, prospective, randomized, multicentered study primarily designed to compare the effectiveness of brolucizumab 3 mg and 6 mg with aflibercept 2 mg in patients with nAMD (N = 1082). Patients were randomized 1:1:1 to receive brolucizumab 3 mg or 6 mg or aflibercept 2 mg for the HAWK study. The Declaration of Helsinki, International Conference on Harmonization E6 Good Clinical Practice Consolidated Guidelines, and other regulations as applicable and complied with the Health Insurance Portability and Accountability Act of 1996 were followed in the study. The study was performed across different centers of North, Central, and South America; Europe; Asia; Australia; and Japan. Before the screening was initiated or any study related procedure was conducted, all the patients submitted their written consent. An independent ethics committee/institutional review board approved all the protocols. A detailed description of the trial oversight, randomization, and sample size is provided by Dugel et al.¹⁷ In the present study, only the patients treated with 6 mg brolucizumab and 2 mg aflibercept (n = 519) were included and assessed in a treatment agnostic manner. The inclusion and exclusion criteria are illustrated in Figure 1. 

Figure 1.

 

Flowchart showing inclusion and exclusion criteria for the study. BCVA, best-corrected visual acuity; GA, geographic atrophy; SD-OCT, spectral domain optical coherence tomography.

Spectral domain OCT macular scans focused on the foveal center point were obtained at each visit (i.e., every 4 weeks) with either the Cirrus (Carl Zeiss Meditec, Dublin, CA, USA) or Spectralis (Heidelberg Engineering, Heidelberg, Germany). Scan patterns consisted of a 512 × 128 macular cube covering a 6 × 6-mm area of the macula with the Cirrus and a 49-line high-speed preset scan covering a 20° × 20° area of the macula with the Spectralis.¹⁶ Complete responders were defined as eyes that achieved 0 mm³ of IRF and SRF by week 16 (including week 16) and maintained the combined value of IRF and SRF <0.001 mm³ for the rest of the 56-week study. Incomplete responders were all the rest of the eyes. 

The entire cohort of 519 patients was divided into a training set (S_tr) comprising 50% of the data set (196 complete responders and 167 incomplete responders) and the remaining 50% (84 complete responders and 72 incomplete responders) were used for independent testing (S_t).¹⁸ The patients were selected in a manner that ensured balance in the number of complete responders and incomplete responders in S_tr and S_t. 

Feature Segmentation and Clinical Feature Extraction

A post hoc analysis of the OCT scans from the HAWK study was conducted at the Cole Eye Institute of the Cleveland Clinic. The macular cube OCT scans were imported into a proprietary multilayer segmentation platform (Cleveland Clinic, Cleveland, OH, USA)¹⁰ that used image processing, ML, and logic to segment fluid, retinal layers, and SHRM enclosure boundaries. Two trained and expert readers reviewed and corrected the accuracy of the segmentation lines as needed in a standardized reading environment. All time points for any given eye were evaluated by the same human reader to minimize interreader and intertimepoint variability. Finally, the scans were evaluated by the senior analyst for segmentation accuracy and consistency.^10,18,19 Following completion of segmentation, quantitative clinical parameters (F_c) were exported for concurrent analysis as a potential predictor of treatment response. These included actual retinal tissue volume, IRF volume, SRF volume, SHRM volume, and sub-RPE volume. Further, macular subretinal fluid index (SRFI = 100 × SRF volume/[total retinal volume]), macular retinal fluid index (RFI = 100 × IRF volume/[total retinal volume – SRF volume]), and macular total retinal fluid index (TRFI = 100 × [IRF + SRF volume]/[total retinal volume]) were calculated, where the volume between the inner limiting membrane (ILM) and the RPE was considered the total retinal volume.^18,19 

Spatial Localization of OCT Compartments

From the image I, we defined subvolumes (of I) corresponding to the fluid, SHRM, and the retinal tissue compartments between ILM to RPE (excluding fluid) and sub-RPE segmentations as I_f, I_s, I₁, and I₂, respectively. 

