segtme-uni2

SegTME — Tumor Microenvironment segmentation using the UNI2 pathology foundation model with a dual-head UperNet + HoverNet-style decoder.

The model simultaneously predicts:

• Semantic segmentation — 6-class tissue map (background, neoplastic, inflammatory, connective, dead, epithelial)
• HV maps — horizontal/vertical nuclear distance gradients for instance segmentation via watershed post-processing

Architecture and model weights are separate: weights are publicly hosted on HuggingFace; the architecture is distributed as this compiled package (code not exposed).

---

Available Models

Model	HuggingFace Repo	Trained on	Best mIoU
M1 — PanNuke	SegTME-UNI2-UperHoVer_PanNuke	PanNuke (pan-cancer nuclei, 7,901 patches)	0.9313
M2 — TCGA-UT-0	SegTME-UNI2-UperHoVer_TCGA-UT-0	TCGA-UT subset 0 (~850 steps/epoch)	0.8197
M3 — TCGA-UT-012345	SegTME-UNI2-UperHoVer_TCGA-UT-012345	TCGA-UT subsets 0–5 combined	0.7724

All three models share the same architecture and hyperparameters; only training data and checkpoint weights differ.

---

Architecture

Input (B, 3, 224, 224) — ImageNet-normalised
       │
       ▼
UNI2 ViT-Giant backbone (depth=24, heads=24, embed_dim=1536, patch=14)
  Multi-scale features extracted at layers 5 / 11 / 17 / 23
  Projected to 256 / 512 / 1024 / 2048 channels via Conv2d
       │
       ├──► UperNet decoder → semantic logits (B, 6, H, W)
       │
       └──► UperNet decoder → HV maps       (B, 2, H, W)

UNI2 backbone — vit_giant_patch14_224 (timm), loaded from MahmoodLab/UNI2-h pretrained weights, 1.1 B parameters.

UNI2UperHoVer — dual-head UperNet decoder; semantic head classifies tissue type per pixel; HV head produces horizontal/vertical nuclear centroid distance fields used for marker-controlled watershed instance segmentation.

---

Installation

# Core (model architecture only)
pip install segtme-uni2

# With large-tile inference pipeline (LargeTilePredictor)
pip install "segtme-uni2[predict]"

Core dependencies: torch, transformers, timm, safetensors, huggingface_hub.
Predict extras: numpy, pillow, torchvision, opencv-python, scipy, scikit-image.

---

Usage

from segtme import UNI2UperHoVer

# Load from HuggingFace (downloads weights automatically)
model = UNI2UperHoVer.from_pretrained("mizjaggy18/SegTME-UNI2-UperHoVer_TCGA-UT-0")
model.eval().cuda()

# Forward pass — input: (B, 3, H, W), ImageNet normalised, recommended 224×224 tiles
import torch
pixel_values = torch.randn(1, 3, 224, 224).cuda()
sem_logits, hv_maps = model(pixel_values)

# sem_logits: (B, 6, H, W)  — per-class tissue logits
# hv_maps:    (B, 2, H, W)  — horizontal / vertical distance gradients

Output classes

Channel	Class
0	Background
1	Neoplastic
2	Inflammatory
3	Connective
4	Dead
5	Epithelial

Recommended inference scale

Resize input tiles so one pixel corresponds to the target MPP before running inference:

Model	Target MPP	Downscale factor
M1 — PanNuke	0.25 µm/px	image_mpp / 0.25
M2 — TCGA-UT-0	0.314 µm/px	image_mpp / 0.314
M3 — TCGA-UT-012345	0.5 µm/px	image_mpp / 0.5

Large-tile inference with LargeTilePredictor

LargeTilePredictor handles the full end-to-end pipeline on images of arbitrary size: upscaling, tiling (224×224, 50% overlap), stitching, watershed instance segmentation, and downscaling back to original resolution.

