Artificial intelligence auxiliary diagnosis and treatment method and device based on deep learning

Abstract

The invention relates to an artificial intelligence auxiliary diagnosis and treatment method and device based on deep learning, and the method comprises the steps: obtaining a to-be-processed computed tomography image corresponding to a to-be-diagnosed and treated user; inputting the to-be-processed computed tomography image into a pre-trained deep learning model for segmentation processing to obtain a hematoma area; the longest axis of the hematoma is determined based on the hematoma area, and a puncture path from the puncture safety area to the longest axis of the hematoma is generated based on comparison between the longest axis of the hematoma and the set puncture safety area so that processing can be conducted based on the puncture path. Therefore, the hematoma region segmentation of the computed tomography image is realized through deep learning, and the puncture path along the longest axis of the hematoma with the starting point located in the safety region is generated, so that the efficiency and effect of artificial intelligence assisted diagnosis and treatment are further improved.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data processing, and particularly to an artificial intelligence-assisted diagnosis and treatment method and device based on deep learning. Background Art

[0002] Spontaneous intracerebral hemorrhage (ICH) is a common and frequently-occurring disease, and is a serious stroke with a high mortality rate and disability rate. Even if the patient survives, most patients will still suffer from long-term nerve damage due to reasons such as cerebral edema and intracranial hypertension.

[0003] Computed Tomography (CT) is the standard imaging tool for the diagnosis, treatment planning, and prognosis evaluation of ICH. Traditionally, radiologists estimate the volume of the hematoma by manually delineating the hematoma area to predict the patient's risk of death and functional recovery. However, this manual segmentation is not only time-consuming and has inter-observer differences, but also requires highly specialized skills. In terms of surgical planning, it completely depends on the personal experience of the surgeon for manual planning. It has high requirements for personal experience and is not accurate enough. Summary of the Invention

[0004] In order to solve the above technical problems or at least partially solve the above technical problems, the present disclosure provides an artificial intelligence-assisted diagnosis and treatment method and device based on deep learning.

[0005] The present disclosure provides an artificial intelligence-assisted diagnosis and treatment method based on deep learning. The method includes: obtaining a to-be-processed computed tomography image corresponding to a user to be diagnosed and treated; inputting the to-be-processed computed tomography image into a pre-trained deep learning model for segmentation processing to obtain a hematoma area; determining the longest axis of the hematoma based on the hematoma area, and comparing the longest axis of the hematoma with a set puncture safety zone to generate a puncture path from the puncture safety zone to the longest axis of the hematoma for processing based on the puncture path.

[0006] The present disclosure provides an artificial intelligence-assisted diagnosis and treatment device based on deep learning. The device includes: an obtaining module, configured to obtain a to-be-processed computed tomography image corresponding to a user to be diagnosed and treated; a processing module, configured to input the to-be-processed computed tomography image into a pre-trained deep learning model for segmentation processing to obtain a hematoma area; a determining module, configured to determine the longest axis of the hematoma based on the hematoma area; and a generating module, configured to compare the longest axis of the hematoma with a set puncture safety zone to generate a puncture path from the puncture safety zone to the longest axis of the hematoma for processing based on the puncture path.

[0007] According to an embodiment of the present disclosure, an electronic device is provided. The electronic device includes: a processor; a memory for storing executable instructions executable by the processor; and the processor is configured to read the executable instructions from the memory and execute the instructions to implement the artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by the embodiments of the present disclosure.

[0008] According to an embodiment of the present disclosure, a computer-readable storage medium is provided. The storage medium stores a computer program, and the computer program is used to execute the artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by the embodiments of the present disclosure.

[0009] Embodiments of the present disclosure further provide a computer program product, including a computer program, and the computer program is executed by a processor to implement the artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by the embodiments of the present disclosure.

[0010] The technical solutions provided by the embodiments of the present disclosure have the following advantages compared with the prior art:

[0011] The present disclosure proposes an artificial intelligence-assisted diagnosis and treatment solution based on deep learning, which obtains a to-be-processed computed tomography (CT) image corresponding to a user to be diagnosed and treated; inputs the to-be-processed CT image into a pre-trained deep learning model for segmentation processing to obtain a hematoma region; determines the longest axis of the hematoma based on the hematoma region, and compares the longest axis of the hematoma with a set puncture safety zone to generate a puncture path from the puncture safety zone to the longest axis of the hematoma for processing based on the puncture path. Thus, through deep learning, the segmentation of the hematoma region in the CT image is realized, and a puncture path along the longest axis of the hematoma with the starting point in the safety zone is generated, further improving the efficiency and effect of artificial intelligence-assisted diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings herein are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure.

