Nucleus segmentation of WSI in PNG results in ValueError: MPP is None. Cannot determine scale in terms of MPP.

• TIA Toolbox version: 1.3.0
• Python version: 3.7
• Operating System: Cent OS

Description

Calling NucleusInstanceSegmentor.predict() with H&E (RGB, PNG) whole-slide image files result in the following error:

ValueError: MPP is None. Cannot determine scale in terms of MPP.

What I Did

I used the IOSegmentorConfig to specify the mpp of my inputs and outputs (H&E images in RGB and PNG format), but when I try to perform the segmentation I get an error saying that the MPP is None. Since PNG files don’t contain any metadata info, I was hoping that by specifying it in IOSegmentorConfig will provide that info.

objIOConfig = IOSegmentorConfig(
    input_resolutions=[
        {'units': 'mpp', 'resolution': 0.325}, 
    ],
    output_resolutions=[
        {'units': 'mpp', 'resolution': 0.325}, 
    ],
    margin=128,
    tile_shape=[1024, 1024],
    patch_input_shape=(256, 256),
    patch_output_shape=(256, 256),
    stride_shape=(164, 164),
    save_resolution={'units': 'mpp', 'resolution': 0.325}
)

# Instantiate the nucleus instance segmentor
objInstSegmentor = NucleusInstanceSegmentor(
    pretrained_model="hovernet_fast-pannuke", 
    num_loader_workers=4, 
    num_postproc_workers=4, 
    batch_size=8, 
    auto_generate_mask=False, 
    verbose=False, 
)

# WSI prediction (if ON_GPU=False, this part will take more than a couple of hours to process)
wsi_output = objInstSegmentor.predict(
    ['TestSet_ROI_6291_wsi_he.png'], 
    ioconfig=objIOConfig, 
    masks=None, 
    save_dir=f'./WSITest/', 
    mode="wsi", 
    on_gpu=ON_GPU, 
    crash_on_exception=True, 
)

The error that I encountered in more detail:

Thank you!

Caleb

[mostafajahanifar]

Dear @Calebium

As you are trying to make predictions on image tiles, it is better to call the predict method in tile mode instead of mode=wsi. If you specify the mode to be wsi, TIAToolbox requires the input to be a WSI structured file with metadata information.

Therefore, an easy solution to your problem is to set the mode in the last method call to tile:

wsi_output = objInstSegmentor.predict(
    ['TestSet_ROI_6291_wsi_he.png'], 
    ioconfig=objIOConfig, 
    masks=None, 
    save_dir=f'./WSITest/', 
    mode="tile",  # NOT `wsi`
    on_gpu=ON_GPU, 
    crash_on_exception=True, 
)

[Extra Suggestion] I noticed you are trying to set the input resolution to “mpp resolution: 0.325” for your input. You should know, this IOConfig is to set the config of image reading with respect to the trained model, not the test data! So, whatever resolution the data has been captured with, if you want to get the best results using TIAToolbox’s HoVerNet, input images to the algorithm should be in “mpp resolution: 0.25”. Now, if your images are saved in a resolution that is different from the images used in the HoVerNet model training, I would suggest you resize the images to “mpp resolution: 0.255” before you process them in order to achieve the best segmentation resolution. However, even without resizing you should obtain acceptable results.

[mostafajahanifar]

Thank you for your suggestions. If I call the predict method in tile mode with a whole slide image, how is it treated differently than calling with the same image in wsi mode? Thanks!

If you use tile mode, TIAToolbox assumes the input is a plain image. If the input is a path to a png image, the toolbox tries to read it with a simple cv2.imread function (which is not capable of reading WSIs) and append a predefined standard metadata to it. Therefore, reading a WSI in tile mode will raise an external exception (error). However, if you read a png in the WSI mode, because in WSI mode toolbox assumes there is some metadata associated with the input it would not change the metadata (does not append predefined standard metadata to WSI) and instead throws warnings that metadata is missing and won’t be able to do operations that are metadata dependant (such as reading from a specific mpp).