A multi-center nasopharyngeal carcinoma MRI image three-dimensional target segmentation training method, system, device and medium

Abstract

The present application relates to the field of medical image processing, aiming at the problems of source center sample imbalance, large scanning protocol difference, target area category imbalance and performance decline of unseen center generalization in multi-center nasopharyngeal carcinoma MRI image target area segmentation, and provides a three-dimensional target area segmentation model of multi-center nasopharyngeal carcinoma MRI image, a training method, equipment and medium thereof. By acquiring multi-center images and labels, establishing case records, calculating protocol signatures, constructing source center quota samplers, performing protocol disturbance enhancement, training three-dimensional segmentation networks, the stability and accuracy of the model in target area segmentation of unseen centers are improved.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, deep learning, medical image segmentation, and automatic delineation of radiotherapy target areas, and particularly to a training method, system, device, and medium for source center balancing and protocol signature-driven three-dimensional target area segmentation for multi-center nasopharyngeal carcinoma magnetic resonance images. Background Technology

[0002] In the process of developing a radiotherapy plan for nasopharyngeal carcinoma, it is necessary to accurately delineate the target areas of the primary tumor and metastatic lymph nodes. Magnetic resonance imaging (MRI) images have good soft tissue contrast and are an important imaging basis for the identification and segmentation of the target area in nasopharyngeal carcinoma. Different MRI sequences have different advantages in target area visualization. For example, T2-weighted images can reflect soft tissue structure and edema-related information, while enhanced T1-weighted images can highlight the tumor boundary and provide basic anatomical information.

[0003] In multi-center clinical settings, MRI images are typically acquired from different hospitals, using different scanners, coils, imaging parameters, and reconstruction procedures. These images exhibit significant differences in slice thickness, in-plane resolution, matrix size, field of view, noise level, tissue contrast, bias field, edge sharpness, and frequency domain texture. These differences can cause scanning protocol shifts, resulting in inconsistent image appearance distribution between training centers and unseen centers. This leads to performance degradation of the segmentation model on unseen centers, particularly manifested as target boundary drift, missed segmentation of small metastatic lymph nodes, and deterioration of the 3D boundary distance metric.

[0004] Existing 3D medical image segmentation methods typically employ procedures such as multi-center data training, intensity normalization, random spatial augmentation, random intensity augmentation, 3D U-shaped networks, Dice loss, cross-entropy loss, and sliding window inference. While these procedures can establish a basic segmentation model, they still have the following shortcomings: First, the number of cases in multi-source centers is usually uneven. If all source center cases are merged and randomly sampled, centers with a larger number of cases will be sampled more frequently during training, making it easier for the model to learn the image appearance and annotation distribution of large sample centers.

[0005] Second, conventional data augmentation typically employs random perturbations with fixed probabilities and ranges, without explicitly utilizing the case's own voxel spacing, slice thickness, anisotropy ratio, and frequency domain energy distribution, nor adaptively adjusting the perturbation type and intensity based on the multi-source center protocol distribution.

[0006] Third, the target area of ​​nasopharyngeal carcinoma, especially the metastatic lymph node target area, is relatively small and exhibits significant class imbalance. Ordinary random 3D image patch sampling is insufficient to stably cover small target areas, resulting in inadequate model learning of small targets.

[0007] Fourth, when the multicenter case list is truncated, sampled, or fine-tuned in a short period of time, ordinary pooling training tends to retain more cases from large sample centers, while small sample centers and centers with large differences in protocols participate insufficiently, affecting subsequent model selection and external center generalization assessment.

[0008] Therefore, a training method for a three-dimensional target segmentation model for multicenter nasopharyngeal carcinoma magnetic resonance images is needed. Without using unseen target center labels, case-level protocol signatures, source center quota sampling, and protocol signature-driven perturbation enhancement can be combined to enable the model to learn more stable cross-center target region representations and improve its generalization ability on unseen target center magnetic resonance images. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a training method, system, device and medium for a three-dimensional target region segmentation model driven by source center balancing and protocol signature. It addresses the problems of uneven source center sample quantity, significant differences in scanning protocols, unbalanced target region categories and decreased generalization performance in multi-center nasopharyngeal carcinoma magnetic resonance images. By using source center quota sampling and protocol signature-driven perturbation enhancement, the stability of the model in target region segmentation of magnetic resonance images without visible centers is improved.

