Systems and methods for identifying progression of hypoxic-ischemic brain injury
A method for identifying the presence or progression of hypoxic ischemic brain injury includes, for each subset of one or more subsets of a three-dimensional medical image of a head of a patient: (i) inputting said each subset into a machine-learning model, (ii) extracting one or more features or feature maps from the machine-learning model, and (iii) constructing, based on the one or more features or feature maps, one of a sequence of vectors. The sequence of vectors is then pooled to obtain a scan-level vector that is used to obtain a score indicating HIBI presence or progression in the patient. For example, the scan-level vector can be inputted into a pre-trained classifier that generates the score based on the scan-level vector. The machine-learning model may be a pre-trained conventional neural network or support vector machine.
1 . A method for identifying the presence or progression of hypoxic ischemic brain injury (HIBI), comprising:
for each subset of one or more subsets of a three-dimensional medical image of a head of a patient:
inputting said each subset into a machine-learning model;
extracting one or more feature maps from the machine-learning model; and
constructing, based on the one or more feature maps, a respective one of a sequence of subset-level feature vectors;
pooling the sequence of subset-level feature vectors to obtain a scan-level feature vector; and
using the scan-level feature vector to obtain a score indicating HIBI presence or progression.
2 . The method of claim 1 , wherein said using the scan-level feature vector includes:
transforming the scan-level feature vector into a reduced-dimensionality scan-level feature vector; and
feeding the reduced-dimensionality scan-level feature vector into a classifier to obtain the score indicating HIBI presence or progression.
3 . The method of claim 2 , wherein the classifier includes a support-vector machine.
4 . The method of claim 2 , wherein said transforming includes multiplying the scan-level feature vector by a projection matrix to obtain the reduced-dimensionality scan-level feature vector.
5 . The method of claim 1 , wherein said constructing comprises:
for each of the one or more feature maps:
determining a mean value of said each of the one or more feature maps; and
appending the mean value to the respective one of the sequence of subset-level feature vectors.
6 . The method of claim 1 , further comprising normalizing the respective one of the sequence of subset-level feature vectors.
7 . The method of claim 1 , wherein said pooling includes max-pooling.
8 . The method of claim 1 , the machine-learning model including a neural network.
9 . The method of claim 8 , wherein said extracting includes extracting the one or more feature maps from one or more max-pooling layers of the neural network.
10 . The method of claim 1 , further comprising identifying, based on the score, one of:
a first endotype indicating absence of HIBI; and
a second endotype indicating presence of HIBI.
11 . The method of claim 10 , further comprising treating the patient based on the first endotype or the second endotype.
12 . The method of claim 1 , further comprising capturing the three-dimensional medical image of the head of the patient.
13 . The method of claim 12 , wherein said capturing occurs within three hours of return of spontaneous circulation following cardiac arrest.
14 . A system for identifying progression of hypoxic-ischemic brain injury (HIBI), comprising:
a processor;
a memory communicably coupled with the processor;
a machine-language model implemented as machine-readable instructions stored in the memory; and
an HIBI predictor implemented as machine-readable instructions that are stored in the memory and, when executed by the processor, control the system to:
for each subset of one or more subsets of a three-dimensional medical image of a head of a patient:
(i) input said each subset into a machine-learning model,
(ii) extract one or more feature maps from the machine-learning model, and
(iii) construct, based on the one or more feature maps, a respective one of a sequence of subset-level feature vectors,
pool the sequence of subset-level feature vectors to obtain a scan-level feature vector, and
use the scan-level feature vector to obtain a score indicating HIBI presence or progression.
15 . The system of claim 14 , wherein the machine-readable instructions that, when executed by the processor, control the system to use the scan-level feature vector include machine-readable instructions that, when executed by the processor, control the system to:
transform the scan-level feature vector into a reduced-dimensionality scan-level feature vector; and
feed the reduced-dimensionality scan-level feature vector into a classifier to obtain the score indicating HIBI presence or progression.
16 . The system of claim 15 , wherein the classifier includes a support-vector machine.
17 . The system of claim 14 , the machine-learning model being a neural network.
18 . The system of claim 17 , the neural network being a VGG neural network.
19 . The system of claim 17 , wherein the machine-readable instructions that, when executed by the processor, control the system to extract include machine-readable instructions that, when executed by the processor, control the system to extract the one or more feature maps from one or more max-pooling layers of the neural network.
20 . The system of claim 14 , the HIBI predictor storing additional machine-readable instructions that, when executed by the processor, control the system to identify, based on the score, one of:
a first endotype indicating absence of HIBI; and
a second endotype indicating presence of HIBI.