MCseg: AI agent-guided workflow search for no-code cell segmentation and transcript attribution in spatial transcriptomics

Cell-level analysis in high-resolution spatial transcriptomics requires precise segmentation and transcript assignment. Current workflows often prioritize boundary purity over transcript capture and demand significant image-analysis expertise. To address these challenges, a downloadable no-code platform called MCseg was developed.

This platform utilizes a fixed segmentation engine generated through an AI-agent-guided search process. During this search, an AI agent iteratively suggested and assessed various combinations of image-processing and segmentation operations, with the Xenium-derived cell boundaries serving as the evaluation standard.

In a lung adenocarcinoma development set, MCseg improved the mean panoptic quality from 0.432 to 0.472 compared to an Optuna-tuned two-diameter Cellpose baseline. Additionally, a reference-guided calibration analysis achieved a score of 0.554. In an independent expert-annotated colorectal cancer region, MCseg outperformed the StarDist-based ENACT workflow in terms of lineage recall and micro-F1 among cells covered by both methods.

When compared to Space Ranger at similar UMI density across 15 colorectal cancer regions, MCseg increased neighborhood expression discordance and reduced lineage-exclusive co-expression. Importantly, the fixed workflow also transferred to fresh-frozen breast cancer samples without the need for tissue-specific architecture search. This demonstrates the potential of an agent-guided approach for achieving reproducible, locally deployable cell-level spatial transcriptomic analysis.