[0010] In a first aspect, the present invention provides a method for training a three-dimensional target region segmentation model for multicenter nasopharyngeal carcinoma MRI images, comprising the following steps: S1. Obtain nasopharyngeal carcinoma magnetic resonance images and their target region annotations from multiple source centers. The magnetic resonance images include at least one of T2-weighted images, T1-weighted images, and enhanced T1-weighted images. The target region annotations include the primary tumor target region and / or the metastatic lymph node target region. S2. Establish a case record for each case, wherein the case record includes at least the case identifier, center identifier, image path, target area annotation path, voxel spacing and image size; S3. Calculate the case-level protocol signature based on the case record. The case-level protocol signature includes at least a spatial resolution signature and a frequency domain energy signature. The spatial resolution signature includes at least one of in-plane voxel area, layer thickness, and anisotropy ratio. The frequency domain energy signature includes at least one of low-frequency energy ratio, high-frequency energy ratio, high-low frequency energy ratio, and their logarithmic values. S4. Construct a source center quota sampler based on the center identifier and the case-level protocol signature. During the training phase, the source center quota sampler first selects a source center, then selects a case within the selected source center, and samples a three-dimensional image patch from the selected case according to the target area foreground constraint. S5. Determine the protocol disturbance type and protocol disturbance intensity based on the case-level protocol signature of the case corresponding to the three-dimensional image block and its position in the distribution of multiple source center protocol signatures; S6. According to the protocol perturbation type and protocol perturbation intensity, perform protocol perturbation enhancement on the three-dimensional image block to generate a protocol perturbation image block corresponding to the target area annotation space; S7. Input the protocol-perturbed image patch into the 3D segmentation network and output the target region category probability; S8. Calculate the training loss based on the target region category probability and the target region annotation; S9. Update the parameters of the three-dimensional segmentation network according to the training loss, and repeat source center quota sampling, protocol perturbation enhancement and network parameter update according to the preset training stage; S10. Obtain a three-dimensional segmentation model for target region segmentation in magnetic resonance images of nasopharyngeal carcinoma without central focus.

[0011] Secondly, the present invention provides a three-dimensional target region segmentation system for multicenter nasopharyngeal carcinoma MRI images, comprising: The case record module is used to acquire nasopharyngeal carcinoma magnetic resonance images and their target region annotations from multiple source centers, and to establish a case record; the magnetic resonance images include at least one of T2-weighted images, T1-weighted images, and enhanced T1-weighted images, and the target region annotations include the primary tumor target region and / or metastatic lymph node target regions; the case record includes at least a case identifier, a center identifier, an image path, a target region annotation path, a voxel spacing, and an image size; The protocol signature calculation module is used to calculate a case-level protocol signature based on the case record. The case-level protocol signature includes at least a spatial resolution signature and a frequency domain energy signature. The spatial resolution signature includes at least one of in-plane voxel area, layer thickness, and anisotropy ratio. The frequency domain energy signature includes at least one of low-frequency energy ratio, high-frequency energy ratio, high-low frequency energy ratio, and their logarithmic values. The source center quota sampling module is used to select source centers during the training phase based on center identifiers and case-level protocol signatures, then select cases within the selected source centers, and sample three-dimensional image patches from the selected cases according to target area foreground constraints. The protocol perturbation enhancement module is used to determine the perturbation type and intensity based on the case-level protocol signature and multi-source center protocol distribution of the selected three-dimensional image block, and to perform protocol perturbation enhancement on the three-dimensional image block according to the protocol perturbation type and protocol perturbation intensity to generate a protocol perturbation image block corresponding to the target area annotation space. The training module is used to input protocol perturbation image patches into a 3D segmentation network and output target region category probabilities; calculate training loss based on the target region category probabilities and the target region annotations, update the parameters of the 3D segmentation network based on the training loss, and repeatedly perform source center quota sampling, protocol perturbation enhancement and network parameter updates according to preset training stages; The inference module is used to perform preprocessing, sliding window prediction, overlapping region fusion, and original space re-attachment on the magnetic resonance images to be segmented. The output module is used to output a three-dimensional segmentation mask of the primary tumor target area and / or metastatic lymph node target area.

[0012] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.

[0013] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0014] One or more technical solutions provided by this invention have at least the following technical effects or advantages: 1. This invention constructs a source center quota sampler based on center identifier and case-level protocol signature, so that different source centers have controllable contributions during training and reduce the dominant role of centers with larger sample sizes in model training.

[0015] 2. This invention introduces case-level protocol signatures, using information such as voxel spacing, slice thickness, in-plane resolution, anisotropy ratio, low-frequency energy, high-frequency energy, and high-low frequency energy ratio in the frequency domain to characterize scanning protocol differences, providing a calculable basis for protocol offset modeling of multicenter magnetic resonance images.

