Trustworthy in silico labeling via semantic visual interpretability of image-to-image translation

Cross-modality image translation promises to provide multiple layers of biological information from a single image input, yet its practical application is stalled by a lack of interpretability and the inability to account for model imperfections. In silico labeling, the inference of organelle localization from label-free images, is a primary example where this black-box nature limits adoption. We present Mask Interpreter, a generalized method for semantic visual interpretability of image-to-image translation models. By uncovering organelle-specific "explanation signatures", Mask Interpreter validates that models rely on authentic biological structures rather than spurious artifacts. Beyond biological validation, it outperforms traditional explainable AI (xAI) approaches, identifies batch effects and localizes prediction errors when ground-truth fluorescence is unavailable. Semantic confidence modeling further provides fine-grained reliability assessment at single-cell resolution, enabling the automated exclusion of artifacts from downstream analyses. By bridging the gap between computational inference and meaningful biological features, Mask Interpreter transforms in silico labeling into a reliable tool for scientific discovery across diverse biomedical imaging modalities.