# paper.pdf ## Metadata - PDFFormatVersion=1.6 - IsLinearized=true - IsAcroFormPresent=false - IsXFAPresent=false - IsCollectionPresent=false - IsSignaturesPresent=false - CreationDate=D:20260527125029+05'30' - ModDate=D:20260527125050+05'30' - Producer=Acrobat Distiller 10.0.0 (Windows) - pdf:keywords="radiomics"; "diabetic retinopathy (RD)"; "fundus photography" - dc:format=application/pdf - dc:title=Radiomics-Based Fundus Photography Analysis in Diabetic Retinopathy - dc:identifier=doi:10.1167/tvst.15.5.27 - dc:description=Translational Vision Science and Technology, 2026. doi:10.1167/tvst.15.5.27 - prism:doi=10.1167/tvst.15.5.27 - prism:url=https://doi.org/10.1167/tvst.15.5.27 - prism:aggregationtype=journal - prism:publicationname=Translational Vision Science and Technology - prism:issn=2164-2591 - dc:creator=Elham Sadeghi,Ryan Chace Williamson,Francesca Corona,Elli Davis,Michael Kozlov,Kiran Kumar Vupparaboina,Sandeep Chandra Bollepalli,José-Alain Sahel,Jay Chhablani,Evan L. Waxman - dc:subject=radiomics,diabetic retinopathy (RD),fundus photography - xmpmm:documentid=uuid:50d88584-e0e9-35e3-a3eb-d1b9c9ad39b0 - xmpmm:instanceid=uuid:cad60a56-82ba-371c-8d70-5ac68b5863a7 ## Contents ### Page 1 RetinaRadiomics-Based Fundus Photography Analysis in Diabetic Retinopathy Elham Sadeghi 1 , Ryan Chace Williamson1 , Francesca Corona1 , Elli Davis 2 , Michael Kozlov3 , Kiran Kumar Vupparaboina 1 , Sandeep Chandra Bollepalli 1 , José-Alain Sahel 1 , Jay Chhablani 1 , and Evan L. Waxman11 Department of Ophthalmology, University of Pittsburgh, Medical Center, Pittsburgh, PA, USA 2 Temple Medical University, Philadelphia, PA, USA 3 SUNY Downstate Health Sciences University, Brooklyn, NY, USACorrespondence: Evan L. Waxman, UPMC Vision Institute, UPMC Mercy Pavilion, 1622 Locust Street, Pittsburgh, PA 15219, USA. e-mail: waxmane@upmc.eduReceived: October 14, 2025Accepted: April 16, 2026Published: May 28, 2026Keywords: radiomics; diabetic retinopathy (RD); fundus photographyCitation: Sadeghi E, Williamson RC, Corona F, Davis E, Kozlov M, Vupparaboina KK, Bollepalli SC, Sahel JA, Chhablani J, Waxman EL. Radiomics-based fundus photography analysis in diabetic retinopathy. Transl Vis Sci Technol. 2026;15(5):27, https://doi.org/10.1167/tvst.15.5.27Purpose: The purpose of this study was to apply radiomics feature extraction from fundus photographs to automatically identify biomarkers that distinguish different stages of diabetic retinopathy (DR).Methods: A total of 52 radiomic features were extracted from fundus photographs representing different DR stages: diabetes without DR (n = 35), mild non-proliferative DR (NPDR; n = 33), moderate NPDR (n = 35), severe NPDR (n = 35), and prolif- erative DR (PDR; n = 34). Additionally, 68 images from 34 eyes with NPDR over 1 year of follow-up were analyzed. Statistical significance was evaluated using 95% bootstrap confidence intervals. Wilcoxon rank-sum tests, linear regression, and multi- class logistic regression were performed using R software (version 4.5.1) and Python software.Results: When comparing eyes without DR to all NPDR eyes, significant differences were detected in 6 of 16 first-order statistics (FOS), 1 of 14 gray-level co-occurrence matrix (GLCM), 13 of 16 Gabor texture (GT), and 1 of 6 Law’s texture energy (LTE) features (P <0.01). Several features demonstrated a linear trend with increasing DR severity, whereas PDR showed a distinct pattern. Across all 5 stages, 14 FOS, 7 GLCM, 11 GT, and 1 LTE features differed significantly (P < 0.05). Overall, 34 of 52 features significantly distin- guished no DR from NPDR, and 17 discriminated across all stages. Longitudinally, 34 NPDR eyes showed no significant 1-year changes.Conclusions: Radiomic features from fundus photographs may help in distinguishing eyes without DR from NPDR, demonstrating strong potential for automated DR classifi- cation and screening applications.Translational Relevance: Radiomic-derived biomarkers from fundus photographs may provide automated, objective support for DR screening and staging.IntroductionDiabetic retinopathy (DR) is a common ocular disease in adults that can lead to blindness. In addition to retinal neurodegeneration, chronic hyperglycemia weakens the vessel walls, resulting in microvascu- lar abnormalities such as microaneurysms, vascular leakage, exudates, and capillary occlusion, leading to fibrous tissue formation and neovascular prolifera- tion.1,2 In 2020, the global prevalence of DR among individuals with diabetes was estimated at 19.7% to 25.0%, affecting approximately 103.1 million adults worldwide, which is expected to reach nearly 700 million by 2045. 