Making AI explain its tiniest discoveries: bacteria in the noise

Explaining why AI detects bacteria is harder than explaining other objects—the targets are tiny and backgrounds are cluttered, so standard explanation methods produce blurry, incoherent maps. SAM-Sode converts these attribution maps into geometry-aware prompts fed to SAM (a foundation model), which reconstructs sharper, morphologically faithful explanations. A dual-constraint mechanism filters out background noise and aligns outputs with expert intuition. Tested on a custom dataset of 2,524 bacteria images plus public benchmarks, the method significantly improves transparency for clinical AI tools.