Radiomics Feature Extraction

Following image acquisition and segmentation, from every voxel within each of the OCT compartments, a total of 481 3D texture features (F_r), from four feature groups (52 Haralick,²⁰ 152 Laws Energy,²¹ 225 Gabor,²² 52 CoLlage features²³), were extracted at baseline in a MATLAB platform (version 2022b; MathWorks, Natick, MA, USA). A detailed description of these features is presented in Supplementary Material Section I. The features corresponding to the subvolumes I_f, I_s, I₁, and I₂ are denoted by F_f, F_s, F₁ , and F₂, respectively, and their combination as F_r. The summary of notations is presented in the Table. Statistics of median, variance, skewness, and kurtosis were computed for each feature, giving rise to a total of 1924 statistical features per compartment. All the feature values were normalized with a zero mean and a standard deviation of 1. 

Table.

 

Summary of Notations

 

Absolute changes in feature statistics between baseline and month 3 were also computed for the individual OCT compartments of the complete responders and incomplete responders to determine the most distinctive delta texture features associated with therapeutic response/durability. For almost all the patients, there was substantial reduction in the fluid following therapy. Therefore, individual assessment of delta fluid features was not possible; rather than considering the fluid compartment separately, entire ILM-RPE retinal tissue/fluid compartment was considered in the present study. The delta texture features for the I_s, I₁, and I₂ subvolumes are denoted as Δ_s, Δ₁ , and Δ₂, respectively, with F_Δr denoting the combination. 

In order to avoid inclusion of redundant features and highly correlated features, feature pruning was done by calculating the correlation of all possible feature pairs using the Spearman correlation coefficient (SCC) followed by elimination of those features with a higher Wilcoxon rank-sum P value from each feature pair with SCC >0.8. 

Statistical Analysis

Three ML models were developed in this study. In the first experiment, to determine the features within F_r that best discriminated between the favorable complete responders and incomplete responders of anti-VEGF therapy on baseline OCT images, 500 iterations of threefold cross-validation on S_tr were used. At each iteration of threefold cross-validation, three feature selection methods, including t-test, Wilcoxon rank sum, and minimum redundancy, maximum relevance, were chosen to select the topmost eight features from F_r and evaluated in conjunction with four different classifiers: linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), random forest, and support vector machine in a cross-validation setting. The ML classifier that yielded the highest area under the receiver operating characteristic curve (AUC) value was considered the best-performing classifier. The ML classifier performance was subsequently validated on S_t. The other performance metrics included accuracy (ACC), sensitivity, and specificity. 

In the second experiment, to determine the features within F_Δr that best discriminated between complete responders and incomplete responders of anti-VEGF therapy, the topmost eight features from F_Δr were selected by three feature selection methods and were evaluated using four different classifiers (similar to that described for experiment 1) on S_tr over 500 runs of threefold cross-validation. The best-performing classifier, selected based on highest AUC on S_tr, was validated on S_t. 

Finally in the third experiment, the combination of F_r, F_Δr, and F_c was evaluated in a similar way to check whether there was a statistically significant improvement in the combined model performance compared to the individual models developed in experiment 1 and experiment 2. For each experiment, the Bonferroni correction²⁴ was applied to correct the P value within all statistical comparisons. 

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

Figure 2.

 

Overall pipeline of the radiomics-based assessment anti-VEGF therapy treatment response in nAMD using baseline OCT scans. (a) OCT scans of the HAWK study were retrospectively collected. (b) Individual fluid and SHRM compartments and the retinal tissue compartments (ILM-RPE, RPE-BM) were partitioned. (c) Texture-based radiomic features were extracted from the individual OCT and retinal tissue compartments using MATLAB V.2022b. For each of the individual compartments, feature statistics of median, standard deviation, skewness, and kurtosis were computed. (d) Top eight features were selected by feature selection and evaluated in conjunction with an ML classifier in a threefold cross-validation setting on the training set. (e) Classifier performance was evaluated on the test set. EZ, ellipsoid zone.

Results

Patient Characteristics

In the overall phase 3 HAWK clinical trial, 1082 patients were randomized. No clinically meaningful differences in demographics and baseline ocular characteristics were observed in the trial. From the 6-mg brolucizumab group and the 2-mg aflibercept group, patients were categized as complete responders (n = 280) and incomplete responders (n = 239) based on whether or not the eyes achieved and maintained fluid resolution and had OCT images of sufficient quality for analysis. 

