An intracranial aneurysm wall inflammation level prediction system based on a multi-scale deformable convolutional network

Abstract

This invention belongs to the field of biomedical engineering technology, specifically a system for predicting the inflammation level of the intracranial aneurysm wall based on a multi-scale deformable convolutional network. The system includes a data acquisition module and an inflammation level prediction module. The data acquisition module acquires high-resolution magnetic resonance imaging (MRI) data of the intracranial aneurysm wall and inflammation level labels at different locations. The inflammation level prediction module trains a model based on the high-resolution MRI data to predict the inflammation level of the aneurysm wall. The prediction model employs a multi-scale deformable convolutional U-shaped network, specifically a four-level encoder-decoder structure, including an encoder, a bottleneck layer, a decoder, and an output head. This invention achieves non-invasive prediction of the inflammation level of the intracranial aneurysm wall from MRI images, providing a better technical means for risk assessment of intracranial aneurysms.

Description

Technical Field

[0001] This invention belongs to the field of biomedical engineering technology, specifically relating to a system for predicting the level of inflammation in the intracranial aneurysm wall from high-resolution magnetic resonance vascular wall imaging. Background Technology

[0002] Intracranial aneurysms are a common cerebrovascular disease, and subarachnoid hemorrhage caused by rupture has extremely high mortality and disability rates. The level of inflammation in the aneurysm wall is considered one of the main influencing factors of aneurysm rupture risk; however, the characteristics of the inflammation level can currently only be obtained through histological examination and cannot be non-invasively predicted using imaging. Automated detection technology based on magnetic resonance angiography can display the imaging properties of the aneurysm wall, providing a possibility for non-invasive prediction of the inflammation level. Summary of the Invention

[0003] The purpose of this invention is to provide a system for predicting the level of inflammation in the intracranial aneurysm wall based on a multi-scale deformable convolutional network, so as to solve the problem that existing technologies are unable to predict the level of inflammation non-invasively.

[0004] The intracranial aneurysm wall inflammation level prediction system based on multi-scale deformable convolutional networks proposed in this invention includes a data reading module and an inflammation level prediction module; wherein:

[0005] The data reading module is used to acquire high-resolution magnetic resonance imaging (MRI) data of the intracranial aneurysm vessel wall and corresponding inflammation level labels. Specifically, the aneurysm wall region is marked using high-resolution MRI data of the vessel wall before intracranial aneurysm resection surgery; the inflammation levels at different locations measured from the aneurysm samples preserved after the aneurysm resection surgery are mapped to the corresponding locations on the MRI image as inflammation level labels.

[0006] The inflammation level prediction module trains a model based on high-resolution magnetic resonance imaging (MRI) of the blood vessel wall to predict the inflammation level of the aneurysm wall. Specifically, a multi-scale deformable convolutional U-shaped network is established. A 128×128×5 image patch centered at each location of the aneurysm is used as the input sample, and the corresponding inflammation level is used as the label. The network is then trained to obtain a model that predicts the inflammation level from the MRI image. The multi-scale deformable convolutional U-shaped network structure employs a 4-level encoder-decoder structure. The encoder compresses the image layer by layer and extracts semantic features, while the decoder restores the spatial resolution layer by layer. The network structure includes an encoder, a bottleneck layer, a decoder, and an output head. The specific structures of each part are as follows:

[0007] (1) Encoder: It contains multi-level encoding modules with channels of 32, 64, 128 and 256 respectively. Each level consists of a double-layer convolutional block and downsampling operation. Each level uses a double-layer convolutional block to calculate the mean and variance of each channel of each sample independently for normalization, and downsampling is performed by 2 × 2 max pooling.

[0008] (2) Bottleneck layer: A lightweight, hollow space pyramid pooling module is used, such as Figure 2 As shown, the module comprises multiple parallel branches, each employing different dilated convolutional kernels and global average pooling to capture multi-scale contextual features. Each branch first compresses the number of channels to one-quarter of the input to reduce computational cost, then concatenates along the channel dimension and fuses them through a 1×1 convolution to obtain the final output. This module enables the network to simultaneously acquire local details and global semantic information at the same layer, enhancing its ability to perceive inflammatory regions of varying sizes.

[0009] (3) Decoder: Contains multi-level decoding modules with channels of 32, 32, 64, and 128 respectively. Each level includes upsampling operations and performs feature fusion with the corresponding encoder through a deformable skip connection feature alignment module and a channel attention mechanism. The deformable skip connection feature alignment module includes an offset prediction network and a deformable convolutional network. [1] Alignment layer, two-layer convolutional fusion layer and channel attention layer [2] By learning this offset, the convolutional kernel can sample at irregular locations. Global average pooling is used to compress the spatial features of each channel into a scalar value, and then a fully connected network is used to learn the importance weights of each channel.

