AI-Guided Radiomics for Spectral-Domain OCT Characterization of Acquired Vitelliform Lesions: A Pilot Study

Objective: To evaluate the ability of artificial intelligence (AI)-driven radiomics applied to structural OCT to characterize hypo-reflective subretinal spaces in acquired vitelliform lesions (aVLs), to distinguish reabsorbed material from subretinal fluid, improving diagnostic precision. Methods: In this retrospective monocentric study, patients with aVLs were compared with a control group of acute central serous chorioretinopathy (CSCR) to provide reference OCT patterns for subretinal fluid. All participants underwent multimodal imaging. OCT scans were manually segmented. Radiomic features were extracted using a standardized pipeline, including image preprocessing, robustness assessment, variance and correlation filtering, and feature selection with a WEKA subset evaluator. A ridge-based Logistic Regression classifier was trained on 75% of the dataset and tested on the remaining 25%. Model performance was assessed using accuracy metrics, ROC AUC, and precision–recall curves. Results: A total of 520 radiomics features were extracted, of which nine were selected for training. The classifier achieved an accuracy of 86.5% in the training set (AUC 0.89; PRC 0.86) and 80% in the independent test set, with excellent AUC (0.92) and PRC (0.92). The negative predictive value reached 92%, indicating a high ability to correctly classify the aVL-associated hypo-reflective phenotype. The positive predictive value was 67%. Conclusions: AI-based radiomics demonstrated promising performance in discriminating, on structural OCT, between hypo-reflective subretinal spaces associated with the resorptive phase of aVLs and subretinal fluid. These preliminary findings support the feasibility of quantitative radiomic characterization of these two OCT phenotypes. Further studies are required to determine whether this approach may also be useful in identifying neovascular complications.