Is contrast enhancement applied on extracted patches ?

[abhi1998goyal]

I notice a clear difference in patches provided in patch directory to those that are created using dump_patches. Is their any normalization applied on these patches after extraction from WSI ?

for st in iter_hest('data_path', id_list=['TENX99']):
patch_save_dir = 'path_dir'
new_patch_size = 256
st.dump_patches(patch_save_dir, name='demo', target_patch_size=new_patch_size, target_pixel_size=0.5
)
256*256 size at 0.5um/pixel images provided in patches directory of hest data is on left and one extracted from .tif using above method is on right side.

[guillaumejaume]

No contrast enhancement nor stain norm has been applied. What you show could be compression artifacts.

[abhi1998goyal]

import h5py
import numpy as np
from PIL import Image

h5_file_path = "/home/demo.h5"

with h5py.File(h5_file_path, "r") as h5f:

    barcodes = h5f['barcode'][:]
    barcodes = barcodes.astype(str)  

    if '221x099' in barcodes:
       barcode_index = np.where(barcodes == '221x099')[0][0]

    img_patch = h5f["img"][barcode_index]  
    img = Image.fromarray(img_patch)
    img.save('/home/TENX99_221x099.png' , format="PNG", compress_level=0)

It could be compression but I don't see why there should be compression. I am doing the same for both the patches/TENX99.h5 and demo.h5 created from TENX99.tif . The picture from patches/TENX99.h5 is far better. Please help !

[guillaumejaume]

tbh, I don't see a huge difference between both images. Could you check the actual difference between both arrays using https://numpy.org/doc/2.1/reference/generated/numpy.allclose.html

[abhi1998goyal]

import h5py
import numpy as np
from PIL import Image
import os
import matplotlib.pyplot as plt

for st in iter_hest('/home/hest_data', id_list=['TENX99']):
    patch_save_dir = 'home/'
    new_patch_size = 224
    st.dump_patches(
    patch_save_dir,
    name='demo',
    target_patch_size=new_patch_size,
    target_pixel_size=0.5
)

h5_file_path = "/home/demo.h5"

with h5py.File(h5_file_path, "r") as h5f:
    barcodes = h5f['barcode'][:]
    barcodes = barcodes.astype(str)  
    if '221x099' in barcodes:
       barcode_index = np.where(barcodes == '221x099')[0][0]
    img_patch = h5f["img"][barcode_index]  


hest_h5_path = os.path.join('/home/hest_data/patches', "TENX99.h5")
with h5py.File(hest_h5_path, "r") as h5f2:
    barcodes = h5f2['barcode'][:]
    barcodes = barcodes.astype(str)
    if '221x099' in barcodes:
       barcode_index1 = np.where(barcodes == '221x099')[0][0]
    img_patch_2=h5f2["img"][barcode_index1]    


are_similar = np.allclose(img_patch, img_patch_2, atol=1) 
print(are_similar)

are_similar = np.allclose(img_patch, img_patch_2, atol=10) 
print(are_similar)

are_similar = np.allclose(img_patch, img_patch_2, atol=60) 
print(are_similar)

are_similar = np.allclose(img_patch, img_patch_2, atol=70) 
print(are_similar)

fig, axes = plt.subplots(1, 2, figsize=(6, 3))

axes[0].imshow(img_patch)
axes[0].set_title("Image from demo.h5")
axes[0].axis("off")

axes[1].imshow(img_patch_2)
axes[1].set_title("Image from TENX99.h5")
axes[1].axis("off")

plt.tight_layout()
plt.show()

diff = 3*np.abs(img_patch.astype(int) - img_patch_2.astype(int)) 

plt.imshow(diff, cmap="hot")  
plt.colorbar()
plt.title("Pixel Differences")
plt.axis("off")
plt.show()

Output: False False False True

[pauldoucet]

Hi @abhi1998goyal,
We slightly modified dump_patches after creating the patches in patches/. The variation you see might be due to a small offset/zoom difference between the old and the new implementation, but I don't think it affects the quality