About patches in WSI

[Yoyiming]

I would like to know what 'in_tissue' in the Visium data, such as 'INT10', represents in adata. If I want to segment the WSI into patches based on adata's 'pxl_row_in_fullres' and 'pxl_col_in_fullres', should I select the positions where 'in_tissue=0' before performing the segmentation?

[pauldoucet]

Hi, in_tissue is a default Spaceranger flag, it doesn't represent anything at the moment as it was just left out during data acquisition.

The actual tissue segmentation is available in tissue_seg/ on HuggingFace. You can automatically load the tissue segmentation when using load_hest, and then patch such that you only keep the patches under tissue as follows:

from hest import load_hest

hest_list = load_hest('hest_data', id_list=['INT10'])

patch_save_dir = 'patches'

st = hest_list[0]

st.dump_patches(
    patch_save_dir,
    name='demo',
    target_patch_size=224, # target patch size in 224
    target_pixel_size=0.5 # pixel size of the patches in um/px after rescaling
)

[Yoyiming]

Thank you very much for your response! It is really helpful.

[Yoyiming]

@pauldoucet Hello, I have one more question that I am a bit confused about. In the adata, does 'pxl_row_in_fullres' and 'pxl_col_in_fullres' correspond to the width and height in the WSI, respectively? Is the following code correct for obtaining patches?

whole_image = cv2.imread(image_path) patch = self.whole_image[x-112:x+112, y-112:y+112]

Where x represents 'pxl_row_in_fullres' and y represents 'pxl_col_in_fullres'.

[Yoyiming]

Hello, I have one more question that I am a bit confused about. In the adata, does 'pxl_row_in_fullres' and 'pxl_col_in_fullres' correspond to the width and height in the WSI, respectively? Is the following code correct for obtaining patches?

whole_image = cv2.imread(image_path) patch = self.whole_image[x-112:x+112, y-112:y+112]

Where x represents 'pxl_row_in_fullres' and y represents 'pxl_col_in_fullres'.

On Mon, Sep 9, 2024 at 5:20 PM Paul Doucet @.***> wrote:

Hi, in_tissue is a default Spaceranger flag, it doesn't represent anything at the moment as it was just left out during data acquisition.

The actual tissue segmentation is available in tissue_seg/ on [HuggingFace].( https://huggingface.co/datasets/MahmoodLab/hest/blob/main/tissue_seg/INT10_contours.geojson ) You can automatically load the tissue segmentation when using load_hest, and then patch such that you only keep the patches under tissue as follows:

from hest import load_hest

hest_list = load_hest('hest_data', id_list=['INT10'])

patch_save_dir = 'patches'

st = hest_list[0]

st.dump_patches( patch_save_dir, name='demo', target_patch_size=224, # target patch size in 224 target_pixel_size=0.5 # pixel size of the patches in um/px after rescaling )

— Reply to this email directly, view it on GitHub https://github.com/mahmoodlab/HEST/issues/47#issuecomment-2337585322, or unsubscribe https://github.com/notifications/unsubscribe-auth/BD6QADQ4TDZIJ2COPPSXRADZVVR4ZAVCNFSM6AAAAABN4CFFS2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGMZXGU4DKMZSGI . You are receiving this because you authored the thread.Message ID: @.***>

[pauldoucet]

Hello, We strongly encourage using st.adata.obsm['spatial'], instead of 'pxl_row_in_fullres' and 'pxl_col_in_fullres' which might not be available for every technology.

You can then use the following:

xy = st.adata.obsm['spatial'] # Nx2 array, first column is x, second is y
coords = xy[0] # get coordinates of first spot
x, y = coords[0], coords[1]
patch = self.wsi.read_region((x-112, y-112), 0, (224, 224))

[Yoyiming]

Thank you for your suggestion! I’ll give it a try.

[pauldoucet]

I'd also suggest having a look at 2-Interacting-with-HEST-1k.ipynb

[Yoyiming]

Thank you! I'll have a look.

[HelloWorldLTY]

Hi, I found that sometimes .obsm['spatial'] contains float data, and thus it will cause an error for the patching code. Do you think it is ok to use int(x) and int(y) to round the location? Thanks.

Furthermore, may I know the reason of having such shape of the selected patch?

(224, 224, 4)

If we intend to use rgb based pathology foundation model to read it, shall we only consider the first three channel, as I find that the last channel is always 255?