RSNABreastCancer

Normalization¶

In [ ]:

from connectome import Transform


class Normalize(Transform):
    __inherit__ = True

    def image(image, padding_value, intensity_sign):
        if padding_value is not None:
            if padding_value > 0:
                return padding_value - image
            return image

        if intensity_sign == 1:
            return image.max() - image

        return image

from connectome import Transform class Normalize(Transform): __inherit__ = True def image(image, padding_value, intensity_sign): if padding_value is not None: if padding_value > 0: return padding_value - image return image if intensity_sign == 1: return image.max() - image return image

Zoom to reduce image size¶

In [ ]:

from connectome import Apply
from scipy.ndimage import zoom

# 0.25 - is the downsample factor. It should probably be tuned via cross-validation
Zoom = Apply(image=lambda x: zoom(np.float32(x), 0.25, order=1))

from connectome import Apply from scipy.ndimage import zoom # 0.25 - is the downsample factor. It should probably be tuned via cross-validation Zoom = Apply(image=lambda x: zoom(np.float32(x), 0.25, order=1))

Artifacts and background removal¶

In [ ]:

from connectome import Transform
from skimage.morphology import label


class GreatestComponent(Transform):
    __inherit__ = True

    def image(image):
        lbl = label(image > 0)
        values, counts = np.unique(lbl, return_counts=True)
        foreground = values != 0
        component = values[foreground][counts[foreground].argmax()]
        # select all the components greater than the background
        #  + the greatest foreground component
        components = set(values[counts > counts[~foreground]]) | {component}
        if len(components) > 1:
            # if there are several components - pick the one with the greatest intensity
            component = max(components, key=lambda c: image[lbl == c].mean())

        return image * (lbl == component)


class CropBackground(Transform):
    __inherit__ = True

    def image(image):
        mask = image > 0
        xs, = mask.any(0).nonzero()
        ys, = mask.any(1).nonzero()
        return image[ys.min():ys.max() + 1, xs.min():xs.max() + 1]

from connectome import Transform from skimage.morphology import label class GreatestComponent(Transform): __inherit__ = True def image(image): lbl = label(image > 0) values, counts = np.unique(lbl, return_counts=True) foreground = values != 0 component = values[foreground][counts[foreground].argmax()] # select all the components greater than the background # + the greatest foreground component components = set(values[counts > counts[~foreground]]) | {component} if len(components) > 1: # if there are several components - pick the one with the greatest intensity component = max(components, key=lambda c: image[lbl == c].mean()) return image * (lbl == component) class CropBackground(Transform): __inherit__ = True def image(image): mask = image > 0 xs, = mask.any(0).nonzero() ys, = mask.any(1).nonzero() return image[ys.min():ys.max() + 1, xs.min():xs.max() + 1]

Data augmentation¶

In [ ]:

from connectome import Transform, impure
import numpy as np


class RandomFlip(Transform):
    __inherit__ = True

    @impure
    def _flip():
        return np.random.binomial(1, 0.5)

    def image(image, _flip):
        if _flip:
            return np.flip(image, axis=1)
        return image

from connectome import Transform, impure import numpy as np class RandomFlip(Transform): __inherit__ = True @impure def _flip(): return np.random.binomial(1, 0.5) def image(image, _flip): if _flip: return np.flip(image, axis=1) return image

Combining it all together¶

In [ ]:

from amid.rsna_bc import RSNABreastCancer
from connectome import Chain

ds = Chain(
    RSNABreastCancer('/path/to/downloaded/folder'),
    Normalize(),
    Apply(image=lambda x: zoom(np.float32(x), 0.25, order=1)),
    GreatestComponent(),
    CropBackground(),

    # aug
    RandomFlip(),
)

from amid.rsna_bc import RSNABreastCancer from connectome import Chain ds = Chain( RSNABreastCancer('/path/to/downloaded/folder'), Normalize(), Apply(image=lambda x: zoom(np.float32(x), 0.25, order=1)), GreatestComponent(), CropBackground(), # aug RandomFlip(), )