Evolutionary Multi-Objective Calibration of MC-Dropout Predictive Uncertainty Quantification for Diabetic Retinopathy

Monte carlo dropout is one of the most widely used methods for uncertainty quantification in deep learning, yet practitioners often rely on fixed heuristic settings for parameters such as the dropout rate and the ensemble size. Therefore, in this work, we formulate the joint tuning of dropout rate p, temperature T, and sample count N as a multi-objective optimisation problem that minimises expected calibration error and computational cost subject to an accuracy-retention constraint. We solve this problem with non-dominated sorting genetic algorithm II and reference vector guided evolutionary algorithm across three architectures-DenseNet121, MobileViT, and TinyViT-for binary diabetic retinopathy classification. Two main findings emerge, (1) calibration gains saturate at low sample counts (N≈10), enabling a 3× reduction in inference cost relative to common heuristics; and (2) optimal dropout rates are strongly architecture-dependent, with convolutional neural networks tolerating moderate perturbation (p≈0.4) while vision transformers require substantially lower rates (p≈0.1), challenging the widespread use of fixed Monte carlo dropout configurations. These findings demonstrate that evolutionary multi-objective optimisation can efficiently navigate the cost-calibration trade-off for uncertainty quantification in medical imaging.