[0016] 3. The protocol perturbation enhancement of the present invention is not a random enhancement within a fixed range, but rather an adaptive determination of the perturbation type and intensity based on the case-level protocol signature and its position in the source-center protocol distribution, so that the enhanced samples cover the sparse region, boundary region and transition region in the multi-center protocol distribution.

[0017] 4. The present invention can set up comparative examples of source center balancing and source center balancing under the same case list, the same network and the same training loss, which facilitates the verification of the contribution of source center quota sampling to small sample centers and protocol difference centers.

[0018] 5. The present invention can be applied to T2-weighted images, T1-weighted images, enhanced T1-weighted images, or multi-sequence magnetic resonance images, wherein T2-weighted images can be used as a preferred embodiment for three-dimensional segmentation of GTVp / GTVn in nasopharyngeal carcinoma.

[0019] 6. This invention eliminates the need for unseen target center labels during the training and model selection phases, which helps improve the model's generalization ability on external central magnetic resonance images and its stability in clinical deployment.

[0020] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the case-level protocol signature distribution for multicenter cases in an embodiment of the present invention; Figure 3 This is a schematic diagram of the system structure in Embodiment 2 of the present invention; Figure 4 This is a schematic diagram of the electronic device in Embodiment 3 of the present invention; Figure 5 This is a schematic diagram of the structure of the medium in Embodiment 2 of the present invention. Detailed Implementation

[0023] The technical solution in this application embodiment follows the following general idea: Addressing the problems of uneven source center sample quantity, significant differences in scanning protocols, imbalanced target categories, and decreased generalization performance in target region segmentation of multi-center nasopharyngeal carcinoma MRI images, this invention proposes a source center balancing and protocol signature-driven three-dimensional target region segmentation model training method. This method acquires multi-center nasopharyngeal carcinoma MRI images and their target region annotations, establishes case records and calculates case-level protocol signatures, constructs a source center quota sampler to balance case sampling from different source centers, adaptively determines the perturbation type and intensity based on the protocol signature to enhance protocol perturbation, and inputs the processed image blocks into a three-dimensional segmentation network for training. Finally, a three-dimensional target region segmentation model with good generalization ability is obtained, effectively improving the stability and accuracy of target region segmentation on MRI images without visible centers.

[0024] Example 1 This embodiment provides a training method for a three-dimensional target region segmentation model of multicenter nasopharyngeal carcinoma MRI images, such as... Figure 1 As shown, the specific process is as follows: S1. Obtain nasopharyngeal carcinoma magnetic resonance images and their target region annotations from multiple source centers. The magnetic resonance images include at least one of T2-weighted images, T1-weighted images, and enhanced T1-weighted images. The target region annotations include the primary tumor target region and / or the metastatic lymph node target region. In a preferred embodiment, the magnetic resonance image is a T2-weighted magnetic resonance image of nasopharyngeal carcinoma, and the target area annotation includes the primary tumor target area GTVp and the metastatic lymph node target area GTVn. The system scans a multicenter case catalog, identifies the T2 images, GTVp annotations, and GTVn annotations, and creates a case record for each case. The case record includes case_id, center_id, image path, label path, sequence type, voxel spacing, image size, number of slices, and geometric consistency check results.

[0025] In one embodiment, after sequence screening and image-label geometric consistency checks, 796 T2 cases suitable for GTVp / GTVn training and validation were obtained, with 493, 197, and 106 cases from the three source centers, respectively. All cases were labeled with both GTVp and GTVn, and their image and label sizes and spatial relationships were checked. In this case list, source center center1 includes sequence families such as T2__TSE, T2__FSE, T2__FRFSE, and T2__BLADE; source center center2 primarily uses the T2__STIR sequence family; and source center center3 uses the T2__FS__FRFSE sequence family. This indicates that the different source centers not only differ in the number of cases but also in sequence naming and acquisition protocols.

[0026] like Figure 2 As shown, each scatter point in the figure represents a case-level protocol signature. The horizontal axis is the logarithm of the in-plane voxel area, log(d_xd_y), used to characterize the in-plane spatial sampling characteristics of the case; the vertical axis is the logarithm of the ratio of high-frequency energy to low-frequency energy in the frequency domain, used to characterize the frequency domain energy relationship of the case image. Different point shapes represent cases from different source centers: center 1 is Sun Yat-sen University Cancer Center, center 2 is Fujian Cancer Hospital, and center 3 is Jiangxi Cancer Hospital.

