US 12,243,297 B2
Expert-level detection of acute intracranial hemorrhage on head CT scans
Jitendra Malik, Berkeley, CA (US); Pratik Mukherjee, San Francisco, CA (US); Esther L. Yuh, San Francisco, CA (US); Wei-Cheng Kuo, Union City, CA (US); and Christian Haene, Berkeley, CA (US)
Assigned to THE REGENTS OF THE UNIVERSITY OF CALIFORNIA, Oakland, CA (US)
Appl. No. 17/597,714
Filed by The Regents of the University of California, Oakland, CA (US)
PCT Filed Jul. 20, 2020, PCT No. PCT/US2020/042811
§ 371(c)(1), (2) Date Jan. 19, 2022,
PCT Pub. No. WO2021/016201, PCT Pub. Date Jan. 28, 2021.
Claims priority of provisional application 62/876,491, filed on Jul. 19, 2019.
Prior Publication US 2022/0254147 A1, Aug. 11, 2022
Int. Cl. G06V 10/82 (2022.01); G06T 7/00 (2017.01); G06V 10/764 (2022.01); G06V 10/774 (2022.01)
CPC G06V 10/82 (2022.01) [G06T 7/0012 (2013.01); G06V 10/764 (2022.01); G06V 10/774 (2022.01); G06T 2207/10081 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30016 (2013.01); G06V 2201/031 (2022.01)] 15 Claims
OG exemplary drawing
 
1. A computer-implemented method, comprising:
during a training phase:
receiving a first plurality of frames from at least one original computed tomography (CT) scan of a target subject, wherein each frame may or may not include a visual indication of a hemorrhage, and further wherein each frame including a visual indication of a hemorrhage has at least one label associated therewith; and
using a fully convolutional neural network (FCN) to train a model by determining, for each of the first plurality of frames, whether at least one sub-portion of the frame includes a visual indication of a hemorrhage and classifying the sub-portion of the frame based on the determining,
wherein the training phase includes:
using three input channels including “flanking” images that are immediately superior and inferior to each frame during evaluation to simulate radiologists' practice of adjudicating tiny hemorrhages by using contextual information in slices adjacent to the image of interest;
modeling x-y axes context by limiting network evaluation on any single pass to a subset or “patch” of the image, which forces the network to make decisions based on more informative local image information; and
a patch classification branch to detach a patch prediction from noisier pixel predictions and to increase patch prediction accuracy; and
during a hemorrhage detection phase:
the hemorrhage detection module receiving a second plurality of frames from a CT scan of a target subject, wherein each frame may or may not include a visual indication of a hemorrhage; and
the hemorrhage detection module determining, for each of the second plurality of frames, whether a plurality of sub-portions of the frame includes a visual indication of a hemorrhage based at least in part on the trained model.