CN110870770A

CN110870770A - ICA-CNN classified fMRI space activation map smoothing and broadening method

Info

Publication number: CN110870770A
Application number: CN201911144802.9A
Authority: CN
Inventors: 林秋华; 牛妍炜
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2019-11-21
Filing date: 2019-11-21
Publication date: 2020-03-10
Anticipated expiration: 2039-11-21
Also published as: CN110870770B

Abstract

An ICA-CNN classified fMRI space activation map smoothing and broadening method belongs to the field of biomedical signal processing. Firstly, three-dimensional space smoothing is applied to an fMRI space activation map obtained by ICA separation, a new sample set is generated by amplification, and then the new sample set is sent to CNN for classification, so that the improvement of classification performance of patients and healthy people is realized. The invention is adopted to classify 82 tested resting state fMRI data into patients and healthy people, three Gaussian filters with different FWHM are adopted to separate ICA to obtain a space activation map for space smoothing, three groups of sample sets are generated, and then the three groups of sample sets are sent to CNN for classification. Compared with the existing multi-model order data augmentation method, the method can improve the classification accuracy by 5.76%; by combining the two methods, the classification accuracy can be improved by 21.33%. Therefore, the invention can independently improve the network classification performance, is easy to combine with other augmentation methods, and obviously improves the classification accuracy.

Description

ICA-CNN classified fMRI space activation map smoothing and broadening method

Technical Field

The invention belongs to the field of biomedical signal processing, and particularly relates to a smoothing and amplifying method for an ICA-CNN classified fMRI space activation map.

Background

Convolutional Neural Networks (CNNs) have the advantages of local sensing, weight sharing, and the like, are prominent in tasks such as recognition, detection, classification, and the like, and are also useful in smart medicine. The resting-state fMRI (rs-fMRI) data has the advantages of non-invasiveness, high spatial resolution, easy acquisition on the test of a patient and the like, and is commonly used for analyzing and diagnosing neurological disorders such as schizophrenia. Therefore, CNNs with rs-fMRI data as training data would represent a unique advantage in the task of classifying healthy persons from patients.

In view of the problem that fMRI of patients is difficult to acquire and the data size is not large, linkawa et al propose an ICA-CNN classification framework for complex fMRI data of patients and healthy persons (patent application No. 201910350137.2). In this framework, two-dimensional slices of spatial components of interest are first separated using Independent Component Analysis (ICA), and then fed into two-dimensional CNN to extract features and classify. Compared with a method of directly utilizing observation data or using three-dimensional CNN classification, the framework reduces the requirement on fMRI data volume, and achieves better classification performance under the condition of limited fMRI data volume.

In fact, in the case where the fMRI data amount is constant, the data expansion can serve to further increase the data amount and improve the ICA-CNN classification effect. In the field of image classification, widely used data augmentation methods include rotation, flipping, noise addition, scaling, texture color transformation, and the like. However, these methods are not suitable for rs-fMRI analysis where the spatial structure cannot be varied arbitrarily. The patent application number 201910350137.2 only provides a data augmentation method with multiple model orders, and has important practical value in exploring other methods suitable for fMRI data augmentation under the ICA-CNN classification framework.

Disclosure of Invention

The invention provides an fMRI space activation map augmentation method embedded into an ICA-CNN classification frame, which can effectively expand fMRI sample size and further improve the classification performance of an ICA-CNN network. The specific scheme is that three-dimensional space smoothing is applied to an fMRI space activation map obtained by ICA separation, a new sample set is generated by amplification, and then the new sample set is sent to CNN for classification, so that the improvement of classification performance of patients and healthy people is realized, as shown in figure 1. The method comprises the following concrete steps:

the first step is as follows: inputting the four-dimensional fMRI observation amplitude data of the tested k

K is 1, …, K is the total number of trials, T is the number of scan time points, X, Y, Z is the three dimensions of brain space.

The second step is that: mixing X^(k)K is 1, …, K, the space dimension is developed into one dimension, namely the size of the space dimension is equal to X multiplied by Y multiplied by Z, then the outside-brain voxels are removed, and only the inside-brain voxels are taken, and the space dimension is obtained

V is the number of endosomes, V<X×Y×Z。

The third step: principal Component Analysis (PCA) (principal Component analysis) pair

K is 1, …, K, and reducing vitamin to obtain

N is the model order and is less than or equal to T.

