We evaluated 11 segmentation methods (Watershed, BIDCell, Cellpose, SCS, SAM, SAM2, mSAM, Baysor, CelloType, Swin-s and Mask R-CNN) across 15 spatial transcriptomics datasets generated using 5 different technologies: Stereo-seq, Xenium, CosMx, STARmap PLUS and MERFISH (Fig. 1a, Supplementary Table 1 and Methods). The datasets span two species (mouse and human), 12 tissue types (brain, breast, embryo, lung, pancreas, skin, colon, stomach, ovarian, kidney, prostate and utero) and a wide range of cell counts, from around 700 cells in an ovarian cancer dataset to large cohorts of more than 800,000 cells in a prostate cancer dataset. We selected this diversity to reflect the breadth of contemporary spatial transcriptomics applications and to challenge segmentation robustness under heterogeneous biological and technical conditions (Supplementary Fig. 1a,b). We also performed cross-technology benchmarking (Fig. 1b) to account for discrepancies between platforms. This design separated algorithm-specific effects from platform-dependent biases. We conducted all the analysis using a unified benchmarking pipeline (Fig. 1c) that applies standardized quality control (QC), patch-based processing and consistent preprocessing across methods to ensure fair and reproducible comparisons. During preprocessing, we established data subsetting criteria for each dataset and method based on technical limitations (Methods). For methods that explicitly integrate transcriptomic information into their...

I actually haven’t worked much on AI for cell biology, most of my experience has been with some diagnostic applications where measuring the properties of beads and circular objects is important. Thus...

Below is cellpose3-cyto3 compared to cellpose4-sam on a resolution test image which consists of circles of different sizes. This is based on a test you can find here which shows how to make the spher...