Multi-level and focus region based fine brain segmentation method, device, equipment and medium

Abstract

The application provides a fine brain region segmentation method, device, equipment and medium based on multiple levels and focus regions. The method comprises the following steps: acquiring a brain magnetic resonance image; performing multi-level segmentation on the brain region through a multi-level brain region segmentation network; according to a fine level brain region segmentation result, assigning different weights to each brain region of the fine level according to segmentation difficulty through a focus region segmentation network; and fusing the output results of the multi-level brain region segmentation network and the output results of the focus region segmentation network to obtain a final segmentation result of all brain regions. The application combines prediction on multiple levels, avoids training a separate network for each brain region, and can have good performance in terms of accuracy and efficiency. Moreover, the application proposes a loss based on focus regions, which helps to improve the segmentation capability of the network in difficult regions. Finally, a fusion module is used to fuse the predictions to obtain more accurate and robust brain region segmentation results.

Description

Technical Field

[0001] This invention relates to the field of brain image segmentation technology, and in particular to a method, apparatus, device, and medium for fine brain region segmentation based on multi-level and focal regions. Background Technology

[0002] Brain region segmentation plays a crucial role in the diagnosis of neurodegenerative diseases, the construction of brain networks, and quantitative analysis of the brain. It utilizes structural magnetic resonance (MR) images of the brain to divide it into non-overlapping regions. Artificial brain segmentation is a highly specialized and time-consuming task; therefore, automated brain region segmentation algorithms are needed. Several major challenges exist in brain region segmentation, such as the complex structure of the brain, the large number and uneven distribution of brain regions, and the indistinct boundaries between regions. Existing methods, such as registration-based methods, are limited by registration errors and are time-consuming. Existing deep learning-based methods, such as U-Net, suffer from inaccurate segmentation when dealing with a large number of brain regions.

[0003] Traditional multi-atlas-based brain parsing (MAP) has been the most common automatic brain segmentation method in the past. It registers multiple brain atlas images onto a target image, then transfers the atlas-labeled brain regions to the target brain image to achieve segmentation. However, MAP relies on image registration, which is computationally inefficient and sensitive to image noise. Currently, many deep-learning-based brain parsing (DLP) methods can achieve end-to-end brain segmentation or replace the core components of MAP (i.e., image registration and label fusion) with deep learning networks. Due to the superior performance of deep learning, DLP methods offer higher computational efficiency and robustness compared to MAP.

[0004] While existing DLP methods have achieved good results, few studies have focused on fine-grained brain region segmentation, i.e., when there are a large number of brain regions to be segmented. On MR images, because the definition of brain regions is partly based on anatomical structure, brain function, metabolism, neurotransmitters, and gene expression, there are often no clear anatomical boundaries between different brain regions. Therefore, as the number of brain regions increases, the segmentation accuracy decreases significantly.

[0005] To address the aforementioned problem of fine-level brain region segmentation, several approaches have been proposed, involving the creation of multiple models at different locations or views of the brain. 2.5D methods, such as FastSurfer, reduce computation by converting 3D MRI images into 2D slices and training dedicated 2D segmentation networks in the coronal, sagittal, and horizontal planes. During prediction, the results from these three directions are fused to obtain the final prediction. This method reduces memory overhead and training time, but the results trained on 2D slices may show inconsistencies in 3D images, and it doesn't consider the three-dimensional structure of brain regions. Furthermore, some have proposed using 8 or 27 networks to train segmentation networks at different locations, but this requires training multiple networks and is less efficient. Other approaches have proposed a coarse-to-fine approach: for each brain region, a coarse segmentation network is first trained to obtain a coarse segmentation result, followed by a fine segmentation network to obtain a refined segmentation result. Transfer learning is then applied to other brain regions using this network, and fine-tuning is performed on the previous network for new target brain regions. While this method can achieve relatively accurate segmentation results, it suffers from inefficiency during both training and testing because it requires training a different network for each brain region. For 125 brain regions, this necessitates 250 different segmentation networks. Similar to multi-model-based methods, the coarse-to-fine approach is also inefficient, as a fine-grained network must be built for each brain region. Furthermore, both methods process each brain region independently, ignoring the interrelationships between them. Summary of the Invention

[0006] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a method, apparatus, device and medium for fine brain region segmentation based on multi-level and focal regions, so as to solve the problems of unclear brain region edges and large number and uneven distribution of brain regions in the prior art.

