Chronic subdural hematoma prognosis analysis method and device based on subdural hematoma segmentation and iterative prediction, processor and medium thereof

Abstract

The present application relates to a kind of chronic subdural hematoma prognosis analysis method based on subdural hematoma segmentation and iterative prediction, wherein the method comprises: collecting patient MRI image for data preprocessing;The corresponding data is carried out data enhancement processing based on geometric transformation, intensity transformation and noise injection;3D-UNet segmentation framework is constructed based on the training data described above, and the voxel level prediction segmentation result is outputted;The data processed above is randomly divided into training set, verification set and test set according to preset proportion, and network training is carried out in the segmentation network combined with loss function;Imageological features are extracted from MRI original data and corresponding segmentation results, and the corresponding clinical indicators are evaluated and predicted using random forest algorithm.The present application also relates to a corresponding device, processor and storage medium thereof.The method, device, processor and storage medium thereof adopted in the present application enable the segmentation network to be applied to image data of different devices and different centers.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more particularly to the field of bioinformatics technology, specifically to a method, apparatus, processor, and computer-readable storage medium for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction. Background Technology

[0002] Chronic subdural hematoma (CSDH) is one of the most common neurosurgical conditions, with its incidence increasing with population aging and related medical conditions. Typical symptoms of CSDH include headache, seizures, memory loss, and confusion. Patients may experience difficulty speaking, swallowing, and walking, as well as weakness or numbness in the limbs and face. CSDH is often associated with head trauma, but can also be related to coagulation disorders, lumbar puncture, or other factors. The elderly are the primary patient group for CSDH, but younger patients and infants can also be affected. Compared to CT scans, MRI is more sensitive in determining the size and internal structure of CSDH, and can observe fresh hemorrhage, hemolysis, and changes in hemoglobin. CSDH can occur on the most curved frontal or occipital convex surfaces, or symmetrically on both frontal and occipital lobes. As CSDH progresses, patients may exhibit drastically different prognoses. Therefore, early quantitative assessment of chronic subdural hematoma is essential. While many algorithms for prognostic analysis of chronic subdural hematoma based on hematoma segmentation exist, they have only been validated in single centers and lack generalization ability. Furthermore, these algorithms directly predict a patient's prognostic indicators using extracted relevant features, while ignoring the correlation between different clinical indicators. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, apparatus, processor and computer-readable storage medium for chronic subdural hematoma prognostic analysis based on subdural hematoma segmentation and iterative prediction that is multi-center, precise and hierarchical.

[0004] To achieve the above objectives, the present invention provides a method, apparatus, processor, and computer-readable storage medium for chronic subdural hematoma prognostic analysis based on subdural hematoma segmentation and iterative prediction, as follows:

[0005] The main feature of this prognostic analysis method for chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction is that the method includes the following steps:

[0006] (1) Collect the patient's MRI images and perform bias field correction and gray level normalization data preprocessing to facilitate subsequent standardized analysis;

[0007] (2) Perform data augmentation processing based on geometric transformation, intensity transformation and noise injection on the preprocessed data to provide it as training data for training the hematoma segmentation model.

[0008] (3) Construct a 3D-UNet segmentation framework, use the training data obtained above to segment chronic subdural hematoma, and output voxel-level predicted segmentation results.

[0009] (4) The data processed above is randomly divided into training set, validation set and test set according to a preset ratio, and the network is trained in the segmentation network by combining the loss function.

[0010] (5) Extract imaging features from the original MRI data and the corresponding segmentation results, and use the random forest algorithm to evaluate and predict the clinical indicators of the corresponding patients with chronic subdural hematoma.

[0011] Preferably, step (1) specifically includes the following steps:

[0012] (1.1) Collect a preset number of MRI images and use 3D Slicer software to annotate the contours of the subdural hematoma to obtain hematoma segmentation labels;

[0013] (1.2) The collected MRI image data are first corrected by bias field using the Python-based SimpleITK library to make the low-frequency gray distribution of the MRI images uniform.

[0014] (1.3) The grayscale of the MRI image data is then normalized to between 0 and 1, and only the grayscale values ​​between 0.5% and 99.5% are retained to complete the data preprocessing.

[0015] Preferably, step (2) specifically includes the following steps:

[0016] (2.1) Data augmentation based on geometric transformation: Each original MRI grayscale image is randomly flipped along the coronal axis with a probability of 0.5 and randomly rotated in the cross section, wherein the rotation angle follows a uniform distribution within 0 to 45°. Each MRI grayscale image is also elastically deformed based on displacement grid with a probability of 0.1, thereby changing the brain structure as a whole.