The fourth step: ICA isolation and extraction of the component of interest. Using Infmax algorithm pairs

And (5) carrying out ICA separation when K is 1, … and K; selecting a spatial activation map of the interested component from N ICA separation components based on the principle of maximum correlation coefficient with a spatial component template

The spatial composition template is selected from the literature (S.M.Smith, P.T.Fox et al, "ceramic of the woven's functional architecture along with actuation and" Proceedings of the National Academy of Sciences of the United States of America, vol.106, No.31Pp.13040-13045,2009).

The fifth step: to pair

K is 1, …, K, and carries out brain extrasomatic zero filling to restore to a three-dimensional space activation map, namely

And a sixth step: the space is smooth. fMRI space activation map of test k

K1, …, K, convolved with a three-dimensional gaussian filter to achieve spatial smoothing, as shown by:

in the formula

Representing a three-dimensional convolution calculation, g (x, y, z) is a three-dimensional zero-mean gaussian function defined as:

where σ is the standard deviation, the width of the filter is determined, and the larger σ is, the larger the smoothing degree is. Full width at Half maximum fwhm (fullbidth at Half maximum) is a common parameter for measuring the size of a gaussian kernel, having a linear relationship with sigma as follows,

the FWHM is typically taken to be 2-3 times the voxel size. Smoothing all tested spatial activation maps with J gaussian filters with different FWHM sizes can expand each tested sample set from the original 1 group to J groups. For test k, the J sets of augmented sample sets were recorded as

As shown in fig. 2. The spatial activation map under each smoothing kernel is then further processed as follows:

wherein λ is₁、λ₂As weight parameters, they satisfy λ₁+λ₂＝1，0<λ₁<1，0<λ₂<1, and usually takes λ₁>λ₂。

The seventh step: will be provided with

K1, …, K, i 1, …, J, unfolded along the Z-axis as a stack of Z two-dimensional slices of size X Y, denoted as

And selecting W slices with large activation information according to the effective activation positions of the interested components, wherein W is less than or equal to Z and is marked as L, … and L + W-1.

Eighth step: cutting W pieces of two-dimensional slices

Voxel values of K1, …, K, i 1, …, J, are normalized to [0,1]Is marked as

Then attaching a corresponding label to each slice

It is indicative of a healthy person,

representing patients with schizophrenia, establishing an augmented CNN sample set

The sample set of each test sample is expanded from original W to W × J, i.e., the expansion is J times of original. All KThe sample set of each test was expanded to kxwxj.

The ninth step: disorganizing the sequence of the tested sample, and proportionally dividing the sample set into a training set, a verification set and a test set; such as: 60% of the sample set is taken as a training set, 20% is taken as a verification set, and 20% is taken as a test set.

The tenth step: and sending the training set into a CNN network for training. The CNN network structure used in the present invention is consistent with patent application No. 201910350137.2, and as shown in fig. 3, includes 2 convolutional layers, 2 max pooling layers, 1 full connection layer, and an output layer. Wherein, the size of the convolution layer core is 3 multiplied by 3, and the number of the convolution layer core is respectively 8 and 16; the size of the pooling layer core is 2 x 2, and the number of the pooling layer cores is 8 and 16 respectively; the full connection layer has 64 nodes, modified linear unit (ReLU) is used as an activation function, and the output layer uses Softmax to give the category to which each slice belongs. The batch processing size is 64, parameter updating is carried out by adopting an Adam algorithm, a two-class cross entropy function is used as a loss function of the network, and meanwhile, L is used₂And (5) regularizing control weight and training a CNN model to have R rounds. At each round, calculating an accuracy ACC using the validation set; and saving the CNN model with the highest verification accuracy in the R round as the CNN model obtained by training. Let CNN verify as

Accuracy ACC is defined as follows:

wherein TP represents true positive, i.e.

TN indicates true negatives, i.e.

FP represents a false positive, i.e

FN indicates false negatives, i.e.

The eleventh step: and inputting the test set into the CNN model obtained by the eleventh training step to obtain the test accuracy ACC.

The twelfth step: and 5-fold cross validation is adopted, namely the ninth step to the tenth step are repeated for 5 times to obtain the test accuracy ACC of 5 times, and the average result is calculated and recorded as the final classification accuracy of the CNN.