[0007] To achieve the above and other related objectives, this application provides a method for fine brain region segmentation based on multi-level and focal region segmentation. The method comprises: acquiring brain MRI images; performing multi-level segmentation of brain regions using a multi-level brain region segmentation network; assigning different weights to each brain region at the fine-level segmentation according to the segmentation difficulty using a focal region segmentation network based on the fine-level segmentation results; and fusing the output results of the multi-level brain region segmentation network with the output results of the focal region segmentation network to obtain the final segmentation result for all brain regions.

[0008] In one embodiment of this application, the multi-level structure includes: a global level, an intermediate level, and a fine-grained level; the loss function of the multi-level brain region segmentation network is defined as follows: in, These represent the segmentation results at the global, intermediate, and fine-grained levels, respectively; Y G Y M Y F These represent the truth values ​​for the global level, intermediate level, and fine level, respectively; α i (i = 1, 2, 3) are the weight parameters for the three levels; loss L MSP Cross-entropy loss is selected.

[0009] In one embodiment of this application, the step of assigning different weights to each brain region at the fine-level according to the segmentation difficulty using a focal region segmentation network based on the fine-level brain region segmentation results includes: training a multi-level brain region segmentation network and performing inference on a validation set to obtain fine-level brain region segmentation results; calculating the average Dice score of each brain region based on the fine-level brain region segmentation results to determine the segmentation difficulty of each brain region; and assigning different weights to each brain region in the focal region segmentation network according to the segmentation difficulty of each brain region.

[0010] In one embodiment of this application, the loss function of the focal region segmentation network is defined as: Where K is the number of brain regions, l k The cross-entropy loss between the segmentation result and the ground truth value of the k-th brain region is w. k It refers to the assigned weights.

[0011] In one embodiment of this application, the weight w k The calculation methods include: Among them, w k ∈[δ,1], 0≤δ<1; x min and x max These are the minimum and maximum Dice metrics across all regions; x k It is the Dice index for the k-th brain region, which reflects the segmentation difficulty of that region.

[0012] In one embodiment of this application, the method includes: x k The smaller the value, the more difficult it is to segment the k-th brain region, so the assigned weight w... k The larger the value, the greater the value; δ is a hyperparameter used to control the normalization of x. k and weight w k The relationship between them; where, the smaller δ is, the better w k The larger the range, the better.

[0013] In one embodiment of this application, the fusion yields the following final segmentation result for all brain regions: Among them, the outputs of the multi-level brain region segmentation network and the focal region segmentation network are O MSP and O FRP ;OMSP (x,k) is the probability that voxel x belongs to the k-th brain region; O FRP (x,k) is the probability that voxel x belongs to the kth brain region.

[0014] To achieve the above and other related objectives, this application provides an electronic device comprising: a multi-level brain region segmentation module for performing multi-level segmentation on brain MRI images; a focal region segmentation module for assigning different weights to each brain region at the fine-level segmentation according to the segmentation difficulty using a focal region segmentation network based on the fine-level brain region segmentation results; and a fusion module for fusing the output results of the multi-level brain region segmentation network with the output results of the focal region segmentation network to obtain the final segmentation results for all brain regions.

[0015] To achieve the above and other related objectives, this application provides a computer device, the device comprising: a memory and a processor; the memory being used to store computer instructions; and the processor executing the computer instructions to implement the method described above.

[0016] To achieve the above and other related objectives, this application provides a computer-readable storage medium storing computer instructions that, when executed, perform the method described above.

[0017] In summary, the present application discloses a method, apparatus, device, and medium for fine brain region segmentation based on multi-level and focal region segmentation. This method acquires brain MRI images and performs multi-level segmentation of brain regions using a multi-level brain region segmentation network. Based on the fine-level brain region segmentation results, a focal region segmentation network assigns different weights to each brain region at the fine level according to the segmentation difficulty. The output results of the multi-level brain region segmentation network and the focal region segmentation network are then fused to obtain the final segmentation results for all brain regions.