[0017] (2.2) Data augmentation based on intensity transformation: Each original MRI grayscale image has a probability of 0.5 to be multiplied by a random floating-point number between 0.5 and 1.5, and the grayscale values ​​of the image are cropped to the same range as the original grayscale values ​​to disturb the contrast of the image.

[0018] (2.3) Data augmentation based on noise injection: Each original MRI grayscale image has a 0.1 probability of being added with random Gaussian noise with a mean of 0 and a standard deviation of 1;

[0019] (2.4) The hematoma segmentation labels mentioned above are all transformed by data augmentation according to steps (2.1) to (2.3), and the grayscale images and hematoma segmentation labels after data augmentation are used for subsequent training of the segmentation network.

[0020] Preferably, step (2.1) specifically includes:

[0021] For each dimension of the MRI grayscale image, a displacement value is generated for each point on its displacement grid, thus obtaining a displacement value array. The displacement value array is added to the coordinate array corresponding to the position of each point on the grid to obtain a displacement matrix. The corresponding coordinates on the displacement matrix are sampled on the original MRI grayscale image, and B-spline interpolation is performed at non-integer coordinate positions to obtain the grayscale values ​​of the deformed image at each coordinate, thus obtaining the final deformed image.

[0022] Preferably, step (3) specifically includes the following steps:

[0023] (3.1) The grayscale image after data augmentation and its hematoma segmentation label are used as training data to train the hematoma segmentation model;

[0024] (3.2) Constructing the 3D-UNet segmentation framework: It consists of a shrinking path and an expanding path, wherein,

[0025] The described shrinkage path consists of four encoding modules, each consisting of two convolutional layers and one max-pooling layer. The main part of each convolutional layer is a 3×3×3 three-dimensional convolutional kernel; and

[0026] Grouping normalization is used before each convolution to handle the internal covariance shift, and a modified linear unit activation function is implemented after the convolution. The shrinking of the path is used to progressively extract abstract features from the input grayscale data.

[0027] The expansion path consists of four upsampling modules. Each upsampling module first uses nearest neighbor interpolation to upsample the grayscale image, making the grayscale image size twice its original size. Then, it is spliced ​​with the corresponding grayscale image in the contraction path along the channel dimension to form a skip connection, so that the information of the contraction path and the expansion path are merged without changing the grayscale image size. Subsequently, the grayscale image is convolved by two 3×3×3 three-dimensional convolution kernels to further extract and fuse the corresponding image data.

[0028] (3.3) The segmentation network described above uses a 3×3×3 three-dimensional convolutional kernel to output voxel-level prediction results.

[0029] Preferably, step (4) specifically includes:

[0030] The data obtained in the above steps are randomly divided into training, validation, and test sets in a ratio of 8:1:1. During training, the training set is divided into 64×64×64 segments with a 32×32×32 interval and input into the training network, retaining only positive data with at least one voxel hematoma. The loss function L used in the training process is as follows:

[0031] L = L KL +L Dice (1)

[0032] Among them, L KL The cross-entropy loss is defined as follows:

[0033]

[0034] in, For a specific voxel x, if this voxel belongs to category c, then otherwise X is the set of all voxels, θ is the trainable parameters in the network, and p c (x; θ) represents the probability that a certain voxel belongs to class c;

[0035] L Dice The Dice loss is defined as follows:

[0036]

[0037] All the loss functions mentioned above are used to update the network parameters through the Adam optimizer, with the learning rate and batch size set to 0.001 and 8, respectively.

[0038] Preferably, step (5) specifically includes the following steps:

[0039] (5.1) The Python-based PyRadiomics library was used to extract radiographic features from the raw MRI data and corresponding segmentation results, including: 16 3D morphological features, 19 first-order features, 24 gray-level co-occurrence matrix features, 16 gray-level size region matrix features, 16 gray-level run matrix features, 5 adjacent gray-level difference matrix features, and 14 gray-level dependence matrix features.

[0040] (5.2) The above features are used as input to the hematoma segmentation model. After training, the output includes OC_MGS, which is the result of Markwalder Grading Scale, representing the prognosis; OC_mRS, which is the result of modified Rankin Scale, representing the prognosis; Radio_Cure, which is the radiological cure status at the endpoint; and Surgery, which is the clinical indicator of whether to switch to surgical treatment at the endpoint.

