SAM-Sode: XAI Framework Enhances Tiny Bacteria Detection Interpretability

A new explainable AI framework named SAM-Sode has been introduced to enhance the interpretability of tiny bacteria detection in clinical diagnostics. Conventional explanation techniques often face challenges with unclear foreground edges and vague feature attribution, largely due to the extreme scarcity of target morphological characteristics and the complexity of background noise. SAM-Sode converts initial feature attribution maps into geometry-aware prompts, utilizing the foundational model SAM3's existing knowledge for spatial refinement and morphological reconstruction. It employs a dual-constraint approach grounded in physical relevance and geometric alignment to perform instance-level denoising, resulting in coherent explanations. This framework meets the essential demand for logically consistent morphological evidence in medical imaging. The findings are published in arXiv preprint 2605.21186.