Use SlidingWindowInferer on image already loaded in Python?

[gdurif]

Hi,

To integrate the HoVer-Net model into a segmentation pipeline that is processing images as Numpy arrays already loaded in memory, is it possible to use the SlidingWindowInferer() on image already loaded into Python?

I want to use the model already trained to do segmentation only (I am not interested in cell type classification), because I do not have annotated data for fine tuning. We have real experimental images that need segmentation and thus no ground truth.

So far, following the README and the tutorial notebooks, I have been able to either:

1. Feed a cropped image to the model (c.f. code below), however:

• I do not know how to post-process the output to get the actual segmentation as I would with a SlidingWindowInferer();
• this would require to write the sliding window algorithm by hand to process the whole image;
• I am not sure which type of normalization I should implement.

import numpy as np

from cellpose.io import imread        # pip install cellpose

import cellseg_models_pytorch as csmp # pip install cellseg-models-pytorch
import torch

# image
image_dir = "."       # edit if necessary
image_name = "TCGA-E2-A14V-01Z-00-DX1"
ext = "png"

# image_path = os.path.join(crop_image_dir, f"{image_name}.{ext}")
image_path = os.path.join(image_dir, f"{image_name}.{ext}")

# load image
image = imread(image_path)

# hover-net model
model = csmp.models.hovernet_base(type_classes=10)

# image conversion (and crop and normalization)
image_tensor = torch.from_numpy(
    (image[np.newaxis, :, :].transpose([0, 3, 1, 2])[:, :, :256, :256] /
     255).astype(np.float32)
)

# forward
res = model(image_tensor)

2. Use the SlidingWindowInferer() to process an image stored on disk (c.f. code below), however:

• the segmentation result is rubbish (maybe related to the following point);
• since I am using the pre-training model, which checkpoint_path am I supposed to provide?

# define the final activations for each model output
out_activations = {"hovernet": "tanh", "type": "softmax", "inst": "softmax"}

# define whether to weight down the predictions at the image boundaries
# typically, models perform the poorest at the image boundaries and with
# overlapping patches this causes issues which can be overcome by down-
# weighting the prediction boundaries
out_boundary_weights = {"hovernet": True, "type": False, "inst": False}

# define the inferer
inferer = csmp.inference.SlidingWindowInferer(
    model=model,
    input_path=f"{image_dir}",
    checkpoint_path=None,
    out_activations=out_activations,
    out_boundary_weights=out_boundary_weights,
    instance_postproc="hovernet",               # THE POST-PROCESSING METHOD
    normalization="percentile",                 # same normalization as in training
    patch_size=(256, 256),
    stride=128,
    padding=80,
    batch_size=1,
    device="cpu"  # or "cuda"
)

# inference
inferer.infer()

# result
inferer.out_masks

In the example, I use the following TCGA-E2-A14V-01Z-00-DX1.png image from the MoNuSeg dataset available under CC BY-NC-SA 4.0 license):

Thanks in advance.

[okunator]

Hi,

To your first points:

• I'm currently working on a release in which you can use numpy arrays or torch tensors as the input to the SlidingWindowInferer but at the moment there is no support for that. The release will be published within a couple weeks
• The normalization of the image should always be the same what was used during training/finetuning the segmentation model.

Second points:

• The checkpoint path should be pointing to the model weights file that you have saved when you trained/fine-tuned the hovernet. You can take a look how to train the hovernet model with pannuke https://github.com/okunator/cellseg_models.pytorch/blob/main/examples/pannuke_nuclei_segmentation_hovernet.ipynb
• When you train with pannuke like in the notebook example linked above, you will receive also cell type classes as outputs but if there is no need for them you can just ignore them and focus on the instance segmentation results 'inst_map'

[gdurif]

Hi @okunator

Thanks for your reply (and sorry I took so long to answer).

Do you have any update regarding the new release ?

Thanks

[okunator]

Hi @gdurif

I'm quite confident I can release within 2 weeks.