[0041] (5.3) Based on the above imaging features, the random forest algorithm is used to predict Radio_Cure, that is, whether radiological cure is achieved at the end point. Then, the above imaging features and the results of Radio_Cure are used as inputs, and the random forest algorithm is used again to predict OC_MGS, that is, the results of Markwalder Grading Scale. This process is repeated until the results of all indicators are predicted.

[0042] The device for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction is characterized in that the device comprises:

[0043] A processor is configured to execute computer-executable instructions;

[0044] The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the above-described method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction.

[0045] The processor that implements the prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction is characterized in that the processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the prognostic analysis method for chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction described above.

[0046] The computer-readable storage medium is characterized in that it stores a computer program that can be executed by a processor to implement the steps of the above-described method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction.

[0047] This invention employs a method, device, processor, and computer-readable storage medium for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction. First, it uses 3D-UNet as the main segmentation framework and designs a data augmentation strategy robust to different scanning devices and centers across multiple centers. Subsequently, the algorithm extracts imaging features using high-quality segmentation results and employs random forest for feature selection and prognostic assessment. During this process, the algorithm uses relatively simple prognostic indicators to iteratively evaluate other more difficult prognostic indicators from easy to difficult, achieving more accurate hierarchical analysis. This allows for more comprehensive and sufficient application in clinical practice and has certain promotional value. Attached Figure Description

[0048] Figure 1 This is a flowchart of the prognostic analysis method for chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction according to the present invention.

[0049] Figure 2 This is a flowchart illustrating the data preprocessing process for this invention.

[0050] Figure 3 This is a schematic diagram illustrating data augmentation for the present invention.

[0051] Figure 4 This is a schematic diagram illustrating the segmentation model training process of this invention. Detailed Implementation

[0052] To more clearly describe the technical content of the present invention, the following description is provided in conjunction with specific embodiments.

[0053] Before describing the embodiments of the present invention in detail, it should be noted that, in the following, the terms “comprising,” “including,” or any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0054] This technical solution collects MRI images from different scanning devices within the same center, as well as from different centers, and performs pixel-by-pixel hematoma annotation. We designed a segmentation framework based on deformation, grayscale, and noise data enhancement, enabling the segmentation network to achieve good segmentation performance on image data from different devices and centers. Furthermore, we designed an iterative prediction-based patient prognostic assessment method. First, based on the initial MRI images and segmentation results for each patient, we predict easily identifiable clinical indicators. Then, based on these existing indicators, we iteratively evaluate other more difficult-to-identify indicators. Our main innovations lie in the following three aspects:

[0055] 1. We collected 257 MRI images from 116 patients from different scanning devices and centers, and performed pixel-by-pixel hematoma annotation to conduct a more precise quantitative analysis of chronic subdural hematoma.

[0056] 2. We designed a data augmentation strategy based on deformation, grayscale, and noise using the 3D-UNet segmentation network. This framework makes the segmentation network more robust and stable, and can obtain relatively accurate segmentation results on MRI images from different devices and centers by training only on single-center, single-device data.

[0057] 3. We designed a hierarchical clinical indicator evaluation algorithm based on iterative prediction. According to the principle of easy to difficult, we first extract imaging features from MRI images and segmentation results, and construct a regressor to obtain the most predictable clinical indicators. Then, based on the already predicted indicators, we iteratively predict other difficult-to-predict indicators to achieve systematic analysis.

[0058] Please see Figure 1 As shown, this method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction includes the following steps:

[0059] (1) Collect the patient's MRI images and perform bias field correction and gray level normalization data preprocessing to facilitate subsequent standardized analysis;

[0060] (2) Perform data augmentation processing based on geometric transformation, intensity transformation and noise injection on the preprocessed data to provide it as training data for training the hematoma segmentation model.

[0061] (3) Construct a 3D-UNet segmentation framework, use the training data obtained above to segment chronic subdural hematoma, and output voxel-level predicted segmentation results.

[0062] (4) The data processed above is randomly divided into training set, validation set and test set according to a preset ratio, and the network is trained in the segmentation network by combining the loss function.

[0063] (5) Extract imaging features from the original MRI data and the corresponding segmentation results, and use the random forest algorithm to evaluate and predict the clinical indicators of the corresponding patients with chronic subdural hematoma.