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0014] Figure 1 It is a flowchart example of an artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by an embodiment of the present disclosure;

[0015] Figure 2 It is an example diagram of an artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by an embodiment of the present disclosure;

[0016] Figure 3 It is an exemplary diagram of another artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by an embodiment of the present disclosure;

[0017] Figure 4 It is a schematic structural diagram of an artificial intelligence-assisted diagnosis and treatment device based on deep learning provided by an embodiment of the present disclosure. Detailed implementation manners

[0018] In order to be able to more clearly understand the above-mentioned objects, features, and advantages of the present disclosure, the solutions of the present disclosure will be further described below. It should be noted that, without conflict, the embodiments of the present disclosure and the features in the embodiments may be combined with each other.

[0019] In the following description, many specific details are set forth in order to fully understand the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only a part of the embodiments of the present disclosure, rather than all the embodiments.

[0020] Specifically, great progress has been made in the field of medical image segmentation by deep learning, which can quickly and accurately complete complex tasks, approaching or even exceeding the level of human experts. It can complete organ segmentation tasks such as brain tumor segmentation; ischemic lesion segmentation; lung tumor segmentation; cardiac anatomy segmentation; pancreatic segmentation.

[0021] In the field of stroke management, lesion segmentation has received increasing attention because it can efficiently process massive amounts of data and assist clinicians in making complex decisions. Although there are already multiple segmentation models based on 2D / 3D CNN (Convolutional Neural Networks) and CRF (Conditional Random Field) for ischemic stroke lesion detection, current methods rarely target intracerebral hemorrhage (ICH) segmentation in CT images. At the same time, the vast majority of existing deep learning artificial intelligence methods are applied in fields such as lesion detection, delineation, and prognosis prediction, and their functions focus on diagnosis, while there are almost no applications in surgical method selection and surgical path planning, contributing little to surgical intervention.

[0022] In view of the above problems, the present disclosure proposes an artificial intelligence-assisted diagnosis and treatment method based on deep learning, which uses a brain tissue mask training strategy for automatic ICH segmentation; then performs three-dimensional reconstruction of the hematoma and automatically generates the longest axis of the hematoma; finally, compares it with the set puncture safety zone to generate a puncture path along the longest axis of the hematoma starting from the puncture safety zone to assist surgeons in performing hematoma puncture and catheterization operations. Thus, through deep learning, the hematoma area in the computed tomography image is segmented, and a puncture path along the longest axis of the hematoma starting from the safety zone is generated, further improving the efficiency and effect of artificial intelligence-assisted diagnosis and treatment.

[0023] Figure 1 FIG. is a flowchart example of an artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by an embodiment of the present disclosure. The relative positions of the lidar and the inertial measurement unit IMU remain unchanged, and it includes:

[0024] Step 101, obtaining a to-be-processed computed tomography image corresponding to a user to be diagnosed and treated.

[0025] Step 102, inputting the to-be-processed computed tomography image into a pre-trained deep learning model for segmentation processing to obtain a hematoma area.

[0026] Step 103, determining the longest axis of the hematoma based on the hematoma area, and comparing the longest axis of the hematoma with the set puncture safety zone to generate a puncture path from the puncture safety zone to the longest axis of the hematoma for processing based on the puncture path.

[0027] Specifically, a deep learning model is pre-trained, and then a computed tomography image corresponding to an authorized user to be diagnosed and treated is obtained and input into the deep learning model for segmentation processing to obtain a hematoma area, determining the longest axis of the hematoma based on the hematoma area, and comparing the longest axis of the hematoma with the set puncture safety zone to generate a puncture path from the puncture safety zone to the longest axis of the hematoma for processing based on the puncture path.

[0028] In some embodiments, multiple computed tomography image samples are obtained, preprocessed to obtain training data and test data, the training data is input into a pre-constructed deep learning network for training to obtain the deep learning model, and the deep learning model is tested based on the test data to obtain the test results corresponding to the deep learning model.

[0029] In some embodiments, preprocessing multiple computed tomography (CT) image samples to obtain training data and test data includes: isotropically resampling the volume data of each CT image sample through trilinear interpolation; adjusting the window width of each CT image sample to a target range; detecting each adjusted CT image sample based on an adaptive binary segmentation algorithm of image gray distribution to maximize the difference in pixel value distributions between the two categories of image foreground and image background in each CT image sample; performing three-dimensional connected component analysis on each detected CT image sample, and expanding a preset distance outward when the image edge is detected by the bounding box to obtain the CT image sample to be divided; and dividing the CT image sample to be divided according to a preset ratio to obtain training data and test data.

[0030] Specifically, the volume data of all CT images are isotropically resampled through trilinear interpolation (for example, the voxel spacing is 1×1×1 mm 3 ), and the window width of the CT image is adjusted to the range of [-100, 200]. The foreground region is automatically detected using an adaptive binary segmentation algorithm of image gray distribution, namely the Otsu threshold method (determining the optimal threshold by maximizing the between-class variance), to maximize the difference in pixel value distributions between the two categories of foreground and background in the image. Subsequently, three-dimensional connected component analysis is performed (a process of grouping points or voxels in three-dimensional space according to their connectivity, for example, the minimum threshold is 500 voxels). The bounding box expands a preset distance, such as 20 mm, outward based on the detected edge to retain the anatomical structure context information. All data are randomly grouped according to the preset ratio, with 80% of the data used for training and 20% of the data used for testing.