[0010] (4) Output header: through a non-negative activation function Ensure that the output is non-negative.

[0011] The training process employs a dual loss function: pixel-level mean squared error loss. , which represents the mean square error between the predicted value and the true label value calculated pixel by pixel within the effective area;

[0012]

[0013] in, and The first The predicted and true values ​​of each pixel, where N is the total number of pixels.

[0014] scalar mean square error loss , which represents the mean square error of comparing the average value of the predicted map within the effective region with the overall inflammation level value of that region;

[0015]

[0016] Total loss:

[0017] .

[0018] Wherein, λ is a weighting factor used to adjust the proportion of local error and overall error. Experimental studies show that a value of 2 has better performance.

[0019] A decoupled weight decay optimizer was used during training. The inflammation level prediction module, after training, can predict the inflammation level of intracranial aneurysm walls based on high-resolution magnetic resonance vascular wall imaging data. Leave-one-out cross-validation was employed, where one sample was used as the test set and all other samples as the training set, and the average of each validation result was used to test the prediction performance.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] (1) For the first time, inflammation level data was used as a label to predict the inflammation level of the aneurysm wall using magnetic resonance imaging.

[0022] (2) The system uses a lightweight hollow spatial pyramid pooling module, which enables the network to acquire local details and global semantic information at the same level, thereby enhancing its ability to perceive inflammatory regions of different sizes.

[0023] (3) The system adopts a deformable jump connection feature alignment module, which solves the problem of misalignment of encoder and decoder feature space on curved and irregular thin-walled structures by traditional jump connection. Attached Figure Description

[0024] Figure 1 This is a diagram of a multi-scale deformable convolutional U-shaped network architecture.

[0025] Figure 2 This is a structural diagram of a lightweight, hollow space pyramid pooling module. Detailed Implementation

[0026] The invention will be further illustrated below with specific test examples, and compared with various existing methods.

[0027] The specific steps of a practical example of this invention are as follows:

[0028] Step 1: Data Acquisition. Data was obtained from retrospective data prior to 2023 from the neurosurgery department of a Shanghai hospital. This included high-resolution magnetic resonance imaging (MRI) of the vessel wall before 9 cases of intracranial aneurysm resection surgery, as well as inflammation level data measured based on aneurysm samples after the resection surgery. The aneurysm wall region was marked using the high-resolution MRI of the vessel wall before intracranial aneurysm resection surgery. Inflammation levels at different locations measured from the aneurysm samples preserved after the resection surgery were mapped to corresponding locations on the MRI images as inflammation level labels.

[0029] Step 2, Data Preprocessing. Effective voxels are extracted from all sample regions in the magnetic resonance vascular wall imaging data. The global grayscale mean and standard deviation are calculated and standardized to ensure a mean of 0 and a standard deviation of 1. For each location, a 5×128×128 image patch centered on that location is extracted as the image input.

[0030] Step 3: Construct a multi-scale deformable convolutional U-shaped network. The overall architecture employs a 4-level encoder-decoder structure based on a U-shaped network. The encoder compresses the input image into a compact semantic feature representation through progressive convolution and pooling operations, while the decoder restores the features to the original spatial resolution through progressive upsampling. The number of parameters is approximately 4.9M. Each level of the encoder uses double-layer convolutional blocks, employing 3 × 3 convolutions, normalizing each instance, using a linear rectified activation function with leakage, and downsampling through 2 × 2 max pooling. The bottleneck layer's lightweight dilated spatial pyramid pooling module has four parallel branches employing dilated convolutions and global average pooling with different dilation rates: 1 × 1 convolution (dilation rate d = 1), 3 × 3 dilated convolution (dilation rate d = 3), 3 × 3 dilated convolution (dilation rate d = 5), and global average pooling. The decoder's deformable skip connection feature alignment module includes an offset prediction network that learns the offset to enable convolutional kernels to sample at irregular locations. The channel attention layer compresses the spatial features of each channel into a scalar value using global average pooling, and then learns the importance weights of each channel through a fully connected network. The output head is activated by a non-negative activation function. This function guarantees non-negative output and smoothly maps any real number to a positive number.

[0031] Step 4: Network Training. A dual loss function is used. Training was performed using a decoupled weight decay optimizer (learning rate 3 × 10⁻⁶). -4The number of 5×128×128 image patches input in each iteration is 64, with a maximum of 150 training rounds. Leave-one-out cross-validation (i.e., taking one sample as the test set and all other samples as the training set in each iteration, and traversing all samples sequentially) is used to test the prediction performance. Mean absolute error, root mean square error, Pearson correlation coefficient, and Spearman rank correlation coefficient are used as evaluation metrics. The comparison results of the method of this invention with various traditional machine learning methods and deep learning methods are shown in Table 1.