3 Early screening for DR is critical for prevent- ing severe complications and minimizing the risk of progression to proliferative DR (PDR) or diabeticCopyright 2026 The Authors tvst.arvojournals.org | ISSN: 2164-2591 1 This work is licensed under a Creative Commons Attribution 4.0 International License. ### Page 2 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 2macular edema. Adoption of advanced screening technologies and timely detection can significantly reduce the global burden of DR, improve clinical outcomes, and avert vision loss in millions of individu- als worldwide.4,5 Deep learning has markedly advanced the screening and diagnosis of DR. Trained on large datasets of retinal images, these neural networks can automatically identify disease-specific patterns and detect abnormalities across the spectrum of DR sever- ity. These algorithms can process vast amounts of data rapidly and continuously improve as they are exposed to more data, thereby reducing the risk of human error and enhancing the efficiency of screening programs. 6,7 In underserved and remote regions with limited access to specialists, these technologies can be integrated into telemedicine platforms, enabling non- specialist healthcare workers to conduct effective initial screenings for DR.4 The US Food and Drug Adminis- tration (FDA)-approved machine learning algorithms in DR are IDx-DR (Digital Diagnostics, formerly IDx Technologies) and EyeArt (Eyenuk, Inc.), which are for detecting more than mild and referable DR. 8,9 Radiomic-based texture analysis uses standard image processing techniques to extract quantitative texture features from medical images and is widely applied in radiology and pathology.10,11 In recent years, its use in ophthalmology has expanded across various imaging modalities and diseases, including optical coherence tomography (OCT) and OCT angiogra- phy (OCTA), fundus photography, and fluorescein angiography, to predict, classify, and monitor the progression of DR, central serous chorioretinopathy (CSCR), age-related macular degeneration (AMD), epiretinal membrane, myopic maculopathy, and optic neuropathies. 12–19 However, most of these efforts have concentrated on retinal layer analysis in OCT scans, with comparatively limited attention given to fundus photographs, despite their everyday use in the follow- up of patients with diabetes mellitus in telemedicine patient care. In this study, we apply radiomic feature extraction to fundus photographs in patients with diabetic melli- tus, with and without DR in different stages, to identify texture-based biomarkers associated with DR. We first identify novel radiomic-derived fundus photographs that distinguish eyes with no DR from those with non-proliferative DR (NPDR). Additionally, we aim to evaluate how image features vary across different stages of DR and how they change over time as the disease progresses. These results may facilitate follow- up of patients in remote or underserved areas who lack access to OCT and wide-field fundus photography for DR screening.Methods Data AcquisitionWe conducted a retrospective study that used 45- degree fundus photographs to collect high-quality images from patients diagnosed with diabetic melli- tus with and without DR. Patients were recruited by the Guerrilla Eye Service, founded in 2005 by the senior author (E.W.) at the University of Pittsburgh School of Medicine. The study adhered to the princi- ples of the Declaration of Helsinki and was approved by the Institutional Review Board of the University of Pittsburgh. The images were initially graded by an ophthalmologist at the time of examination. During study enrollment, a second ophthalmologist indepen- dently re-evaluated the images, and any cases with discrepancies in DR staging were excluded from the study. We included 172 fundus photographs by quality assessment, from patients with diabetes mellitus: 35 with no DR, 33 with mild NPDR, 35 with moder- ate NPDR, 35 with severe NPDR, and 34 with PDR (Fig. 1). Additionally, we analyzed 34 eyes with NPDR that had 2 fundus images obtained over a 1-year follow- up period (Fig. 2). Each participant received a compre- hensive ophthalmic examination, including detailed medical history, visual acuity measurement, intraocu- lar pressure assessment, slit-lamp biomicroscopy, and dilated fundus examination to confirm the presence or absence of pathology. Exclusion criteria were applied to eliminate eyes with other vitreoretinal pathologies, including