PiMorph: proposal checkpoints, benchmark results and training tiles for endothelial cell complexes

Abstract

Proposal checkpoints, benchmark results and training tiles for PiMorph, which reconstructs endothelial monolayers from fluorescence microscopy as embedded cell complexes: cells, gaps, cell-cell contacts and multicellular vertices with exact incidence, plus a posterior over legal complexes for every field. The code, the tests and the documentation are at github.com/okezue/PiMorph; this record holds what is too large for git or what a reader needs to reproduce the tables without retraining. New in version 1.2.0 pimorph_cellpose_ft_cpsam_confluent.tar, pimorph_cellpose_ft_cpsam_haec.tar, pimorph_cellpose_ft_cpsam_qbam.tar (1.2 GB each): three Cellpose-SAM models fine-tuned with the official recipe on exactly the fields that trained PiMorph's v6_pool (confluent pool, HAEC) and on 1,032 registered iPSC-RPE QBAM tiles, with training metadata and logs. They make the headline comparison symmetric: on the held-out cultured endothelial monolayer the fine-tuned Cellpose-SAM reaches vertex F1 0.695, adjacency F1 0.937 and panoptic quality 0.862 against 0.680, 0.867 and 0.816 for PiMorph's own proposal model, while PiMorph keeps the higher vertex F1 on FlyWing (0.875 against 0.852) and HAEC (0.352 against 0.331) and one checkpoint holds both culture densities. Cellpose is now also a proposer inside PiMorph, so these weights drive the constrained decoder, energy and posterior directly at no loss of accuracy. pimorph_healthy2_qbam.tar (0.9 GB): the 168 QBAM bright-field absorbance tiles of 84 well-timepoints of the NIST/NEI Healthy-2 iPSC-RPE maturation series (Schaub, Hotaling, Bharti et al. 2020, DOI 10.18434/T4/1503229) with their transepithelial resistance table, used to test the structural part of PiMorph's barrier proxy against 216 measured TER values. The test is negative: the predicted permeability index rises with TER (Spearman +0.52) and is indistinguishable from cell density, so the proxy remains flagged as not validated; the junction-coverage term awaits a dataset that pairs a junction stain with TER on the same wells. pimorph_results.tar now also contains the symmetric comparison, the Cellpose-proposer runs, the laser-ablation validation of force inference (Lang et al. 2019: boundary-cable tension excess in 15 of 15 fields, Spearman 0.64 with recoil velocity), the EpiCure event-calculus runs, and the study documents as of 2026-09-21. The checkpoint and training-tile archives are unchanged from version 1.1.0. Files Archives above about 1 GB are stored as 75 MB parts (.part00, .part01, ...); scripts/fetch_zenodo.py in the repository downloads, verifies and joins them, or cat name.tar.part?? > name.tar does the same by hand. parts_md5.txt lists the whole-file checksums. pimorph_models.tar (536 MB): eight proposal checkpoints, pimorph_proposals_v0_synth.pt to pimorph_proposals_v6_pool.pt (multi-head UNet, 19.4 M parameters; heads for boundary, signed distance, seeds, vertices, gaps and boundary uncertainty), one model card per checkpoint, and the training log and configuration of every run. v6_pool is the recommended checkpoint for confluent monolayers with a bright membrane or junction channel. pimorph_results.tar (997 MB): every benchmark output of the repository. Held-out per-image and summary tables for human aortic endothelial cells (HAEC, 86 fields), mCellSeg (40 DIC images), cultured human corneal endothelial monolayers (hCEC, 5 fields with about 3,000 manually traced tricellular vertices each), alizarine-stained corneal endothelium (10), FlyWing E-cadherin epithelium (10) and NIH-NEI RPE monolayers (20 tiles); decoder tuning tables on the training splits; the re-test of the three shear-stress network findings with 32-hypothesis posteriors on all 102 S-BIAD1540 EGM2 fields; the legacy EndoPiGraph per-field outputs; the multi-junction (VE-cadherin, claudin-5, F-actin on S-BIAD1169), dynamics (TissueMiner, Cell Tracking Challenge, EpiCure) and mechanics reports; figures; and the result documents. pimorph_training_tiles.tar (1,584 MB): 200 synthetic validation tiles, 177 pseudo-label tiles from S-BIAD1540 and VE-strat, and 1,200 consensus pseudo-label tiles cut from 300 real PECAM-1 HUVEC monolayer fields (Cellpose-SAM primary, PiMorph v4 second opinion, disputed boundaries marked in the ignore mask). Tiles are npz files with the image channels, the instance labels and the six training targets. README.md: contents, the checkpoint lineage with the held-out numbers, and reproduction commands. Headline results in the tables On the cultured endothelial monolayer test fields the v6_pool checkpoint through the constrained decoder reaches vertex F1 0.680, adjacency F1 0.867, panoptic quality 0.816 and boundary F1 0.916 with a calibrated vertex count (3,014 predicted against 2,974 true); Cellpose-SAM masks passed through the same complex extractor give 0.635, 0.830, 0.790 and 0.880 on the same fields. On corneal endothelium in situ every method reaches vertex F1 0.98 to 0.99; on FlyWing PiMorph has the better vertices (0.875 against 0.849) and Cellpose-SAM the better adjacency (0.966 against 0.911). On the sub-confluent HAEC fields vertex F1 is 0.352 against 0.039, bounded by the one-pixel ambiguity of whether two cells touch in the reference. The shear re-test on all 102 EGM2 fields keeps one of the three earlier findings (the reticular junction fraction rises from 0.112 static to 0.207 at 6 dyn cm-2, Mann-Whitney p = 2.6e-9, consistent in all three biological replicates, distinct high-shear regime), reduces one to a corollary (the all-reticular 3-clique increase is fully explained by the reticular fraction; conditional-null enrichment z 0.003 against 0.025, p = 0.95) and does not replicate the third (area-degree correlation 0.722 against 0.746, p = 0.43). About 95% of contact-graph 3-cliques are realized by a multicellular vertex of the complex. Reconstruction ambiguity is negligible for all three statistics (median 90% credible half-width 0.001 to 0.006). Not included Third-party images and their ground truth stay at their original sources under their own licences; docs/DATASET_HUNT_2026-09-18.md in the repository lists every source, licence and download URL and the loaders read them in place. Training tiles cut from those images are not redistributed (the hCEC set is CC BY-NC-ND 4.0). The 8,000 synthetic training tiles are regenerated exactly by pimorph synth with the seeds given in the repository. Code The complex extractor, the constrained decoder, the posterior, the neural proposal model and its training, the benchmark harness with every loader, the dynamics, mechanics, 3-D and multi-junction modules and the test suite are at github.com/okezue/PiMorph. scripts/fetch_zenodo.py downloads and checksum-verifies the files of this record; scripts/publish_zenodo.py creates its versions.