[0064] like Figure 1As shown, we collected 257 MRI images from 116 patients across different scanning devices and centers, and used 3D Slicer software to annotate the contours of subdural hematomas to obtain segmentation labels. These data then underwent the four-stage process described above for training the neural network and optimizing the algorithm. For data preprocessing, we first used the Python-based SimpleITK library for N4 bias field correction to homogenize the low-frequency grayscale distribution. Then, we normalized the MRI data to the range of 0-1, retaining only the grayscale values ​​between 0.5% and 99.5% for subsequent standardized analysis. This preprocessing workflow is as follows: Figure 2 As shown in the figure; then we perform data augmentation based on deformation, grayscale and noise; then we use the augmented data to train a 3D-UNet segmentation network to obtain hematoma segmentation results; then we use the segmentation results to extract imaging features and use random forest to iteratively evaluate clinical indicators.

[0065] In a preferred embodiment of the present invention, step (1) specifically includes the following steps:

[0066] (1.1) Collect a preset number of MRI images and use 3D Slicer software to annotate the contours of the subdural hematoma to obtain hematoma segmentation labels;

[0067] (1.2) The collected MRI image data are first corrected by bias field using the Python-based SimpleITK library to make the low-frequency gray distribution of the MRI images uniform.

[0068] (1.3) The grayscale of the MRI image data is then normalized to between 0 and 1, and only the grayscale values ​​between 0.5% and 99.5% are retained to complete the data preprocessing.

[0069] In a preferred embodiment of the present invention, step (2) specifically includes the following steps:

[0070] (2.1) Data augmentation based on geometric transformation: Each original MRI grayscale image is randomly flipped along the coronal axis with a probability of 0.5 and randomly rotated in the cross section, wherein the rotation angle follows a uniform distribution within 0 to 45°. Each MRI grayscale image is also elastically deformed based on displacement grid with a probability of 0.1, thereby changing the brain structure as a whole.

[0071] (2.2) Data augmentation based on intensity transformation: Each original MRI grayscale image has a probability of 0.5 to be multiplied by a random floating-point number between 0.5 and 1.5, and the grayscale values ​​of the image are cropped to the same range as the original grayscale values ​​to disturb the contrast of the image.

[0072] (2.3) Data augmentation based on noise injection: Each original MRI grayscale image has a 0.1 probability of being added with random Gaussian noise with a mean of 0 and a standard deviation of 1;

[0073] (2.4) The hematoma segmentation labels mentioned above are all transformed by data augmentation according to steps (2.1) to (2.3), and the grayscale images and hematoma segmentation labels after data augmentation are used for subsequent training of the segmentation network.

[0074] In a preferred embodiment of the present invention, step (2.1) specifically comprises:

[0075] For each dimension of the MRI grayscale image, a displacement value is generated for each point on its displacement grid, thus obtaining a displacement value array. The displacement value array is added to the coordinate array corresponding to the position of each point on the grid to obtain a displacement matrix. The corresponding coordinates on the displacement matrix are sampled on the original MRI grayscale image, and B-spline interpolation is performed at non-integer coordinate positions to obtain the grayscale values ​​of the deformed image at each coordinate, thus obtaining the final deformed image.

[0076] like Figure 3 As shown, the data augmentation methods adopted in this technical solution are as follows:

[0077] A major problem in deep learning is data scarcity, which limits model performance. To alleviate this problem and improve the robustness and generalization ability of our model, we employ data augmentation strategies to augment the training data before inputting the original training data into the 3D-UNet segmentation network, without requiring additional annotations. For example... Figure 3As shown, our data augmentation strategy mainly includes three aspects: geometric transformation-based data augmentation, intensity transformation-based data augmentation, and noise injection-based data augmentation. For geometric transformation-based data augmentation, each original MRI grayscale image has a 0.5 probability of being randomly flipped along the coronal axis and randomly rotated within the transverse plane, with the rotation angle following a uniform distribution within the range of 0–45°. Each image also has a 0.1 probability of undergoing elastic deformation based on a displacement grid to alter the overall brain structure. Specifically, for each dimension of the image, a displacement value is generated for each point on the displacement grid, resulting in a displacement value array. This array is added to the corresponding coordinate array (i.e., the position of each point on the grid) to obtain a displacement matrix. Then, samples are taken from the original image according to these coordinates, and B-spline interpolation is performed at non-integer coordinate positions to obtain the grayscale values ​​of the deformed image at each coordinate, ultimately yielding the deformed image. For intensity-based data augmentation, each image has a 0.5 probability of being multiplied by a random floating-point number between 0.5 and 1.5, and the grayscale values ​​are cropped to the same range as the original grayscale to disrupt the contrast. For noise-injection-based data augmentation, each image has a 0.1 probability of being added with random Gaussian noise with a mean of 0 and a standard deviation of 1. Correspondingly, the aforementioned hematoma segmentation labels also undergo transformations including flipping, rotating, and deforming, with the specific implementation methods consistent with those for the grayscale images. The augmented grayscale images and their segmentation labels are then used to train the segmentation network.

