WO2021058843A1 - Método y sistema para la segmentación automática de hiperintensidades de sustancia blanca en imágenes de resonancia magnética cerebral - Google Patents
Método y sistema para la segmentación automática de hiperintensidades de sustancia blanca en imágenes de resonancia magnética cerebral Download PDFInfo
- Publication number
- WO2021058843A1 WO2021058843A1 PCT/ES2020/070069 ES2020070069W WO2021058843A1 WO 2021058843 A1 WO2021058843 A1 WO 2021058843A1 ES 2020070069 W ES2020070069 W ES 2020070069W WO 2021058843 A1 WO2021058843 A1 WO 2021058843A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- layers
- image
- convolutional
- neural network
- convolutional neural
- Prior art date
Links
- 206010072731 White matter lesion Diseases 0.000 title claims abstract description 35
- 210000004556 brain Anatomy 0.000 title claims abstract description 29
- 238000000034 method Methods 0.000 title claims abstract description 26
- 230000011218 segmentation Effects 0.000 title abstract description 17
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 83
- 230000001575 pathological effect Effects 0.000 claims abstract description 12
- 238000012549 training Methods 0.000 claims description 30
- 238000013528 artificial neural network Methods 0.000 claims description 17
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 16
- 230000009467 reduction Effects 0.000 claims description 16
- 230000004913 activation Effects 0.000 claims description 12
- 230000017105 transposition Effects 0.000 claims description 12
- 238000002595 magnetic resonance imaging Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000001914 filtration Methods 0.000 claims description 2
- 238000007670 refining Methods 0.000 claims description 2
- 238000012935 Averaging Methods 0.000 claims 1
- 238000004513 sizing Methods 0.000 claims 1
- 230000006870 function Effects 0.000 description 7
- 238000012545 processing Methods 0.000 description 7
- 230000003902 lesion Effects 0.000 description 6
- 238000012544 monitoring process Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 6
- 208000037451 Leukoaraiosis Diseases 0.000 description 5
- 210000004885 white matter Anatomy 0.000 description 4
- 238000003745 diagnosis Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000004590 computer program Methods 0.000 description 2
- 238000002790 cross-validation Methods 0.000 description 2
- 230000010339 dilation Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 239000012530 fluid Substances 0.000 description 2
- 238000002075 inversion recovery Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 208000024827 Alzheimer disease Diseases 0.000 description 1
- 101100400452 Caenorhabditis elegans map-2 gene Proteins 0.000 description 1
- 206010012289 Dementia Diseases 0.000 description 1
- 208000027418 Wounds and injury Diseases 0.000 description 1
- 230000005856 abnormality Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 208000029028 brain injury Diseases 0.000 description 1
- 210000003710 cerebral cortex Anatomy 0.000 description 1
- 210000004289 cerebral ventricle Anatomy 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000006378 damage Effects 0.000 description 1
- 238000013434 data augmentation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000036449 good health Effects 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 208000014674 injury Diseases 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000002610 neuroimaging Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 210000003625 skull Anatomy 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 208000024891 symptom Diseases 0.000 description 1
- 210000001519 tissue Anatomy 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G06N3/0455—Auto-encoder networks; Encoder-decoder networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves
- A61B5/055—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves involving electronic [EMR] or nuclear [NMR] magnetic resonance, e.g. magnetic resonance imaging
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H30/00—ICT specially adapted for the handling or processing of medical images
- G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30016—Brain
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30096—Tumor; Lesion
Definitions
- the present invention relates to the technical field of computer-aided analysis of medical images and more specifically to the processing by means of convolutional neural networks of images obtained by magnetic resonance, for an automatic segmentation of the images based on the volume of leukoaraiosis, especially useful for the study of radiologists-neurologists on possible brain injuries.
- magnetic resonance imaging is a commonly used preliminary diagnostic tool to detect abnormalities in tissues. Specifically, in the field of neuro-radiology, it is used by specialists, among other things, to diagnose white matter lesions.
- the volume of leukoaraiosis is used in medicine precisely to detect the diffuse loss of density in various regions of the white matter of the brain.
- it is not a diagnosis, but rather a radiological finding related to alterations in the white matter (hyperintensities in this case) and which is detected by images obtained by, among other techniques, magnetic resonance imaging.
- leukoaraiosis is a descriptive neuroimaging term that appears both in people in good health and in different clinical and pathological situations, including Alzheimer's disease and others. dementias, although there are no specific associated symptoms.
