Efficient Uncertainty Quantification in Medical Imaging via Mamba State Space Models

Background/Objectives: Reliable uncertainty quantification (UQ) is a prerequisite for deploying automated systems in safety-critical medical imaging workflows, yet existing approaches either sacrifice computational efficiency or provide poorly calibrated confidence estimates. We present UQ-Mamba, a lightweight architecture that embeds uncertainty quantification natively into a Mamba state space model via linearized error propagation. Methods: UQ-Mamba yields per-prediction approximate epistemic and aleatoric uncertainty estimates in a single deterministic forward pass at only 9.5% additional inference overhead. We note that these components are heuristic approximations derived under three explicit assumptions (diagonal covariance, first-order linearization, and scalar mean-activation reduction) and have not been empirically validated as true Bayesian posteriors. By propagating learnable log-variance parameters through the SSM state transition matrix, UQ-Mamba bridges the gap between parameter efficiency and principled calibration without requiring stochastic sampling or multiple forward passes. Results: Evaluated across four medical imaging modalities—CT organ classification, colorectal histopathology, dermoscopy, and chest radiography—UQ-Mamba achieves 89.71% accuracy with ECE = 0.0217 on OrganMNIST using only 466K parameters (3.3× lower ECE than ResNet-50 at 50× fewer parameters; note that UQ-Mamba optimizes NLL, whereas ResNet-50 uses standard cross-entropy, which is a confounding factor in the ECE comparison), improves the Mamba baseline by 2.42 percentage points on PathMNIST (ECE = 0.1188 after temperature scaling), achieves 68.88% test accuracy with ECE = 0.0597 on HAM10000 dermoscopy (matching EfficientNet-B0 at 9× fewer parameters), and reaches mAUC = 0.8196 on CheXpert chest radiographs. Conclusions: Ablation studies confirm that the SSM propagation mechanism is necessary for meaningful uncertainty decomposition. These results establish uncertainty-aware SSMs as a promising proof-of-concept direction for calibrated, parameter-efficient medical image classification, with potential relevance to resource-constrained deployment settings pending further clinical validation.