[0027] In one specific drawing implementation, kernel density estimation is performed on the case-level protocol signatures of the three centers. The outer envelope in the figure represents approximately 90% of the main density coverage area corresponding to the center, and the inner envelope represents approximately 50% of the core high-density area corresponding to the center; the center marker indicates the median position of the corresponding center in the two-dimensional protocol space. The 50% and 90% figures are merely one implementation method for displaying the protocol distribution range in this embodiment and do not constitute a limitation on the scope of protection of this invention.

[0028] The three source centers exhibit differences in their location and distribution within the case-level protocol space. Center 1 shows a wider distribution of protocol signatures and includes multiple localized high-density regions, suggesting that the same center may contain multiple scanning or reconstruction protocols. Some cases in Center 2 have smaller in-plane voxel areas and generally lower high-to-low frequency energy ratios, forming separate protocol subgroups. Most cases in Center 3 are relatively concentrated in the in-plane voxel area direction, but still exhibit some variation in the frequency energy direction.

[0029] The above results indicate that source center identifiers cannot fully describe the actual scanning protocols of a case. Protocol mixing may exist within the same source center, and there may be local overlap or transitional regions between different source centers. Therefore, this invention not only determines the source of training samples based on center identifiers, but also determines the source center sampling weights, protocol perturbation types, and protocol perturbation strengths based on the case-level protocol signatures and their positions in the multi-source center protocol distribution.

[0030] Before training, the system can establish protocol distribution statistics based on the protocol signatures of all source center cases. These statistics include at least one of the following: center location, variance, quantile range, probability density range, and protocol coverage envelope. During training, the system compares the protocol signature of selected cases with these statistics to determine whether the case is located within a central protocol distribution, in an overlapping region of multiple central protocol distributions, in a protocol distribution boundary region, or in a sparse protocol region. Based on the determination result, the system performs protocol signature-driven perturbation scheduling.

[0031] S2. Establish a case record for each case, wherein the case record includes at least the case identifier, center identifier, image path, target area annotation path, voxel spacing and image size; The system constructs training, validation, and test sets based on the center identifiers in the case records. In one partitioning embodiment, the training set consists of 657 cases, with 419 cases from the three source centers (center1, center2, and center3), 167 cases from center2, and 71 cases from center3; the validation set consists of 117 cases, with 74 cases from the three source centers (center1, center2, and center3), 30 cases from center2, and 13 cases from center3; an additional 22 cases from center3 are set up as quality validation or test cases. Cases without a defined center or test case are not included in model parameter updates and optimal model selection.

[0032] S3. Calculate the case-level protocol signature based on the case record. The case-level protocol signature includes at least a spatial resolution signature and a frequency domain energy signature. The spatial resolution signature includes at least one of in-plane voxel area, layer thickness, and anisotropy ratio, and may also include the number of layers and coverage area. The frequency domain energy signature includes at least one of low-frequency energy ratio, high-frequency energy ratio, high-low frequency energy ratio, and their logarithmic values. The system can represent case-level protocol signatures as follows: φ(v)=[log__voxel__area,log__anisotropy__ratio,log__fft__ratio]; Where v represents a case.

[0033] Of the 796 T2 cases, center1 had a median of 36 layers, a median layer thickness of 6.0 mm, a median in-plane pixel pitch of approximately 0.469 mm, and a median anisotropy ratio of approximately 12.80; center2 had a median of 36 layers, a median layer thickness of 6.0 mm, a median in-plane pixel pitch of approximately 0.240 mm, and a median anisotropy ratio of approximately 25.04; and center3 had a median of 16 layers, a median layer thickness of approximately 6.0 mm, a median in-plane pixel pitch of approximately 0.469 mm, and a median anisotropy ratio of approximately 12.80. Center3 had a higher proportion of short-coverage cases (72 cases), medium-coverage cases (25 cases), and long-coverage cases (9 cases). These data demonstrate that case-level protocol signatures can reflect differences in in-plane resolution, layer thickness coverage, and anisotropy between source centers.

[0034] In the statistical output of an existing protocol, the system further calculates the log_voxel_area and frequency domain high-low frequency relationships for multiple centers or sequences. For example, the median log_voxel_area of ​​center1 is approximately -1.207, with a range of approximately -1.689 to -0.940 from 5% to 95%; the median log_voxel_area of ​​center2 is approximately -2.858, with a range of approximately -2.858 to -1.434 from 5% to 95%. The median log_fft_ratio of center1 is approximately -3.356, and that of center2 is approximately -5.477. The above statistics are used to illustrate that protocol signatures can distinguish different centers or protocol clusters; specific thresholds can be re-estimated based on the actual dataset.