- the present invention describes, in a first aspect, a method for segmenting hyperintensities of white matter present in magnetic resonance brain images comprising: providing a magnetic resonance brain image to a set of convolutional neural networks, previously trained; identify voxels in the image that contain white matter hyperintensities; determining, for each of the convolutional neural networks and for each voxel, a probability that the identified hyperintensity corresponds to a previously defined pathological hyperintensity; average all the probabilities determined for each voxel; comparing the averaged probabilities for each voxel with a previously established threshold; and generating an image mask with the voxels that exceed the threshold.
- a pre-processing of the provided image comprising: dimensioning each 2D section of the image to a size of 256x256; scale the intensity values of each 2D section between 0 and 1, according to the equation: rh, j ⁇ min (/)
- I is the 2D section to be scaled, l, the intensity value in row i and column j, min (l) is the minimum value of the section and max ( l) is its maximum value; and standardize each voxel by subtracting the mean of a training data set and dividing by the standard deviation of the training data set, according to the following equation: where, I 1 is the scaled intensity; m is the mean intensity of the scaled training data set and s is the standard deviation of the scaled training data set.
- the set of convolutional neural networks comprises a standard convolutional neural network, a residual convolutional neural network, a dilated convolutional neural network, and a dilated residual convolutional neural network, all based on an encoder-architecture architecture. decoder.
- the combination of these neural networks produces better results than those offered when used in isolation.
- the following steps are contemplated: filtering the magnetic resonance brain image by means of convolution operations, in some convolutional layers of each of the convolutional neural networks; reducing the image, by intermediate reduction layers sandwiched between the convolutional layers arranged in an encoding path; augmenting the image, by transposition convolutional layers sandwiched between the convolutional layers arranged in a decoding path; refining the output of the convolutional neural network using layers of deep monitoring; and defining a map of the probability of the presence of hyperintensity, by a sigmoid activation function.
- a total probability map representing, averaged for each voxel, the probabilities determined by the convolutional neural networks for each voxel.
- the probability threshold that determines whether or not each voxel is included in the image mask is set to 50%.
- the training of the convolutional neural networks of the present invention comprises, in one of its preferred embodiments, applying a Tversky loss function.
- the imbalance is advantageously balanced between the number of voxels that exceed the previously established threshold and those that do not.
- a second aspect of the present invention relates to a system for segmenting white matter hyperintensities present in magnetic resonance brain images comprising:
- an arrangement of convolutional neural networks configured to receive a magnetic resonance brain image, identify the voxels in the image that contain white matter hyperintensities and determine, for each of these networks, a probability for each voxel that the identified hyperintensity will be corresponds to a previously defined pathological hyperintensity;
- a processor module configured to average all the probabilities determined for each voxel, compare the probabilities averaged for each voxel with a previously established threshold, and generate an image mask with the voxels that exceed the threshold.
- the arrangement of neural networks comprises, according to one of the preferred embodiments of the present invention, at least one standard neural network comprising: convolutional layers configured to filter the magnetic resonance brain image by means of convolution operations; reduction layers sandwiched between convolutional layers in an encoding path, configured to reduce the image; transposition layers sandwiched between the convolutional layers in a decoding path, configured to magnify the image; a bottleneck section between the encoding path and the decoding path; a sigmoid activation layer, configured to generate a probability map; and deep monitoring layers, configured to combine the output of the convolutional layers of the decoding path.
- the neural network arrangement comprises at least one residual convolutional neural network, which in turn comprises: residual blocks configured to filter the magnetic resonance brain image through convolution operations; reduction layers sandwiched between residual blocks in an encoding path, configured to reduce the image; transposition layers sandwiched between residual blocks in a decoding path, configured to magnify the image; a bottleneck section between the encoding path and the decoding path; a sigmoid activation layer, configured to generate a probability map; and deep monitoring layers, configured to combine the output of the convolutional layers of the decoding path.
- the neural network arrangement comprises at least one dilated convolutional neural network, which in turn comprises: convolutional layers configured to filter the magnetic resonance brain image by means of convolution operations; reduction layers sandwiched between convolutional layers in an encoding path, configured to reduce the image; transposition layers sandwiched between the convolutional layers in a decoding path, configured to magnify the image; a bottleneck section between the encoding path and the decoding path, wherein said section comprises stretched convolutional layers with different dilation rates; a sigmoid activation layer, configured to generate a probability map; and deep monitoring layers, configured to combine the output of the convolutional layers of the decoding path.
