The invention replaces the original convolution block of U-Net with a basic module to ensure stable training when the depth of the network is significantly increased. The extended convolution is used at the top of the encoder to extend the reception range of the network, with only 3 downsamplings.

In order to enhance the performance of the network by exploring remote context information, a dual-rechecking module is embedded between the encoder and the decoder. The input size for the ResUNet model is 80X 80, which is sufficient to accommodate most intracranial aneurysms. During training, a 3D image patch having the above size is randomly sampled from the entire CTA volume.

To balance the number of training samples containing and not containing aneurysms, the sampled patch contains a 50% probability of containing an aneurysm. Data enhancements (e.g., rotation, scaling and flipping) are applied to the CTA data prior to patch sampling. Before arriving at the network, the input is clipped to [0, 900] Huo Ensi Field units (Hu) and then normalized to [ -1,1]. The model trains the network to optimize the weighted sum of the binary cross-entropy loss and the die (Dice) loss.

The Adam optimizer was used by setting the momentum and weight attenuation coefficients to 0.9 and 0.0001, respectively. The present invention employs a multivariate learning rate policy in which after each iteration, the initial learning rate is multiplied by

. The initial learning rate is 0.0001 and the training period is 100. At each cycle, the model first randomly selected images of 300 patients from the training set, and then randomly cropped 50 patches containing positive and negative examples from the images of each patient. A total of about 15,000 patches are used per epoch to train the modelAnd (4) molding. During model training, sub-volumes of 16 slices are randomly drawn from the volume. The dataset was preprocessed to find the contours of the skull, then each volume was cropped in the axial plane around the skull before resizing each slice to 208 x 208 pixels. The slice is then cropped to 192 × 192 pixels (random cropping is used during training, center cropping is used during testing), with the final input size for each example being 16 × 192 × 192; the same transformation is applied to the segment labels. The segmentation output is trained to optimize a weighted combination of voxel binary cross entropy and Dice loss.

The model crops the input into [ -300, 700] Huo Ensi field units, normalizes to [ -1,1], and starts from zero. The model was trained on 3 Graphics Processing Units (GPUs), with a minimum of 2 cases per GPU. The parameters of the model were optimized using a random gradient descent optimizer with a momentum of 0.9, a weight peak learning rate of 0.1 for random initialization, and a peak learning rate of 0.01 for pre-training weights. For normalization, the loss of all trainable parameters is increased by a weight decay of 0.001 and a random depth dip is used in the encoder block. To control class imbalance, 3 methods were used. First, an auxiliary penalty is added after the encoder, and a focus penalty is used to encourage larger parameter updates for misclassified positive samples. Second, the sampling frequency of the abnormal training example is higher than that of the normal example, so the abnormal case accounts for 30% of the training iteration number. The parameters of the decoder are not updated in the training iteration, where the segmentation labels consist of only background voxels.

To generate a segmented prediction of the entire volume, only the segmented outputs of the successive 16-slice subvolumes need to be connected together. If the number of slices is not evenly divisible by 16, the last input volume will be filled with 0 and the corresponding output volume will be truncated to the original size.

The present application also provides a CTA image data processing apparatus, as shown in fig. 8, including:

a receiving module to receive CTA image data, the CTA image data including one or more original images;

and the correction module is used for processing the original image based on a preset three-dimensional CNN network model, removing a bone graph and a vein graph in the original image and obtaining a corrected image only with an artery graph.

Example 4

The present invention provides a CTA image data processing method, which is the third processing step inembodiment 1, as shown in fig. 9, and includes:

step S210, receiving CTA image data, wherein the CTA image data comprises one or more original images;

step S220, processing the original image to obtain all nodes in the original image, wherein the nodes are points larger than a preset voxel value;

step S230, connecting every two adjacent nodes in the original image to generate a node frame graph;

step S240, a triangle path in the node frame graph is obtained to obtain a triangle path histogram, where the triangle path is a path where a plurality of nodes form a triangle.