[0031] Specifically, the purpose of Z-score normalization is to represent the original data points as multiples of the standard deviation by which they deviate from the mean, so that data of different scales can be compared. The normalization in the processing of image I can be expressed as formula (1).

[0032]

[0033] where μ I is the average pixel value of the image, and σ I is the standard deviation of the image pixel values.

[0034] In some embodiments, the deep learning network includes a contracting path and an expanding path; training data is input into a pre-built deep learning network for training to obtain a deep learning model, including: inputting a computed tomography image of the training data into the contracting path for multi-scale feature transformation, hierarchical downsampling, and dynamic channel expansion to obtain a feature map to be processed; inputting the feature map to be processed into the expanding path for transposed convolutional upsampling, cross-layer connection, and feature fusion to obtain a fused feature map; inputting the fused feature map into a classification function to obtain a classification result, and continuously adjusting the network parameters of the deep learning network based on the classification result, the pre-annotated result, and a preset loss function to obtain a deep learning model.

[0035] Specifically, the deep learning network is a U-net architecture, and its core structure consists of a symmetric contracting path and an expanding path, which achieve multi-scale feature fusion through cross-layer connection. The network topology and parameter configuration can be systematically described as follows.

[0036] Specifically, for feature extraction in the contracting path, the contracting path, as the encoder part of the network, specifically includes: a multi-scale feature transformation unit. Each level of the feature transformation unit consists of two cascaded convolutions and activation combinations, which can be formally represented by formula (2).

[0037]

[0038] Among them, and respectively represent the input and output feature maps of the l-th layer; W l (1) and W l (2) are 3×3×3 convolutional kernel parameter matrices; σ(x) represents the ReLU activation function: σ(x) = max(0, x). * represents a three-dimensional convolution operation, and a non-padding method with a stride of 1 is adopted.

[0039] Thus, the preservation of spatial resolution is ensured, and at the same time, the expression ability of the model is enhanced through double non-linear transformation. In particular, the choice of non-padding convolution stems from the uncertainty of the edge information of medical images, avoiding the bias that may be introduced by artificial padding.

[0040] Specifically, the contracting path also includes a hierarchical downsampling strategy. The downsampling module uses a max-pooling operation, which can be formally expressed by formula (3).

[0041]

[0042] Specifically, the significance of the above hierarchical downsampling strategy is that after each downsampling, the receptive field area of the network increases by 4 times and the volume increases by 8 times, enabling deep features to capture a wider range of anatomical structure relationships. At the same time, the decreasing spatial dimension of the feature map (halving in each dimension) allows computing resources to be concentrated on deeper feature learning. In 3D medical image processing, this strategy is particularly important because it can effectively capture long-range dependencies between voxels, especially for the segmentation tasks of anatomical tissues with complex topological structures.

[0043] Specifically, the contracting path also includes a dynamic channel expansion mechanism. To compensate for the loss of spatial information, this architecture implements a channel expansion strategy, which can be formally expressed as formula (4).

[0044] C l+1 = 2 × C l (4).

[0045] Where C l represents the number of feature channels in the l-th layer. The theoretical basis for channel expansion lies in the information preservation theory, that is, by increasing the feature dimension to balance the information loss caused by spatial downsampling. In multi-modal medical image segmentation, this mechanism is particularly important because it can retain complementary information from different modalities, such as the mutually complementary pathological features in T1, T2, and FLAIR sequences.

[0046] Specifically, in the feature reconstruction process of the expanding path, as the decoder part of the network, its core function is to re-project the abstract semantic features back to the original spatial resolution, while fusing multi-scale context information to achieve accurate voxel-level segmentation.

[0047] The expanding path specifically includes adaptive transposed convolution upsampling. The upsampling module uses parameterized learned transposed convolution instead of a fixed interpolation method, which can be formally expressed as formula (5).

[0048]

[0049] Where represents the transposed convolution operation, is a learnable 2×2×2 transposed convolution kernel. This parameterized upsampling has obvious advantages compared to bilinear or nearest neighbor interpolation. The transposed convolution parameters are learned through backpropagation, enabling the upsampling process to adapt to specific data. The learned upsampling filter can perform adaptive upsampling according to the surrounding anatomical structure features, and can better maintain the continuity and topological integrity of the anatomical structure during the reconstruction process. At the same time, the halving of the number of channels (C l = C l+1 / 2) realizes a gradual conversion from abstract features to specific spatial features while controlling the computational complexity.

[0050] The extended path specifically also includes skip connections, which are the core innovation of the U-shaped network. By fusing the high-resolution features in the encoder with the upsampled features in the decoder, it can be formally expressed as formula (6).