- the neural network arrangement comprises at least one residual-dilated convolutional neural network, which in turn comprises: residual blocks configured to filter by means of operations of brain magnetic resonance imaging convolution; reduction layers sandwiched between residual blocks in an encoding path, configured to reduce the image; transposition layers sandwiched between residual blocks in a decoding path, configured to magnify the image; a bottleneck section between the encoding path and the decoding path, wherein said section comprises stretched convolutional layers with different dilation rates; a sigmoid activation layer, configured to generate a probability map; and deep monitoring layers, configured to combine the output of the convolutional layers of the decoding path.
- the arrangement of neural networks is composed of the standard convolutional neural network, the residual convolutional neural network, the dilated convolutional neural network and the residual-dilated convolutional neural network, all of them configured to a 256x256 image input size, and where all convolutional layers have 3x3 receptive field followed by batch normalization layer and rectified linear unit layer.
- a final aspect of the present invention relates to a computer-implemented method for segmenting white matter hyperintensities present in magnetic resonance brain images, comprising: providing a magnetic resonance brain image to a set of previously trained convolutional neural networks; identify voxels in the image that contain white matter hyperintensities; determining, for each of the convolutional neural networks and for each voxel, a probability that the identified hyperintensity corresponds to a previously defined pathological hyperintensity; average all the probabilities determined for each voxel; comparing the averaged probabilities for each voxel with a previously established threshold; and generating an image mask with the voxels that exceed the threshold.
- the convolutional neural networks are implemented in a central processing unit CPU. In a specific embodiment of the present invention, at least one of the convolutional neural networks is implemented in a graphics processing unit GPU supporting the central processing unit.
- the present invention has technical characteristics that involve a multitude of advantageous effects.
- the arrangement of neural networks described in the present invention makes it possible to discern between physiological hyperintensities and hyperintensities that could correspond to white matter lesions, which enables a specialist to concentrate his work on the areas segmented by the present invention and included in the image mask, or even to obtain additional quantitative parameters based on the segmentation provided to describe the lesion load (such as volume or number of lesions) of people with leukoaraiosis.
- the benefits extend for example to the optimization of the time of the specialist as in an increase of his precision.
- the novel combination of convolutional neural networks of the present invention offers greater robustness to segmentation than any other of the known solutions. It allows images to be processed by extracting characteristics at different levels, specifically focusing on the complex task of segmenting white matter lesions, which vary their distribution from one person to another and require a special configuration such as that of the present invention.
- FIG. 1 represents a block diagram of an embodiment of the invention.
- - Figure 2 represents in detail an embodiment of a standard convolutional neural network used in the present invention.
- - Figure 3 represents in detail an embodiment of a residual convolutional neural network used in the present invention.
- FIG. 4 represents in detail an embodiment of a dilated convolutional neural network used in the present invention.
- FIG. 5 represents in detail an embodiment of a residual-dilated convolutional neural network used in the present invention.
- the present invention discloses a method and a system for the automatic segmentation of brain magnetic resonance images, which can be used to measure the volume of leukoaraiosis and which can subsequently be used by neurologist-radiologists to detect possible white matter lesions.
- the automatic segmentation method and system is an aid to the diagnosis of specialists, through image processing, as detailed below, to finally provide in its output an image mask generated from the individual processing of each voxel of individual images, where the criteria established to determine that a voxel belongs to injury or does not seek to differentiate, from among all the voxels in the image, those that correspond to pathological hyperintensities from those that do not, which will be of great help in the subsequent studies, such as the identification of white matter lesions.
- the input images come from magnetic resonance imaging (MRI) of the brain. From these magnetic resonance images, the FLAIR (inversion-recovery in fluid attenuation) sequence is used to detect voxels belonging to lesions in the white matter of the brain.
- These white matter hyperintensities (WMH) are segmented by the present invention by generating a probability map by different CNNs that are combined and thresholdized to detect voxels comprising pathological hyperintensities.
- Figure 1 represents the basic blocks of the invention, according to one of its embodiments, where image 1 is processed by a set of convolutional neural networks 11, 12, 13 and 14, abbreviated as CNN, each configured with an architecture concrete and previously trained.
- CNN convolutional neural networks
- Each of the CNNs provides at its output a 16, 17, 18 and 19 probability map of the presence / absence of a white matter lesion in each voxel of the image.
- a processor module 5 the probability maps associated with each of the CNNs are averaged to obtain a final probability map 2, which establishes the threshold 3 to generate the image masks 4 provided at the output.
- a pre-processing of the input images is contemplated, such that each 2D section of each 3D image is reshaped to a common size of 256x256, to fit the input size of the convolutional neural network architectures ( CNN) used by applying a bicubic interpolation.
