The traditional level set algorithm selects the position of the initial contour randomly and lacks the processing of edge information. Therefore, it cannot accurately extract the edge of the brain tissue. In order to solve this problem, this paper proposes a level set algorithm that fuses partition and Canny function. Firstly, the idea of partition is fused, and the initial contour position is selected by combining the morphological information of each region, so that the initial contour contains more brain tissue regions, and the efficiency of brain tissue extraction is improved. Secondly, the canny operator is fused in the energy functional, which improves the accuracy of edge detection of rhesus monkey brain tissue while retaining the advantage of the traditional level set algorithm in processing an uneven gray image. Experimental results show that the algorithm can accurately extract the brain tissue of rhesus monkeys with an accuracy of up to 86%.

Most brain research is based on MRI images [

At present, brain extraction methods are divided into three categories: region-based method, graph-based method, and deep learning-based method. The method based on the region [

To sum up, facing the existing problems in the current research results, this paper uses the level set method [

Therefore, the paper, with a view to the study of the extraction of the monkey brain, gives the Level set of fused Regional morphology and the Level set of the Canny functional function (RMCA-Level Set). Based on the level set algorithm, the construction method of its initial contour and energy function is improved. The main contents are as follows:

To solve the problem of location selection of the initial contour, the idea of partition is introduced and the contour building model combining regional morphological information is proposed. The initial contour construction is completed by combining the morphological information of each region, which can speed up the iteration speed of the initial contour and improve the accuracy of monkey brain tissue extraction.

To solve the problem that the effect of edge segmentation of brain tissue is not ideal, a level set algorithm integrating Canny energy functional is proposed. In LBF energy functional, Canny edge detection operator is integrated to enhance the segmentation of edge regions, which not only extends the advantages of traditional level set algorithm in processing gray-scale inhomogeneous images, but also strengthens the detection of target edge, so as to achieve the purpose of accurately extracting the brain tissue of macaque monkeys.

The level set algorithm is a method to solve the curve evolution. The idea is that the low-dimensional curve is taken as the zero level set of the high-dimensional surface, and the initial contour is segmenting through the evolution iteration of the curve, that is, the initial curve where the level set function starts to iterate. The more target regions the initial contour contains, the shorter the iteration time is, and the more accurate the result is [

As shown in

Image binarization is the process of changing the gray value of pixel point to 0 or 255 according to the set threshold. The threshold

where, ω1 = number of background pixels/total pixels, ω2 = number of foreground pixels/total pixels.

The connected area of the image was counted, and the small non-brain tissue area was deleted by threshold setting. To compare the area of the small connected area of the rhesus monkey image and accurately select the threshold,

At the same time, the area of the connected area of the layers 230, 260, 280, 290, 310, 320, 330, 330, 340, 350, 360, 370, 380, 390 and 410 in no.32126 and no.32127 rhesus macaque MRI images were calculated. In

To make the initial contour contain more target areas and improve the accuracy of monkey brain extraction, the image needs to be partitioned. Because the four local areas can maximize the coverage of the target area with the least amount of calculation, and the more target regions are included, the faster and more accurate the curve evolves. For any image A, divide it into four rectangular local areas as the seed areas for constructing the initial contour. The target image

The 300th layer image of rhesus monkey no. 32127 was divided into four local areas, and the centroid points of each connected region in each local area were calculated respectively. The results were shown in

According to the morphological characteristics of the images, the monkey brain focusses on the image center and around the brain tissue adhesion. Therefore, the Euclidean distance between each centroid point and the partition point is calculated based on the centroid points of each connected region. In each local area, the centroid point with the smallest Euclidean distance is selected as the seed point to construct the initial contour, and the construction of the initial contour is completed [

To improve the efficiency of monkey brain edge detection, the edge detection operator was fused in the traditional level set LBF energy functional [

The horizontal set function shall conform to the characteristics of the symbolic distance function, which is defined as follows:

Among,

Introduce the level set function into the energy functional, and define the following LBF energy function:

Among,

When canny detects the edge, the noise is first removed by the Gaussian filter. The generation equation of the Gaussian filter with the size of:

After the filtering operation is completed, the convolution operation is carried out, that is, all pixels in the image should be weighted and summed with their surrounding pixels. The gradient amplitude and direction of all pixels in the image are obtained according to the following formula.

where,

where,

Also, the length term constraint function

In the evolution process, in order to prevent the horizontal set function from losing the characteristic of sign distance function and causing instability in the evolution process, the distance penalty term function was added in the energy functional to shorten the reinitialization process of the level set, so that the evolution process could be carried out smoothly.