Although the triangular path histogram is a set of voxel-based features, it is derived from the graph structure extracted from a given image. The graphical structure may simply be extracted from a binary image of the target structure (e.g. the vessel system).

The triangular path histogram feature set in the graph is defined at each node (i.e., each voxel) in a given graph based on a three-dimensional histogram of shortest path distances between the node and each of its neighboring node pairs. The feature vector effectively encodes the local graph network pattern around the node. Since the triangular path histogram in the graph does not use any three-dimensional thinning algorithm, the triangular path histogram does not have the problem of wrong node identification. Although the triangular path histogram features in the figures are particularly good at describing branch vessel structures, they may also describe a bump-like or node-like structure.

The triangle path histogram in the graph is powerful, and the triangle path histogram feature set in a single graph is enough to correctly detect the cerebral aneurysm by using a single support vector machine classifier. The triangular path histogram feature in the graph has robustness to non-rigid transformation, and the branched vessel structure and the protrusion-shaped structure can be effectively coded. A triangle path histogram characteristic set in a graph is defined on each node of any undirected graph.

In step S240, the method further includes:

an undirected graph G with a vertex set as V and two vertexes

The shortest path length between is

The shortest path is a path along the direction of the edge of the graph G;

presetting natural number

The triplet of (2);

triple at node

The node at which the triangular path histogram feature value is defined as the shortest path distance satisfies

The conditions of (a) are:

triangle path histogram eigenvalue

Is defined as follows:

wherein

Is the number of elements of a given set

Triangle path histogram feature value

By a series of triplets

) Is defined as

The sequence of triples is determined to satisfy

And

all distance combinations of conditions. For example, suppose

,

For example, suppose

,

And sequentially obtaining the triangular path histogram feature calculation model in the graph, as shown in fig. 10.

The invention also employs a multi-resolution strategy. A given binary volume will be rescaled by a factor of 0.5. Then, using the contracted volumes, the triangle path histogram feature in the graph is calculated. These features are then returned to the original voxels.

The present invention also uses two well-known Hessian-derived voxel-based features (shape index and point enhancement filters) to evaluate the effectiveness of the cooperative use of two different types of features, namely a grayscale-based and a graph-based feature set.

Processing the original image to obtain all nodes in the original image, wherein the nodes which are points larger than a preset voxel value comprise:

configuring the node enhancement filter to the following formula:

wherein

Is the eigenvalue of the Hessian matrix;

the shape index is configured as:

wherein k1 and k2 are the principal curvatures;

based on the above steps, the features of the voxels are processed.

Each feature vector is mapped to a high-dimensional space by a kernel calculation before it is evaluated by a linear support vector machine classifier. The present invention solves this problem using an explicit feature mapping approach. In the explicit feature mapping method, the triangular path histogram feature vector in each graph adopts an index

Approximate finite feature mapping.

Meanwhile, feature vectors based on the Hessian matrix are mapped by utilizing Gaussian kernels. The purpose of this explicit feature mapping is to reduce computation time significantly, while using two different kernels (i.e. exponentials)

Nuclei and gaussian nuclei). Specifically, the triangle path histogram variables in all histogram-based maps are first multiplied by a factor such that the standard deviation of each feature becomes 1. Furthermore, all the grayscale-based features are linearly normalized such that their mean and standard deviation are 0 and 0, respectively

。

The parameters control the weights between the triangle path histogram features and the grayscale-based features in the histogram-based map. It is further noted that the mean and standard deviation are calculated in advance from all training data sets.

The present invention uses a linear support vector machine as a voxel-based classifier. A base truth label roll is prepared prior to the training phase. Using these ground truth label volumes, foreground voxels are divided into positive and negative classes. To avoid an imbalance of the positive and negative sample size, the negative samples are randomly down-sampled so that only 0.5% of the negative samples remain in each case. And finally, collecting all positive and negative voxels from all training samples and training the support vector machine.