- CNN convolutional neural network architectures
- Min-Max scaling the intensity values of each 2D section are scaled between 0 and 1: where, I is the section to be scaled, l, the intensity value in row i and column j, min (l) is the minimum value of the section and max (l) is its maximum value; and 2. standardization: the training data set is normalized to have a mean of zero and a variance of one. For that, each voxel is standardized by subtracting the mean of the training data set and dividing by the deviation training data set standard.
- G is the min-max scaled intensity
- m is the mean intensity of the scaled training data set
- s is the standard deviation of the scaled training data set.
- Standard CNN 11 allows features to be extracted at different levels, ranging from simpler features based on edges and contrasts to more complex ones based on the shape of the image and the region of interest to segment.
- CNN with residual convolutional blocks 12 allows more complex characteristics to be obtained without the need to increase the number of convolutional layers.
- CNN with dilated convolutional filters 13 that allows increasing the region over which characteristics are extracted. Therefore, this architecture is designed and optimized to extract different types of information from the image and increase the precision in the segmentation of white matter lesions.
- a decoder-encoder architecture has an encoding path in which, given an input image, a high-dimensional feature vector is generated by adding features at different levels; and a decoding path, in which the previously generated feature vector is decoded at multiple levels to, finally, generate a probability map.
- some bridge connections are arranged, combining both local and contextual information by concatenating features from the encoding path with features from the decoding path.
- some layers of deep monitoring are included. These layers allow for faster convergence, which means less training time and better edge delineation.
- all convolutional neural network architectures in line with the preprocessing described above, have an input layer with a size of (256, 256, 1) with the values corresponding to (# rows, # columns, # channels).
- the number of channels is set to 1 due to working with monochrome images.
- Figure 2 represents in detail the architecture of the standard convolutional neural network 11 with an encoder-decoder configuration, which includes characteristics common to the other CNN networks.
- the architecture of the encoding path depicted comprises convolutional layers 21 alternating with intermediate reduction layers 22. All convolutional layers have a 3x3 receptive field followed by a batch normalization layer (BN) and a rectified linear unit layer (ReLU ).
- the BN layers normalize the output of the convolutional layers to zero mean and variance one, increasing the stability of the convolutional neural network (CNN).
- ReLU is an activation function that introduces non-linearities into the model.
- the "max-pooling" reduction layer 22 with a 2x2 core is arranged after every two convolutional layers 21, thus reducing the image size in half. After each "max-pooling" reduction layer, the number of filters in the convolutional layers doubles.
- the two convolutional layers of the bottleneck section 25 between the encoder and the decoder layers have 512 filters for a reduced 16x16 image.
- the decoding path has a similar architecture, the only difference being the replacement of the reduction layers 22 by convolutional transposition layers 23, thus doubling the size of the image. In the same way, the number of filters in each convolutional layer after each transposed layer is reduced by half.
- a layer concatenation operator 27 is introduced.
- the layer concatenation operator 27 is also connected to the encoding path, so that it takes at its input a set of matrices, all of them of the same size (except in the concatenation axis, which in the present embodiment is corresponding to the third dimension, that is, the channel) and returns an array resulting from concatenating the previous ones. For example, with two inputs with dimensions [256, 256, 32], the output results from dimensions [256, 256, 64]
- deep supervision layers 26 are included, made up of convolutional filters with a receptive field of 1x1 that combine the output of the second convolutional layer 21 of each of the convolutional blocks of the reduction path by means of a layer sum operator 28
- a layer addition operator 28 takes as input a set of matrices, all of them of the same size and returns to its output a matrix of the same size corresponding to the sum of all of them.
- the sum is made along the third dimension, that is, the dimension corresponding to the channel. For example, with two inputs with dimensions [256, 256, 32], the output results from the same dimensions [256, 256, 32]
- sigmoid activation functions 24 are included.
- Figure 3 represents in detail the architecture of the residual convolutional neural network 12, where the standard convolutional blocks 21 are replaced by residual blocks 30.
- the precision increases as the number of layers increases, however, there is a limit where training accuracy begins to degrade. Thanks to the residual blocks, the training of some layers can be omitted, therefore, it allows to design deeper architectures without incorporating additional layers.
- an operator 27 of layer concatenation Following each of the residual blocks 30, both in the encoding and in the decoding path, an operator 27 of layer concatenation.
- Figure 4 represents in detail the architecture of the dilated convolutional neural network 13, where some dilated convolutional layers 40 are included in the bottleneck section 25 between the encoder and the decoder layers. Included in this embodiment are six dilated convolutional layers 40 with respective dilatation rates of 1, 2, 4, 6, 8, 16, and 32. These dilated layers add multiscale contextual information by increasing the receptive field without losing resolution.
