Reliable MD Prediction from OCT
Abstract
Purpose
MD values quantify visual loss. Through OCT images, physicians can obtain MD values to assess patients' visual status. This task is crucial for early patient screening. However, current AI-based MD prediction methods lack interpretability—a common limitation of deterministic neural networks. In medically sensitive scenarios, this hinders practical deployment of such models.This study proposes an evidence-based learning approach for MD prediction. The method actively models uncertainty, with the network simultaneously modeling both the outcome and its uncertainty. This provides partial interpretability while delivering strong performance.
Methods
We design an evidential learning–based model for MD prediction that outputs a predictive distribution rather than a single point estimate, thereby improving interpretability and reliability in clinical settings. Since MD is a continuous regression target, we assume its observations follow a Gaussian likelihood and impose conjugate priors on the unknown parameters , resulting in a Normal–Inverse-Gamma (NIG) evidential distribution. For each input image, the network simultaneously predicts the four NIG hyperparameters , where represents the predicted MD mean, captures aleatoric (data) uncertainty, and characterizes epistemic uncertainty. During training, we maximize the marginal likelihood (equivalently, the negative log-likelihood of a Student-t evidence distribution) and introduce an error-aware evidence regularization term to suppress incorrect yet overconfident predictions. As a result, the proposed model achieves strong MD regression performance while explicitly providing uncertainty estimates to support early screening and risk-aware clinical decision making.
Results
Baseline MAE: 1.62 Our method MAE: 1.47. Baseline GM: 1.15 Our method GM: 1.20
Conclusion
This framework accurately predicts glaucomatous visual field defects (MD) from OCT.