Deep learning-based prediction of one-year functional and structural outcomes in treatment-naïve DMO after anti-VEGF therapy.

Abstract

Purpose: To evaluate whether baseline structural optical coherence tomography (OCT) scans contain prognostic information that can be used by deep learning models to predict one-year functional and anatomical outcomes in treatment-naïve diabetic macular oedema (DMO).

Design: Retrospective cohort study.

Participants: Eighty-seven eyes (74 patients) with treatment-naïve DMO receiving intravitreal anti-VEGF therapy.

Methods: Baseline OCT volumes were analysed using a deep learning regression model. The images were homogenised - preprocessed and transformed into a common reference frame - to streamline the learning task. The model predicted 12-month changes in best-corrected visual acuity (BCVA) and central subfield thickness (CST).

Main outcome measures: Performance was assessed using 5-fold cross-validation and compared between original and preprocessed datasets. Receiver operating characteristic (ROC) analysis evaluated classification of responders versus non-responders, while occlusion sensitivity analysis identified image regions contributing most to predictions.

Results: Mean BCVA improved from 0.40 ± 0.40 to 0.33 ± 0.30 logMAR, and CST decreased from 390.2 ± 77.1 µm to 342.2 ± 76.3 µm (both p < 0.001). Prediction of BCVA change achieved an R² of 0.575, improving to 0.661 after preprocessing. CST prediction improved from R² = 0.447-0.633. Classification performance increased from AUC 0.65 to 0.71 for BCVA response and from 0.93 to 0.97 for CST response. Key predictive regions included the outer retina and retinal pigment epithelium.

Conclusions: Deep learning analysis of baseline OCT can predict one-year outcomes in DMO. Image homogenisation enhances performance, and responder identification suggests the potential utility of this approach for prognostic assessment and risk stratification, warranting further validation before clinical implementation.