Embodiment 4 of the present invention further comprises the steps of:

multiplying all triangle path histogram variables in the histogram-based graph by a factor to change the standard deviation of each feature to 1;

all voxels are linearly normalized for gray-based features such that the mean and standard deviation of the voxels are 0 and 0, respectively

；

Wherein,

the parameters control the weights between the triangle path histogram features and the grayscale-based features in the histogram-based map.

The average value and/or the standard deviation are preset and/or obtained from all previous training data sets.

Processing the original image by criteria of window spacing [0, 450] and [ -50, 650], removing a CTA image of bone;

searching an initial region of the blood vessel by using a window with a preset threshold value, and reserving the maximum connectivity region as a final region of the blood vessel;

and obtaining corresponding cutting intervals to carry out normalization processing on the original image cutting based on the histogram of the brightness of the voxels in the final region.

The image characteristics of the triangular path histogram and the Hessian matrix in the graph have synergistic effect. The triangular path histogram feature in the graph in the algorithm model can cooperate with various existing image features. The invention simultaneously utilizes the triangular path histogram in the graph and the characteristic based on deep learning. Another important feature is that the triangle path histogram image features in the graph are invariant to translation and mirroring and robust to rotation and small local deformations. This is because the feature set comes from only one graph structure, which does not change significantly when small deformations occur locally.

Compared with the artificial data amplification technology frequently used in the deep learning-based method, the method based on the triangular path histogram in the graph does not need any data amplification (such as rotation and non-rigid deformation), so that the characteristic set of the triangular path histogram in the graph can successfully distinguish the vascular structure from the non-vascular structure in the detection task. Meanwhile, the triangular path histogram in the graph effectively captures the topological branch pattern of the human organ. In the inference stage, a segmented prediction of the entire image is generated by merging uniformly sampled predictions. Two adjacent patches may have a 1/8 overlap.

To detect intracranial aneurysms in some low contrast images cropped with a default window interval of [0, 900], two additional intervals [0, 450] and [ -50, 650] are used to normalize the source image. The setting is automatically selected according to the brightness distribution. Given a CTA image with bone removed, an initial region of the blood vessel is found using a threshold (e.g., 150 Hu), and then the maximum connectivity region is retained as the final region of the blood vessel. The histogram of the intensity of voxels in this region is analyzed to find the appropriate clipping interval. The present invention calculates the three distributions of [0, 200], [200, 300] and [300, 500] intervals, which correspond to [0, 450], [ -50, 650] and [0, 900]. Finally, a cropping interval corresponding to the dominant distribution interval is selected to normalize the source image.

The present invention also provides a CTA image data processing apparatus, as shown in fig. 11, including:

the processing module is used for processing the original image to obtain all nodes in the original image, wherein the nodes are points larger than a preset voxel value;

the generating module is used for connecting every two adjacent nodes in the original image to generate a node frame graph;

the acquisition module is used for acquiring a triangular path in the node frame graph to obtain a triangular path histogram, wherein the triangular path is a path of a triangle formed by a plurality of nodes.

In one embodiment, the obtaining module is further configured to perform the following steps, including:

an undirected graph G with V set of vertexes, and two vertexes

The shortest path length between is

The shortest path is a path along the direction of the edge of the graph G;

in advanceSetting natural number

The triplet of (a);

triple unit

At a node

The conditions of (a) are:

triangle path histogram feature value

Is defined as:

wherein

Is the number of elements of a given set

Triangle path histogram feature value

By a series of triplets

) Is defined as

Compared with the traditional aneurysm identification method, the method has the advantages that:

1. the invention can efficiently screen out the aneurysm through the CTA image, the detection speed, the diagnosis effective rate, the sensitivity and the specificity are all superior to those of common radiologists, the daily work efficiency of the radiologists can be improved, and misdiagnosis is avoided.