AMD, uveitis, retinal vascular occlusion, high myopia, epireti- nal membrane, or any condition that could affect the fundus images. Images with artifacts or shadows that could interfere with image acquisition or texture analy- sis were also excluded. All the fundus images had the same dimensions, field of view, and were of consistently high quality.Feature ExtractionsRadiomic feature extraction was performed on the whole 45-degree fundus images from CenterVue DRSplus (CenterVue S.p.A., Padua, Italy), and each image was captured at a resolution of approximately 10 megapixels (approximately 77–80 pixels per degree of retina), sufficient for detailed analysis of the poste- rior pole using the Pyfeats python-based radiomics library.20 All color fundus images were converted to grayscale prior to analysis by using black-and-white filter using the Windows Photos app on Windows 11. ### Page 3 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 3Figure 1. Fundus photography of the right eyes of five patients with diabetes with no diabetic retinopathy (1), mild non-proliferative diabetic retinopathy (NPDR) (2), moderate NPDR (3), severe NPDR (4), and proliferative diabetic retinopathy (PDR) (5). The gray-scale images were used to extract radiomics features.Figure 2. Fundus photography of the right eye with moderate non-proliferative diabetic retinopathy, at baseline (1) and 1 year later (2). The gray-scale images were used to extract radiomics features. ### Page 4 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 4In these images, brighter areas represent highly reflec- tive structures, such as the optic disc and blood vessels, whereas darker regions correspond to less reflective retinal areas. This processing standardized the images for analysis (see Fig. 1). A total of 52 texture features were extracted from each image, encompassing 4 categories: first-order statistics (FOS), gray-level co- occurrence matrix (GLCM), Gabor texture (GT), and Law’s texture energy (LTE)21–23 :• FOS features (n = 16): Quantified the inten- sity distribution of pixels, including percentiles (10th, 25th, 75th, and 90th), coefficient of varia- tion, energy, entropy, histogram width, kurto- sis, mean, median, mode, skewness, and variance. These measures describe the global brightness and variability within the segmented choroid.• GLCM features (n = 14): Calculated from co- occurrence matrices at a fixed pixel offset and included angular second moment, contrast, corre- lation, sum of squares variance, inverse difference moment, sum average, sum variance, sum entropy, entropy, difference variance, difference entropy, information measures of correlation, and maximal correlation coefficient. These capture spatial dependencies and textural regularity between neighboring pixels.• GT (Gabor) features (n = 16): Obtained by normalizing pixel values to the range −0.5 to 0.5 and applying Gabor filters with 4 orientations (0 degrees, 45 degrees, 90 degrees, and 135 degrees) and 2 spatial frequencies (0.1 and 0.4 cycles/pixel). The mean and standard deviation of the filtered outputs were computed for each orientation– frequency pair, representing multi-scale, multi- directional texture information.• LTE features (n = 6): Derived using Law’s texture energy method with a 3 × 3 convolution mask, capturing localized energy patterns and directional textures.Statistical AnalysisWe first compared eyes without DR (group 1= 35 images) to eyes with any stage of NPDR (groups 2–4 = 138 images) to better characterize the differences between normal fundus photographs and those with NPDR. All extracted features were then compared across the five groups to assess trends in feature changes with increasing disease severity. Additionally, we analyzed a subset of 34 eyes with NPDR that had 2 fundus images obtained over a 1 year follow up period. Although their disease stage remained unchanged, these eyes exhibited increased hemorrhages and exudates, allowing us to investi- gate radiomic feature changes associated with subtle progression in fundus appearance. All numeric extracted features were compared between early and advanced DR using two-sided Wilcoxon rank-sum tests, with Bonferroni correc- tion applied for multiple comparisons. Features were assigned to one of four predefined families (FOS, GLCM, Gabor, and LTE), and the two features with the lowest adjusted P values in each family were selected for visualization. Selected features were standardized using z-scores, and 95% confi- dence intervals for the median were estimated using 1000 bootstrap resamples. Boxplots