Mouse Nuclei
Find me on Github The mouse nuclei dataset is composed of a train and test dataset.
In [1]:
Copied!
# Imports necessary to execute the code
import torch
import matplotlib.pyplot as plt
import numpy as np
from careamics import CAREamist
from careamics.config import create_n2v_configuration
from careamics.utils.metrics import scale_invariant_psnr
from careamics_portfolio import PortfolioManager
from PIL import Image
# Imports necessary to execute the code import torch import matplotlib.pyplot as plt import numpy as np from careamics import CAREamist from careamics.config import create_n2v_configuration from careamics.utils.metrics import scale_invariant_psnr from careamics_portfolio import PortfolioManager from PIL import Image
In [2]:
Copied!
use_gpu = "yes" if len([torch.cuda.get_device_properties(i) for i in range(torch.cuda.device_count())]) > 0 else "no"
print(f"Using GPU: {use_gpu}")
use_gpu = "yes" if len([torch.cuda.get_device_properties(i) for i in range(torch.cuda.device_count())]) > 0 else "no" print(f"Using GPU: {use_gpu}")
Using GPU: yes
Import the dataset¶
The dataset can be directly downloaded using the careamics-portfolio package, which uses pooch to download the data.
In [ ]:
Copied!
# download file
portfolio = PortfolioManager()
files = portfolio.denoiseg.MouseNuclei_n20.download()
files.sort()
# load images
train_data = np.load(files[1])["X_train"]
print(f"Train data shape: {train_data.shape}")
# download file portfolio = PortfolioManager() files = portfolio.denoiseg.MouseNuclei_n20.download() files.sort() # load images train_data = np.load(files[1])["X_train"] print(f"Train data shape: {train_data.shape}")
Data description¶
In [4]:
Copied!
portfolio.denoiseg.MouseNuclei_n20.description
portfolio.denoiseg.MouseNuclei_n20.description
Out[4]:
'A dataset depicting diverse and non-uniformly clustered nuclei in the mouse skull, consisting of 908 training and 160 validation patches. The test set counts 67 additional images'
Visualize data¶
In [5]:
Copied!
indices = [34, 293, 571, 783]
fig, ax = plt.subplots(2, 2, figsize=(8, 8))
ax[0, 0].imshow(train_data[indices[0]], cmap="gray")
ax[0, 0].set_title(f"Image {indices[0]}")
ax[0, 0].set_xticks([])
ax[0, 0].set_yticks([])
ax[0, 1].imshow(train_data[indices[1]], cmap="gray")
ax[0, 1].set_title(f"Image {indices[1]}")
ax[0, 1].set_xticks([])
ax[0, 1].set_yticks([])
ax[1, 0].imshow(train_data[indices[2]], cmap="gray")
ax[1, 0].set_title(f"Image {indices[2]}")
ax[1, 0].set_xticks([])
ax[1, 0].set_yticks([])
ax[1, 1].imshow(train_data[indices[3]], cmap="gray")
ax[1, 1].set_title(f"Image {indices[3]}")
ax[1, 1].set_xticks([])
ax[1, 1].set_yticks([])
plt.show()
indices = [34, 293, 571, 783] fig, ax = plt.subplots(2, 2, figsize=(8, 8)) ax[0, 0].imshow(train_data[indices[0]], cmap="gray") ax[0, 0].set_title(f"Image {indices[0]}") ax[0, 0].set_xticks([]) ax[0, 0].set_yticks([]) ax[0, 1].imshow(train_data[indices[1]], cmap="gray") ax[0, 1].set_title(f"Image {indices[1]}") ax[0, 1].set_xticks([]) ax[0, 1].set_yticks([]) ax[1, 0].imshow(train_data[indices[2]], cmap="gray") ax[1, 0].set_title(f"Image {indices[2]}") ax[1, 0].set_xticks([]) ax[1, 0].set_yticks([]) ax[1, 1].imshow(train_data[indices[3]], cmap="gray") ax[1, 1].set_title(f"Image {indices[3]}") ax[1, 1].set_xticks([]) ax[1, 1].set_yticks([]) plt.show()
Train with CAREamics¶
The easiest way to use CAREamics is to create a configuration and a CAREamist.
Create configuration¶
The configuration can be built from scratch, giving the user full control over the various parameters available in CAREamics. However, a straightforward way to create a configuration for a particular algorithm is to use one of the convenience functions.