[0078] Through these data augmentation strategies, we expect the trained network to generalize to other devices and centers beyond the device where the training data was located.

[0079] In a preferred embodiment of the present invention, step (3) specifically includes the following steps:

[0080] (3.1) The grayscale image after data augmentation and its hematoma segmentation label are used as training data to train the hematoma segmentation model;

[0081] (3.2) Constructing the 3D-UNet segmentation framework: It consists of a shrinking path and an expanding path, wherein,

[0082] The described shrinkage path consists of four encoding modules, each consisting of two convolutional layers and one max-pooling layer. The main part of each convolutional layer is a 3×3×3 three-dimensional convolutional kernel; and

[0083] Grouping normalization is used before each convolution to handle the internal covariance shift, and a modified linear unit activation function is implemented after the convolution. The shrinking of the path is used to progressively extract abstract features from the input grayscale data.

[0084] The expansion path consists of four upsampling modules. Each upsampling module first uses nearest neighbor interpolation to upsample the grayscale image, making the grayscale image size twice its original size. Then, it is spliced ​​with the corresponding grayscale image in the contraction path along the channel dimension to form a skip connection, so that the information of the contraction path and the expansion path are merged without changing the grayscale image size. Subsequently, the grayscale image is convolved by two 3×3×3 three-dimensional convolution kernels to further extract and fuse the corresponding image data.

[0085] (3.3) The segmentation network described above uses a 3×3×3 three-dimensional convolutional kernel to output voxel-level prediction results.

[0086] The enhanced data (grayscale images and segmentation labels) was used as training data to train the hematoma segmentation model. UNet is one of the most popular semantic segmentation fully convolutional networks (FCNs). Its unique U-shaped architecture allows the model to combine hierarchical features to achieve voxel-level dense prediction. However, when applied to 3D data, 2D UNet cannot identify anatomical structures orthogonal to the 2D plane. To effectively capture information in all three dimensions, we employed a 3D U-shaped model for chronic subdural hematoma segmentation. Figure 4 As shown, our model consists of a contraction path and an expansion path. The contraction path comprises four encoding modules, each consisting of two convolutional layers and a max-pooling layer. The main part of each convolutional layer is a 3×3×3 three-dimensional convolutional kernel. We apply group normalization before each convolution to handle the internal covariance shift, and implement a modified linear unit (ReLU) activation function after the convolution. This path can progressively extract abstract features from the input data (grayscale image). The expansion path consists of four upsampling modules. Each upsampling module first upsamples the feature map using nearest neighbor interpolation, doubling its size. Then, it concatenates the corresponding feature map from the contraction path along the channel dimension to form a skip connection, merging information from the contraction and expansion paths without changing the feature map size. Subsequently, it convolves the feature map with two 3×3×3 three-dimensional convolutional kernels to further extract and fuse this information. Finally, our network outputs voxel-level predictions using 3×3×3 three-dimensional convolutional kernels.

[0087] In a preferred embodiment of the present invention, step (4) specifically comprises:

[0088] The data obtained in the above steps are randomly divided into training, validation, and test sets in a ratio of 8:1:1. During training, the training set is divided into 64×64×64 segments with a 32×32×32 interval and input into the training network, retaining only positive data with at least one voxel hematoma. The loss function L used in the training process is as follows:

[0089] L = L KL +L Dice (1)

[0090] Among them, L KL The cross-entropy loss is defined as follows:

[0091]

[0092] in, For a specific voxel x, if this voxel belongs to category c, then otherwise X is the space spanned by all voxels, θ is the trainable parameters in the network, and p c (x; θ) represents the probability that a certain voxel belongs to class c;

[0093] L Dice The Dice loss is defined as follows:

[0094]

[0095] All the loss functions mentioned above are used to update the network parameters through the Adam optimizer, with the learning rate and batch size set to 0.001 and 8, respectively.

