TUNE++: Topology-Guided Uncertainty Estimation for Reliable 3D Medical Image Segmentation

Deep learning models for medical image segmentation lack mechanisms to assess their own reliability, leading to two critical failures: they provide no uncertainty estimates to distinguish confident predictions from error-prone ones, and often produce anatomically implausible segmentations or incorrect connectivity that violate known structural constraints. We observe that uncertainty and topology are intrinsically linked and anatomically complex regions naturally exhibit higher prediction uncertainty, while uncertain predictions require stronger enforcement of structural constraints. Building on this insight, we propose TUNE++, a unified framework that jointly learns segmentation, uncertainty quantification, and topology preservation through a novel Topology-Uncertainty aware Paired Attention (TUPA) mechanism. Our method decomposes uncertainty into aleatoric and epistemic components while simultaneously enforcing anatomical correctness through persistent homology-based constraints. A key innovation is our topology-uncertainty alignment loss that minimizes the discrepancy between predicted total uncertainty and a topological complexity score computed from organ boundaries, multi-organ junction counts, and critical points extracted from persistence diagrams, teaching the model to be uncertain precisely where anatomical structure is geometrically complex. Our empirical results demonstrate that joint modeling of TUNE++ produced enhanced segmentation accuracy, well-calibrated uncertainty estimates that successfully identify errors, substantial reduction in topological violations, and learned confidence that correlates strongly with anatomical complexity.