[0018] It has the following beneficial effects:

[0019] The method described in this application combines multi-level predictions while avoiding the cumbersome process of training separate networks for each brain region, achieving good performance in both accuracy and efficiency. Furthermore, it proposes a focus-region-based loss, which helps improve the network's segmentation ability in difficult regions. Finally, a fusion module is used to combine the predictions, resulting in more accurate and robust brain region segmentation results. Attached Figure Description

[0020] Figure 1 The diagram shown is a flowchart illustrating a fine brain region segmentation method based on multi-level and focal regions in one embodiment of this application.

[0021] Figure 2The diagram shown is an example of a framework based on a deep learning network for fine brain segmentation based on multi-level and focal regions, as described in one embodiment of this application.

[0022] Figure 3 The diagram shows a model illustrating brain region segmentation results using different methods in one embodiment of this application.

[0023] Figure 4 The diagram shows a box plot of DSCs in an embodiment of this application, using different methods in an average of 101 brain regions and 12 typical brain regions.

[0024] Figure 5 The diagram shown is a schematic representation of an electronic device according to one embodiment of this application.

[0025] Figure 6 The diagram shown is a structural schematic of a computer device according to one embodiment of this application. Detailed Implementation

[0026] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0027] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Although the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components, the form, quantity and proportion of each component can be arbitrarily changed in actual implementation, and the layout of the components may also be more complex.

[0028] Throughout this specification, when it is said that a part is "connected" to another part, this includes not only "direct connection" but also "indirect connection" by placing other elements in between. Furthermore, when it is said that a part "includes" a certain constituent element, unless otherwise stated otherwise, this does not exclude other constituent elements, but rather means that other constituent elements may also be included.

[0029] The terms "first," "second," and "third," etc., used herein are for the purpose of describing various parts, components, regions, layers, and / or segments, but are not limited thereto. These terms are used only to distinguish one part, component, region, layer, or segment from others. Therefore, the following description of a first part, component, region, layer, or segment may refer to a second part, component, region, layer, or segment without departing from the scope of this application.

[0030] Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” and “including” indicate the presence of the stated feature, operation, element, component, item, kind, and / or group, but do not exclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are interpreted as inclusive, or mean any one or any combination thereof. Therefore, “A, B, or C” or “A, B, and / or C” means “any one of the following: A; B; C; A and B; A and C; B and C; A, B, and C.” Exceptions to this definition only arise when the combination of elements, functions, or operations is inherently mutually exclusive in some way. Typically, corporate technology contracts involve related technological achievements, and the signing of technology contracts is a direct manifestation of the transformation of these technological achievements. Therefore, some technology contracts, after being recognized by government or organizational units, can obtain certain subsidies or tax exemptions and other benefits or policies.

[0031] like Figure 1 The figure shows a flowchart illustrating a fine brain region segmentation method based on multi-level and focal regions according to an embodiment of this application. As shown, the method includes:

[0032] Step S101: Obtain brain MRI images and segment the brain regions in multiple levels using a multi-level brain region segmentation network.

[0033] In some examples, addressing the issues of indistinct brain region boundaries and a large number of brain regions, this invention proposes a multi-level segmentation algorithm. Specifically, this invention defines brain regions at different levels: fine-grained, intermediate, and global. From the global level to the fine-grained level, brain regions are progressively divided into more specific sub-regions. To this end, this invention proposes a multi-task deep learning network to predict semantic segmentation maps at these three levels in parallel.

[0034] As mentioned earlier, fine-grained brain region segmentation involves numerous brain regions with indistinct boundaries, making it challenging. In fact, based on anatomical knowledge, fine-level brain regions can be merged into intermediate-level regions, such as lobes or subcutaneous nuclei, while intermediate-level regions can be further merged into global-level regions, such as white matter and gray matter. Therefore, this invention proposes not to directly segment a large number of brain regions at the fine-grained level, but rather to segment at multiple levels, such as intermediate and global levels.

[0035] like Figure 2 The diagram illustrates an example framework of a deep learning network for fine-grained brain region segmentation based on multi-level and focal regions, as described in one embodiment of this application. As shown, the multi-level brain region segmentation network employs a multi-task structure, with each task responsible for segmenting a brain region at a specific scale: global, intermediate, and fine-grained levels. All tasks share the same feature extraction network but have different prediction heads. Thus, under the constraint of large-scale (i.e., global and intermediate levels) segmentation, the features learned by the feature extraction network contain information from multi-level brain regions with hierarchical relationships. Compared to features learned from single-level brain region segmentation, these features are more effective at distinguishing different brain regions.