with overlaid bootstrapped confidence intervals were generated for the selected features. To evaluate the discriminatory ability of radiomics features, least absolute shrinkage and selection opera- tor (LASSO)-regularized logistic regression was used to distinguish eyes in group 1 from those in groups 2 to 4. Model performance was assessed using repeated nested stratified cross-validation (5 outer folds and 5 inner folds, repeated 50 times). The regularization parameter was selected in the inner loop based on the area under the receiver operating characteristic curve (AUC) using the λ1 se selection rule. Discrimination was quantified using mean AUC derived from out-of- fold predictions, with empirical 95% uncertainty inter- vals estimated across repetitions. A Youden threshold was identified from out-of-fold predictions within each repetition, and sensitivity and specificity were averaged across repetitions. Linear regression analysis and multi- class logistic regression were done for the prediction of DR stage. All analyses were performed in R software (version 4.5.1) using the pROC and glmnet packages, as well as Python software.ResultsOverall, 172 fundus photographs from patients with diabetes mellitus were analyzed, including 35 with no DR, 33 with mild NPDR, 35 with moderate NPDR, 35 with severe NPDR, and 34 with PDR. In addition, 34 eyes with NPDR that had 2 fundus images obtained over a 1 year follow-up period were analyzed. Because the radiomic extraction is fully automated, running the pipeline multiple times on the same image yields identical features, ensuring complete model stability and reproducibility.No DR Versus NPDRTo investigate whether radiomic features could distinguish images without DR from those with ### Page 5 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 5Figure 3. The difference between selected radiomics features from fundus photographs with no diabetic retinopathy (35 images) with green color versus eyes with non-proliferative diabetic retinopathy (138 images) with purple color.NPDR, we compared 35 fundus photographs without DR to 138 photographs with any stage of NPDR (Fig. 3).• FOS: Of the 16 extracted features, 6 were signif- icantly different between the 2 groups, includ- ing FOS_mean (P = 0.001), FOS_median (P <0.001), FOS_mode (P = 0.007), FOS_coeffi- cient_of_variation (P = 0.002), FOS_25 percentile (P < 0.001), and FOS_75 percentile (P = 0.007).• GLCM: Among the 14 extracted features, GLCM_sum_average_mean showed a significant difference (P = 0.001).• GT (Gabor) features: Of the 16 features, 13 demon- strated significant differences between the groups, including:• GT_th_0_0_freq_0_1_mean (P = 0.001), GT_th _0_0_freq_0_4_mean (P < 0.001), GT_th_0_0 _freq_0_4_std (P = 0.005),• GT_th_1_0_freq_0_1_mean (P = 0.001), GT_th _1_0_freq_0_4_mean (P < 0.001), GT_th_1_0 _freq_0_4_std (P = 0.002),• GT_th_2_0_freq_0_1_mean (P < 0.001), GT_th _2_0_freq_0_1_std (P = 0.034), GT_th_2_0_freq _0_4_mean (P < 0.001), GT_th_2_0_freq_0_4 _std (P = 0.015),• GT_th_3_0_freq_0_1_mean (P = 0.001), GT_th _3_0_freq_0_4_mean (P < 0.001), and GT_th _3_0_freq_0_4_std (P = 0.002).• LTE: Of the 6 extracted features, LTE_ss_3 was significantly different between groups (P = 0.047). The LASSO-regularized logistic regression model used to discriminate eyes in group 1 from those in groups 2 to 4 achieved a mean AUC of 0.742 (95% empirical interval = 0.701–0.775), with a sensitivity of 0.692 (95% empirical interval = 0.592–0.775) and a specificity of 0.755 (95% empirical interval = 0.664– 0.879).Different Stages of DRTo evaluate whether radiomic features could differ- entiate between stages of DR, we analyzed 35 fundus photographs without DR (group 1), 33 with mild NPDR (group 2), 35 with moderate NPDR (group 3), 35 with severe NPDR (group 4), and 34 with PDR (group 5). A linear trend of feature changes was ### Page 6 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 6Figure 4. Differences in selected radiomics features across fundus photographs from eyes with no diabetic retinopathy (DR; group 1, n =35), mild NPDR (group 2, n = 33), moderate NPDR (group 3, n = 35), severe NPDR (group 4, n = 35), and proliferative DR (PDR) (group 5, n =34).observed with increasing disease severity from no DR to severe NPDR, whereas PDR exhibited a distinct pattern across all images (Fig. 4).