[0051]

[0052] Based on the multi-scale feature complementarity principle mentioned above, it has important theoretical significance in medical image segmentation. It promotes a more stable gradient optimization process and alleviates the vanishing gradient problem in deep networks. High-resolution features provide precise spatial localization information, making up for the loss of boundary details in the downsampling process and promoting the recovery of fine boundaries. At the same time, it completes multi-scale context integration, effectively combines local details with global semantics, and promotes the recognition of segmentation structures at different scales. This multi-scale feature fusion mechanism significantly improves the segmentation accuracy of lesions of different sizes, especially for lesion areas with blurred boundaries and variable shapes.

[0053] The extended path specifically also includes a cascaded refinement module. The feature refinement module further integrates and refines the fused features through two consecutive convolutional operations, which can be formally expressed as formula (7).

[0054]

[0055] The above cascaded design enables the network to enhance the expression of fused features through non-linear transformation, effectively fusing multi-source information. Thus, it eliminates the discontinuity of image features and smoothly connects the features filtered by the encoder and decoder. It improves the recognition ability of the boundaries of the parts to be segmented and enhances boundary sensitivity. Experiments show that compared with single convolution, this cascaded refinement design significantly improves the segmentation accuracy, especially for areas with complex boundaries.

[0056] Specifically, the deep learning network also has a network topology structure and global characteristics, which reflect the structured inductive bias in the deep network. By constraining the topology structure, it guides the network to learn the inherent characteristics of medical images, specifically including a classification mapping and decision-making mechanism. At the end of the network, 1×1×1 convolution is used to achieve the mapping from features to the semantic space, which can be formally expressed as formula (8).

[0057] Y = SoftMax(W final *F final +b final ) (8).

[0058] Among them, Y represents the pixel-level probability distribution map, and the SoftMax function ensures the probability normalization of multi-class segmentation. This design enables the network to achieve an effective mapping from the high-dimensional feature space to the low-dimensional semantic space while maintaining the spatial structure. Compared with the fully connected layer, the 1×1×1 convolution preserves the spatial relationship, significantly reduces the number of parameters, and improves the computational efficiency and generalization ability.

[0059] The network topology structure and global characteristics specifically also include a symmetric architecture. The symmetric design of the network follows the information flow symmetry principle. The encoding process goes from concrete to abstract, with the spatial resolution decreasing and the number of channels increasing. The decoding process goes from abstract to concrete, with the spatial resolution increasing and the number of channels decreasing. This symmetry ensures the balance between the feature extraction and reconstruction processes, forming a complete closed-loop of "feature abstraction to context integration to fine reconstruction". This architecture is superior to the asymmetric design in medical image segmentation, especially when dealing with multi-tissue interfaces and complex morphological structures.

[0060] The network topology structure and global characteristics specifically also include the seamless segmentation and large-scale processing strategy. To achieve seamless processing of medical volume data of any size, this architecture adopts an overlap-tile strategy. By ensuring that each downsampling operation is applied to a feature map with an even dimension, fully deterministic boundary processing is achieved, avoiding the boundary inconsistency problem in the traditional sliding window method, which can be formally expressed as formula (9).

[0061]

[0062] Among them, represents the input block with overlap, f() represents the forward propagation of the network, and Stitch() represents the fusion operation of the overlapping area. Through block overlap and special fusion strategies, the global consistency and coherence of the segmentation results are ensured.

[0063] Specifically, after using the U-shaped network as the underlying architecture, the system independently configures, designs, and trains three U-net models: 2D U-Net, 3D U-Net, and two cascaded 3D U-Net models (where the first one generates a preliminary segmentation at a low resolution, and then is refined by the second model). The main differences from the original U-Net architecture are: using padding convolution to achieve the same input and output shapes, using instance normalization instead of batch normalization, and using Leaky ReLU instead of the standard ReLU. This solves the situation where some image data is always silent (i.e., the input is always negative) during the training process.

[0064] Specifically, according to the shape of the preprocessed training data, the system adaptively adjusts the specific instance of the U-Net. Specifically, nnU-Net automatically sets the batch size, patch size, and the number of pooling operations for each axis under a specific memory budget. To maximize the capture of spatial context information, the system prefers a large patch size rather than a large batch size (the minimum batch size is 2). The pooling operation along each axis continues until further pooling would reduce the spatial size of that axis to less than 4 voxels. All U-Net architectures use 30 convolutional filters in the first layer and double this number after each pooling operation. If the selected patch size covers less than 25% of the voxels of a typical training case, a 3D U-Net cascade model is additionally configured and trained based on the downsampled version of the training data. This cascade design aims to enable nnU-Net to obtain sufficient context information when the patch size is too small to cover the full resolution.

[0065] Specifically, all U-Net architectures are trained using five-fold cross-validation. One training epoch is defined as processing 250 batches. The loss function is the sum of the cross-entropy loss and the Dice loss. The optimizer uses the Stochastic Gradient Descent (SGD) algorithm for stochastic gradient descent, which can be formally expressed as formula (10).

[0066]

[0067] Where, θt represents the model parameters at the t-th iteration, η is the learning rate, represents the gradient of the loss function calculated on a randomly selected sample (x (i) , y (i) ). This disclosure adopts a variant of SGD with momentum, and its update rule is modified to formula (11) and formula (12).