The bone and vein images of the CTA image can influence the diagnosis of the aneurysm, and the image preprocessing algorithm can effectively remove the bone and vein images, improve the effective diagnosis rate and avoid false positive.

3. The prevalence rates of aneurysms of different patients are very different, from common people (3% -7%) to subarachnoid hemorrhage patients (85%), the HCNN model can effectively deal with diagnosis of the aneurysms with large differences, and the detection result is stable.

The present invention allows radiologists to know the status of an aneurysm by giving them an assessment of sensitivity and specificity. Sensitivity indicates the number of negative results in total aneurysm positive cases, specificity indicates the number of positive results in total aneurysm negative cases, and accuracy indicates the number of positive and negative results in all tested cases.

To determine the robustness of the results and whether the results were due to the radiologist used, the present invention performed a sensitivity analysis and a T-test for differences in sensitivity, specificity, and accuracy. The quantitative variables are expressed as mean ± SD if the data is normally distributed, and median and quartile spacing are used when non-normally distributed data is used. The classification variables are expressed as frequency or percentage.

To evaluate the performance of the algorithmic tests, the models were evaluated in each cohort, respectively, for accuracy in correctly displaying the patient, patient level sensitivity, specificity, diagnostic efficacy, and variability was evaluated using 95% wilson score confidence intervals. In the test data set containing 357 aneurysms, 92.9% of the aneurysms were successfully detected by the present invention. Due to the flexible decision zone approach, the model performs well in detecting aneurysms of various sizes. The proposed bounding box can automatically fit the size of the aneurysm. The invention has 96.7 percent of total sensitivity for the aneurysm with the diameter of more than 3 millimeters.

The readable storage medium may be a computer storage medium or a communication medium. Communication media includes any medium that facilitates transfer of a computer program from one place to another. Computer storage media may be any available media that can be accessed by a general purpose or special purpose computer. For example, a readable storage medium is coupled to a processor such that the processor can read information from, and write information to, the readable storage medium. Of course, the readable storage medium may also be an integral part of the processor. The processor and the readable storage medium may reside in an Application Specific Integrated Circuits (ASIC). Additionally, the ASIC may reside in user equipment. Of course, the processor and the readable storage medium may also reside as discrete components in a communication device. The readable storage medium may be a read-only memory (ROM), a random-access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.

The present invention also provides a program product comprising execution instructions stored in a readable storage medium. The at least one processor of the device may read the execution instructions from the readable storage medium, and the execution of the execution instructions by the at least one processor causes the device to implement the methods provided by the various embodiments described above.

In the above embodiments of the terminal or the server, it should be understood that the Processor may be a Central Processing Unit (CPU), other general-purpose processors, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of a method disclosed in connection with the present invention may be embodied directly in a hardware processor, or in a combination of the hardware and software modules within the processor.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A CTA image data identification method, comprising:

step 1, receiving a preprocessed image, wherein the preprocessed image only contains an artery graph;

step 2, predicting each voxel in the artery graph to obtain the probability that each voxel in the artery graph is an aneurysm;

step 3, marking the aneurysm at the position where the probability of the voxel being the aneurysm is larger than a preset value;

the above three steps are processed by an HCNN model, which is a CNN having an encoder-decoder structure;

the decoder is used for expanding the code into a full-resolution partition volume;

the HCNN model is trained by the following steps, including:

2. The CTA image data recognition method according to claim 1, further comprising:

3. A CTA image data recognition apparatus, comprising:

the marking module is used for marking the aneurysm, wherein the probability of the voxel being the aneurysm is larger than a preset value;

the receiving module, the predicting module and the predicting module form an HCNN model, and the HCNN model is a CNN with an encoder-decoder structure;

the HCNN model is trained by the following units, including:

and the extension unit is used for carrying out horizontal turning on the projection image to extend the training data set and then carrying out training again.

4. A readable storage medium, characterized in that a computer program is stored in the readable storage medium, which computer program, when being executed by a processor, is adapted to carry out the method of one of the claims 1-2.