[0032] Table 1. Comparison results of the present invention with other methods

[0033] .

[0034] With various deep learning architectures (3D U-shaped network) [3] Residual network [4] Two-dimensional U-shaped network [5] The performance comparison results between the system established in this invention and traditional machine learning methods (ridge regression, Lasso regression, support vector regression, random forest, gradient boosting regression) show that the system has significant advantages in correlation indicators, providing system-level support for predicting the inflammation level of intracranial aneurysms through magnetic resonance imaging.

[0035] References

[0036] [1] DAI J, QI H, XIONG Y, et al. Deformable convolutional networks[C] / / Proceedings of the IEEE International Conference on Computer Vision. Piscataway, NJ: IEEE, 2017: 764-773.

[0037] [2] HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and PatternRecognition. Piscataway, NJ: IEEE, 2018: 7132-7141.

[0038] [3] ÇIÇEK Ö, ABDULKADIR A, LIENKAMP S S, et al. 3D U-Net: learningdense volumetric segmentation from sparse annotation[C] / / OURSSELIN S,JOSKOWICZ L, SABUNCU M R, et al, eds. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016. Cham: Springer, 2016: 424-432.

[0039] [4] HE K, ZHANG X, REN S, SUN J. Deep residual learning for imagerecognition[C] / / Proceedings of the IEEE Conference on Computer Vision andPattern Recognition. Piscataway, NJ: IEEE, 2016: 770-778.

[0040] [5] RONNEBERGER O, FISCHER P, BROX T. U-Net: convolutional networksfor biomedical image segmentation[C] / / NAVAB N, HORNEGGER J, WELLS W M, et al,eds. Medical Image Computing and Computer-Assisted Intervention – MICCAI2015. Cham: Springer, 2015: 234-241。

Claims

1. A system for predicting the level of inflammation in the wall of intracranial aneurysms based on multi-scale deformable convolutional networks, characterized in that, Includes the following modules: Data reading module: used to acquire high-resolution magnetic resonance imaging data of intracranial aneurysms and corresponding inflammation level labels; specifically including: using high-resolution magnetic resonance imaging data of intracranial aneurysms before resection surgery to mark the aneurysm wall region; using the inflammation level measured at different locations of aneurysm samples saved after aneurysm resection surgery to correspond to the corresponding locations on the magnetic resonance image as inflammation level labels. Inflammation level prediction module: Training a model based on high-resolution magnetic resonance vascular wall imaging data to predict the inflammation level of aneurysm wall; specifically including: establishing a multi-scale deformable convolutional U-shaped network, using 128×128×5 image blocks centered on each location of the aneurysm as input samples, and using the corresponding inflammation level as labels to train the network, thereby obtaining a model that predicts the inflammation level from magnetic resonance images.

2. The intracranial aneurysm wall inflammation prediction system based on multi-scale deformable convolutional networks according to claim 1, characterized in that, The multi-scale deformable convolutional U-shaped network structure in the inflammation level prediction module adopts a 4-level encoder-decoder structure, including an encoder, a bottleneck layer, a decoder, and an output head. The specific structure of each part is as follows: (1) Encoder: It contains multi-level encoding modules with channels of 32, 64, 128 and 256 respectively. Each level consists of a double-layer convolutional block and downsampling operation. Each level uses a double-layer convolutional block to calculate the mean and variance of each channel of each sample independently for normalization, and downsampling is performed by 2 × 2 max pooling. (2) Bottleneck layer: A lightweight dilated spatial pyramid pooling module with 256 channels is adopted, which contains multiple parallel branches and captures multi-scale contextual features with different dilated convolution kernels and global average pooling respectively. Each branch first compresses the number of channels to 1 / 4 of the input to reduce the amount of computation, and then concatenates them along the channel dimension and fuses them into the final output through 1×1 convolution; (3) Decoder: It contains multi-level decoding modules with channels numbered 32, 32, 64 and 128 respectively. Each level contains upsampling operations and performs feature fusion with the corresponding encoder through a deformable skip connection feature alignment module and a channel attention mechanism. The deformable skip connection feature alignment module includes an offset prediction network, a deformable convolutional alignment layer, a two-layer convolutional fusion layer and a channel attention layer. (4) Output header: through a non-negative activation function Ensure that the output is non-negative.

3. The intracranial aneurysm wall inflammation prediction system based on multi-scale deformable convolutional networks according to claim 1, characterized in that, The training process employs a dual loss function: pixel-level mean squared error loss. This represents the mean squared error between the predicted value and the true label value calculated pixel by pixel within the effective region; scalar mean squared error loss. , represents the mean squared error of comparing the average value of the predicted map within the effective region with the overall inflammation level value of that region; total loss Where λ is a weighting factor used to adjust the proportion of local error and overall error.

Images