[0036] In one embodiment of this application, the loss function of the multi-level brain segmentation network (MSP) is defined as follows:

[0037]

[0038] in, These represent the segmentation results at the global, intermediate, and fine-grained levels, respectively; Y G Y M Y F These represent the truth values ​​for the global level, intermediate level, and fine level, respectively; α i (i = 1, 2, 3) represent the weight parameters for the three levels, where the weight parameters are set according to the relative difficulty and importance of the segmentation task at each level; loss L MSP Cross-entropy loss is selected.

[0039] Preferably, α1, α2, and α3 are set to 0.1, 0.2, and 0.7, respectively.

[0040] Step S102: Based on the brain region segmentation results at the fine level, use the focal region segmentation network to assign different weights to each brain region at the fine level according to the difficulty of segmentation.

[0041] In some embodiments, to address the problem of uneven brain region distribution, this invention proposes a loss function for focal region enhancement. This loss function weights different brain regions according to the difficulty of brain region segmentation, that is, assigning greater weight to difficult regions, i.e., focal regions, and less weight to easy regions, thereby achieving enhancement of difficult regions.

[0042] Because the anatomical boundaries of most brain regions are not clearly visible on MR images, and due to the uneven distribution of brain region volume, segmenting some brain regions at a fine scale is difficult. Therefore, this invention proposes an auxiliary focal region segmentation network (FRP) (see...). Figure 2 The second part focuses on brain regions that are difficult to segment on a fine scale.

[0043] In one embodiment of this application, step S102 includes:

[0044] A. By training a multi-level brain region segmentation network and performing inference on the validation set, fine-level brain region segmentation results can be obtained.

[0045] B. Calculate the average Dice score for each brain region based on the segmentation results at the fine-level hierarchy, and use this score to determine the segmentation difficulty of each brain region;

[0046] C. Based on the difficulty of segmenting each brain region, different weights are assigned to each brain region in the focal region segmentation network.

[0047] Specifically, firstly, a multi-level brain region segmentation network (MSP) is trained, and inference is performed on a validation set to obtain fine-level brain region segmentation results. Based on this, the average Dice score for each brain region can be calculated. The difficulty level of brain region segmentation can be determined based on the average Dice score. Then, in the loss function of the focal region segmentation network (FRP), a larger weight is assigned to the loss of relatively difficult brain regions. In this way, the segmentation accuracy of difficult brain regions can be improved.

[0048] In some examples, the loss function of the Focus Region Segmentation Network (FRP) is defined as:

[0049]

[0050] Where K is the number of brain regions, l k The cross-entropy loss between the segmentation result and the ground truth value of the k-th brain region is w. k It refers to the assigned weights.

[0051] Here, l k It is the cross-entropy loss between the segmentation result of the k-th brain region and the ground truth value. Weight w k It can be calculated in the following way.

[0052]

[0053] Among them, w k ∈[δ,1], 0≤δ<1; x min and x max These are the minimum and maximum Dice metrics across all regions; x k It is the Dice index for the k-th brain region, which reflects the segmentation difficulty of that region.

[0054] Specifically, x k The smaller the value, the more difficult it is to segment the k-th brain region; therefore, it should receive more attention, i.e., a larger weight should be used for it. k Here, δ is a hyperparameter used to control the normalization of x. k and weight w k The relationship between δ and w. The smaller δ is, the better. k The larger the range, the better. To maintain healthy competition between brain regions, some regions are not assigned a weight of 0. Therefore, in the experiment, δ was set to 0.2 to ensure stability and discriminative power.

[0055] It should be noted that since the MSP network in the first part provides the brain region segmentation difficulty information needed by the FRP network in the second part, the MSP network needs to be trained first. Furthermore, to provide a good initialization for the FRP network, the network parameters from the previously trained MSP network (i.e., the fine-level feature extraction network and prediction head in the MSP) are passed to the subsequent FRP network. Throughout the training of the entire framework (i.e., MSP+FRP), the MSP network remains fixed, and only the parameters of the FRP network are updated.

[0056] Step S103: Fuse the output of the multi-level brain region segmentation network with the output of the focal region segmentation network to obtain the final segmentation result of all brain regions.