The invention has the beneficial effects that: the invention focuses on the classification problem in the field of computer-aided diagnosis, and provides a data augmentation method aiming at fMRI data characteristics and embedded in an ICA-CNN classification framework, which is used for training a CNN network and improving classification results. For example, the invention is adopted to classify 82 rs-fMRI data to be tested into schizophrenic patients and healthy people, ICA separation is carried out on fMRI observation data through Infmax to obtain a spatial activation map, the spatial activation map of a Default Mode Network (DMN) is extracted, and the FWHM is respectively 6mm³、8mm³、10mm³The three Gaussian filters carry out three-dimensional spatial smoothing on all tested DMN spatial activation graphs to generate three groups of sample sets, and finally, the three groups of sample sets are sent to the CNN for training and testing. Compared with the multi-model order data amplification method provided by the patent 201910350137.2, the method can improve the classification accuracy by 15.76% under the condition of the same sample number. If the method is combined with a multi-model order augmentation method, three model orders of N-20, 60 and 100 are set for fMRI observation data, DMN space activation graphs are respectively extracted, space smoothing is applied to the DMN space activation graphs obtained by each model order to generate three groups of sample sets, and then all the three groups of sample sets are sent to a CNN network for classification. Compared with the method only using the multiple model order augmentation, the combination method can improve the classification accuracy by 21.33%. Therefore, the invention not only can independently improve the network classification performance, but also can be easily combined with other augmentation methods, thereby obviously improving the classification accuracy.

Drawings

FIG. 1 illustrates the location and usage steps of the present invention in the ICA-CNN classification framework.

FIG. 2 is a schematic representation of the present invention for spatially smooth augmentation of fMRI spatial activation maps.

Fig. 3 is a CNN network architecture used in the present invention.

Detailed Description

An embodiment of the present invention will be described in detail below with reference to the accompanying drawings. There were 82 rs-fMRI magnitude data tested, including 42 schizophrenic patients and 40 healthy people. Each test contained 146 scans, each scan sharing whole-brain data of voxels X × Y × Z53 × 63 × 46 153594, with an intra-brain voxel V62336 and voxel size of 3 × 3 × 3mm³。

The first step is as follows: inputting all 82 tested four-dimensional fMRI observation amplitude data

k＝1,…,82。

The second step is that: mixing X^(k)K is 1, …,82, the spatial dimension is one-dimensional, i.e. the size of the spatial dimension is 53 × 63 × 46 or 153594, then the extra-brain voxels are removed and only the intra-brain voxels are taken, and the method is obtained

The third step: using PCA pairs

k is 1, … and 82, dimension reduction is carried out, the model order N is 50, and the obtained product is obtained

And (4) carrying out ICA separation when k is 1, … and 82; based on the literature (S.M. Smith, P.T. Fox et al, "Correstondence of the blue's functional architecture during activation and rest," proceedings of the National Academy of Sciences of the United States of America, vol.106, No.31, pp.13040-13045,2009), selecting DMN space activation map from 50 ICA separation components

The fifth step: to pair

k is 1, …,82, and carries out brain extrasomatic zero filling to restore to three-dimensional space activation map, namely

And a sixth step: the space is smooth. fMRI space activation map of test k

k is 1, …,82, and is convolved with a three-dimensional gaussian filter to achieve spatial smoothing, as shown in equation (1). According to the definition of the formula (2), 3 three-dimensional Gaussian filters are generated, and the FWHM is 6mm³、8mm³、10mm³. Smoothing all tested DMN spatial activation maps by using the 3 Gaussian filters, and expanding each tested sample set from 1 group to 3 groups. For test k, 3 sets of the augmented sample set were recorded as

k is 1, …,82, i is 1,2, 3. Then, the space activation map under each smooth kernel is further processed by the formula (3) to obtain

k is 1, …,82, i is 1,2,3, and λ is taken₁＝0.8，λ₂＝0.2。

The seventh step: will be provided with

k 1, …,82,

i

1,2,3, unfolded along the z-axis into a stack of 46 two-dimensional slices of size 53 × 63, noted

According to the effective activation position of the interested component, 25 slices with large activation information quantity are selected, specifically, L is 12, z is 12, … and 36.

Eighth step: 25 two-dimensional slices

Voxel values with k 1, …,82,

i

1,2,3 are normalized to [0,1]Is marked as

Then attaching a corresponding label to each slice

It is indicative of a healthy person,

Each sample set tested was expanded from W25 to W × J25 × 3 to 75, i.e., 3 times the original sample set. The number of samples of all 82 samples tested was increased to 82 × W × J, 6150.

The ninth step: the sequence of the tested samples is disturbed, 60% of sample sets are taken as training sets, 20% are taken as verification sets, and 20% are taken as testing sets.

