MCseg: reference-guided AI-agent development of a fixed cell segmentation and transcript attribution workflow for spatial transcriptomics
Year: 2026
Type: preprint
Abstract: Cell-level analysis of high-resolution spatial transcriptomics depends on how tissue images are segmented and how transcripts are assigned to cells, yet paired reference measurements for developing and evaluating such workflows are scarce and cannot be assumed for each sample requiring analysis. We developed MCseg using a single paired tissue section as a development reference: an AI agent iteratively composed and scored combinations of pretrained Cellpose models and image-processing operations against Xenium-derived cell boundaries, and the retained workflow was then fixed so that it can be deployed without the reference assay or the agent. In the lung adenocarcinoma development set, the fixed workflow increased mean panoptic quality from 0.432 to 0.472 relative to an Optuna-tuned two-diameter Cellpose baseline, with improvement in four of six regions. In an expert-annotated colorectal cancer region not used during development, MCseg showed higher lineage recall and micro-F1 (0.805 versus 0.732) than the StarDist-based ENACT workflow among 10,275 reference cells covered by both methods. Across 15 regions from one colorectal cancer section, MCseg showed higher neighborhood expression discordance and lower lineage-exclusive co-expression than Space Ranger at similar UMI density, while capturing fewer transcripts. The fixed workflow could also be applied to fresh-frozen breast cancer, a distinct tissue and preservation condition, without a new architecture search. These results show how a limited paired reference can be used to develop a fixed, reusable workflow that no longer depends on the reference assay; MCseg distributes this workflow as a locally run, no-code platform. Availability and implementation MCseg is available under the MIT License at https://github.com/ddmanyes/MCseg . The repository contains the segmentation workflow, transcript-attribution scripts, analysis code, AI-agent prompts, and development logs. Bullet points A paired spatial reference AI-agent search for a fixed segmentation workflow MCseg improves agreement with Xenium boundaries over a tuned Cellpose baseline MCseg reduces lineage mixing relative to Space Ranger in colorectal cancer A local no-code tool runs the fixed workflow without the reference assay or an AI agent
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7214371405