[0057] As described above, within the framework of this application, the MSP network can make good predictions for most brain regions, while the FRP network focuses on difficult brain regions. Therefore, in order to obtain more accurate and reliable final segmentation results, this invention applies a fusion module (MM) to combine the results of the MSP and FRP networks to achieve high-confidence segmentation of all brain regions.

[0058] Specifically, for an input image, the outputs of the MSP and FRP networks are O MSP and O FRP The final segmentation result output by the merging module is:

[0059]

[0060] Among them, the outputs of the multi-level brain region segmentation network and the focal region segmentation network are O MSP and O FRP ;O MSP (x,k) is the probability that voxel x belongs to the k-th brain region; O FRP (x,k) is the probability that voxel x belongs to the kth brain region.

[0061] Thus far, the first part learns highly discriminative brain region features based on the MSP network, the second part uses the FRP network to enhance difficult brain regions, and finally the merging module can produce high-precision final segmentation results.

[0062] The method proposed in this invention combines multi-level predictions while avoiding the cumbersome process of training separate networks for each brain region, achieving good performance in both accuracy and efficiency. Furthermore, it introduces a focus-region-based loss, which helps improve the network's segmentation ability in difficult regions. Finally, a fusion module is used to combine the predictions, resulting in more accurate and robust brain region segmentation results.

[0063] In one or more embodiments, specific implementations of the method described in this application may be as follows:

[0064] For a brain MRI image to be labeled, the first step is data preprocessing, including: skull removal, bias field correction, and resampling to 1×1×1mm. 3 The data was cropped or padded to a size of 256×256×256. Data labeling used DK-Atlas, dividing the entire brain into 101 regions as a fine-level labeling hierarchy. Then, based on the anatomical definitions of the brain regions, certain regions were merged to obtain intermediate-level labels for 26 brain regions, such as lobes, subcortical areas, and nuclei. These regions were further merged into global labels for three tissue categories: white matter, gray matter, and cerebrospinal fluid. At this point, all the necessary data was obtained.

[0065] The implementation details are as follows: all data is randomly divided into training, validation, and test sets. During training, the 3D T1 image is randomly cropped into 128×128×32 3D image patches. The optimization algorithm is the Adam algorithm with an initial learning compensation of 0.0001, and the step size decreases by a factor of 0.5 at the 30th, 50th, and 70th training iterations.

[0066] To evaluate the proposed framework, this application compares it with three state-of-the-art methods, including U-Net, ResU-Net, and V-Net. The Dice Similarity Coefficient (DSC) is used to quantitatively evaluate the accuracy of brain region segmentation.

[0067] Typical segmentation results obtained by different methods are as follows: Figure 3 As shown in the figure, the first row is a magnified view of a portion of the brain; the second row is a cross-sectional view of the segmented brain; and the third row is a three-dimensional view of the segmented brain. By comparing the brain regions indicated by the arrows in the first and third rows, it can be seen that the method proposed in this invention achieves more accurate segmentation results at the boundaries of the regions. Furthermore, this method achieves accurate results not only in large regions but also in small, difficult-to-segment regions.

[0068] Table 1. Quantitative comparison of results from our method and three advanced methods.

[0069]

[0070] Quantitative results are shown in Table 1, which presents the mean ± standard deviation of DSC (%) for an average of 101 brain regions and three typical brain regions. Bold text indicates best results. On an average of 101 brain regions, the method of this application outperformed U-Net, ResU-Net, and V-Net by 7.93%, 6.56%, and 4.07%, respectively. The method of this application also outperformed other methods on the gyri of the cerebral cortex, such as the left precentral gyrus, or on subcortical nuclei, such as the left hippocampus and left nucleus accumbens.

[0071] In addition, a pairwise significance test, namely a t-test, was performed on the mean DSC of all brain regions, and the results showed that the improvement of the proposed method was statistically significant (p-values ​​< 0.01).

[0072] Figure 4 The image shows box plots of DSC (%) for different methods in an average of 101 brain regions and 12 typical brain regions, where "L" and "R" refer to the left and right hemispheres, respectively. It can be seen that the method proposed in this invention outperforms the other three methods, especially in difficult regions where DSC is relatively low, where this method shows a greater improvement in accuracy.