The tenth step: and (5) sending the training set into a CNN network for training, and training the CNN model R for 50 rounds. At each round, calculating the accuracy rate ACC by using the verification set according to the formula (3); and saving the CNN model with the highest verification accuracy in 50 rounds.

The eleventh step: and (4) inputting the test set into the CNN model obtained by the eleventh training step, and calculating according to the formula (3) to obtain the test accuracy ACC.

The twelfth step: and 5-fold cross validation is adopted, namely the tenth step to the twelfth step are repeated for 5 times to obtain the test accuracy ACC for 5 times, and the average result is calculated to obtain the final classification accuracy of 0.8522.

Claims

1. A kind of ICA-CNN classified fMRI space activation map smooth augmentation method, apply three-dimensional space smoothing to fMRI space activation map that ICA separates, augment and produce the new sample set, then send into CNN to classify, realize the improvement to patient and classification performance of the healthy person; the method is characterized by comprising the following steps:

K is 1, …, K is the total number of trials, T is the number of scan time points, X, Y, Z is the three dimensions of brain space;

V is the number of endosomes, V<X×Y×Z；

The third step: PCA pairs Using principal component analysis

K is 1, …, K, and reducing vitamin to obtain

N is the model order, and N is less than or equal to T;

the fourth step: ICA separation and interesting component extraction, and Infmax algorithm pair

K is 1, …, K, ICA separation is carried out, and a space activation map of the interested component is selected from N ICA separation components based on the principle that the correlation coefficient with a space component template is maximum

The fifth step: to pair

And a sixth step: spatial smoothing, fMRI spatial activation map of the examined k

in the formula

wherein, σ is standard deviation, determines the width of the filter, the larger σ is, the larger smoothing degree is, the full width at half maximum FWHM is a common parameter for measuring the size of Gaussian kernel, and has a linear relation with σ as follows,

smoothing all tested space activation maps by using J Gaussian filters with different FWHM sizes, and increasing each tested sample set from 1 group to J groups, and recording J groups of increased sample sets as tested k

The spatial activation map under each smoothing kernel is then further processed as follows:

in the formula of₁、λ₂Is a weight parameter;

the seventh step: will be provided with

Selecting W slices with large activation information according to the effective activation positions of the interested components, wherein W is less than or equal to Z and is recorded as L, … and L + W-1;

eighth step: cutting W pieces of two-dimensional slices

Voxel values of K1, …, K, i 1, …, J, are normalized to [0,1]Is marked as

Then attaching a corresponding label to each slice

It is indicative of a healthy person,

Each sample testedThe sample set is expanded from the original W to W multiplied by J, namely the expansion is multiplied by J times, and all K tested sample sets are expanded to K multiplied by W multiplied by J;

the ninth step: disorganizing the sequence of the tested sample, and proportionally dividing the sample set into a training set, a verification set and a test set;

the tenth step: sending a training set into a CNN network for training, wherein the CNN network structure comprises 2 convolutional layers, 2 maximum pooling layers, 1 full-connection layer and an output layer, the size of each convolutional layer core is 3 multiplied by 3, the number of each convolutional layer core is 8 and 16, the size of each pooling layer core is 2 multiplied by 2, the number of each pooling layer core is 8 and 16, 64 nodes are arranged in the full-connection layer, a modified linear unit ReLU is used as an activation function, the output layer gives the category to which each slice belongs by using Softmax, the batch processing size is 64, parameter updating is performed by using an Adam algorithm, a two-classification cross entropy function is used as a loss function of the network, and meanwhile, L₂Regularizing control weight, training CNN models in R rounds, calculating ACC by using a verification set in each round, storing the CNN model with the highest verification accuracy in the R rounds as the CNN model obtained by training, and making the CNN verification result be

Accuracy ACC is defined as follows:

wherein TP represents true positive, i.e.

TN indicates true negatives, i.e.

FP represents a false positive, i.e

FN indicates false negatives, i.e.

The eleventh step: inputting the test set into the CNN model obtained by the eleventh training step to obtain a test accuracy ACC;

2. The ICA-CNN classified fMRI spatial activation map smooth augmentation method according to claim 1, wherein FWHM is 2-3 times of fMRI voxel size, λ₁、λ₂Satisfy lambda₁+λ₂＝1，0<λ₁<1，0<λ₂<1, and λ₁>λ₂。

3. The ICA-CNN classified fMRI spatial activation map smoothing and augmenting method according to claim 1 or 2, wherein in the ninth step, 60% of the sample set is used as a training set, 20% is used as a verification set, and 20% is used as a test set.