import fastai.medical.imaging

If you use these transformations which all build on TorchIO, please cite the following paper:

Pérez-García et al., TorchIO:a Python library for efficient loading,preprocessing, augmentation and patch-based sampling of medical images in deep learning. Link: https://arxiv.org/abs/2003.04696

Spike artifact

import matplotlib.pyplot as plt
from ipywidgets import interact, interactive, fixed, interact_manual
import ipywidgets as widgets

spikeTransform[source]

spikeTransform(image, intensityFactor, spikePosition=[0.1, 0.1])

get_spike_series[source]

get_spike_series(image, model, start=0, end=2.5, step=0.5, spikePosition=[0.1, 0.1], **kwargs)

fname = "example/b0/images/val_example_0.png"
img = lambda: Image.open(fname).resize((256,256)).convert ("RGB")
trueMask = lambda: Image.open(fname.replace("image","mask"))

spikeTransform(img(), 1, spikePosition=[.5,.5])

def interactiveSpikeTransform(intensity, x, y):
    plt.imshow(spikeTransform(img(), intensity, [x,y])) #figsize=(8,8))

interactive_plot = interactive(
    interactiveSpikeTransform,
    intensity=widgets.FloatSlider(min=0, max=3, step=.1, value=1, continuous_update=True),
    x=widgets.FloatSlider(min=0, max=.99, step=.05, value=.5, continuous_update=True),
    y=widgets.FloatSlider(min=0, max=.99, step=.05, value=.5, continuous_update=True)
)
interactive_plot

plt.imshow(spikeTransform(img(), .8, [.4,.3]), cmap="bone") #figsize=(8,8), 

<matplotlib.image.AxesImage at 0x7f7e1d4c0760>

def label_func(x):
    pass
def acc_seg(input, target):
    pass
def diceComb(input, targs):
    pass
def diceLV(input, targs):
    pass
def diceMY(input, targs):
    pass

trainedModel = Fastai2_model('chfc-cmi/transversal-cmr-seg', 'b0_transversal_5_5', force_reload=False)

Using cache found in /home/markus/.cache/torch/hub/chfc-cmi_transversal-cmr-seg_master

Spike intensity

series = get_spike_series(img(), trainedModel, truth=trueMask(), tfm_y=False)

plot_series(series)

spikePosition = [.45,.55]
plot_series(get_spike_series(img(), trainedModel, truth=trueMask(), spikePosition=spikePosition))

spikePosition = [.5,.52]
plot_series(get_spike_series(img(), trainedModel, truth=trueMask(), spikePosition=spikePosition))

eval_spike_series[source]

eval_spike_series(image, mask, model, step=0.1, start=0, end=2.5, spikePosition=[0.1, 0.1], param_name='intensity', **kwargs)

results = eval_spike_series(img(), trueMask(), trainedModel)
plt.plot(results['intensity'], results['c1'], label='LV')
plt.plot(results['intensity'], results['c2'], label='MY')
plt.legend()
plt.xlabel('relative intensity')
plt.ylabel('dice')
_ = plt.title(f'spike position: {[.1,.1]}')

Spike position

spikePosTransform[source]

spikePosTransform(image, spikePositionX, spikePositionY=0.1, intensityFactor=0.5)

get_spike_pos_series[source]

get_spike_pos_series(image, model, start=0.1, end=0.9, step=0.1, intensityFactor=0.5, spikePositionY=0.1, **kwargs)

series = get_spike_pos_series(img(), trainedModel, intensityFactor=0.1, spikePositionY=0.1)
plot_series(series, nrow=1, figsize=(20,15), param_name='Spike X Position')

eval_spike_pos_series[source]

eval_spike_pos_series(image, mask, model, step=0.1, start=0.1, end=0.9, intensityFactor=0.1, param_name='Spike X Position', **kwargs)

intensityFactor = 0.2
results = eval_spike_pos_series(img(), trueMask(), trainedModel, step=0.05, intensityFactor=intensityFactor)
plt.plot(results['Spike X Position'], results['c1'], label='LV')
plt.plot(results['Spike X Position'], results['c2'], label='MY')
plt.legend()
plt.xlabel('X position of the spike')
plt.ylabel('Dice')
_ = plt.title(f'Spike intensity = {intensityFactor}')

gif_series(
    get_spike_pos_series(img(),trainedModel, start=0.1, end=0.9 ,step=0.1, intensityFactor=0.2),
    "example/kaggle/spike_pos.gif",
    param_name="Spike X Position",
    duration=400
)

Bias Field ($B_0$)

The bias field function is still experimental as it lacks configurability. To make it deterministic all coefficients are set to the exact same value. This represents only one specific possibility of field inhomogeneity.

image = img()
image = np.array(image)[:,:,0]
image = np.expand_dims(np.array(image), 0)
image = np.expand_dims(np.array(image), 0)
image.shape

(1, 1, 256, 256)

coefficients = RandomBiasField().get_params(3,[-.3,-.3])

bf = BiasField.generate_bias_field(image, order=3, coefficients=coefficients)

plt.imshow(bf[0])

<matplotlib.image.AxesImage at 0x7f7e18d295e0>

image[0][0].shape
image = image/255

image = torch.from_numpy(image)
result = image[0][0] * bf[0]
result = torch.stack((result,result,result))
result = np.array((torch.clamp(result,0,1)))
result = result*255
result = result.astype(np.uint8)
result = np.moveaxis(result, 0, 2)
result = Image.fromarray(result)
plt.imshow(result)

<matplotlib.image.AxesImage at 0x7f7e18ba4ca0>

biasfieldTransform[source]

biasfieldTransform(image, coef, order=3)

get_biasfield_series[source]

get_biasfield_series(image, model, start=0, end=-0.6, step=-0.2, order=3, **kwargs)

plot_series(get_biasfield_series(img(),trainedModel,step=-.1),nrow=1)

biasfieldTransform(img(),-.5)

eval_biasfield_series[source]

eval_biasfield_series(image, mask, model, step=-0.05, start=0, end=-0.55, order=3, **kwargs)

results = eval_biasfield_series(img(), trueMask(), trainedModel)
plt.plot(results['coefficient'], results['c1'], label='LV')
plt.plot(results['coefficient'], results['c2'], label='MY')
plt.legend()
plt.xlabel('coefficient')
_ = plt.ylabel('dice')