# Requires: pip install "segtme-uni2[predict]"
from segtme import UNI2UperHoVer, LargeTilePredictor

# Load from HuggingFace (downloads weights automatically)
model = UNI2UperHoVer.from_pretrained("mizjaggy18/SegTME-UNI2-UperHoVer_TCGA-UT-0")

predictor = LargeTilePredictor(
    model,
    input_mpp=0.5,    # µm/px of input image (0.5 = 20× scanner)
    model_mpp=0.314,  # M2 training resolution (default)
)

result = predictor.predict("path/to/image.png")

result.sem        # uint8  (H, W)    6-class semantic mask
result.inst       # uint16 (H, W)    watershed instance map (0 = background)
result.class_img  # uint8  (H, W, 3) RGB instances coloured by class
result.outline    # uint8  (H, W, 3) original H&E + class-coloured boundaries

# Save outputs
from PIL import Image
import cv2
Image.fromarray(result.sem).save("pred_sem.png")
Image.fromarray(result.inst).save("pred_inst.png")
Image.fromarray(result.outline).save("pred_outline.png")

Load from a local safetensors checkpoint (no HuggingFace download):

from segtme import UNI2UperHoVer, UNI2UperHoVerConfig, LargeTilePredictor
from safetensors.torch import load_file

config = UNI2UperHoVerConfig()
model  = UNI2UperHoVer(config)
model.load_state_dict(
    load_file("checkpoint-212500/model.safetensors"), strict=True)

predictor = LargeTilePredictor(model, input_mpp=0.5)
result = predictor.predict("image.png")

Model MPP reference for model_mpp:

Model	model_mpp	Best for
M1 — PanNuke	0.25	40× / ~0.25 µm/px inputs
M2 — TCGA-UT-0	0.314 (default)	20× / 0.5 µm/px inputs
M3 — TCGA-UT-012345	0.5	Variable 0.5–1.0 µm/px inputs

Low-level tiling helpers

The standalone functions underlying LargeTilePredictor are also exported for users who need to compose the pipeline manually:

from segtme import (
    pad_to_multiple, tile_image, stitch_semantic, stitch_hv,
    predict_tile_dual, clean_mask, run_watershed,
    map_instances_to_classes, build_outline_image,
    REGION_COLORS, REGION_LABELS,
)

---

Training Curriculum

Three-stage curriculum training, each stage initialised fresh (no weight inheritance):

Stage	Model	Dataset	Epochs	Steps	mIoU
1	M1	PanNuke	249	24,651	0.9313
2	M2	TCGA-UT subset 0	250	212,500	0.8197
3	M3	TCGA-UT subsets 0–5	100	335,100	0.7724

All stages: initial LR 5×10⁻⁵, linear decay, AdamW optimiser. Backbone: UNI2-h (frozen or fine-tuned depending on stage).

---

Citation

If you use this model in your work, please cite:

@article{wanahmad2026segtme,
  title={SegTME-UNI2: A Foundation Model-Based Framework for Generalisable Multiclass Cell Segmentation and LLM-Driven Tumour Microenvironment Characterisation in Histopathology},
  author={Wan Ahmad, Wan Siti Halimatul Munirah and Samidi, Faris Syahmi and Ahmmed, Mohammad Badal and Thiviyanathan, Vimal Angela and Thavaraj, Selvam James and Abdul Majeed, Anwar P.P.},
  journal={arXiv preprint arXiv:2606.17702},
  year={2026},
  doi={10.48550/arXiv.2606.17702}
}

---

Links

• Paper (arXiv): https://arxiv.org/abs/2606.17702
• PyPI: https://pypi.org/project/segtme-uni2/
• HuggingFace (M1): https://huggingface.co/mizjaggy18/SegTME-UNI2-UperHoVer_PanNuke
• HuggingFace (M2): https://huggingface.co/mizjaggy18/SegTME-UNI2-UperHoVer_TCGA-UT-0
• HuggingFace (M3): https://huggingface.co/mizjaggy18/SegTME-UNI2-UperHoVer_TCGA-UT-012345
• UNI2 backbone: https://huggingface.co/MahmoodLab/UNI2-h