Mechanisms for brain analysis
AssigneeKoninklijke Philips N.V.
InventorEvan Schwab
Life of this patent
Estimating the term…
Abstract
A method for localizing structural connectivity biomarkers in neurological diseases, includes dividing a diffusion magnetic resonance imaging brain volume into a set of connected brain regions; extracting three-dimensional voxels along fiber connections which structurally connect the connected brain regions, wherein the brain regions comprise bundles of neurons; applying a deep neural network to diffusion magnetic resonance imaging features extracted from the three-dimensional voxels for each set of fiber connections which structurally connect brain regions; outputting a disease classification based on applying the deep neural network; and applying multi-instance learning to predict whether each fiber connection is indicative of a healthy brain or a diseased brain.
Claims17
A method for localizing structural connectivity biomarkers in neurological diseases, comprising:
- dividing a diffusion magnetic resonance imaging brain volume into a set of connected brain regions;
- extracting three-dimensional voxels along fiber connections which structurally connect the connected brain regions, wherein the connected brain regions comprise bundles of neurons;
- applying a deep neural network to diffusion magnetic resonance imaging features extracted from the three-dimensional voxels for each set of fiber connections which structurally connect brain regions;
- defining, for multi-instance learning, each instance as a fiber connection that structurally connects a pair of segmented brain regions delineated by a brain atlas;
- producing, via a Sigmoid layer of the deep neural network, a probabilistic class score for each fiber connection indicating a positive or negative label for the presence or absence of a specified neurological disease;
- aggregating the per-fiber-connection probabilistic class scores in a final multi-instance learning layer to produce a brain-level classification, wherein the brain-level classification is based on the aggregation, and outputting the brain-level classification;
- and
- outputting a set of fiber-connection locations and the corresponding probabilistic class scores to provide a connection-wise probabilistic localization of a disease biomarker.
A system for localizing structural connectivity biomarkers in neurological diseases, comprising:
- a memory that stores instructions;
- a display;
- and
- a processor that executes the instructions, wherein based on the processor executing the instructions, the system is caused to:
- divide a diffusion magnetic resonance imaging brain volume into a set of connected brain regions;
- extract three-dimensional voxels along fiber connections which structurally connect the connected brain regions, wherein the connected brain regions comprise bundles of neurons;
- apply a deep neural network to diffusion magnetic resonance imaging features extracted from the three-dimensional voxels for each set of fiber connections which structurally connect brain regions;
- apply multi-instance learning in which each instance corresponds to a fiber connection that structurally connects a pair of segmented brain regions delineated by a brain atlas, including producing, via a Sigmoid layer of the deep neural network, a probabilistic class score for each fiber connection indicating a positive or negative label for the presence or absence of a specified neurological disease, aggregating the per-fiber-connection probabilistic class scores in a final multi-instance learning layer to produce a brain-level classification, and output the brain-level classification;
- and
- output a set of fiber-connection locations and the corresponding probabilistic class scores to provide a connection-wise probabilistic localization of a disease biomarker.
A controller for localizing structural connectivity biomarkers in neurological diseases, comprising:
- a memory that stores instruction;
- and
- a processor that executes the instructions, wherein based on the processor executing the instructions, the controller is caused to:
- divide a diffusion magnetic resonance imaging brain volume into a set of connected brain regions;
- extract three-dimensional voxels along fiber connections which structurally connect the connected brain regions, wherein the connected brain regions comprise bundles of neurons;
- apply a deep neural network to diffusion magnetic resonance imaging features extracted from the three-dimensional voxels for each set of fiber connections which structurally connect brain regions;
- apply multi-instance learning in which each instance corresponds to a fiber connection that structurally connects a pair of segmented brain regions delineated by a brain atlas, including producing, via a Sigmoid layer of the deep neural network, a probabilistic class score for each fiber connection indicating a positive or negative label for the presence or absence of a specified neurological disease, aggregating the per-fiber-connection probabilistic class scores in a final multi-instance learning layer to produce a brain-level classification, and output the brain-level classification;
- and
- output a set of fiber-connection locations and the corresponding probabilistic class scores to provide a connection-wise probabilistic localization of a disease biomarker.
Dependencies are inferred from the claim text.