• FOS: Of the 16 extracted features, 14 were signif- icantly different among the 5 groups, includ- ing FOS_mean (P = 0.001), FOS_variance (P = 0.021), FOS_median (P < 0.001), FOS_mode (P = 0.010), FOS_skewness (P < 0.001), FOS_kurtosis (P = 0.007), FOS_energy (P <0.001), FOS_entropy (P < 0.001), FOS_coefficient of variation (P = 0.006), FOS_10 percentile (P < 0.001), FOS_25 percentile (P < 0.001), FOS_75 percentile (P < 0.001), FOS_90 percentile (P = 0.016), and FOS_histogram width (P =0.021).• GLCM: Among the 14 extracted features, 7 features were statistically significant among all groups including GLCM_ASM_mean (P = 0.005), GLCM_sum of squares variance_mean (P = 0.024), GLCM_sum average_mean (P = 0.001), GLCM_sum variance_mean (P = 0.023), GLCM_sum entropy_mean (P <0.001), GLCM_entropy_mean (P = 0.001), and GLCM_information 2_mean (P = 0.011).• GT (Gabor) features: Of the 16 features, 11 demon- strated significant differences among the groups, including:• GT_th_0.0_freq_0.1_mean (P = 0.002), GT_th _0.0_freq_0.4_mean (P = 0.001), GT_th_0.0 _freq_0.4_std (P = 0.012),• GT_th_1.0_freq_0.1_mean (P = 0.002), GT_th _1.0_freq_0.4_mean (P = 0.001), GT_th_1.0 _freq_0.4_std (P = 0.009),• GT_th_2.0_freq_0.1_mean (P = 0.003), GT_th _2.0_freq_0.4_mean (P = 0.002),• GT_th_3.0_freq_0.1_mean (P = 0.002), GT_th _3.0_freq_0.4_mean (P = 0.001), GT_th_3.0 _freq_0.4_std (P = 0.007).• LTE: Of the six extracted features, LTE_LL_3 was significantly different among the five DR stages (P = 0.022).Linear Regression Analysis for the Prediction of DR Stage From Individual Features in no DR and all NPDRIn this analysis, all DR stages were put as the X variable and each individual feature as the Y variable, and we determined how many were significant for each ### Page 7 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 7Table 1. Linear Regression Analysis for the Predic- tion of Diabetic Retinopathy (DR) Stage From Individual Features in no DR and all Non-Proliferative DR Features (Linear RegressionR2 Results) Mean CI, Low CI, High Number of SignificantP values FOS 0.088 0.051 0.124 11 out of 16 LTE 0.047 0.022 0.069 4 out of 6 GLCM 0.035 0.017 0.060 6 out of 14 GT 0.105 0.079 0.129 13 out of 16 FOS, first-order statistics; GLCM, gray-level co-occurrence matrix; GT, Gabor texture; LTE, Law’s texture energy; CI, confi- dence interval.Table 2. Multi-Class Logistics Regression to Predict Diabetes Stages Mean CI, Low CI, High All features 0.366 0.322 0.417 All features shuffle 0.214 0.162 0.258 FOS features 0.395 0.346 0.444 LTE features 0.326 0.238 0.403 GT features 0.268 0.218 0.317 GLCM features 0.326 0.244 0.383feature group. The R-squared values are small, but in the first analysis above, we show that combined, they allow for significant classification (Table 1).Multi-Class Logistics Regression to Predict Diabetes StagesThe results show classification accuracy that has non-overlapping confidence intervals with the shuffle control for classification using all features and classi- fication using FOS features alone. The others (LTE, GC, and GLCM) had overlapping confidence intervals, suggesting that they are not significantly better than chance (Table 2).Linear Regression Results for the Prediction of DR Stage From Individual Features in all DR StagesAll DR stages were put as the X variable and each individual feature as the Y variable, and we determined how many were significant for each feature group. The R-squared values are small, but in the first analysis above, we show that combined, they allow for signifi- cant classification (Table 3).Table 3. Linear Regression Results for the Prediction of Diabetic Retinopathy (DR) Stage From Individual Features in all DR Stages Features (Linear RegressionR2 Results) Mean CI, Low CI, High Number of SignificantP Values FOS 0.024 0.013 0.038 6 out of 16 LTE 0.024 0.010 0.040 4 out of 6 GLCM 0.026 0.013 0.038 6 out of 14 GT 0.007 0.004 0.012 1 out of 16Table 4. Linear Regression Analysis for the Prediction of Diabetic Retinopathy (DR) Differences From Individ- ual Features in Longitudinal Data Features Visit 1 vs. Visit 2 (Linear Regression R2 Results) Mean CI, Low CI, High Train 0.776 0.733 0.818 Test 0.528 0.453 0.608 First visit 2.705 2.147 3.206 Second visit 2.470 1.852 3.029 Second first visit −0.235 −0.794 0.294Longitudinal AnalysisTo evaluate whether radiomic features could capture subtle progression over time, we analyzed 34 eyes with NPDR that had fundus photographs obtained at baseline and after 1 year of follow-up. Among the 52 extracted features, 3 showed statis- tically significant differences between the 2 time points: GLCM_difference variance_mean (P = 0.020), GLCM_difference entropy_mean (P = 0.044), and GT_th_0.0_freq_0.1_std (P = 0.033). However, none of these associations remained significant after