[0068]

[0069] θ t+1 = θ t + v t+1 (12).

[0070] Where, μ ∈ [0, 1] is the momentum coefficient, which is set to 0.99 in this disclosure, and Vt is the cumulative momentum vector. This design can effectively suppress the oscillation in the parameter update process, accelerate convergence, and help escape from local minima.

[0071] As an example, the initial learning rate is set to 0.01, and the weight decay is set to 3e -5When the exponential moving average of the training loss has not improved for 30 consecutive epochs, the learning rate is reduced to 0.2 times the original. When the learning rate is lower than 10^-6 or exceeds 1000 training epochs, the training stops. Finally, the present disclosure uses a sliding window method to predict cases, with an overlap of half the patch size between windows. This increases the prediction weight for positions closer to the center relative to the edges. During testing, data augmentation is achieved by mirror flipping along all axes. The overall process is as Figure 2 shown.

[0072] The embodiments of the present disclosure adopt two complementary quantitative evaluation metrics: the Dice Similarity Coefficient (DSC) and the 95% Hausdorff Distance (95HD) to comprehensively evaluate the segmentation performance of the model. These two metrics evaluate the segmentation quality from the perspectives of overlap and spatial distance respectively, and jointly provide a multi-dimensional quantitative analysis of the segmentation results.

[0073] In some embodiments, determining the longest axis of the hematoma based on the hematoma region includes: obtaining the total number of voxels in the hematoma region and the volume of each voxel; multiplying the total number of voxels by the volume of each voxel to obtain the hematoma volume; determining the longest axis of the hematoma based on the hematoma volume.

[0074] Specifically, the present disclosure uses a direct calculation method of the segmentation volume to calculate the hematoma volume. Based on the complete three-dimensional voxel space, by identifying and calculating all voxels belonging to the target structure, and then multiplying the number of voxels by the actual volume of a single voxel to obtain the true volume of the target. The calculation of the single voxel volume can be expressed as formula (13).

[0075] V voxel = dx·dy·dz (13).

[0076] Wherein, dx, dy, and dz represent the voxel resolutions of the voxels in three directions (unit: mm).

[0077] As an example of a scenario, the classical nnUNet architecture is adopted. This architecture uses a symmetric design and consists of a contracting path (left side) and an expanding path (right side). The contracting path follows the structure of a traditional convolutional network, and through repeated application of two consecutive 3×3 convolutions (without padding), followed by a rectified linear unit (ReLU) activation function after each convolution, and then 2×2 max pooling (stride of 2) for downsampling. At each downsampling stage, the number of feature channels is doubled.

[0078] In the expansion path, at each stage, the feature map is first upsampled, and then the number of feature channels is halved using a 2×2 "transposed convolution". The upsampled feature map is concatenated with the corresponding cropped feature map in the contraction path, followed by two consecutive 3×3 convolutions, each followed by a ReLU activation. The cropping operation is to compensate for the loss of edge pixels during each convolution. In the last layer, a 1×1 convolution is used to map each 64-dimensional feature vector to the desired number of classes. The entire network consists of 23 convolutional layers.

[0079] To achieve seamless stitching of the output segmentation map (as shown in Figure 3 ), it is necessary to select an appropriate input patch size to ensure that all 2×2 max-pooling operations are performed on layers where both the x and y directions are even numbers.

[0080] Using this U-Net architecture, automatic hematoma segmentation, automatic skull segmentation, and automatic eyeball segmentation can be achieved.

[0081] For example, for image preprocessing and semantic segmentation, the dataset is randomly divided into a training set (80%, 277 patients) and a test set (20%, 70 patients) according to patient-level identifiers to prevent data leakage. All CT volumes are isotropically resampled using trilinear interpolation (1×1×1mm 3 voxel spacing), and the CT image window width is adjusted to [-100, 200]. The gray values are normalized using global Z-score normalization. The foreground region is automatically detected by Otsu thresholding, followed by 3D connected component analysis (minimum 500-voxel threshold). The bounding box is extended 20mm outside the detected edges to preserve the anatomical context. The nnU-Net architecture is implemented in PyTorch, using stochastic gradient descent optimization (momentum = 0.99, initial learning rate = 0.01, cosine annealing). Input patches (1792×768×32 voxels) are sampled using weighted foreground sampling. The model performance is evaluated through 5-fold cross-validation and using composite metrics, including the Dice similarity coefficient (DSC) and the Hausdorff distance (HD). The segmentation output is converted to the NIfTI format using the NiBabel package. Three-dimensional reconstruction is generated by the ray casting algorithm in Mayavi, combined with depth shading and anatomical label overlay. The visualization process includes: multi-planar viewer: simultaneously navigate the axial / coronal / sagittal planes; isosurface rendering: threshold-based mesh generation (HU > 40); measurement tool: automatic hematoma volume calculation.

