Validation of an automated AI-based micro-CT organ segmentation workflow against expert annotations and its impact on fluorescence quantification

Abstract Objective Automated, artificial intelligence (AI)-based, organ segmentation has the potential to streamline preclinical imaging workflows, but its suitability must be evaluated not only by geometric accuracy, but also by impact on downstream quantitative analyses. We validated a commercially available AI-based organ segmentation workflow for whole-body micro-computed tomography (micro-CT) data regarding segmentation accuracy, reproducibility, processing time, and effects on fluorescence tomography (FLT) quantification. Materials and methods Heart, lungs, liver, and kidneys were segmented in micro-CT scans from 27 mice by three independent experts and by a commercially available automated deep learning tool applied to standard and iteratively reconstructed CT datasets. Segmentation performance was evaluated using the Sørensen-Dice similarity coefficient (DSC) and organ volumes, while downstream effects were assessed using FLT overlay quantification across five fluorescent probes. Results AI-based segmentations showed organ-dependent agreement with experts, with lower DSC scores for heart and lungs reflecting systematic boundary differences. Repeated AI segmentations yielded identical results, demonstrating reproducibility while reducing segmentation time from approximately 30 min per mouse to about 5 min, with AI inference completed within seconds. While segmentation geometry and organ volumes differed from expert annotations in an organ-dependent manner, these differences showed no linear relationship with downstream FLT quantification. Iterative CT reconstruction improved the agreement of fluorescence measurements with expert-derived values. Conclusion Geometric segmentation metrics alone are insufficient to predict downstream fluorescence quantification, underscoring the need for task-based evaluation strategies in preclinical multimodal imaging. AI-based segmentation offers a reproducible and time-efficient alternative to manual annotation, enabling practical, scalable, and generalizable image analysis. Relevance statement Reproducible, AI-based organ segmentation can substantially improve the efficiency and scalability of quantitative preclinical imaging workflows, supporting more robust translation of multimodal imaging biomarkers toward clinical research and future patient-focused applications. Key Points In hybrid micro-CT/FLT imaging, geometric segmentation metrics did not reliably predict downstream fluorescence quantification outcomes. AI segmentation assessment should shift from benchmark accuracy metrics toward task-based validation reflecting real downstream imaging applications across different workflows. A commercially available AI-based organ segmentation tool demonstrates fully reproducible segmentation and substantially reduced analysis time compared with expert annotations. AI applied to iteratively reconstructed CT data improves the agreement of fluorescence quantification with expert annotations compared with standard reconstruction.