Hierarchical deep learning approach for segmentation of glioblastoma tumor niches on digital histopathology

1 . A method, comprising:

providing one or more digitized biopsy images of a patient having a glioma;

sampling a first plurality of patches from the one or more digitized biopsy images;

providing the first plurality of patches to a first deep learning algorithm, wherein the first deep learning algorithm is configured to identify one or more necrotic regions and one or more non-necrotic regions by identifying the first plurality of patches as necrotic patches or non-necrotic patches;

sampling a second plurality of patches from the one or more non-necrotic regions; and

providing the second plurality of patches to a second deep learning algorithm, wherein the second deep learning algorithm is configured to identify glioblastoma multiforme (GBM) histopathological indicators within the one or more non-necrotic regions by identifying the second plurality of patches as GBM histopathological indicator patches or cellular tumor patches.

2 . The method of claim 1 , further comprising:

generating a prognosis for the patient from at least the GBM histopathological indicators.

3 . The method of claim 1 , wherein the GBM histopathological indicators comprise one or more of microvascular proliferation regions, hyperplastic blood vessel regions, and pseudopalisading cell regions.

4 . The method of claim 1 , further comprising:

sampling non-overlapping patches from the one or more digitized biopsy images to form the first plurality of patches;

discarding one or more of the first plurality of patches that fail to meet a selection criteria;

augmenting data within the first plurality of patches after discarding one or more of the first plurality of patches; and

providing the first plurality of patches to the first deep learning algorithm after augmenting the data within the first plurality of patches.

5 . The method of claim 4 , wherein augmenting the data within the first plurality of patches comprises performing one or more of horizontal flips, vertical flips, random shifts, color transformations, and random rotations on the first plurality of patches.

6 . The method of claim 1 , further comprising:

transforming a color of the one or more digitized biopsy images from an RGB color to a CIELAB color space prior to providing the one or more digitized biopsy images to the first deep learning algorithm.

7 . The method of claim 1 , wherein the first plurality of patches respectively have a first size and the second plurality of patches respectively have a second size that is larger than the first size.

8 . A non-transitory computer-readable medium storing computer-executable instructions that, when executed, cause a processor to perform operations, comprising:

forming an imaging data set comprising a digitized biopsy image of a patient having a glioma;

sampling a first plurality of patches from the digitized biopsy image;

segmenting the digitized biopsy image by operating upon the first plurality of patches using a first deep learning algorithm to form a plurality of intermediate patches that identify one or more necrotic regions and one or more non-necrotic regions within the digitized biopsy image;

reassembling the plurality of intermediate patches to form an intermediate digitized biopsy image comprising the one or more non-necrotic regions;

sampling a second plurality of patches from the intermediate digitized biopsy image; and

segmenting the one or more non-necrotic regions by operating upon the second plurality of patches using a second deep learning algorithm to identify the second plurality of patches as GBM histopathological indicator patches comprising a first glioblastoma multiforme (GBM) histopathological indicator or cellular tumor patches.

9 . The non-transitory computer-readable medium of claim 8 , further comprising:

mapping a color of the digitized biopsy image from an RGB color to a CIELAB color space prior to providing the digitized biopsy image to the first deep learning algorithm.

10 . The non-transitory computer-readable medium of claim 8 , further comprising:

transforming a color of the digitized biopsy image from an RGB color to an L-Channel of a CIELAB color space, wherein the first GBM histopathological indicator is a pseudopalisading cell region; or

transforming the color of the digitized biopsy image from the RGB color to an A-Channel of the CIELAB color space, wherein the first GBM histopathological indicator is a microvascular proliferation region.

11 . The non-transitory computer-readable medium of claim 8 , further comprising:

segmenting the non-necrotic regions using a third deep learning algorithm to identify a second GBM histopathological indicator, the second GBM histopathological indicator being different than the first GBM histopathological indicator.

12 . The non-transitory computer-readable medium of claim 11 ,

wherein the first GBM histopathological indicator and the second GBM histopathological indicator are respectively a microvascular proliferation region, a hyperplastic blood vessel region, or a pseudopalisading cell region.

13 . The non-transitory computer-readable medium of claim 8 , wherein the first plurality of patches respectively have a first size and the second plurality of patches respectively have a second size that is larger than the first size.

14 . The non-transitory computer-readable medium of claim 8 , wherein the second plurality of patches comprise a region surrounding one or more of the GBM histopathological indicators.

15 . A prognostic apparatus, comprising:

a memory configured to store an imaging data set comprising a digitized whole slide image (WSI) of a patient having a glioma;

one or more deep learning classifier circuits configured to:

segment the digitized WSI using a first deep learning algorithm to identify one or more necrotic regions and one or more non-necrotic regions; and

segment the one or more non-necrotic regions using a second deep learning algorithm to identify glioblastoma multiforme (GBM) histopathological indicators; and

a sampling circuit configured to:

sample a first plurality of patches from the WSI, wherein the first plurality of patches are provided to the first deep learning algorithm to identify the one or more necrotic regions and the one or more non-necrotic regions; and

sample a second plurality of patches from the non-necrotic regions, wherein the second plurality of patches are provided to the second deep learning algorithm to identify the GBM histopathological indicators and the one or more non-necrotic regions.

16 . The prognostic apparatus of claim 15 , wherein the first plurality of patches respectively have a first size and the second plurality of patches respectively have a second size that is larger than the first size.

17 . The method of claim 1 , further comprising:

taking a tissue sample from the patient; and

digitizing one or more slices of the tissue sample to generate the one or more digitized biopsy images.

18 . The non-transitory computer-readable medium of claim 8 , wherein the GBM histopathological indicator patches comprise a hyperplastic blood vessel.

19 . The prognostic apparatus of claim 15 , wherein the non-necrotic regions comprise cellular tumor regions.

20 . The prognostic apparatus of claim 15 , wherein the digitized WSI comprises a digitized H&E stained tissue slide.