[0073] Table 2 Results of the ablation study

[0074]

[0075] Furthermore, Table 2 presents the results of ablation experiments using the proposed method. The table shows the mean ± standard deviation of DSC (%) for an average of 101 brain regions and three typical brain regions. The baseline network is a 3DV-Net that directly outputs fine-level labels. "MSP", "FRP", and "MM" represent the multi-scale brain region segmentation network, the focal region-based brain region segmentation network, and the merging module, respectively. As shown in the table, all three components played a positive role. The multi-scale brain region segmentation network (MSP) significantly improved upon the baseline network not only on an average of all brain regions but also on other typical brain regions. The focal region-based segmentation (FRP) network also performed better than the baseline network + MSP. Finally, by integrating the outputs of the MSP and FRP networks, the merging module (MM) improved the final segmentation accuracy.

[0076] In summary, the method described in this application combines multi-level predictions while avoiding the cumbersome process of training separate networks for each brain region, achieving good performance in both accuracy and efficiency. Furthermore, it proposes a focus-region-based loss, which helps improve the network's segmentation ability in difficult regions. Finally, a fusion module is used to combine the predictions, resulting in more accurate and robust brain region segmentation results.

[0077] like Figure 5 The diagram shown illustrates a module schematic of an electronic device according to an embodiment of this application. As shown, the device 500 includes:

[0078] The multi-level brain region segmentation module 501 is used to acquire brain MRI images and perform multi-level segmentation of brain regions through a multi-level brain region segmentation network.

[0079] The focal region segmentation module 502 is used to assign different weights to each brain region at the fine level according to the segmentation difficulty based on the fine-level brain region segmentation results using the focal region segmentation network.

[0080] The fusion module 503 is used to fuse the output results of the multi-level brain region segmentation network with the output results of the focal region segmentation network to obtain the final segmentation results of all brain regions.

[0081] It should be noted that the information interaction and execution process between the modules / units of the above-mentioned device are based on the same concept as the method embodiments described in this application, and the resulting technical effects are the same as those in the method embodiments of this application. For details, please refer to the description in the method embodiments shown above in this application, and it will not be repeated here.

[0082] It should also be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these units can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, the fusion module 503 can be a separately established processing element, or it can be integrated into a chip in the above device. Alternatively, it can be stored as program code in the memory of the above device, and its functions can be called and executed by a processing element of the above device. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element mentioned here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0083] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more digital signal processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together to form a system-on-a-chip (SOC).

[0084] like Figure 6 The figure shows a schematic diagram of the structure of a computer device according to an embodiment of this application. As shown, the computer device 600 includes: a memory 601 and a processor 602; the memory 601 is used to store computer instructions; the processor 602 executes the computer instructions to implement... Figure 1 The method described.

[0085] In some embodiments, the number of memories 601 in the computer device 600 may be one or more, and the number of processors 602 may be one or more. Figure 6 Each example is taken as an instance.

[0086] In one embodiment of this application, the processor 602 in the computer device 600 will perform as follows: Figure 1 The steps described involve loading one or more instructions corresponding to the process of an application into memory 601, and then having processor 602 run the application stored in memory 601, thereby achieving the following: Figure 1 The method described.

[0087] The memory 601 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. The memory 601 stores an operating system and operating instructions, executable modules or data structures, or subsets thereof, or extended sets thereof. The operating instructions may include various operation instructions used to implement various operations. The operating system may include various system programs used to implement various basic services and handle hardware-based tasks.

[0088] The processor 602 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0089] In some specific applications, the various components of the computer device 600 are coupled together through a bus system, which may include, in addition to a data bus, a power bus, a control bus, and a status signal bus, etc. However, for clarity, in... Figure 6 All kinds of buses are referred to as bus systems.

[0090] In one embodiment of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the following: Figure 1 The method described.

[0091] At any possible level of technical detail, this application can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this application.

[0092] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, (but not limited to) electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0093] The computer-readable program described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer-readable program instructions from the network and forwards those instructions to the computer-readable storage medium stored in the respective computing / processing device.

[0094] The computer program instructions used to perform the operations of this application may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, integrated circuit configuration data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), are personalized by utilizing the status information of the computer-readable program instructions. These electronic circuits can execute the computer-readable program instructions to implement various aspects of this application.