In [6]:
Copied!
config = create_n2v_configuration(
experiment_name="mouse_nuclei_n2v",
data_type="array",
axes="SYX",
patch_size=(64, 64),
batch_size=16,
num_epochs=10,
)
print(config)
config = create_n2v_configuration( experiment_name="mouse_nuclei_n2v", data_type="array", axes="SYX", patch_size=(64, 64), batch_size=16, num_epochs=10, ) print(config)
{'algorithm_config': {'algorithm': 'n2v',
'loss': 'n2v',
'lr_scheduler': {'name': 'ReduceLROnPlateau',
'parameters': {}},
'model': {'architecture': 'UNet',
'conv_dims': 2,
'depth': 2,
'final_activation': 'None',
'in_channels': 1,
'independent_channels': True,
'n2v2': False,
'num_channels_init': 32,
'num_classes': 1,
'use_batch_norm': True},
'n2v_config': {'masked_pixel_percentage': 0.2,
'name': 'N2VManipulate',
'remove_center': True,
'roi_size': 11,
'strategy': 'uniform',
'struct_mask_axis': 'none',
'struct_mask_span': 5},
'optimizer': {'name': 'Adam', 'parameters': {}}},
'data_config': {'axes': 'SYX',
'batch_size': 16,
'data_type': 'array',
'patch_size': [64, 64],
'train_dataloader_params': {'num_workers': 4,
'pin_memory': True,
'shuffle': True},
'transforms': [{'flip_x': True,
'flip_y': True,
'name': 'XYFlip',
'p': 0.5},
{'name': 'XYRandomRotate90', 'p': 0.5}],
'val_dataloader_params': {'num_workers': 4,
'pin_memory': True}},
'experiment_name': 'mouse_nuclei_n2v',
'training_config': {'accumulate_grad_batches': 1,
'check_val_every_n_epoch': 1,
'checkpoint_callback': {'auto_insert_metric_name': False,
'mode': 'min',
'monitor': 'val_loss',
'save_last': True,
'save_top_k': 3,
'save_weights_only': False,
'verbose': False},
'gradient_clip_algorithm': 'norm',
'max_steps': -1,
'num_epochs': 10,
'precision': '32'},
'version': '0.1.0'}
Train¶
A CAREamist can be created using a configuration alone, and then be trained by using the data already loaded in memory.
In [ ]:
Copied!
# instantiate a CAREamist
careamist = CAREamist(source=config)
# train
careamist.train(
train_source=train_data,
)
# instantiate a CAREamist careamist = CAREamist(source=config) # train careamist.train( train_source=train_data, )
In [8]:
Copied!
loss_dict = careamist.get_losses()
plt.plot(loss_dict["train_epoch"], loss_dict["train_loss"], loss_dict["val_epoch"], loss_dict["val_loss"])
plt.legend(["Train loss", "Val loss"])
plt.title("Losses")
loss_dict = careamist.get_losses() plt.plot(loss_dict["train_epoch"], loss_dict["train_loss"], loss_dict["val_epoch"], loss_dict["val_loss"]) plt.legend(["Train loss", "Val loss"]) plt.title("Losses")
{'epoch': [], 'learning_rate': [], 'step': [], 'train_loss_epoch': [], 'train_loss_step': [], 'val_loss': []}
Out[8]:
Text(0.5, 1.0, 'Losses')
Predict with CAREamics¶
Prediction is done with the same CAREamist used for training. Because the image is large we predict using tiling.
In [ ]:
Copied!
prediction = careamist.predict(source=train_data)
prediction = careamist.predict(source=train_data)
Visualize the prediction¶
In [10]:
Copied!
# download ground truth
files = portfolio.denoiseg.MouseNuclei_n0.download()
files.sort()
# load images
gt_data = np.load(files[1])["X_train"]
print(f"GT data shape: {gt_data.shape}")
# download ground truth files = portfolio.denoiseg.MouseNuclei_n0.download() files.sort() # load images gt_data = np.load(files[1])["X_train"] print(f"GT data shape: {gt_data.shape}")
GT data shape: (908, 128, 128)
In [11]:
Copied!
indices = [389, 621]
for i in indices:
# compute psnr
psnr_noisy = scale_invariant_psnr(gt_data[i], train_data[i])
psnr_denoised = scale_invariant_psnr(gt_data[i], prediction[i].squeeze())
# plot images
fig, ax = plt.subplots(1, 3, figsize=(10, 10))
ax[0].imshow(train_data[i], cmap="gray")
ax[0].set_title(f"Noisy Image\nPSNR: {psnr_noisy:.2f}")
ax[0].set_xticks([])
ax[0].set_yticks([])
ax[1].imshow(prediction[i].squeeze(), cmap="gray")
ax[1].set_title(f"Denoised Image\nPSNR: {psnr_denoised:.2f}")
ax[1].set_xticks([])
ax[1].set_yticks([])
ax[2].imshow(gt_data[i], cmap="gray")
ax[2].set_title("GT Image")
ax[2].set_xticks([])
ax[2].set_yticks([])
plt.show()
indices = [389, 621] for i in indices: # compute psnr psnr_noisy = scale_invariant_psnr(gt_data[i], train_data[i]) psnr_denoised = scale_invariant_psnr(gt_data[i], prediction[i].squeeze()) # plot images fig, ax = plt.subplots(1, 3, figsize=(10, 10)) ax[0].imshow(train_data[i], cmap="gray") ax[0].set_title(f"Noisy Image\nPSNR: {psnr_noisy:.2f}") ax[0].set_xticks([]) ax[0].set_yticks([]) ax[1].imshow(prediction[i].squeeze(), cmap="gray") ax[1].set_title(f"Denoised Image\nPSNR: {psnr_denoised:.2f}") ax[1].set_xticks([]) ax[1].set_yticks([]) ax[2].imshow(gt_data[i], cmap="gray") ax[2].set_title("GT Image") ax[2].set_xticks([]) ax[2].set_yticks([]) plt.show()
Compute metrics¶
In [12]:
Copied!