- Figure 5 represents in detail the architecture of the residual-dilated convolutional neural network 14, where the two previous architectures are combined by replacing the standard convolutional blocks 21 with residual blocks 30 and incorporating the dilated convolutional layers 40 in the connection 25 between the routes of encoding and decoding.
- the outputs are averaged to obtain the final probability map. 2.
- This final probability map is the basis for setting the thresholds in the next thresholding stage 3, since it sets the threshold that determines whether a voxel will be part of the final mask 4 or not.
- all voxels with a probability greater than 50% are set as lesions, that is, they are segmented as part of a lesion image mask, while voxels with probabilities less than 50% are set as the background.
- a mask with the white matter lesions is provided at the output, which will be very useful especially for neurologists-radiologists, by concentrating the most relevant parts for their specialist analysis, since the voxels corresponding to brain parenchyma and physiological hyperintensities have been leaked.
- the set of neural networks has to be trained.
- the necessary initial manual annotations are carried out by specialist neuroradiologists.
- This initial segmentation consists of manually selecting voxels corresponding to white matter lesions, excluding physiological hyperintensities, such as the ependymal layer.
- the training process of each CNN architecture is carried out, according to one of the embodiments of the invention, over 200 epochs with a batch size set to 30.
- a cross-validation of 5 iterations is performed. Therefore, the training dataset is divided into 5 subsets, using, in each iteration, 4 of them to train and the remaining subset to validate. The error obtained throughout the 5 iterations is averaged to calculate the final training error.
- Adam's optimization algorithm is used to update the weights of the networks iteratively during the training process.
- the initial learning rate is set to 1e-5.
- the remaining hyperparameters (b1, b2, and x) are kept at their default values (0.9, 0.999, and 10-7 respectively), which are known to give good results.
- WHM white matter hyperintensities
- DCL Loss of DICE coefficient
- DC DICE coefficient
- Binary Cross Entropy (BCE) ⁇ measures how far the predicted probability is from the actual label. It is widely used in classification models. Where y is the true label (0 for the background and 1 for the region of interest) and p (y) is the probability that the voxel is the y label for all N voxels.
- Tversky focal loss (FTL) ⁇ is an extension of the TL to focus on difficult classes detected with less probability, that is, in small regions, since they do not contribute significantly to the loss.
- the network obtained using TL is selected, which is the one used to perform the external validation.
- different on-the-fly data augmentation techniques can be applied to each training batch.
- the entire set formed by the neural networks, probability maps and the thresholding module that generates the final mask at its output can be evaluated to determine its performance and ensure that the precision is as required.
- the following parameters are calculated:
- ASSD Average symmetric surface distance
- This optional evaluation is carried out once the cross validation has been carried out, which guarantees the proper functioning of the architecture designed during the training process, so if the evaluation shows that there is an error, it would be caused by the differences between the set. training and test data.
- the solution is either to expand the data growth by applying more techniques or to increase the training data set to make it more robust.
- the magnetic resonance imaging white matter hyperintensity segmentation method and system of the present invention can be integrated into medical imaging storage and processing infrastructures.
- the method of the present invention can be implemented by a computer, in a processor, microprocessor, server, web server, or cloud.
- the algorithms and models described for implementing the neural networks and other processes of the present invention can be implemented as part of an operating system or a specific application, component, program, object, module, or sequence of instructions referred to as "computer programs.”
- Computer programs typically comprise one or more instructions stored in various memory and storage devices that, when read and executed by one or more processors of a computer, cause the computer to perform necessary operations.
- Examples of computer-readable media include, but are not limited to, writable-type media, such as volatile and non-volatile memory devices, floppy and other removable disks, hard drives, optical discs (for example, compact disc read-only memory (CD ROM), digital versatile discs (DVD), etc.), among others, and transmission media such as digital and analog communication links.