[0096] In a preferred embodiment of the present invention, step (5) specifically includes the following steps:

[0097] (5.1) The Python-based PyRadiomics library was used to extract radiographic features from the raw MRI data and corresponding segmentation results, including: 16 3D morphological features, 19 first-order features, 24 gray-level co-occurrence matrix features, 16 gray-level size region matrix features, 16 gray-level run matrix features, 5 adjacent gray-level difference matrix features, and 14 gray-level dependence matrix features.

[0098] (5.2) The above features are used as input to the hematoma segmentation model. After training, the output includes OC_MGS, which is the result of Markwalder Grading Scale, representing the prognosis; OC_mRS, which is the result of modified Rankin Scale, representing the prognosis; Radio_Cure, which is the radiological cure status at the endpoint; and Surgery, which is the clinical indicator of whether to switch to surgical treatment at the endpoint.

[0099] (5.3) Based on the above imaging features, the random forest algorithm is used to predict Radio_Cure, that is, whether radiological cure is achieved at the end point. Then, the above imaging features and the results of Radio_Cure are used as inputs, and the random forest algorithm is used again to predict OC_MGS, that is, the results of Markwalder Grading Scale. This process is repeated until the results of all indicators are predicted.

[0100] like Figure 4 As shown, we used PyTorch (v1.8.0) to design and implement the segmentation network, and trained it on a GeForce RTX 3090 GPU with 24GB of VRAM. We randomly divided the above data into training, validation, and test sets in an 8:1:1 ratio. During training, we segmented the training set into 64×64×64 segments with a 32×32×32 interval and input them into the network, retaining only positive data with at least one voxel hematoma. The loss function used during training is as follows:

[0101] L = L KL +L Dice (1)

[0102] Where L KL It is the cross-entropy loss, defined as follows:

[0103]

[0104] in It is the label of a specific voxel x. If this voxel belongs to category c, then otherwise X is the set of all voxels, θ is the trainable parameters in the network, and p c (x; θ) represents the probability that a certain voxel belongs to class c.

[0105] And L Dice It is the Dice loss, defined as follows:

[0106]

[0107] All the loss functions mentioned above are used to update the network parameters through the Adam optimizer, with the learning rate and batch size set to 0.001 and 8, respectively.

[0108] To predict outcomes related to clinical indicators, including OC_MGS (Markwalder Grading Scale result, representing prognosis), OC_mRS (modified Rankin Scale result, representing prognosis), Radio_Cure (radiological cure at the endpoint), and Surgery (whether conversion to surgery at the endpoint), we first used the Python-based PyRadiomics library to extract radiographic features from the raw MRI data and corresponding segmentation results (including 16 3D morphological features, 19 first-order features, 24 gray-level co-occurrence matrix (GLCM) features, 16 gray-level size-zone matrix (GLSZM) features, 16 gray-level run-length matrix (GLRLM) features, 5 adjacent gray-level difference matrix (NGTDM) features, and 14 gray-level dependence matrix (GLDM) features). Then, using these features as input and the corresponding clinical indicators as output, we employed a random forest algorithm for prognostic prediction. The dataset partitioning was the same as described above. Random forest is a classification algorithm that is robust to overfitting. We constructed 100 decision trees to form the prediction model. At each node of the trees, several candidate features were randomly selected from all discriminative features. In the testing step, given a test patient, each decision tree votes on the mutation status, and the patient's prognosis is determined by accumulating the votes of all decision trees. We first use the random forest algorithm described above to predict Radio_Cure based on the aforementioned imaging features, i.e., whether radiological cure is achieved at the endpoint. Then, we use the aforementioned imaging features and the Radio_Cure result as input, and again use the random forest algorithm as described before to predict OC_MGS, i.e., the Markwalder Grading Scale result. This iterative process continues until we predict the results of all indicators. For example, assuming we have already predicted the Radio_Cure result using random forest, we use this result together with the imaging features as input and again use random forest to predict the OC_MGS result. Through this iterative process, we can gradually predict other indicators, such as OC_mRS and Surgery. By using this hierarchical clinical indicator assessment algorithm based on iterative prediction, we can more accurately predict the clinical indicator outcomes of patients with chronic subdural hematoma and provide targeted treatment recommendations.

