Uncertainty-Aware Lesion Localization

Supervised lesion segmentation depends on dense pixel-level annotations whose per-image expert cost bounds the scale of available data, whereas image-level diagnostic labels are recorded routinely but carry no spatial information. Recovering lesion extent from such weak supervision is difficult because localization derived from a single classifier inherits that architecture's inductive bias: it concentrates on the most discriminative sub-region, covers extent incompletely, and varies across stochastic realizations, so no measure indicates where its evidence can be trusted. We present an uncertainty- and reliability-aware weakly supervised localization framework. Architecture-diverse classifiers trained on image-level labels alone yield complementary localization evidence; stochastic inference produces repeated spatial explanations from which mean evidence and pixel-wise localization uncertainty are estimated. Validation-derived model reliability and this spatial uncertainty jointly set spatially adaptive fusion weights, determining which model to trust and where, before thresholding into a binary mask. On a controlled non-medical setting with true masks and on TBX11K box-reference localization, adaptive fusion improved overlap over single-architecture localization, equal-weight averaging, and hard consensus, supporting annotation-efficient mask generation for expert review.