apply- ing Bonferroni correction (Fig. 5). The linear regression analysis for the prediction of DR differences from individual features in longitudinal data showed no significant difference between the early and late stages (Table 4).DiscussionIn this study, we extracted and compared radiomic features from fundus photographs of patients with DR. Our analysis showed that radiomics could distin- guish fundus photographs with no DR from those with NPDR. Significant differences were also observed ### Page 8 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 8Figure 5. In the follow-up analysis of 34 eyes with NPDR, 3 radiomic features demonstrated significant differences between baseline (green) and 1 year of follow-up (purple); however, these differences did not remain significant after Bonferroni correction.across five stages of DR, including no DR, mild, moderate, and severe NPDR, and PDR. However, radiomics was not sensitive enough to detect subtle changes in fundus photographs in eyes with NPDR in a 1-year follow-up. PDR exhibited features that were distinct from the other stages. In this study, radiomics was used as a handcrafted feature extrac- tion framework, in which predefined quantitative descriptors were computed from fundus images. These features were subsequently used as inputs to conven- tional machine learning classifiers for feature selection and classification. Importantly, no end-to-end feature learning or deep learning-based representation learn- ing was performed, allowing a clear separation between feature engineering and model training. The rising number of DR cases poses a public health challenge, making regular screening essential. Whereas manual diagnosis is time-consuming, expert- dependent, and subject to interobserver variabil- ity, artificial intelligence provides a reliable tool to support clinicians and enhance diagnostic consistency; however, current FDA-approved algorithms remain limited by their time requirements, licensing costs, and narrow disease scope.8,9,19 Radiomics uses mathemat- ical formulas to analyze grayscale histograms, region- of-interest shapes, and texture-defining matrices. 24 It offers several advantages over traditional deep learning for feature extraction, including the ability to operate with small datasets, limited computing power, and standardized, interpretable features that allow in-depth study and improved model transparency. 16 Addition- ally, its predefined and potentially nonlinear feature transformations are computationally efficient yet flexi- ble. They can be combined with other machine learn- ing classifiers to create interpretable, sample-efficient models applicable across diverse settings.16 The identi- fication of DR stages through radiomics has the potential to significantly improve diagnostic accuracy and streamline disease staging, particularly within telemedicine with extensive patient data. Traditional evaluation of fundus photographs by expert ophthal- mologists provides valuable insights for screening but is limited by its time-consuming nature and reliance on manual interpretation. In contrast, radiomics enables quantitative analysis of image heterogeneity with minimal assumptions about structure, offering a faster and more detailed assessment of the underlying archi- tecture. Our findings suggest that radiomic features can capture fundus changes associated with DR, indicat- ing potential utility in supporting telemedicine-based patient care. Radiomics, although initially established in oncol- ogy and showing considerable promise, remains relatively underexplored in ophthalmology due to the complexity of ocular imaging and the diversity of disease-specific modalities. 19,25 Radiomics has been applied to retinal diseases primarily for diagnostic screening, treatment response prediction, differential diagnosis, staging, and longitudinal monitoring, with particular emphasis on AMD, DR, CSCR, and other macular disorders. Studies to date have leveraged diverse imaging modalities, including OCT, OCTA, fundus photography, and ultra-widefield fluorescein angiography, to extract quantitative features that enhance disease characterization. 