[0109] In practical applications, for the segmentation algorithm, we compared the segmentation results of 3D-UNet without using a data augmentation strategy for multi-device, multi-center data. We found that at the same center, a certain improvement was achieved on different devices (Dice coefficient: 0.8255 vs. 0.8146), and the improvement was more significant on data from different centers (0.8225 vs. 0.7774). Furthermore, we found that the segmentation results using the data augmentation strategy of this technical solution are more robust to more difficult scenarios, the segmented hematoma morphology is smoother, and it is closer to the gold standard, which helps with subsequent quantitative analysis.

[0110] This device enables prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction, wherein the device comprises:

[0111] A processor is configured to execute computer-executable instructions;

[0112] The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the above-described method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction.

[0113] The processor that implements chronic subdural hematoma prognostic analysis based on subdural hematoma segmentation and iterative prediction is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the aforementioned chronic subdural hematoma prognostic analysis method based on subdural hematoma segmentation and iterative prediction.

[0114] The computer-readable storage medium contains a computer program that can be executed by a processor to implement the steps of the above-described method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction.

[0115] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0116] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution device.

[0117] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0118] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0119] In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0120] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

[0121] This invention employs a method, device, processor, and computer-readable storage medium for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction. First, it uses 3D-UNet as the main segmentation framework and designs a data augmentation strategy robust to different scanning devices and centers across multiple centers. Subsequently, the algorithm extracts imaging features using high-quality segmentation results and employs random forest for feature selection and prognostic assessment. During this process, the algorithm uses relatively simple prognostic indicators to iteratively evaluate other more difficult prognostic indicators from easy to difficult, achieving more accurate hierarchical analysis. This allows for more comprehensive and sufficient application in clinical practice and has certain promotional value.

[0122] In this specification, the invention has been described with reference to specific embodiments thereof. However, it will be apparent that various modifications and variations can be made without departing from the spirit and scope of the invention. Therefore, the specification and drawings should be considered illustrative rather than restrictive.

Claims

1. A method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction, characterized in that, The method includes the following steps: (1) Collect the patient's MRI images and perform bias field correction and grayscale normalization data preprocessing to facilitate subsequent standardized analysis; (2) Perform data augmentation processing based on geometric transformation, intensity transformation and noise injection on the preprocessed data so that it can be used as training data for training the hematoma segmentation model. (3) Construct a 3D-UNet segmentation framework, use the acquired training data to segment chronic subdural hematoma, and output voxel-level predicted segmentation results. (4) The processed data is randomly divided into training set, validation set and test set according to a preset ratio, and the network is trained in combination with loss function in the segmentation network; (5) Extract imaging features from the original MRI data and the corresponding segmentation results, and use the random forest algorithm to evaluate and predict the clinical indicators of the corresponding patients with chronic subdural hematoma. Step (5) specifically includes the following steps: (5.1) The PyRadiomics library based on Python was used to extract radiographic features from the raw MRI data and corresponding segmentation results, specifically including: 16 3D morphological features, 19 first-order features, 24 gray-level co-occurrence matrix features, 16 gray-level size region matrix features, 16 gray-level run matrix features, 5 adjacent gray-level difference matrix features and 14 gray-level dependence matrix features; (5.2) The above features are used as inputs to the hematoma segmentation model. After training, the output includes OC_MGS, which is the result of Markwalder Grading Scale, representing the prognosis; OC_mRS, which is the result of modified Rankin Scale, representing the prognosis; Radio_Cure, which is the radiological cure status at the endpoint; and Surgery, which is the clinical indicator of whether to switch to surgical treatment at the endpoint. (5.3) Based on the above imaging features, the random forest algorithm is used to predict Radio_Cure, that is, whether radiological cure is achieved at the end point. Then, the above imaging features and the results of Radio_Cure are used as inputs, and the random forest algorithm is used again to predict OC_MGS, that is, the result of Markwalder Grading Scale. This process is repeated until the results of all indicators are predicted.

2. The method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction according to claim 1, characterized in that, Step (1) specifically includes the following steps: (1.1) Collect a preset number of MRI images and use 3D Slicer software to annotate the contours of the subdural hematoma to obtain hematoma segmentation labels; (1.2) The collected MRI image data are first subjected to bias field correction using the Python-based SimpleITK library to make the low-frequency gray distribution of the MRI images uniform. (1.3) The grayscale of the MRI image data is then normalized to between 0 and 1, and only the grayscale values ​​between 0.5% and 99.5% are retained to complete the data preprocessing.

