cellvit-histo-hgsc-pan-v1

CellVit Model for High-Grade Serous Ovarian Cancer Panoptic Segmentation

Model

• histolytics implementation of panoptic CellVit: https://arxiv.org/abs/2306.15350
• Backbone encoder: pre-trained samvit_base_patch16 from pytorch-image-models https://github.com/huggingface/pytorch-image-models

USAGE

1. Install histolytics and albumentations

pip install histolytics
pip install albumentations

2. Load trained model

from histolytics.models.cellvit_panoptic import CellVitPanoptic

model = CellVitPanoptic.from_pretrained("hgsc_v1_efficientnet_b5")

3. Run inference for one image

from albumentations import Resize, Compose
from histolytics.utils import FileHandler
from histolytics.transforms import MinMaxNormalization

model.set_inference_mode(mixed_precision=False)

# Resize to multiple of 32 of your own choosing
transform = Compose([Resize(1024, 1024), MinMaxNormalization()])

im = FileHandler.read_img(IMG_PATH)
im = transform(image=im)["image"]

prob = model.predict(im)
out = model.post_process(prob)
# out = {"nuc": [(nuc instances (H, W), nuc types (H, W))], "tissue": [tissues (H, W)], "cyto": None}

3.1 Run inference for image batch

import torch
from histolytics.utils import FileHandler

model.set_inference_mode()

# dont use random matrices IRL
batch = torch.rand(8, 3, 1024, 1024)

prob = model.predict(im)
out = model.post_process(prob)
# out = {
#  "nuc": [
#    (nuc instances (H, W), nuc types (H, W)),
#    (nuc instances (H, W), nuc types (H, W)),
#    .
#    .
#    .
#    (nuc instances (H, W), nuc types (H, W))    
#  ],
#  "tissue": [
#    (nuc instances (H, W), nuc types (H, W)),
#    (nuc instances (H, W), nuc types (H, W)),
#    .
#    .
#    .
#    (nuc instances (H, W), nuc types (H, W))    
#  ],
#  "cyto": None,
}

4. Visualize output

from matplotlib import pyplot as plt
from skimage.color import label2rgb

fig, ax = plt.subplots(1, 4, figsize=(24, 6))
ax[0].imshow(im)
ax[1].imshow(label2rgb(out["nuc"][0][0], bg_label=0)) # inst_map
ax[2].imshow(label2rgb(out["nuc"][0][1], bg_label=0)) # type_map
ax[3].imshow(label2rgb(out["tissue"][0], bg_label=0)) # tissue_map

Dataset Details

Semi-manually annotated HGSC Primary Omental samples from the (private) DECIDER cohort. Data acquired in the DECIDER project, funded by the European Union’s Horizon 2020 research and innovation programme under grant agreement No 965193.

Contains:

• 198 varying sized image crops at 20x magnification.
• 98 468 annotated nuclei
• 699 744 885 pixels of annotated tissue region

Dataset classes

nuc_classes = {
    0: "background",
    1: "neoplastic",
    2: "inflammatory",
    3: "connective",
    4: "dead",
    5: "macrophage_cell",
    6: "macrophage_nuc",
}

tissue_classes = {
    0: "background",
    1: "stroma",
    2: "omental_fat",
    3: "tumor",
    4: "hemorrage",
    5: "necrosis",
    6: "serum",
}

Dataset Class Distribution

Nuclei:

• connective nuclei: 46 100 (~47%)
• neoplastic nuclei: 22 761 (~23%)
• inflammatory nuclei 19 185 (~19%)
• dead nuclei 1859 (~2%)
• macrophage nuclei and cytoplasms: 4550 (~5%)

Tissues:

• stromal tissue: 28%
• tumor tissue:29%
• omental fat: 20%
• hemorrhage 5%
• necrosis 13%
• serum 5%

Model Training Details:

First, the image crops in the training data were tiled into 224x224px patches with a sliding window (stride=32px).

Rest of the training procedures follow this notebook: [link]

Citation

histolytics:

@article{
}

CellVit original paper:

@article{CellViT,
    title = {CellViT: Vision Transformers for precise cell segmentation and classification},
    journal = {Medical Image Analysis},
    volume = {94},
    pages = {103143},
    year = {2024},
    issn = {1361-8415},
    doi = {https://doi.org/10.1016/j.media.2024.103143},
    url = {https://www.sciencedirect.com/science/article/pii/S1361841524000689},
    author = {Fabian Hörst and Moritz Rempe and Lukas Heine and Constantin Seibold and Julius Keyl and Giulia Baldini and Selma Ugurel and Jens Siveke and Barbara Grünwald and Jan Egger and Jens Kleesiek},
}

Licence

These model weights are released under the Apache License, Version 2.0 (the "License"). You may obtain a copy of the License at:

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Additional Terms

While the Apache 2.0 License grants broad permissions, we kindly request that users adhere to the following guidelines: Medical or Clinical Use: This model is not intended for use in medical diagnosis, treatment, or prevention of disease of real patients. It should not be used as a substitute for professional medical advice.