psnrs = np.zeros(gt_data.shape[0])
for i in range(gt_data.shape[0]):
psnrs[i] = scale_invariant_psnr(gt_data[i], prediction[i].squeeze())
print(f"PSNR: {np.mean(psnrs):.2f} ± {np.std(psnrs):.2f}")
psnrs = np.zeros(gt_data.shape[0]) for i in range(gt_data.shape[0]): psnrs[i] = scale_invariant_psnr(gt_data[i], prediction[i].squeeze()) print(f"PSNR: {np.mean(psnrs):.2f} ± {np.std(psnrs):.2f}")
PSNR: 32.03 ± 2.15
Export the model¶
The model is automatically saved during training (the so-called checkpoints) and can be loaded back easily, but you can also export the model to the BioImage Model Zoo format.
In [13]:
Copied!
# create a cover image
index_cover = 5
height, width = 128, 128
# create image
cover = np.zeros((height, width))
# normalize train and prediction
norm_train = (train_data[index_cover] - train_data[index_cover].min()) / (
train_data[index_cover].max() - train_data[index_cover].min()
)
pred = prediction[index_cover].squeeze()
norm_pred = (pred - pred.min()) / (pred.max() - pred.min())
# fill in halves
cover[:, : width // 2] = norm_train[:height, : width // 2]
cover[:, width // 2 :] = norm_pred[:height, width // 2 :]
# plot the single image
plt.imshow(cover, cmap="gray")
# save the image
im = Image.fromarray(cover * 255)
im = im.convert("L")
im.save("MouseNuclei_Noise2Void.jpeg")
# create a cover image index_cover = 5 height, width = 128, 128 # create image cover = np.zeros((height, width)) # normalize train and prediction norm_train = (train_data[index_cover] - train_data[index_cover].min()) / ( train_data[index_cover].max() - train_data[index_cover].min() ) pred = prediction[index_cover].squeeze() norm_pred = (pred - pred.min()) / (pred.max() - pred.min()) # fill in halves cover[:, : width // 2] = norm_train[:height, : width // 2] cover[:, width // 2 :] = norm_pred[:height, width // 2 :] # plot the single image plt.imshow(cover, cmap="gray") # save the image im = Image.fromarray(cover * 255) im = im.convert("L") im.save("MouseNuclei_Noise2Void.jpeg")
In [14]:
Copied!
general_description = (
"This model is a UNet trained with the Noise2Void algorithm to denoise images. "
"The training data consists of images of non-uniformly clustered nuclei in the "
"mouse skull. The notebook used to train this model is available on the CAREamics "
"documentation website at the following link: "
"https://careamics.github.io/0.1/applications/Noise2Void/Mouse_Nuclei/."
)
print(general_description)
general_description = ( "This model is a UNet trained with the Noise2Void algorithm to denoise images. " "The training data consists of images of non-uniformly clustered nuclei in the " "mouse skull. The notebook used to train this model is available on the CAREamics " "documentation website at the following link: " "https://careamics.github.io/0.1/applications/Noise2Void/Mouse_Nuclei/." ) print(general_description)
This model is a UNet trained with the Noise2Void algorithm to denoise images. The training data consists of images of non-uniformly clustered nuclei in the mouse skull. The notebook used to train this model is available on the CAREamics documentation website at the following link: https://careamics.github.io/0.1/applications/Noise2Void/Mouse_Nuclei/.
In [ ]:
Copied!
# Export the model
careamist.export_to_bmz(
path_to_archive="mousenuclei_n2v_model.zip",
friendly_model_name="MouseNuclei_N2V",
input_array=train_data[[indices[0]]].astype(np.float32),
authors=[{"name": "CAREamics authors", "affiliation": "Human Technopole"}],
data_description=portfolio.denoiseg.MouseNuclei_n20.description,
general_description=general_description
)
# Export the model careamist.export_to_bmz( path_to_archive="mousenuclei_n2v_model.zip", friendly_model_name="MouseNuclei_N2V", input_array=train_data[[indices[0]]].astype(np.float32), authors=[{"name": "CAREamics authors", "affiliation": "Human Technopole"}], data_description=portfolio.denoiseg.MouseNuclei_n20.description, general_description=general_description )