[0095] In summary, this application provides a method, apparatus, device, and medium for fine brain region segmentation based on multi-level and focal region segmentation. This involves acquiring brain MRI images and performing multi-level segmentation of brain regions using a multi-level brain region segmentation network. Based on the fine-level brain region segmentation results, a focal region segmentation network is used to assign different weights to each brain region at the fine level according to the segmentation difficulty. The output results of the multi-level brain region segmentation network and the focal region segmentation network are then fused to obtain the final segmentation results for all brain regions.

[0096] This application effectively overcomes the various shortcomings of the prior art and has high industrial application value.

[0097] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this invention should still be covered by the claims of this application.

Claims

1. A method for multi-level and focal region brain region segmentation, characterized in that, The method includes: Brain region images are acquired, and the brain regions are segmented at multiple levels using a multi-level brain region segmentation network. The multiple levels include: global level, intermediate level, and fine-grained level. The loss function of the multi-level brain region segmentation network is defined as follows: ； And among them, , , These represent the segmentation results at the global, intermediate, and fine-grained levels, respectively. , , These represent the truth values ​​for the global level, intermediate level, and fine level, respectively. These are the weight parameters for three levels; loss. Choose cross-entropy loss; Based on the results of fine-level brain region segmentation, a focal region segmentation network is used to assign different weights to each brain region at the fine level according to the difficulty of segmentation. The outputs of the multi-level brain region segmentation network and the focal region segmentation network are fused to obtain the final segmentation results for all brain regions; wherein, the fusion to obtain the final segmentation results for all brain regions is as follows: ； Furthermore, the outputs of the multi-level brain region segmentation network and the focal region segmentation network are respectively... and ; It is a voxel Belongs to the The probability of each brain region; It is a voxel Belongs to the The probability of each brain region.

2. The method according to claim 1, characterized in that, The process involves assigning different weights to each brain region at the fine-level segmentation stage based on the segmentation difficulty, using a focal region segmentation network, according to the results of fine-level brain region segmentation. This includes: By training a multi-level brain region segmentation network and performing inference on a validation set, fine-level brain region segmentation results can be obtained. The average Dice score of each brain region is calculated based on the segmentation results of each brain region at a fine level, thereby determining the segmentation difficulty of each brain region; Based on the difficulty of segmenting each brain region, different weights are assigned to each brain region in the focal region segmentation network.

3. The method according to claim 2, characterized in that, The loss function of the focal region segmentation network is defined as: ； in, It refers to the number of brain regions. It is the first Cross-entropy loss between the segmentation results and the ground truth values ​​for each brain region It refers to the assigned weights.

4. The method according to claim 3, characterized in that, The weight The calculation methods include: ； in, , ; and These are the minimum and maximum Dice metrics across all regions; It is the first The Dice index for each brain region reflects the difficulty of segmenting that region.

5. The method according to claim 4, characterized in that, The method includes: The smaller, the first The more difficult it is to segment individual brain regions, the more weights need to be allocated. The larger; These are hyperparameters used to control normalization. and weight The relationship between them; among them, The smaller, The larger the range, the better.

6. An electronic device, characterized in that, The device includes: A multi-level brain region segmentation module is used to acquire brain region images and perform multi-level segmentation of the brain regions through a multi-level brain region segmentation network; wherein, the multi-level includes: global level, intermediate level, and fine level; the loss function of the multi-level brain region segmentation network is defined as follows: ； And among them, , , These represent the segmentation results at the global, intermediate, and fine-grained levels, respectively. , , These represent the truth values ​​for the global level, intermediate level, and fine level, respectively. These are the weight parameters for three levels; loss. Choose cross-entropy loss; The focal region segmentation module is used to assign different weights to each brain region at the fine level according to the segmentation difficulty based on the fine-level brain region segmentation results using the focal region segmentation network. The fusion module is used to fuse the output results of the multi-level brain region segmentation network with the output results of the focal region segmentation network to obtain the final segmentation results for all brain regions; wherein, the fusion results in the final segmentation results for all brain regions as follows: ； Furthermore, the outputs of the multi-level brain region segmentation network and the focal region segmentation network are respectively... and ; It is a voxel Belongs to the The probability of each brain region; It is a voxel Belongs to the The probability of each brain region.

7. A computer device, characterized in that, The device includes: a memory and a processor; the memory is used to store computer instructions; the processor executes the computer instructions to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The device stores computer instructions that, when executed, perform the method as described in any one of claims 1 to 5.

Images