13,16,17,19 Several studies have demonstrated the potential of radiomics in DR detection. Baffa et al. applied a radiomics-based approach to fundus images of ### Page 9 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 9patients with mild DR and controls, using a deep neural network classifier, and achieved 94% accuracy and 93.34% sensitivity, highlighting the promise of computer vision in ophthalmology.26 Carrera-Escalé et al. evaluated multiple classifiers using OCT, OCTA, and fundus photography for diagnosing DM, DR, and referable DR, finding that OCTA-based models performed best for DR and referable DR, whereas OCT with logistic regression performed best for DM. 13 Shamsan et al. developed 3 approaches, each combining radiomic features from Dense-121 or Alex models with handcrafted features, achieving early DR detection with 97.92% sensitivity, 99.1% accuracy, 99.4% specificity, and 99.06% precision.27 Addition- ally, Soren et al. applied radiomics analysis to ultra- wide OCTA and observed significant differences in radiomic features across increasing DR severity, from no DR to PDR. 28 Although most published studies focus on early detection of DR from healthy eyes using wide-field imaging, OCT, or OCTA, in real- world telemedicine practice, such modalities are not universally available, and patient follow-up is most commonly performed using standard fundus photog- raphy. Our results demonstrate that radiomic feature extraction from fundus photographs can distinguish patients with no DR from those with NPDR, which may help identify patients with diabetes in telemedicine follow-up who require closer monitoring and timely intervention. We also observed progressive changes in multiple features in different stages of DR from no DR to severe NPDR, although subtle changes during follow-up may not be detectable with this approach. The distinct features observed in PDR compared with other stages may be attributable to massive retinal hemorrhages, laser scars, or tractional retinal detach- ment, all of which can affect radiomic features. The main limitations of this study include its retro- spective design, relatively small sample size, single device, single-region setting, and variability in image quality due to intraocular lens status, media opacity, or subtle cataract changes, which may alter fundus image texture, lack of wide-field imaging, and short follow-up period, during which subtle fundus changes occurred, but no DR stage progression was observed. Neverthe- less, as a proof-of-concept demonstration of radiomics for identifying pathology in fundus photographs, several measures, such as using standardized radiomics features and a simple Wilcoxon rank-sum test, were implemented to reduce model complexity and overfit- ting; however, larger and longer-term studies, compar- ing the performance of additional machine learning models to provide a more comprehensive assessment of radiomic-based classification will be needed to validate and generalize these findings.ConclusionsOverall, this study demonstrates that radiomic texture analysis can identify features distinguishing diabetic patients with and without NPDR in fundus images and reveal significant differences across DR stages, from no DR to mild, moderate, severe NPDR, and PDR; however, it was not sensitive enough to detect subtle changes during short-term follow-up. These findings suggest that radiomic features can capture fundus alterations associated with DR, with possible relevance for telemedicine applications.AcknowledgmentsSupported by NIH Core Grant P30 EY08098 to the Department of Ophthalmology, The Eye and Ear Foundation of Pittsburgh, and an unrestricted grant from Research to Prevent Blindness, New York, NY. Disclosure: E. Sadeghi, None; R.C. Williamson, Bausch + Lomb (R); F. Corona, None; E. Davis, None; M. Kozlov, None; K.K. Vupparaboina, Netra- Mind Innovations (O); S.C. Bollepalli, NetraMind Innovations (O); J.-A. Sahel, NetraMind Innova- tions (O), Pixium Vision (O), GenSight Biologics (O), Sparing Vision (O), Prophesee (O), and Chronolife (O);J. Chhablani, NetraMind Innovations (O), Allergan (C), Novartis (C), Salutaris (C), OD-OS (C), Erasca (C), B&L (C), Iveric Bio (C), Ocular Therapeutics (F), AcuViz (F), AbbVie (F), Springer (R), Elsevier (R);E.L. Waxman, NoneReferences1. Afarid M, Sadeghi E, Johari M, Namvar E, Sanie-Jahromi F. Evaluation of the effect of garlic tablet as a complementary treatment for patients with diabetic retinopathy. J Diabetes Res. 2022;2022(1):6620661. 2. Akhtar S, Aftab S. A framework for diabetic retinopathy detection using transfer learning and data fusion. Int J Inform Technol Comput Sci (IJITCS). 2024;16(6):61–73. 3. Teo ZL, Tham Y-C, Yu M, et al. Global preva- lence of diabetic retinopathy and projection of bur- den through 2045: systematic review and meta- analysis. Ophthalmology. 2021;128(11):1580–1591. 