No-Reference Hallucination Assessment for AI-Reconstructed Fluorescence Microscopy Image

Artificial intelligence (AI) has revolutionized fluorescence microscopy image restoration, enabling high-resolution imaging with cost-effective computations. However, the inherent biases of AI models in training sets may lead to hallucinations, such as hallucinating artificial structures that do not exist in the original sample or removing real biological features, compromising the scientific authenticity and reliability of imaging-based discoveries. To address this critical challenge, we present the first systematic investigation and formal definition of AI-reconstructed hallucinations in fluorescence microscopy. We introduce HallAssess, a no-reference assessment method explicitly tailored for the reliable identification and quantification of AI-reconstructed hallucinations. By shifting the assessment from the high-quality (HQ) image domain to the low-quality (LQ) image domain, HallAssess effectively transforms a no-reference problem into a full-reference one. Our approach enables accurate hallucination quantification without requiring ground-truth HQ images by re-degrading AI-reconstructed images using an imaging model that simulates real-world image degradation processes, and then comparing them with the original LQ inputs. We validate HallAssess across multiple imaging modalities (SIM, confocal), diverse AI models, and common fluorescence microscopy image restoration tasks such as denoising and super-resolution. The results demonstrate its effectiveness in detecting AI-reconstructed hallucinations. Furthermore, we provide an open-access platform featuring an interactive web demo and a dynamic leaderboard, allowing researchers to evaluate hallucinations in fluorescence microscopy image restoration results and benchmark state-of-the-art methods under a standardized framework. This work provides a foundational tool for ensuring the reliability of AI-assisted imaging in life sciences, particularly in cell biology, where accurate interpretation of subcellular structures is essential for understanding cellular function and disease mechanisms.