[0082] In some embodiments, a first anatomical landmark boundary, a second anatomical landmark boundary, and a third anatomical landmark boundary are determined; a puncture safety zone is determined based on the first anatomical landmark boundary, the second anatomical landmark boundary, and the third anatomical landmark boundary.

[0083] Specifically, Region 1 (frontal lobe approach area): This triangular area is defined by three anatomical landmarks: The first anatomical landmark boundary P1: The mid-sagittal plane is translated 20 mm towards the hematoma side; the axial plane located 5 mm above the frontal sinus; Adjustment: If the frontal sinus level is lower than the supraorbital margin, the supraorbital margin plane is used as the lower boundary. The second anatomical landmark boundary P2: Project from P1 towards the vertex along the cranial surface (vertical displacement of 30 mm). The third anatomical landmark boundary: Project from P1 towards the lateral side along the cranial surface (horizontal displacement of 30 mm).

[0084] In some embodiments, the target center is confirmed through surface curvature analysis, and a circular area with a preset radius centered on the target center is used as the puncture safety zone.

[0085] Specifically, Region 2 (parietal lobe approach area): Centered on the maximum convex point of the parietal tuber (confirmed through surface curvature analysis, such as mean curvature > 0.4 mm-1), for example, a circular area with a radius of 15 mm.

[0086] Specifically, these two safety zones are selected based on their ability to avoid key neuroanatomical sensitive structures to the greatest extent, including important functional areas, complex bony landmarks, major cerebral arteries, and major draining veins / sinuses.

[0087] Specifically, for the development of the correction plane, eye segmentation and landmark point registration, eye segmentation is achieved through nnUNet using the public dataset SegRap2023 (192 cases for training and 48 cases for validation). After segmentation, the bilateral horizontal axis of eye symmetry is defined as the corrected axial plane, and the vertical midline is used as the corrected sagittal reference line.

[0088] Specifically, the marked safety zones and correction planes are integrated into iterative learning. The custom loss function includes a combination of Dice loss and cross-entropy (CE) loss, and Focal Loss is added as an enhancement on the original nnUNet framework. The calculation formula for the hybrid loss function is: L = 0.5LDice + 0.3LCE + 0.2Lfocal.

[0089] Specifically, candidate trajectories are generated through voxel space vector analysis between the hematoma centroid (Euclidean coordinates) and the entry point of the predefined safety zone. The optimal trajectory is selected according to two criteria: the maximum hematoma crossing length (the path with the most voxels) and the minimum distance from adjacent critical structures (such as safety margin ≥ 3 mm).

[0090] Specifically, feasible trajectories are identified by the user based on a trajectory evaluation scoring system from four aspects, and low-risk trajectories (average score ≥ 15) are retained for surgical consideration. The specific scoring criteria are based on the principles of neurosurgical path planning. The reviewer consistency between the calculation plan and the evaluation is evaluated through intraclass correlation analysis (ICC = 0.91, 95% CI 0.87 - 0.94).

[0091] Based on the above description, intracerebral hemorrhage (ICH) is a neurosurgical emergency that brings significant mortality and long-term functional disability. Minimally invasive puncture and catheter drainage of hematoma remains one of the main methods for minimally invasive surgical treatment of spontaneous intracerebral hemorrhage, which can rapidly reduce intracranial pressure with very little trauma. This result has been verified through comparison based on stereotactic catheter drainage in clinical trials, showing a reduction in the socioeconomic burden and improvement in long-term functional preservation. However, due to the pathological and anatomical complexity of the hematoma morphology - multilobular structures, septal areas, and dendritic projections often obscure the spatial relationship with important subcortical white matter pathways and perforating blood vessels - the optimization of the surgical trajectory is severely affected, especially in deep hemorrhages involving the basal ganglia. Although a semi-automatic hematoma segmentation method has been developed using dedicated software, due to the professional requirements of multimodal image fusion and computational planning, this method is difficult to popularize in primary hospitals in China. And in resource-limited settings, this capacity gap is particularly evident because 68% of ICH cases globally occur in such areas. Emerging AI solutions have the technical potential to solve this problem. Although deep learning architectures have a high accuracy in hemorrhage detection in neuroimaging diagnosis, their application in surgical guidance or intervention is relatively scarce. In the past, attempts have been made to plan the puncture path after classifying the hematoma into intraparenchymal and intraventricular types, and the path mainly points to the centroid of the hematoma. However, due to the fact that this method does not follow the established principles of neurosurgical path planning, its clinical applicability is limited. Based on this, the concept of a safe puncture area for the hematoma is introduced to clarify the target endpoint and design the puncture trajectory.

[0092] Therefore, the present disclosure uses deep learning for ICH segmentation based on CT scan images; by sampling pixels only in the brain tissue region for training, the irrelevant background in the CT scan is eliminated, improving the convergence speed and segmentation accuracy; using balanced class pixel sampling to solve the problem of class imbalance in the dataset; combining atrous convolution, Dice loss function, and 3D CRF to improve the segmentation performance of the model; automatically defining the safe areas (frontal, parietal, and calvarial parts) for puncture operations, and generating a puncture path along the longest axis of the hematoma with the starting point located within the safe area.