- writable-type media such as volatile and non-volatile memory devices, floppy and other removable disks, hard drives, optical discs (for example, compact disc read-only memory (CD ROM), digital versatile discs (DVD), etc.), among others, and transmission media such as digital and analog communication links.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Physics & Mathematics (AREA)
- Molecular Biology (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Medical Informatics (AREA)
- Radiology & Medical Imaging (AREA)
- Public Health (AREA)
- Epidemiology (AREA)
- Primary Health Care (AREA)
- Heart & Thoracic Surgery (AREA)
- Pathology (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- High Energy & Nuclear Physics (AREA)
- Veterinary Medicine (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
Description
Claims
Priority Applications (5)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
AU2020352676A AU2020352676A1 (en) | 2019-09-23 | 2020-01-30 | Method and system for the automatic segmentation of white matter hyperintensities in brain magnetic resonance images |
US17/762,628 US20220343142A1 (en) | 2019-09-23 | 2020-01-30 | Method and system for the automatic segmentation of white matter hyperintensities in magnetic resonance brain images |
EP20867430.9A EP4020322A4 (en) | 2019-09-23 | 2020-01-30 | METHOD AND SYSTEM FOR AUTOMATIC SEGMENTATION OF WHITE SUBSTANCE HYPERINTENSITIES IN BRAIN MAGNETIC RESONANCE IMAGES |
BR112022005437A BR112022005437A2 (pt) | 2019-09-23 | 2020-01-30 | Método implementado por computador e sistema para segmentação de hipersensibilidades da substância branca presentes nas imagens cerebrais decorrentes da ressonância magnética |
JP2022543800A JP7462055B2 (ja) | 2019-09-23 | 2020-01-30 | 脳の磁気共鳴画像における白質高信号域の自動セグメント化のための方法およびシステム |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
ESP201930818 | 2019-09-23 | ||
ES201930818A ES2813777B2 (es) | 2019-09-23 | 2019-09-23 | Metodo y sistema para la segmentacion automatica de hiperintensidades de sustancia blanca en imagenes de resonancia magnetica cerebral |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021058843A1 true WO2021058843A1 (es) | 2021-04-01 |
Family
ID=75109007
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/ES2020/070069 WO2021058843A1 (es) | 2019-09-23 | 2020-01-30 | Método y sistema para la segmentación automática de hiperintensidades de sustancia blanca en imágenes de resonancia magnética cerebral |
Country Status (7)
Country | Link |
---|---|
US (1) | US20220343142A1 (es) |
EP (1) | EP4020322A4 (es) |
JP (1) | JP7462055B2 (es) |
AU (1) | AU2020352676A1 (es) |
BR (1) | BR112022005437A2 (es) |
ES (1) | ES2813777B2 (es) |
WO (1) | WO2021058843A1 (es) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113096142A (zh) * | 2021-04-30 | 2021-07-09 | 北京理工大学 | 基于联合嵌入空间的白质神经束自动分割方法 |
Families Citing this family (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11495353B2 (en) | 2018-05-10 | 2022-11-08 | Mohamed Anver Jameel | Method, apparatus, and computer readible media for artificial intelligence-based treatment guidance for the neurologically impaired patient who may need neurosurgery |
ES2828728A1 (es) * | 2019-11-27 | 2021-05-27 | Fundacion Para La Investigacion Del Hospital Univ La Fe De La Comunidad Valenciana | Metodo para obtener un biomarcador de imagen que cuantifica la calidad de la estructura trabecular de los huesos |
CN113159147B (zh) * | 2021-04-08 | 2023-09-26 | 平安科技(深圳)有限公司 | 基于神经网络的图像识别方法、装置、电子设备 |
CN114419066B (zh) * | 2022-01-14 | 2022-12-13 | 深圳市铱硙医疗科技有限公司 | 脑白质高信号分割方法、装置、设备及存储介质 |
CN115514343B (zh) * | 2022-05-13 | 2023-08-11 | 浙江腾腾电气有限公司 | 电网波形滤波系统及其滤波方法 |
CN115115628B (zh) * | 2022-08-29 | 2022-11-22 | 山东第一医科大学附属省立医院(山东省立医院) | 一种基于三维精细化残差网络的腔隙性脑梗死识别系统 |
CN117556715B (zh) * | 2024-01-12 | 2024-03-26 | 湖南大学 | 基于信息融合的典型环境下智能电表退化分析方法及系统 |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108171711A (zh) * | 2018-01-17 | 2018-06-15 | 深圳市唯特视科技有限公司 | 一种基于完全卷积网络的婴幼儿脑部磁共振图像分割方法 |