Experiment 1: Distinguishing Therapeutic Response/Durability Based on Texture-Based Radiomic Features at Baseline

The texture-based radiomic features that were most consistently selected by 500 iterations of threefold cross-validation are presented in Supplementary Table S1. Out of these, the top eight features within F_r were identified, with six corresponding to the sub-RPE compartment and the other two features corresponding to the SHRM compartment. The Laws energy feature, “skewness-Laws E3S3L3,” pertaining to the sub-RPE compartment was found to be the most discriminating feature from F_r. The feature E3S3L3 represents Laws energy–based textural patterns of edges (sharp transitions in intensity, such as boundaries, or contours) in horizontal, spots (regions with small, localized variations in intensity, such as speckles or fine texture details) in vertical, and levels (regions with a relatively constant or uniform intensity, such as smooth surfaces) in diagonal directions, respectively, by using a 3D convolution filter. Figure 3 presents the feature map for the sub-RPE Laws E3S3L3 feature for one case of a complete responder and one case of an incomplete responder. Higher expression of the Laws E3S3L3 feature was observed for the complete responders, which is reflective of the greater degree of heterogeneity within the sub-RPE texture for the complete responder (Fig. 3c) as compared to the incomplete responder (Fig. 3d). 

Figure 3.

 

Illustration of the discriminability of the “Laws E3S3L3” texture feature within the sub-RPE compartment on baseline OCT scans: segmentation of sub-RPE (RPE-BM) compartment for one case of (a) complete responder and (b) one case of an incomplete responder. (c, d) Zoomed-in sub-RPE compartment for (a) and (b), respectively. (e, f) Visualization of the heatmap of the most discriminating feature (Laws E3S3L3) expression on baseline OCT scan for (c) complete responder and (d) incomplete responder, respectively. Prevalence of warmer color tones in the feature expression of the complete responder is reflective of a higher order of heterogeneity within the sub-RPE compartment texture for the complete responder.

The box-and-whisker plot of the top-performing feature “skewness-Laws E3S3L3,” presented in Figure 4a, illustrates a statistically significant difference between the two groups of patients with a P value of 7.3139e-48. 

Figure 4.

 

The box-and-whisker plot of the most discriminating feature (a) baseline sub-RPE skewness Laws E3S3L3 (identified from experiment 1) and (b) SHRM delta skewness Laws L3E3S3 (identified from experiment 2). The plot on the left corresponds to the feature values from the complete responders (n = 280) and that on the right corresponds to the feature values from the incomplete responders (n = 239).

Out of the four ML classifiers, the QDA classifier (M₁) yielded highest AUC of 0.82 ± 0.02 on S_tr using the top identified features from F_r and AUC of 0.78 (95% confidence interval [CI], 0.62–0.85) on S_t. The corresponding ACC, sensitivity, and specificity values were 77%, 81%, and 79%, respectively. The different performance metrices yielded by other ML classifiers are presented in Supplementary Table S2. 

Experiment 2: Distinguishing Treatment Response/Durability Based on Delta Texture Features

The F_Δr features that were most consistently selected by 500 iterations of threefold cross-validation are presented in Supplementary Table S1. The topmost eight features from F_Δr included five features from the SHRM compartment and three from the sub-RPE compartment. The feature “delta skewness-Laws L3E3S3,” pertaining to the SHRM compartment, was identified as the most discriminating delta texture feature. The feature L3E3S3 represents patterns of levels, edges, and spots extracted by a 3 × 3 × 3 convolution filter rotated in the horizontal, vertical, and diagonal directions, respectively, by using a 3D convolution filter. The feature map of the SHRM Laws L3E3S3 feature is presented in Figure 5 at baseline and posttherapy for one case of a complete responder and one case of an incomplete responder. As observed from Figure 5c, the feature value was highly expressed at baseline for the complete responders as compared to the incomplete responders. For the complete responders, the heterogeneity within the texture (reflected by warmer color tones at baseline in Fig. 5c) was substantially reduced after three cycles of anti-VEGF therapy (reflected by prevalence of cooler color tones in the posttherapy image). The box-and-whisker plot of the feature skewness-Laws L3E3S3 within the SHRM compartment, presented in Figure 4b, shows a statistically significant difference (P = 4.4648e-07) between the two groups of patients. 

Figure 5.

 

Illustration of the discriminability of the “Laws L3E3S3” texture feature within the SHRM compartment on baseline and posttherapy OCT scans: segmentation of SHRM compartment for one case of (a) a complete responder and one case of (b) an incomplete responder for baseline and posttherapy (month 3) OCT scans. (c, d) Zoomed-in SHRM compartment for (a) and (b), respectively. Visualization of the heatmap of the most discriminating delta texture feature (Laws L3E3S3) expression on baseline and posttherapy OCT scans for one case of (e) a complete responder and one case of (f) an incomplete responder. The heterogeneity within the texture is reflected by warmer color tones, whereas the cooler color tones represent the texture is more homogeneous. The Laws energy feature captures the textural alteration within the SHRM compartment following therapy for the complete responder.

