Author: Javier Zarranz-Ventura
Co-authors: Laura Carrera-Escale, Anass Benali, Ann-Christin Rathert, Ruben Martin-Pinardel, Anibal Ale-Chilet, Marina Barraso, Carolina Bernal-Morales, Sara Marin-Martinez, Silvia Feu-Basilio, Josep Rosines-Fonoll, Teresa Hernandez, Irene Vila, Rafael Castro, Cristian Oliva, Irene Vinagre, Emilio Ortega, Marga Gimenez, Alfredo Vellido, Enrique Romero
Abstract
Purpose:To evaluate the diagnostic accuracy of machine learning (ML) techniques applied to radiomic features extracted from different retinal imaging techniques for diabetes mellitus (DM), diabetic retinopathy (DR) and referable DR (R-DR) diagnosis.
Setting:
Cross sectional, image dataset analysis from a previous prospective optical coherence tomography angiography (OCTA) trial (ClinicalTrials.gov NCT03422965).
Methods:
All DM patients and controls underwent a comprehensive ocular examination that included a full battery of retinal imaging. Fundus retinographies (FR), structural OCT (OCT) 6x6mm macular cube scans and OCTA 3x3mm and 6x6mm macular cube scans were consecutively captured by an SD-OCT device (Cirrus HD-OCT Model 5000, Carl Zeiss Meditec). Two-dimensional OCTA images from the superficial capillary plexus (SCP) and the deep capillary plexus (DCP) were analyzed. Radiomic features were extracted from all retinal images in each study eye. Logistic regression (LR), linear discriminant analysis (LDA), support vector classifier (SVC)-linear, SVC-rbf and random forest (RF) ML models were created to evaluate their diagnostic accuracy for DM and DR (all grades) diagnosis in all images types. Models´ performance was described by their Area-Under-the-Curve (AUC) mean and standard deviation (SD).
Results:
A dataset of 763 eyes (465 individuals) was created. For DM diagnosis, the greatest AUC was observed for OCT (AUC 0.82, SD 0.03) and the combination of the three techniques (AUC 0.87, 0.03). For DR diagnosis, the greatest AUC within the DM cohort was observed for OCTA (0.77, 0.03), especially in the 3x3 mm superficial capillary plexus OCTA scan (AUC 0.76, 0.04). The combination of imaging techniques did not improve the performance of the models (0.77, 0.02). For R-DR detection, the greatest AUC was observed for OCTA (0.87, 0.12), especially in the 3x3 mm DCP OCTA scan (AUC 0.86, 0.08). Again, the combination of imaging techniques did not improve the performance of the models (0.87, 0.12). The addition of clinical variables (age, sex, etc.) improved most models AUC for DM (OCT: 0.86, 0.03) and DR diagnosis (OCTA: 0.82, 0.04), as well as for detection of R-DR (OCTA: 0.92, 0.02).
Conclusions:
Radiomics extracted from different retinal images allow identification of DM individuals, DR patients and R-DR patients using standard ML classifiers. OCT was the best test for DM diagnosis, and OCTA for DR and R-DR diagnosis. The combination of retinal image techniques (FR+OCT+OCTA) improved the performance of the models for DM diagnosis, but did not improve it for DR and R-DR diagnosis, as compared to single OCTA images. For all three classification tasks, the addition of clinical variables improved most models. This pioneering study demonstrates that radiomics-based ML techniques may be an option for DM, DR and R-DR diagnosis in the community.