Summary
A bioRxiv preprint describes MCseg, a no-code tool whose segmentation workflow was selected through AI-agent-guided searches. In a lung adenocarcinoma dataset, its fixed-parameter workflow scored higher on panoptic quality than an Optuna-tuned Cellpose baseline.
Researchers describe MCseg, a downloadable no-code platform for cell segmentation and transcript assignment in spatial transcriptomics. In a bioRxiv preprint, they report that its fixed-parameter workflow achieved a mean panoptic quality of 0.472 in a lung adenocarcinoma development set, compared with 0.432 for an Optuna-tuned, two-diameter Cellpose baseline.
Spatial transcriptomics measures gene activity while retaining information about where RNA molecules occur within tissue. To analyse the data cell by cell, software must identify cell boundaries and assign detected transcripts to the right cells. Those assignments shape the gene-expression profiles used to compare cell types and their surroundings.
How MCseg was developed and tested
The authors used an AI agent to search for combinations of image-processing and segmentation operations. The agent proposed and evaluated workflows against cell boundaries derived from Xenium data; the resulting search produced MCseg’s fixed segmentation engine. This is a workflow-selection approach: the reported system uses the selected engine rather than requiring each user to assemble a segmentation pipeline manually.
The lung result compares that fixed-parameter workflow with the tuned Cellpose baseline using panoptic quality, a metric that combines how well objects are detected with how closely their predicted boundaries match reference boundaries. The paper’s abstract also reports a separate reference-guided calibration analysis with a score of 0.554.
Other tests assessed transcript and cell-lineage assignments. In an independent, expert-annotated colorectal cancer region, MCseg had higher lineage recall and micro-F1 than the StarDist-based ENACT workflow among cells covered by both methods. Across 15 colorectal cancer regions, the authors report greater neighbourhood expression discordance and lower lineage-exclusive co-expression with MCseg than with Space Ranger, at similar UMI density. These are distinct comparisons: the first evaluates segmentation against reference boundaries, while the latter measures patterns in expression assignments.
The authors also report transferring the fixed workflow to fresh-frozen breast cancer data without a tissue-specific architecture search. This suggests the same selected workflow could be applied beyond the tissue used in its development, though the reported tests cover particular cancer datasets and regions.
What the results mean
MCseg addresses a practical challenge in cell-level spatial analysis: segmentation choices can affect which transcripts are attributed to each cell, while developing image-analysis workflows can require specialist effort. A no-code platform built from a searched and then fixed workflow could make such analyses easier to reproduce and use.
The evidence is a computational evaluation reported in a preprint, not a clinical trial or a test of patient outcomes. The abstract gives results from specific tissue datasets and regions; the findings therefore describe performance in those evaluations rather than establishing a universal advantage across spatial transcriptomics experiments.