WO2018140596A2 (en) * | 2017-01-27 | 2018-08-02 | Arterys Inc. | Automated segmentation utilizing fully convolutional networks |
US20180260951A1 (en) * | 2017-03-08 | 2018-09-13 | Siemens Healthcare Gmbh | Deep Image-to-Image Recurrent Network with Shape Basis for Automatic Vertebra Labeling in Large-Scale 3D CT Volumes |
WO2018229490A1 (en) * | 2017-06-16 | 2018-12-20 | Ucl Business Plc | A system and computer-implemented method for segmenting an image |
CN109410167A (zh) * | 2018-08-31 | 2019-03-01 | 深圳大学 | 一种3d乳腺图像的分析方法及相关产品 |
CN109872328A (zh) * | 2019-01-25 | 2019-06-11 | 腾讯科技(深圳)有限公司 | 一种脑部图像分割方法、装置和存储介质 |
WO2019109410A1 (zh) * | 2017-12-06 | 2019-06-13 | 深圳博脑医疗科技有限公司 | 用于分割 mri 图像中异常信号区的全卷积网络模型训练方法 |
CN109886273A (zh) * | 2019-02-26 | 2019-06-14 | 四川大学华西医院 | 一种cmr图像分割分类系统 |
CN109993809A (zh) * | 2019-03-18 | 2019-07-09 | 杭州电子科技大学 | 基于残差U-net卷积神经网络的快速磁共振成像方法 |
CN109993735A (zh) * | 2019-03-29 | 2019-07-09 | 成都信息工程大学 | 基于级联卷积的图像分割方法 |
CN110189334A (zh) * | 2019-05-28 | 2019-08-30 | 南京邮电大学 | 基于注意力机制的残差型全卷积神经网络的医学图像分割方法 |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10210246B2 (en) * | 2014-09-26 | 2019-02-19 | Oracle International Corporation | Techniques for similarity analysis and data enrichment using knowledge sources |
US11357398B2 (en) * | 2017-01-31 | 2022-06-14 | Nidek Co., Ltd. | Image processing device and non-transitory computer-readable recording medium |
US10223610B1 (en) * | 2017-10-15 | 2019-03-05 | International Business Machines Corporation | System and method for detection and classification of findings in images |
-
2019
- 2019-09-23 ES ES201930818A patent/ES2813777B2/es active Active
-
2020
- 2020-01-30 EP EP20867430.9A patent/EP4020322A4/en active Pending
- 2020-01-30 WO PCT/ES2020/070069 patent/WO2021058843A1/es unknown
- 2020-01-30 AU AU2020352676A patent/AU2020352676A1/en active Pending
- 2020-01-30 BR BR112022005437A patent/BR112022005437A2/pt unknown
- 2020-01-30 JP JP2022543800A patent/JP7462055B2/ja active Active
- 2020-01-30 US US17/762,628 patent/US20220343142A1/en active Pending
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018140596A2 (en) * | 2017-01-27 | 2018-08-02 | Arterys Inc. | Automated segmentation utilizing fully convolutional networks |
US20180260951A1 (en) * | 2017-03-08 | 2018-09-13 | Siemens Healthcare Gmbh | Deep Image-to-Image Recurrent Network with Shape Basis for Automatic Vertebra Labeling in Large-Scale 3D CT Volumes |
WO2018229490A1 (en) * | 2017-06-16 | 2018-12-20 | Ucl Business Plc | A system and computer-implemented method for segmenting an image |
WO2019109410A1 (zh) * | 2017-12-06 | 2019-06-13 | 深圳博脑医疗科技有限公司 | 用于分割 mri 图像中异常信号区的全卷积网络模型训练方法 |
CN108171711A (zh) * | 2018-01-17 | 2018-06-15 | 深圳市唯特视科技有限公司 | 一种基于完全卷积网络的婴幼儿脑部磁共振图像分割方法 |
CN109410167A (zh) * | 2018-08-31 | 2019-03-01 | 深圳大学 | 一种3d乳腺图像的分析方法及相关产品 |
CN109872328A (zh) * | 2019-01-25 | 2019-06-11 | 腾讯科技(深圳)有限公司 | 一种脑部图像分割方法、装置和存储介质 |
CN109886273A (zh) * | 2019-02-26 | 2019-06-14 | 四川大学华西医院 | 一种cmr图像分割分类系统 |
CN109993809A (zh) * | 2019-03-18 | 2019-07-09 | 杭州电子科技大学 | 基于残差U-net卷积神经网络的快速磁共振成像方法 |
CN109993735A (zh) * | 2019-03-29 | 2019-07-09 | 成都信息工程大学 | 基于级联卷积的图像分割方法 |
CN110189334A (zh) * | 2019-05-28 | 2019-08-30 | 南京邮电大学 | 基于注意力机制的残差型全卷积神经网络的医学图像分割方法 |
Non-Patent Citations (4)
Title |
---|
BERNAL JOSE ET AL., DEEP CONVOLUTIONAL NEURAL NETWORKS FOR BRAIN IMAGE ANALYSIS ON MAGNETIC RESONANCE IMAGING: A REVIEW, XP085638002 * |
KARIMI DAVOOD ET AL.: "Reducing the Hausdorff Distance in Medical Image Segmentation With Convolutional Neural Networks", IEEE TRANSACTIONS ON MEDICAL IMAGING, vol. 39, no. 2, 19 July 2019 (2019-07-19), Piscataway, Nj, Us, pages 499 - 513, XP011769481, ISSN: 0278-0062, DOI: 10.1109/TMI.2019.2930068 * |
See also references of EP4020322A4 * |