Therefore, the total energy functional of the model in this paper is:

where,

The standard gradient descent flow is used to minimize the energy functional so that the

The gradient descent flow equation obtained by keeping

where,

The process of solving the minimum value of the horizontal set energy functional is the process of solving the minimum value of the partial differential equation. This process meets the requirements of using the finite-difference method. The image space belongs to the equipartition grid, so the finite-difference method is generally used for calculation. It should be noted that in the case of fixed space step h, the time step

where

The evolution equation of the level set function

where,

UC Davis Dataset. The data set collected data from 19 rhesus monkeys using the Siemens Skyra3T scanner. The data included structures T1 and T2, as well as task-state fMRI and dMRI. In this experiment, NMR data with structures T1 were used. All the 19 monkeys were females, ranging in age from 18.5 to 22.5 years. The weight distribution was 7.28-14.95kg. Scanning sequence parameters: voxel resolution = 0.3 × 0.3 × 0.3 mm, TE = 6.93 ms, TR = 15 ms, TI = 1100 ms, Flip Angle = 8°.

Mountsinai-P Dataset. Data of 9 macaques were collected by the Philips 3T scanner. The data included structures T1, T2, and dMRI. In this experiment, NMR data with structures T1 were used. The data set included 8 males and 1 female, with an age distribution of 3.4-8 years and a weight distribution of 4.7-7.42kg. The scanning sequence parameters were as follows: voxel resolution = 0.5 × 0.5 × 0.5 mm, TE = 6.93 ms, TR = 15 ms, TI = 1100 ms, Flip Angle = 8°.

Princeton Dataset. The data set used Simens Prisma VE11C 3T scanner to collect data from two rhesus monkeys. The data included structures T1, T2, dMRI, and task-state fMRI. In this experiment, NMR data with structures T1 were used. All the two macaques in the data set were male, with an age distribution of 3 years and a weight distribution of 4.7-5.5kg. Scanning sequence parameters were as follows: voxel resolution = 0.5 × 0.5 × 0.5 mm, TE = 2.32 ms, TR = 2700 ms, TI = 850 ms, Flip Angle = 9°.

Uminn Dataset. The data set was used to collect the data of 2 rhesus monkeys with Simens 7T scanner. The data included structures T1, T2, dMRI, and task-state fMRI. In this experiment, NMR data with structures T1 were used. All the two macaques in the data set were female, and their age distribution was over 10 years old. The scanning sequence parameters were as follows: voxel resolution = 0.3 × 0.3 × 0.3 mm, TE = 3.65 ms, TR = 2500 ms, TI = 1100 ms, Flip Angle = 7°.

Experiment one: The ce-LevelSet of the model in this paper was applied to UC Davis, Mountsinai-P, Princeton and Uminn, respectively, to verify the feasibility of the model in this paper. The default parameters of the model in this paper are: time step

Dataset | DSC mean value | JS mean value |
---|---|---|

UC Davis | 0.745 | 0.752 |

Mountsinai-P | 0.734 | 0.745 |

Princeton | 0.768 | 0.758 |

Uminn | 0.753 | 0.763 |

Experiment two: BET algorithm [

As can be seen from

BET, watershed model, shown in

In order to more accurately compare the differences between the extraction results of different models, DSC and Jaccard similarity coefficient are used to carry out quantitative analysis on the experimental results. The DSC and Jaccard similarity coefficients range from 0 to 1 The closer the value is to 1, the higher the similarity is and the more accurate the segmentation result of the model is. On the contrary, the closer the similarity coefficient value is to 0, the lower the similarity is, and the worse the segmentation effect of the model is.

DSC | JS | |||
---|---|---|---|---|

Mean value | Standard deviation | Mean value | Standard deviation | |

BET | 0.651 | 0.192 | 0.659 | 0.225 |

Watershed algorithm | 0.402 | 0.233 | 0.413 | 0.233 |

LBF model | 0.645 | 0.119 | 0.579 | 0.107 |

Unet | 0.689 | 0.103 | 0.648 | 0.093 |

Model of the paper |

Meanwhile, in this paper, the brain tissue of Monkey No. 32125 to 32136 was extracted using the above five models, and the DSC value and JS value of the extracted results were shown in

As can be seen from

The automatic extraction of the macaque brain is the primary problem encountered in the research of the macaque brain. Therefore, this paper puts forward the RMCA-Level Set algorithm. First, the idea of partition is fused, and the initial contour is constructed by combining the mentality information of each region. The initial contour is constructed by binarization, deletion of small connected regions, and calculation of the center of mass by regions, etc., so as to improve the efficiency of brain tissue extraction. Secondly, to solve the problem of low edge extraction accuracy, a canny operator was fused into LBF energy functional to improve the extraction accuracy of the edge of brain tissue in this paper. Numerous experiments show that this model can be used to extract the brain tissue of the rhesus monkey.

Thank you for the National Natural Science Foundation of China, Natural Science Foundation of Shanxi Province, Key Research and Development Projects of Shanxi Province.