4. Akhtar S, Aftab S, Ali O, et al. A deep learning based model for diabetic retinopathy grading. Sci Rep. 2025;15(1):3763. ### Page 10 Radiomics Features in Diabetic Retinopathy TVST | May 2026 | Vol. 15 | No. 5 | Article 27 | 105. Ghosh S, Chatterjee A, eds. Transfer-ensemble learning based deep convolutional neural networks for diabetic retinopathy classification. 2023 3rd International Conference on Advancement in Elec- tronics & Communication Engineering (AECE); 2023: IEEE. Available at: https://ieeexplore.ieee. org/document/10428233. 6. Singh LK, Khanna M, Thawkar S. A novel hybrid robust architecture for automatic screening of glaucoma using fundus photos, built on feature selection and machine learning-nature driven com- puting. Expert Syst. 2022;39(10):e13069. 7. Singh LK, Khanna M, Thawkar S, Singh R. Nature-inspired computing and machine learn- ing based classification approach for glaucoma in retinal fundus images. Multimed Tools Appl. 2023;82(27):42851–42899. 8. US Food and Drug Administration (FDA). De Novo Classification Request for IDx-DR. 2018. Available at: https://www.accessdata.fda.gov/cdrh _docs/reviews/DEN180001.pdf. 9. US Food and Drug Administration (FDA). 510(k) Summary for EyeArt. 2020. Available at: https://www.accessdata.fda.gov/cdrh_docs/pdf20/ K200667.pdf. 10. Coroller TP, Agrawal V, Narayan V, et al. Radiomic phenotype features predict pathological response in non-small cell lung cancer. Radiother Oncol. 2016;119(3):480–486. 11. Zhang G, Xu L, Zhao L, et al. CT-based radiomics to predict the pathological grade of bladder cancer.Eur Radiol. 2020;30(12):6749–6756. 12. Banerjee I, de Sisternes L, Hallak JA, et al. Predic- tion of age-related macular degeneration disease using a sequential deep learning approach on lon- gitudinal SD-OCT imaging biomarkers. Sci Rep. 2020;10(1):15434. 13. Carrera-Escalé L, Benali A, Rathert A-C, et al. Radiomics-based assessment of OCT angiography images for diabetic retinopathy diagnosis. Ophthal- mol Sci. 2023;3(2):100259. 14. Savastano MC, Vagni M, Carlà MM, et al. Radiomic feature extraction from OCT angiogra- phy of idiopathic epiretinal membranes and corre- lation with visual acuity: a pilot study. Ophthalmol Sci. 2025;5(3):100716. 15. Meng Z, Chen Y, Li H, et al. Machine learning and optical coherence tomography-derived radiomics analysis to predict persistent diabetic macular edema in patients undergoing anti-VEGF intrav- itreal therapy. J Transl Med. 2024;22(1):358. 16. Williamson RC, Vupparaboina KK, Bollepalli SC, et al. Radiomics-based OCT analysis of choroid reveals biomarkers of central serous chori- oretinopathy. Transl Vis Sci Technol. 2025;14(4):23. 17. Williamson RC, Selvam A, Sant V, et al. Radiomics-based prediction of anti-VEGF treatment response in neovascular age-related macular degeneration with pigment epithelial detachment. Transl Vis Sci Technol. 2023;12(10):3. 18. Williamson R, Vupparaboina KK, Bollepalli SC, et al. Radiomics-based OCT analysis of unaf- fected fellow eyes uncovers biomarkers of central serous chorioretinopathy. Invest Ophthalmol Vis Sci. 2025;66(8):1724. 19. Zhang H, Zhang H, Jiang M, et al. Radiomics in ophthalmology: a systematic review. Eur Radiol. 2025;35(1):542–557. 20. Giakoumoglou N. PyFeats: open source soft- ware for image feature extraction. GitHub Reposi- tory. 2021. Available at: https://github.com/giakou moglou/pyfeats. 21. Haralick RM, Shanmugam K, Dinstein IH. Tex- tural features for image classification. IEEE Trans Syst Man Cybern. 2007(6):610–621. 22. Laws KI, ed. Rapid texture identification. Image processing for missile guidance. Bellingham, WA: SPIE Digital Library; 1980. 23. Gabor D. Theory of communication. J Institution Electric Eng. 1947;94(73):58. 24. Prinzi F, Orlando A, Gaglio S, Vitabile S. Breast cancer classification through multivariate radiomic time series analysis in DCE-MRI sequences.Expert Syst Appl. 2024;249:123557. 25. Lambin P, Leijenaar RT, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749–762. 26. Baffa MDFO, Martins JVG, Coelho AM, Felipe JC. Radiomic features for diabetic retinopa- thy early detection using deep neural networks. 17th International Conference on Signal-Image Technology & Internet-based Systems (SITIS), Bangkok, Thailand. IEEE Xplore. 2023:281– 286. 27. Shamsan A, Senan EM, Ahmad Shatnawi HS. Predicting of diabetic retinopathy development stages of fundus images using deep learning based on combined features. PLoS One. 2023;18(10): e0289555. 28. Soren VN, Prajwal H, Sundaresan V. Automated grading of diabetic retinopathy and radiomics analysis on ultra-wide optical coherence tomog- raphy angiography scans. Image Vis Comput. 2024;151:105292.