QDA classifier (M₂) yielded the highest AUC of 0.86 ± 0.02 on S_tr using the top identified features from F_Δr and AUC of 0.79 (95% CI, 0.65–0.88) on S_t using the topmost features from F_Δr. The corresponding response prediction ACC, sensitivity, and specificity values were 78%, 82%, and 80%, respectively. The performance metrices yielded by other ML classifiers are reported in Supplementary Table S3. 

Experiment 3: Texture-Based Baseline and Delta Features Combined with Clinical Parameters Are Associated with Response/Durability of Anti-VEGF Therapy in nAMD

Finally, in the third experiment, F_r, Δ_oct, and F_c were combined and evaluated using M₃ to discriminate between the two groups of patients. The most consistent eight features (within combined F_r, F_Δr, and F_c) identified by 500 iterations of threefold cross-validation are presented in Supplementary Table S1 from the combination of F_r, F_Δr, and F_c included four delta texture features from the SHRM compartment. Among the additional four features, two belonged to the baseline SHRM compartment and two were from the baseline sub-RPE compartment. All eight features belonged to the Laws energy feature family. Using these top-performing eight features, the QDA classifier (M₃) attained the highest AUC of 0.89 ± 0.09 on S_tr (compared to the other classifiers) and AUC of 0.84 (95% CI, 0.73–0.96) on S_t. Supplementary Table S2 presents the performance metrices yielded by other ML classifiers. M₃ achieved a statistically significant improvement (DeLong test)²⁵ in the AUC value on S_t with P values of 0.042 and 0.043 compared to using only F_r and only F_Δr, respectively. The other performance metrics included an ACC of 88%, a sensitivity of 82%, and a specificity of 79%. 

Discussion

Emerging evidence^6,17,26 suggests that the therapeutic response and treatment requirement for nAMD widely vary across patient regimen. For some patients with nAMD, substantial and immediate reduction in fluid volume following anti-VEGF treatment is observed, whereas retinal fluid still persists irrespective of the fact that the patients are exposed to regular injections of anti-VEGF drugs.^6,27 Optimal treatment is therefore needed to tailor to an individual using precision medicine tools. The main difficulty in the disease management pipeline is that currently, there are no well-established predictors or biomarkers for therapeutic efficacy and durability of anti-VEGF therapy in nAMD. Identification of effective diagnostic biomarkers may provide a more accurate assessment of the potential benefit from treatment with anti-VEGF agents and provide more insight into the mechanisms of action of anti-VEGF agents. 

Previous works^28–31 have attempted to uncover a connection between clinical features and traditional OCT parameters and therapeutic durability. OCT images provide excellent high-resolution visualization of fluid and retinal tissue compartments that provide detailed information for developing important biomarkers for therapeutic response and durability of treatment.²⁸ As reported by Bogunovic et al.,²⁹ features associated with SRF area, IRF volume, and central retinal layer thickness were found to be discriminate, early in the course of therapy (i.e., at month 1 or 2), for the patients with nAMD of the HARBOR study. Characterization of IRF/SRF volume on posttherapy OCT scans also identified the presence of varying degrees of fluid across patients with nAMD.^30,31 

Prior work¹⁵ from our group reported an ML model using OCT-derived textural radiomic features from baseline and posttreatment OCT scans that were associated with treatment response to anti-VEGF therapy and therapeutic durability in patients with nAMD. The major limitation of this model is that it was developed on a relatively small data set of 81 patients and was not validated on independent test set. The small training set in this previous study raises the possible concern about model overfitting and lack of generalizability to new unseen test sets. Consequently, in this present study, we sought to evaluate the robustness and generalizability of the radiomic-based quantitative OCT biomarkers in terms of their association with anti-VEGF treatment response and durability on a much larger clinical trial data set of 519 patients with nAMD. Since the clinical trial data set incorporates data from multiple sites, validation of the radiomic model on a clinical trial data set ensures reliability, reproducibility, and robustness of the model. 

