Apparatus to analyse diffusion magnetic resonance imaging data
An apparatus includes an input unit, a processing unit, and an output unit. The input unit is configured to provide the processing unit with at least one diffusion magnetic resonance imaging dMRI image of a patient's brain. The processing unit is configured to: 1) determine an estimate of an orientation of neurons at each voxel in the dMRI image; 2) determine a plurality of fiber tracts in the at least one dMRI image; 3) select a plurality of voxels along at least one fiber tract of the plurality of fiber tracts; and 4) determine a neurological disease classification.
1. An apparatus to analyze diffusion magnetic resonance imaging data, the apparatus comprising:
an input unit;
a processing unit; and
an output unit;
wherein, the input unit is configured to provide the processing unit with set of diffusion-weighted diffusion magnetic resonance imaging (dMRI) images of a patient's brain;
wherein, the processing unit is configured to determine an estimate of an orientation of neurons at each voxel of a brain volume in the dMRI image set, the determination comprising:
determining a spherical probability distribution function at each voxel of the brain volume relating to probabilities of diffusion in a given direction at each voxel based on the set of dMRI images, and
determining the estimated orientation of neurons at each voxel by utilizing a peak in the spherical probability distribution function at each voxel;
wherein three or more regions of the brain volume in the set of dMRI images are segmented;
wherein the processing unit is configured to determine at least one diffusivity feature for each voxel of the brain volume based on the set of dMRI images;
wherein, the processing unit is configured to determine a plurality of fiber tracts in the brain volume in the dMRI image set using the estimated orientation of neurons at each voxel of the brain volume;
wherein the processing unit is configured to identify a plurality of pairs of the segmented regions which are connected by fiber tracts to one another, and for each pair of connected segmented regions:
select a plurality of voxels along fiber tracts connecting the pair of regions,
provide the at least one diffusivity feature of each of the selected voxels as inputs to a multi-input deep neural network which comprises a respective sparse 3D convolutional neural network for each pair of segmented regions, wherein each respective sparse neural network outputs a respective feature vector for the pair of regions based on the input features,
wherein, the processing unit is configured to determine a neurological disease classification, wherein the multi-input deep neural network concatenates the feature vectors generated for each of the pairs of brain regions in a further fully connected layer, and wherein an output of the multi-input deep neural network is the classification of the neurological disease; and
wherein, the output unit is configured to output the neurological disease classification,
wherein the multi-input deep neural network is trained on a basis of a plurality of dMRI image sets of reference patients' brains and information relating to neurological disease information for each of the reference patients.
2. The apparatus of claim 1 , wherein the processing unit is configured to determine at least one connectivity weight for each connected pair of segmented regions based on the number of fiber tracts that connect the two connected segmented regions.
3. The apparatus of claim 1 , wherein the processing unit is configured to determine at least one connectivity weight for each pair of connected segmented regions based on at least one diffusivity feature associated with each of a plurality of voxels along fiber tracts that connect the two connected segmented regions.
4. The apparatus of claim 1 , wherein the processing unit is configured to select voxels to be input to the multi-input deep neural network for active or non-zero voxels of the plurality of pixels along the at least one fiber tract.
5. An imaging system-, comprising:
a magnetic resonance image unit;
an apparatus according to claim 1 ;
wherein, the magnetic resonance imaging unit is configured to acquire the set of dMRI images, and provide the set of dMRI images to the apparatus,
and wherein, the apparatus is configured to output a neurological disease classification for the patient on the basis of the set of dMRI images.
6. A method of analyzing diffusion magnetic resonance imaging data, the method comprising:
a) providing a processing unit with a set of diffusion-weighted diffusion magnetic resonance imaging (dMRI) images of a patient's brain wherein three or more regions of the brain volume in the dMRI image set are segmented;
b) determining by the processing unit an estimate of an orientation of neurons at each voxel of a brain volume in the dMRI image set by
determining a spherical probability distribution function at each voxel of the brain volume relating to probabilities of diffusion in a given direction at each voxel based on the set of diffusion weighted images, and
determining an estimated orientation of neurons at each voxel by utilizing a peak in the spherical probability distribution function at each voxel;
determining by the processing unit at least one diffusivity feature for each voxel of the brain volume based on the set of diffusion-weighted images;
c) determining by the processing unit a plurality of fiber tracts in the brain volume in the dMRI image set, the determining comprising utilizing the estimated orientation of neurons at each voxel of the brain volume;
d) identifying by the processing unit a plurality of pairs of the segmented regions which are connected by fiber tracts to one another, and for each pair of connected segmented regions:
selecting by the processing unit a plurality of voxels along fiber-tracks connecting the pair or regions and;
providing the at least one diffusivity feature of each of the selected voxels as inputs to a multi-input deep neural network which comprises a respective sparse 3D convolutional neural network for each pair of segmented regions, wherein each respective sparse neural network outputs a respective feature vector for the pair of regions based on the input features;
e) determining by the processing unit a neurological disease classification, wherein the multi-input deep neural network concatenates the feature vectors generated for each of the pairs of brain regions in a further fully connected layer, and wherein an output of the multi-input deep neural network is the classification of the neurological disease; and
f) outputting by an output unit the neurological disease classification.
7. The method of claim 6 , further comprising: determining, by the processing unit, at least one connectivity weight for each connected pair of segmented regions based on the number of fiber tracts that connect the two connected segmented regions.
8. The method of claim 6 , further comprising: determining, by the processing unit, at least one connectivity weight for each pair of connected segmented regions based on at least one diffusivity feature associated with each of a plurality of voxels along fiber tracts that connect the two connected segmented regions.
9. The method of claim 6 , further comprising: selecting, by the processing unit, voxels to be input to the multi-input deep neural network for active or non-zero voxels of the plurality of pixels along the at least one fiber tract.
10. A computer programing comprising machine executable instructions stored on a non-transitory computer readable medium for execution by a processing unit, wherein when executed causes the processing unit to perform the method of claim 6 .