[0093] Corresponding to the aforementioned artificial intelligence-assisted diagnosis and treatment method based on deep learning, an embodiment of the present disclosure provides an artificial intelligence-assisted diagnosis and treatment device based on deep learning, Figure 4 which is a schematic structural diagram of an artificial intelligence-assisted diagnosis and treatment device provided by an embodiment of the present disclosure. This device can be implemented by software and / or hardware, and is generally integrated in an electronic device, such as Figure 4As shown, the mine road network includes multiple loading areas, and there are multiple electric shovels in each of the loading areas. The relative positions of the lidar and the inertial measurement unit IMU remain unchanged. The artificial intelligence-assisted diagnosis and treatment device 400 based on deep learning includes the following modules:

[0094] An acquisition module 401, configured to acquire a to-be-processed computed tomography image corresponding to a user to be diagnosed and treated;

[0095] A processing module 402, configured to input the to-be-processed computed tomography image into a pre-trained deep learning model for segmentation processing to obtain a hematoma region;

[0096] A determination module 403, configured to determine the longest axis of the hematoma based on the hematoma region;

[0097] A generation module 404, configured to compare the longest axis of the hematoma with a set puncture safety zone, and generate a puncture path from the puncture safety zone to the longest axis of the hematoma for processing based on the puncture path.

[0098] In some embodiments, the device further includes: an acquisition sample module, configured to acquire a plurality of computed tomography image samples; a preprocessing module, configured to preprocess the plurality of computed tomography image samples to obtain training data and test data; a training module, configured to input the training data into a pre-constructed deep learning network for training to obtain the deep learning model; a testing module, configured to test the deep learning model based on the test data to obtain a test result corresponding to the deep learning model.

[0099] In some embodiments, the preprocessing module is specifically configured to: perform isotropic resampling on the volume data of each computed tomography image sample by trilinear interpolation; adjust the window width of each computed tomography image sample to a target range; perform detection on each adjusted computed tomography image sample based on an adaptive binary segmentation algorithm for image gray distribution to maximize the pixel value distribution difference between the two categories of the image foreground and the image background in each computed tomography image sample; perform three-dimensional connected component analysis on each detected computed tomography image sample, and expand a preset distance outward when the image edge is detected in the bounding box to obtain a to-be-partitioned computed tomography image sample; partition the to-be-partitioned computed tomography image sample according to a preset ratio to obtain the training data and the test data.

[0100] In some embodiments, the deep learning network includes a contraction path and an expansion path; the training module is specifically configured to: input the computed tomography image of the training data into the contraction path for multi-scale feature transformation, hierarchical downsampling, and dynamic channel expansion to obtain a feature map to be processed; input the feature map to be processed into the expansion path for transposed convolutional upsampling, cross-layer connection, and feature fusion to obtain a fused feature map; input the fused feature map into a classification function to obtain a classification result, and continuously adjust the network parameters of the deep learning network based on the classification result, the pre-annotated result, and a preset loss function to obtain the deep learning model.

[0101] In some embodiments, the determining module 403 is specifically configured to: obtain the total number of voxels in the hematoma region and the volume of each voxel; multiply the total number of voxels by the volume of each voxel to obtain the hematoma volume; determine the longest axis of the hematoma based on the hematoma volume.

[0102] In some embodiments, the device further includes: a first safety zone determining module, configured to determine a first anatomical landmark boundary, a second anatomical landmark boundary, and a third anatomical landmark boundary; determine the puncture safety zone based on the first anatomical landmark boundary, the second anatomical landmark boundary, and the third anatomical landmark boundary.

[0103] In some embodiments, the device further includes: a second safety zone determining module, configured to confirm the target center through surface curvature analysis; use a circular region with a preset radius centered at the target center as the puncture safety zone.

[0104] The artificial intelligence-assisted diagnosis and treatment device based on deep learning provided by the embodiments of the present invention can execute the artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by any embodiment of the present invention, and has corresponding functional modules and beneficial effects for executing the method.

[0105] According to an embodiment of the present disclosure, an electronic device is provided, the electronic device includes: a processor; a memory for storing executable instructions executable by the processor; the processor is configured to read the executable instructions from the memory and execute the instructions to implement the artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by the embodiment of the present disclosure.

[0106] The embodiment of the present disclosure further provides a storage medium containing computer-executable instructions, and the computer-executable instructions are used to implement the artificial intelligence-assisted diagnosis and treatment method provided by the embodiment of the present invention when executed by a computer processor.

[0107] Of course, for the storage medium containing computer-executable instructions provided by the embodiments of the present invention, the computer-executable instructions are not limited to the method operations described above, and can also execute the related operations in the artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by any embodiment of the present invention.

[0108] Embodiments of the present disclosure also provide a computer program product, including a computer program, where the computer program, when executed by a processor, implements the artificial intelligence-assisted diagnosis and treatment method based on deep learning provided by the embodiments of the present disclosure.