We did three experiments in the present study. In the first experiment, we evaluated the role of baseline textural features extracted from fluid, SHRM, and different retinal tissue and sub-RPE compartments to discriminate between patients with nAMD who achieved and maintained retinal fluid resolution (complete responders) and those who did not achieve or maintain retinal fluid resolution (incomplete responders) with anti-VEGF therapy. Multiple studies have indicated the presence of SRF and IRF, and SHRM is associated with worse visual outcomes and response to anti-VEGF therapy in patients with nAMD.^3,7,32 In the present study, the texture-based radiomic features within the sub-RPE compartment of baseline OCT scans were found to be most implicated in response to therapy. Higher expression of Laws energy feature was observed within the sub-RPE texture of the eyes that were complete responders to anti-VEGF therapy. This may be strongly related to the perfusion of the neovascular net in the sub-RPE space and the associated advanced vascular atrophy³³ or the accumulation of cellular components within the sub-RPE compartment. This is consistent with the previous findings,^10,33 where sub-RPE features were found to be associated with therapeutic response in patients with nAMD. 

In the second experiment, we sought to evaluate the role of changes in texture features (delta texture) across different fluid, SHRM, and retinal tissue compartments between baseline and posttherapy OCT scans to discriminate between complete responders and incomplete responders to anti-VEGF therapy. The SHRM compartment was identified to contain the most discriminating delta features for distinguishing between the two groups of patients. The Laws energy descriptor was overly expressed for the complete responders, and a reduction in the heterogeneity was also observed posttherapy among the complete responders compared with the incomplete responders. This may be related to the reduction in fluid component within the SHRM compartment induced by anti-VEGF therapy.³⁴ This is very much in line with our prior work¹⁵ on the OSPREY data set. This finding is also consistent with previous findings^10,32 where the SHRM volume was found to be reduced significantly with anti-VEGF therapy. 

In-depth characterization of different higher-order clinical parameters, such as retinal tissue volume, intra- and subretinal fluid volume, SHRM volume, and sub-RPE volume, has been shown to be associated with anti-VEGF therapy treatment response.^10,35 Substantial reduction in sub-RPE volume and SHRM volume was evident in posttreatment OCT scans of patients with nAMD.¹⁰ Reduction in fluid volume has also been reported as the primary indicator of anti-VEGF therapy treatment response for patients with nAMD.^10,35 In the third experiment, we sought to investigate whether the combination of these clinical parameters with baseline and delta texture features improved the predictive performance of the classifier. In our experiment, the fusion of baseline and texture features with different clinical parameters yielded a statistically significant improvement (DeLong test²⁵) in the QDA classifier performance on the independent test set in their association with treatment response compared to the use of only baseline texture features and only delta texture features (P = 0.042 and 0.043, respectively). These findings further appear to validate the role of these features in association with treatment response and durability. These findings also suggest that the combination of different baseline and texture features requires further evaluation as potential predictive biomarkers for predicting therapeutic response. 

In the present study, the texture-based radiomic features within the sub-RPE compartment of baseline OCT scans were found to be most implicated in response to therapy. Higher expression of Laws energy feature was observed within the sub-RPE texture of the eyes that were complete responders to anti-VEGF therapy. The increased signal may reflect textural variability that is linked to structural features that are directly associated with the underlying pathology. This may be strongly related to the perfusion of the neovascular net in the sub-RPE space and the associated advanced vascular atrophy³³ or the accumulation of cellular components within the sub-RPE compartment. This is consistent with the previous findings,^10,33 where sub-RPE features were found to be associated with therapeutic response in patients with nAMD. Further analysis and validation are needed to be better understand the potential of phenotyping the sub-RPE compartment and neovascular pathology utilizing radiomics techniques and quantitative textural features. 

Additionally, Laws texture features capture the heterogeneity in the texture in the form of spots, edges, waves, ripples, and speckles. There might be discrepancy in the texture microarchitecture between the complete and incomplete responders of anti-VEGF therapy due to underlying disease pathophysiology. The most discriminating sub-RPE texture features and SHRM delta texture features identified in the present study might be capturing the discrepancies in microenvironment of different OCT compartments and the associated heterogeneity in texture between the two treatment response groups. 

As the prevalence of nAMD is growing rapidly all over the world, artificial intelligence (AI) models have huge potential for early identification of the optimal therapeutic to minimize risk of irreversible vision loss. One of the major concerns of translating AI models into clinical deployment is the need for a multipronged validation strategy. In that respect, the unique aspect of the present study is the post hoc validation of the model on a large completed clinical trial data set. Since the phase 3, multicenter HAWK clinical trial incorporates images from multiple sites with variations in imaging centers and image acquisition devices, the AI model validated on a clinical trial data set provides a unique opportunity for identification and evaluation of the new radiomics features in association with treatment response and might have broader implications for patient care and therapeutic efficacy. This also sets the stage of validating the model performance on a prospective interventional trial setting. 

