Yinoo/TissueLab-WholeCell-v4-EMA750

TissueLab WholeCell v4 — EMA750 weights

Public, ungated distribution of the selected v4 EMA step 750 adaptation for TissueLab whole-cell segmentation. Downloads do not require a Hugging Face account, access to the training server, or access to a private code repository. The inference implementation is distributed separately through TissueLab-Models.

This adaptation requires both files below. The base remains hosted by its original publisher at a fixed revision. This repository contains the adaptation, its manifest/checksums, and attribution; it does not contain training images, datasets, optimizer state, or the private application code.

File	Bytes	Source
vitamin_p_flex.pth	2,023,371,074	Pinned VitaminP base
v4_ema_step_000750.pt	478,122,551	v4 adaptation

Download and verify

Choose a local directory readable by the model-worker service account. After installing the current TissueLab WholeCell package, use:

export WHOLECELL_WEIGHTS_DIR="$HOME/.cache/tissuelab/wholecell"
python -m wholecell_seg download --model v4
python -m wholecell_seg verify --model v4

The package pins both downloads to immutable revisions and verifies byte size and SHA256 before loading. Set WHOLECELL_WEIGHTS_DIR in the task manager's environment too, so workers use that directory after restart.

For a download without installing the package, start in a new directory and run:

curl -fL --retry 3 -o vitamin_p_flex.pth \
  https://huggingface.co/idso-fa1-pathology/VitaminP/resolve/6598c2fd890b498fd48728beedaeaba018f313b1/vitamin_p_flex.pth
curl -fL --retry 3 -o v4_ema_step_000750.pt \
  https://huggingface.co/Yinoo/TissueLab-WholeCell-v4-EMA750/resolve/main/v4_ema_step_000750.pt
curl -fL --retry 3 -o SHA256SUMS \
  https://huggingface.co/Yinoo/TissueLab-WholeCell-v4-EMA750/resolve/main/SHA256SUMS
sha256sum -c SHA256SUMS

Both hashes are also recorded in manifest.json:

ca36503fa88e7c7689b4f3737ab0f897a3149db12acb1eb722d1e49f11e8ef85  vitamin_p_flex.pth
7b3ccdbdad887acd9cff0cd43a59b8c910ec3781eb8a0c4ae04106c7676cf976  v4_ema_step_000750.pt

Model and evaluation scope

The model accepts RGB H&E and measured microns per pixel. It uses a frozen DINOv2-L backbone with adapted feature projection/ASPP, cell and nucleus decoders, and a small boundary head. The v4 pipeline uses 512-pixel context at 0.5 microns/pixel and central 224-pixel decoded outputs. Nuclear-marker and horizontal/vertical-field watershed produce instances; the full-slide runtime stitches overlapping predictions.

The adaptation was trained on 5,198 Xenium-derived H&E tiles and selected using 298 development tiles, with stain-supported cell/nucleus reference masks. These references are not manual membrane ground truth. On those same development tiles, whole-cell PQ was 0.325543, versus 0.365601 for the explicit official VitaminP-Flex pipeline. The v4 registered release gate failed. The thresholded auxiliary boundary head had zero development recall at 0.5. These weights are supplied for research and engineering evaluation; public availability does not establish superior segmentation, unseen-domain accuracy, or clinical readiness. Upstream pretraining overlap is unknown.

The artifact contains the selected adaptation tensors, step number, and protocol hash. It is not a standalone Transformers model and requires the separately distributed matching inference implementation and base checkpoint.

Attribution

Based on VitaminP, UT MD Anderson IDSO-FA1-Pathology. Upstream code commit: de637b2102f9be6fa064879d5a29d4491061ec8e; base weight revision: 6598c2fd890b498fd48728beedaeaba018f313b1. The upstream MIT notice is preserved in LICENSE_VitaminP.txt. This artifact distribution does not change the licensing or access settings of the separately distributed TissueLab code.