XU BOTIAN ET AL.: "Orchestral fully convolutional networks for small lesion segmentation in brain MRI", IEEE 15TH INTERNATIONAL SYMPOSIUM ON BIOMEDICAL IMAGING (ISBI 2018, 4 April 2018 (2018-04-04), pages 889 - 892, XP033348292, DOI: 10.1109/ISBI.2018.8363714 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113096142A (zh) * | 2021-04-30 | 2021-07-09 | 北京理工大学 | 基于联合嵌入空间的白质神经束自动分割方法 |
Also Published As
Publication number | Publication date |
---|---|
EP4020322A1 (en) | 2022-06-29 |
US20220343142A1 (en) | 2022-10-27 |
ES2813777B2 (es) | 2023-10-27 |
EP4020322A4 (en) | 2023-10-25 |
JP2023514964A (ja) | 2023-04-12 |
AU2020352676A1 (en) | 2022-04-21 |
BR112022005437A2 (pt) | 2022-06-21 |
ES2813777A1 (es) | 2021-03-24 |
JP7462055B2 (ja) | 2024-04-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
ES2813777B2 (es) | Metodo y sistema para la segmentacion automatica de hiperintensidades de sustancia blanca en imagenes de resonancia magnetica cerebral | |
ES2701095T3 (es) | Un sistema y método para anotar imágenes propagando información | |
Milošević et al. | Automated estimation of chronological age from panoramic dental X-ray images using deep learning | |
ES2595210T3 (es) | Método y aparato para procesar imágenes médicas | |
Klein et al. | Early diagnosis of dementia based on intersubject whole-brain dissimilarities | |
CN111932492B (zh) | 一种医学图像处理方法、装置及计算机可读存储介质 | |
CN113096137B (zh) | 一种oct视网膜图像领域适应分割方法及系统 | |
Ellingsen et al. | Segmentation and labeling of the ventricular system in normal pressure hydrocephalus using patch-based tissue classification and multi-atlas labeling | |
KR102043829B1 (ko) | 병변 발생 시점 추정 방법, 장치 및 프로그램 | |
Li et al. | BrainK for structural image processing: creating electrical models of the human head | |
Dai et al. | Bagging ensembles for the diagnosis and prognostication of Alzheimer's disease | |
JP6785976B2 (ja) | 脳画像正規化装置、方法およびプログラム | |
Ngo et al. | Joint-embeddings reveal functional differences in default-mode network architecture between marmosets and humans | |
Sampath et al. | Classification of Alzheimer disease stages exploiting an ANFIS classifier | |
Jain et al. | A review of deep learning-based disease detection in Alzheimer's patients | |
Shetty et al. | Detection and Prediction of Alzheimer's disease using Deep learning: A review | |
KR20220133834A (ko) | 학습 모델을 이용한 데이터 처리 방법 | |
Fujibayashi et al. | Nonlinear registration as an effective preprocessing technique for Deep learning based classification of disease | |
Baydargil et al. | A parallel deep convolutional neural network for Alzheimer's disease classification on PET/CT brain images | |
Khatami et al. | Parcellation-Free prediction of task fMRI activations from dMRI tractography | |
Rachmadi et al. | Evaluation of four supervised learning schemes in white matter hyperintensities segmentation in absence or mild presence of vascular pathology | |
Dey | A subpixel residual U-Net and feature fusion preprocessing for retinal vessel segmentation | |
Mukhopadhyay | Task-Agnostic Continual Hippocampus Segmentation for Smooth Population Shifts | |
Jiménez Mesa | Enhancing diagnostic accuracy in neuroimaging through machine learning: advancements in statistical classification and mapping | |
Yee | 3D convolutional neural networks for Alzheimer’s disease classification |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 20867430 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2022543800 Country of ref document: JP Kind code of ref document: A |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 2020867430 Country of ref document: EP Effective date: 20220322 |
|
REG | Reference to national code |
Ref country code: BR Ref legal event code: B01A Ref document number: 112022005437 Country of ref document: BR |
|
ENP | Entry into the national phase |
Ref document number: 2020352676 Country of ref document: AU Date of ref document: 20200130 Kind code of ref document: A |
|
ENP | Entry into the national phase |
Ref document number: 112022005437 Country of ref document: BR Kind code of ref document: A2 Effective date: 20220323 |