The present study has some limitations that must be acknowledged. A detailed analysis of the sensitivity of the segmentation method on derived features needs to be addressed. Additionally, the nature of our study was retrospective rather than prospective. Independent prospective validation of the radiomic model is required to validate the texture-based radiomic features predictive of therapeutic response and durability of anti-VEGF therapy on patients with nAMD. Furthermore, the classification of the groups was based solely on IRF and SRF response without considering other OCT features (e.g., SHRM) or visual acuity,¹⁷ which is the most important outcome measures for eyes with nAMD. No external validation was performed for the OCT devices or segmentation software used. Any error in the segmentation of the inner and outer retinal boundaries (by the segmentation software) is likely to affect the accuracy of retinal thickness map and needs manual correction. This might have an impact on volumetric assessment of OCT scans and voxel size. Since radiomics features are sensitive to voxel size, any error in retinal thickness map and difference in voxel size could have a potential impact on the predictive capability of the radiomics features.³⁶ 

In conclusion, this study identified and evaluated OCT-derived texture-based radiomic features predictive of anti-VEGF therapy treatment response and therapeutic durability in patients with nAMD. In this analysis, we identified two biomarkers that characterize heterogeneity within the sub-RPE compartment and alteration of texture within the SHRM compartment that are most implicated to treatment response in eyes with nAMD. In future, the predictability of these features needs validation on a prospective interventional trial. 

Acknowledgments

Supported by NIH-NEI P30 Core Grant (IP30EY025585) (Cole Eye Institute), unrestricted grants from the Research to Prevent Blindness (Cole Eye Institute), and Cleveland Eye Bank Foundation awarded to the Cole Eye Institute (Cole Eye), K23-EY022947-01A1 (JPE). Research reported in this publication was also supported by the National Cancer Institute under award numbers R01CA268287A1, U01CA269181, R01CA26820701A1, R01CA249992-01A1, R01CA202752-01A1, R01CA208236-01A1, R01CA216579-01A1, R01CA220581-01A1, R01CA257612-01A1, 1U01CA239055-01, 1U01CA248226-01, and 1U54CA254566-01; National Heart, Lung, and Blood Institute (1R01HL15127701A1, R01HL15807101A1); National Institute of Biomedical Imaging and Bioengineering (1R43EB028736-01); National Center for Research Resources under award number 1 C06 RR12463-01; VA Merit Review Award IBX004121A from the US Department of Veterans Affairs Biomedical Laboratory Research and Development Service; the Office of the Assistant Secretary of Defense for Health Affairs, through the Breast Cancer Research Program (W81XWH-19-1-0668); the Prostate Cancer Research Program (W81XWH-20-1-0851); the Lung Cancer Research Program (W81XWH-18-1-0440, W81XWH-20-1-0595); the Peer Reviewed Cancer Research Program (W81XWH-18-1-0404, W81XWH-21-1-0345, W81XWH-21-1-0160); the Kidney Precision Medicine Project (KPMP) Glue Grant; and sponsored research agreements from Bristol Myers-Squibb, Boehringer-Ingelheim, Eli-Lilly, and Astrazeneca. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health, the US Department of Veterans Affairs, the Department of Defense, or the US government. 

Disclosure: S. Sil Kar, None; H. Cetin, None; S.K. Srivastava, Regeneron (F, C), Gilead (F), Bausch and Lomb (C), Novartis (C); A. Madabhushi, Astrazeneca (F), Bristol Myers-Squibb (F), Boehringer-Ingelheim (F), Eli-Lilly, Picture Health (F, C), Inspirata (F), Elucid Bioimaging (F), Aiforia (C), Caris (C), Roche (C), Biohme (C), Castle Biosciences (C), SimBioSys (C); J.P. Ehlers, Aerpio (F, C), Alcon (F, C), Thrombogenics/Oxurion (F, C), Regeneron (F, C), Genentech (F), Novartis (F, C), Allergan (F, C), Iveric BIO (F, C), Stealth (F, C), Roche (F), Adverum (C), Apellis (C), Allegro (C), Genentech/Roche (C), Leica (C, P), Zeiss (C), Santen (C), Janssen (C), RegenxBIO (C)