[0109] From the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software and necessary general-purpose hardware. Of course, it can also be implemented by hardware, but in many cases the former is a better implementation. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a flash memory (FLASH), a hard disk, or an optical disc of a computer, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0110] It should be noted that in the embodiments of the above search device, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of mutual distinction and do not limit the protection scope of the present invention.

[0111] It should be noted that in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article or device including a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0112] The above are only specific embodiments of the present disclosure, enabling those skilled in the art to understand or implement the present disclosure. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure will not be limited to these embodiments described herein, but rather to the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. An artificial intelligence-assisted diagnosis and treatment method based on deep learning, characterized in that: The method comprises: Obtaining a to-be-processed computer tomography image corresponding to the user to be diagnosed and treated; Inputting the to-be-processed computed tomography image into a pre-trained deep learning model for segmentation processing to obtain a hematoma area; The longest axis of the hematoma is determined based on the hematoma area, and based on the comparison between the longest axis of the hematoma and the set puncture safety zone, a puncture path from the puncture safety zone to the longest axis of the hematoma is generated for processing based on the puncture path.

2. The artificial intelligence-assisted diagnosis and treatment method based on deep learning according to claim 1, characterized in that: The method further comprises: acquiring a plurality of computed tomography image samples; Preprocessing the plurality of computed tomography image samples to obtain training data and test data; Inputting the training data into a pre-built deep learning network for training to obtain the deep learning model; The deep learning model is tested based on the test data to obtain a test result corresponding to the deep learning model.

3. The artificial intelligence-assisted diagnosis and treatment method based on deep learning according to claim 2, characterized in that: The preprocessing of the plurality of computer tomography image samples to obtain training data and test data comprises: isotropically resampling the volume data of each of the computed tomography image samples by trilinear interpolation; adjusting the window width of each of the computed tomography image samples to a target range; Using an adaptive binary segmentation algorithm based on image grayscale distribution, each of the adjusted computed tomography image samples is detected to maximize the difference in pixel value distribution between the image foreground and image background categories in each of the computed tomography image samples; Performing a three-dimensional connected component analysis on each of the CT image samples after detection, and expanding outwards by a preset distance when the image edge is detected by the bounding box, to obtain CT image samples to be divided; The computed tomography image samples to be divided are divided according to a preset ratio to obtain the training data and the test data.

4. The artificial intelligence-assisted diagnosis and treatment method based on deep learning according to claim 2 is characterized in that: The deep learning network includes a contraction path and an expansion path; the training data is input into a pre-built deep learning network for training to obtain the deep learning model, including: Inputting the computed tomography image of the training data into the contraction path to perform multi-scale feature transformation, hierarchical downsampling and dynamic channel expansion to obtain a feature map to be processed; Inputting the feature map to be processed into the expansion path for transposed convolution up-sampling, cross-layer connection and feature fusion to obtain a fused feature map; The fused feature map is input into a classification function to obtain a classification result, and the network parameters of the deep learning network are continuously adjusted based on the classification result, the pre-labeled result and a preset loss function to obtain the deep learning model.

5. The artificial intelligence-assisted diagnosis and treatment method based on deep learning according to claim 1, characterized in that: Determining the longest axis of the hematoma based on the hematoma area comprises: Obtaining the number of all voxels and the volume of each voxel in the hematoma area; The hematoma volume is obtained by multiplying the number of all voxels by the volume of each voxel; Determine the longest axis of the hematoma based on the hematoma volume.

6. The artificial intelligence-assisted diagnosis and treatment method based on deep learning according to claim 1 is characterized in that: The method further comprises: Determine the first anatomical landmark boundary, the second anatomical landmark boundary, and the third anatomical landmark boundary; The puncture safety zone is determined based on the first anatomical landmark boundary, the second anatomical landmark boundary, and the third anatomical landmark boundary.

7. The artificial intelligence-assisted diagnosis and treatment method based on deep learning according to claim 1 is characterized in that: The method further comprises: Confirm the target center through surface curvature analysis; A circular area with a preset radius at the center of the target is used as the puncture safety zone.

8. An artificial intelligence-assisted diagnosis and treatment device based on deep learning, characterized in that: The device comprises: An acquisition module, used to acquire a to-be-processed computer tomography image corresponding to the user to be diagnosed and treated; A processing module, used for inputting the to-be-processed computed tomography image into a pre-trained deep learning model for segmentation processing to obtain a hematoma area; A determination module, used for determining the longest axis of the hematoma based on the hematoma area; A generation module is used to generate a puncture path from the puncture safety zone to the longest axis of the hematoma based on the comparison between the longest axis of the hematoma and the set puncture safety zone, so as to perform processing based on the puncture path.

9. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing instructions executable by the processor; The processor is used to read the executable instructions from the memory and execute the instructions to implement the deep learning-based artificial intelligence-assisted diagnosis and treatment method described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and the computer program is used to execute the deep learning-based artificial intelligence-assisted diagnosis and treatment method described in any one of claims 1 to 7.

Images