3. The method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction according to claim 2, characterized in that, Step (2) specifically includes the following steps: (2.1) Data augmentation based on geometric transformation: Each original MRI grayscale image is randomly flipped along the coronal axis with a probability of 0.5 and randomly rotated in the cross section, wherein the rotation angle follows a uniform distribution within 0 to 45°. Each MRI grayscale image is also elastically deformed based on displacement grid with a probability of 0.1, thereby changing the brain structure as a whole. (2.2) Data augmentation based on intensity transformation: Each original MRI grayscale image has a probability of 0.5 to be multiplied by a random floating-point number between 0.5 and 1.5, and the grayscale values ​​of the image are cropped to the same range as the original grayscale values ​​to disturb the contrast of the image; (2.3) Data augmentation based on noise injection: Each original MRI grayscale image has a 0.1 probability of being added with random Gaussian noise with a mean of 0 and a standard deviation of 1; (2.4) The hematoma segmentation labels mentioned above are all transformed by data augmentation according to steps (2.1) to (2.3), and the grayscale images and hematoma segmentation labels after data augmentation are used for subsequent training of the segmentation network.

4. The method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction according to claim 3, characterized in that, The specific steps (2.1) are as follows: For each dimension of the MRI grayscale image, a displacement value is generated for each point on its displacement grid, thus obtaining a displacement value array. The displacement value array is added to the coordinate array corresponding to the position of each point on the grid to obtain a displacement matrix. The corresponding coordinates on the displacement matrix are sampled on the original MRI grayscale image, and B-spline interpolation is performed at non-integer coordinate positions to obtain the grayscale values ​​of the deformed image at each coordinate, thus obtaining the final deformed image.

5. The method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction according to claim 3, characterized in that, Step (3) specifically includes the following steps: (3.1) The grayscale image after data augmentation and its hematoma segmentation label are used as training data to train the hematoma segmentation model; (3.2) Constructing the 3D-UNet segmentation framework: It consists of a shrinking path and an expanding path, wherein, The described shrinkage path consists of four encoding modules, each consisting of two convolutional layers and one max-pooling layer. The main part of each convolutional layer is a 3×3×3 three-dimensional convolutional kernel; and Grouping normalization is used before each convolution to handle the internal covariance shift, and a modified linear unit activation function is implemented after the convolution. The shrinking of the path is used to progressively extract abstract features from the input grayscale data. The expansion path consists of four upsampling modules. Each upsampling module first uses nearest neighbor interpolation to upsample the grayscale image, making the grayscale image size twice its original size. Then, it is spliced ​​with the corresponding grayscale image in the contraction path along the channel dimension to form a skip connection, so that the information of the contraction path and the expansion path are merged without changing the grayscale image size. Subsequently, the grayscale image is convolved by two 3×3×3 three-dimensional convolution kernels to further extract and fuse the corresponding image data. (3.3) The segmentation network described above uses a 3×3×3 three-dimensional convolutional kernel to output voxel-level prediction results.

6. The method for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction according to claim 5, characterized in that, The specific steps (4) are as follows: The data obtained in the above steps are randomly divided into training, validation, and test sets in a ratio of 8:1:

1. During training, the training set is divided into 64×64×64 segments with a 32×32×32 interval and input into the training network, retaining only positive data with at least one voxel hematoma. The loss function used in the training process is... as follows: in, The cross-entropy loss is defined as follows: in, For a specific voxel The label, if this voxel belongs to category c, then ,otherwise ; For the set of all voxels, These are the trainable parameters in the network. This represents the probability that a certain voxel belongs to class c; The Dice loss is defined as follows: All the loss functions mentioned above are used to update the network parameters through the Adam optimizer, with the learning rate and batch size set to 0.001 and 8, respectively.

7. A device for prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction, characterized in that, The device includes: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the chronic subdural hematoma prognostic analysis method based on subdural hematoma segmentation and iterative prediction as described in any one of claims 1 to 6.

8. A processor for implementing prognostic analysis of chronic subdural hematoma based on subdural hematoma segmentation and iterative prediction, characterized in that, The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the chronic subdural hematoma prognostic analysis method based on subdural hematoma segmentation and iterative prediction as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the chronic subdural hematoma prognostic analysis method based on subdural hematoma segmentation and iterative prediction as described in any one of claims 1 to 6.

Images