DeepBrainSeg.helpers package¶
Submodules¶
DeepBrainSeg.helpers.dcm2niftii module¶
DeepBrainSeg.helpers.helper_mask module¶
DeepBrainSeg.helpers.npanalysis module¶
DeepBrainSeg.helpers.postprocessing module¶
-
DeepBrainSeg.helpers.postprocessing.class_wise_cc(pred)[source]¶ Applies connected components on class wise slices
- args
- pred dimension: width, height, depth
- returns
- tensor of same size as logits (uint8)
-
DeepBrainSeg.helpers.postprocessing.connected_components(voxels, threshold=0.8)[source]¶ This clusters entire segmentations into multiple clusters and considers significant cluster for further analysis
- args
voxels: np.uint8 height x width x depth threshold: number of pixels in cluster to
consider it as significant- returns
- tensor with same size as voxels
-
DeepBrainSeg.helpers.postprocessing.densecrf(logits)[source]¶ applies coditional random fields on predictions The idea is consider the nbr voxels in making class prediction of current pixel
refer CRF and MRF papers for more theoretical idea
- args
- logits: Nb_classes x Height x Width x Depth
- returns
- tensor of size Height x Width x Depth
DeepBrainSeg.helpers.preprocessing module¶
-
DeepBrainSeg.helpers.preprocessing.clip(x, q=90)[source]¶ used to eliminate intensity outliers
q: percentile for max_value
-
DeepBrainSeg.helpers.preprocessing.normalize(img, mask=None)[source]¶ - squeezes the intensity values between 0-1
- (x - min(x)) / (max(x) - min(x))
img: ndarray, image used for analysis mask: ndarray, brain mask
-
DeepBrainSeg.helpers.preprocessing.resample3D(img, outputSize=None, interpolator=3)[source]¶ Resample 3D images Image, interpolates or subsamples the volume based on the arguments provided
img: ndarray, 3D volume outputSize: tuple (x,y,z), required dimension
of the image in output- interpolator: int (0, …, 5), type of interpolator
For Labels use nearest neighbour interpolation For image can use any:
sitkNearestNeighbor = 1, sitkLinear = 2, sitkBSpline = 3, sitkGaussian = 4, sitkLabelGaussian = 5,
-
DeepBrainSeg.helpers.preprocessing.standardize(img, mask=None, median=False)[source]¶ standardizes the volume by offsetting the intensities w.r.t mean or median on brain intensities and normalize with the standard deviation of intensity values
img: ndarray, image used for analysis mask: ndarray, brain mask median: boolean, arguments allowes us to use
median for offsetting intensity values
DeepBrainSeg.helpers.utils module¶
-
DeepBrainSeg.helpers.utils.adjust_classes(volume)[source]¶ readjusts class values in the form of brats data. Removes brain class and reassigns class 3 to class 4
- args
- volume: segmentaiton mask with Height x Width x Depth
- returns
- tensor with same dimension as input volume
-
DeepBrainSeg.helpers.utils.adjust_classes_air_brain_tumour(volume)[source]¶ Converts ABL classes to AL class as brain is introduced class, it is merged with air in this funciton
- args
- volume: segmentation mask (H x W x D)
- returns
- tensor with same dimension as volume
-
DeepBrainSeg.helpers.utils.apply_argmax_to_logits(logits)[source]¶ Applies argmax to logits along dimension 0
- args
- logits dimensions: nclasses x Height x Width x Depth
- returns
- tensor with dimension Height x Width x Depth (uint8)
-
DeepBrainSeg.helpers.utils.bbox(vol, pad=18)[source]¶ estimates closest bounding rect for the tumor
- args
vol: segmentation mask (H x W X D) with tumor pad: int; number of zeros to be padded to the
estimated bounding box- returns
- tuple(<int> of len 6):
- xmin, xmax, ymin, ymax, zmin, zmax
-
DeepBrainSeg.helpers.utils.combine_logits_AM(x)[source]¶ Method to ensemble the predictions of multiple models considers AM over logits
estimates $$rac{Sum x_i}{len(x)}$$
- args
- x: list of all predictions
- returns
- tensor of dimension same as x[0]
-
DeepBrainSeg.helpers.utils.combine_logits_GM(x)[source]¶ Method to ensemble the predictions of multiple models considers GM over logits
estimates $$(Prod x_i)^{1/len(x)}$$- args
- x: list of all predictions
- returns
- tensor of dimension same as x[0]
-
DeepBrainSeg.helpers.utils.combine_mask_prediction(mask, pred)[source]¶ Conbines final inner class predictions with localized predictions
- args
- mask: localized prediction tensor pred: inner class prediction tensor
- returns
- tensor of same dimension as mask or pred
-
DeepBrainSeg.helpers.utils.combine_predictions_AM(x)[source]¶ Method to ensemble the predictions of multiple models considers AM over predictions
estimates $$rac{Sum x_i}{len(x)}$$
- args
- x: list of all predictions
- returns
- tensor of dimension same as x[0]
-
DeepBrainSeg.helpers.utils.combine_predictions_GM(x)[source]¶ Method to ensemble the predictions of multiple models considers GM over predictions
estimates $$(Prod x_i)^{1/len(x)}$$- args
- x: list of all predictions
- returns
- tensor of dimension same as x[0]
-
DeepBrainSeg.helpers.utils.convert5class_logitsto_4class(logits)[source]¶ Merges brain class with background
- args
- logits: 5 x H x W x D
- returns
- tensor of dim 4 x H x W x D
-
DeepBrainSeg.helpers.utils.convert_image(image)[source]¶ Rearranging image channels accordingly
- args
- image: W x H x C
- returns
- image: C x W x H
-
DeepBrainSeg.helpers.utils.get_brain_mask(t1)[source]¶ Provides an alternative for ANTs used to extract brain mask given raw image Does the task of skull stripping
t1: T1 MRI volume: height x width x depth
returns: boolean mask
-
DeepBrainSeg.helpers.utils.get_dice_score(prediction, ground_truth)[source]¶ Calculates dice coefficient for each class
calculates $$rac{2*A cap B}{A cup B}$$
- args
- prediction: ndarray with predicted segmentation classes ground_truth: ndarray with expected segmentation classes
- returns
- tuple(<float> of len 3) whole tumor dice,
- enhancing tumor dice, and tumor core dice
-
DeepBrainSeg.helpers.utils.get_enhancing_tumor(data)[source]¶ gets enhancing tumor region from a given segmentation mask
- args
- data: ndarray (uint8)
- returns
- boolean array with same dimension as data
-
DeepBrainSeg.helpers.utils.get_tumor_core(data)[source]¶ gets tumor core region from a given segmentation mask
- args
- data: ndarray (uint8)
- returns
- boolean array with same dimension as data
-
DeepBrainSeg.helpers.utils.get_whole_tumor(data)[source]¶ gets whole tumor region from a given segmentation mask
- args
- data: ndarray (uint8)
- returns
- boolean array with same dimension as data
-
DeepBrainSeg.helpers.utils.imshow(*args, **kwargs)[source]¶ Handy function to show multiple plots in on row, possibly with different cmaps and titles
Usage: imshow(img1, title=”myPlot”) imshow(img1,img2, title=[‘title1’,’title2’]) imshow(img1,img2, cmap=’hot’) imshow(img1,img2,cmap=[‘gray’,’Blues’])