[0035] S4. Construct a source center quota sampler based on the center identifier and the case-level protocol signature. During the training phase, the source center quota sampler first selects a source center, then selects a case within the selected source center, and samples a three-dimensional image patch from the selected case according to the target area foreground constraint. If the number of source centers is M, the initial sampling probability of each source center can be set to 1 / M; alternatively, the case list can be grouped by center first, and then cases can be selected according to the center rotation order until the preset number of training or validation cases is reached.

[0036] In the control example of passive center balancing, if the training cases are truncated according to the order of the case list, the short-term training with 16 or 128 cases only includes center1; with 512 cases, it includes 419 center1 cases and 93 center2 cases, but not center3. Correspondingly, if the validation set is truncated sequentially with 24 or 64 cases, it also only includes center1. This control demonstrates that when the case list is arranged by center and the number of center cases is unbalanced, ordinary sequential truncating easily excludes small sample centers.

[0037] In the embodiment of source center balancing, the center distribution of 16 short-range training cases is 6:5:5; the center distribution of 128 training cases is approximately 43:43:42; and the center distribution of 512 training cases after all training cases in center3 have been taken is 274:167:71. The distribution of validation cases is 8:8:8 when 24 cases are selected, and 26:25:13 when 64 cases are selected. These results demonstrate that source center balancing sampling can preferentially cover multiple source centers with a limited number of training cases, reducing the dominant role of large sample centers in the training and validation process.

[0038] Within the selected cases, the system samples 3D image patches based on target foreground constraints. When a GTVp or GTVn foreground voxel exists in the case, the system can select sampling anchor points near GTVp or GTVn according to a preset probability and crop 3D image patches around the sampling anchor points; otherwise, it randomly crops from the entire image space. For small target regions such as GTVn, a higher foreground sampling ratio can be set to increase the frequency of small target regions appearing during training.

[0039] S5. Determine the protocol disturbance type and protocol disturbance intensity based on the case-level protocol signature of the case corresponding to the three-dimensional image block and its position in the distribution of multiple source center protocol signatures; When determining the protocol perturbation type and intensity, the protocol perturbation type can include layer thickness direction perturbation, in-plane resolution perturbation, intensity perturbation, noise perturbation, bias field perturbation, and frequency domain amplitude perturbation. The perturbation intensity can be determined by the distance between the case protocol signature and the source center protocol distribution, the center distribution variance, and a preset perturbation upper limit, so that the enhanced sample covers the internal region, boundary region, and transition region in the multi-center protocol distribution.

[0040] In one implementation, layer thickness perturbation includes blurring, sharpening, or downsampling followed by upsampling along the ≈ direction; in-plane resolution perturbation includes scaling or resampling along the in-plane direction; frequency domain amplitude perturbation must simultaneously satisfy the dual constraints of spatial resolution signature and frequency domain energy signature, adjusting low-frequency or high-frequency amplitude components under case-level protocol signature constraints. All of the above perturbations maintain label space consistency, meaning that the perturbated image patch maintains a correspondence with the original label image patch in terms of spatial location and category definition.

[0041] S7. Input the protocol-perturbed image patch into the 3D segmentation network and output the target region category probability; S8. Calculate the training loss based on the target region category probability and the target region annotation; This embodiment employs a 3D segmentation network with an encoder-decoder structure. The network can use in-plane convolutional kernels in shallow layers to learn in-plane anatomical details, and 3D convolutional kernels in deeper layers to fuse 3D context, while delaying information compression in the layer thickness direction during downsampling. Training losses include Dice loss and cross-entropy loss, and class weights can be set based on the volume differences between GTVp and GTVn. It should be understood that the 3D segmentation network can also be replaced with other 3D U-shaped networks, residual networks, or convolutional-attention hybrid networks; the core lies in source center balanced sampling and protocol signature-driven enhancement training mechanisms.

[0042] S9. Update the parameters of the three-dimensional segmentation network according to the training loss, and repeat source center quota sampling, protocol perturbation enhancement and network parameter update according to the preset training stage; In a training statistics embodiment employing case-level protocol signatures and source-center balanced sampling, the system uses the same case list, the same network, and the same loss function, only changing the case selection method to form a control with and without source-center balanced sampling. During the short-term fine-tuning phase, the training case center distribution is approximately 43:43:42, and the validation case center distribution is approximately 8:8:8; during the longer fine-tuning phase, the training case center distribution is 274:167:71, and the validation case center distribution is 26:25:13.

[0043] When using image block-level validation, the foreground Dice after a relatively long fine-tuning training phase is approximately 0.820, GTVpDice is approximately 0.702, and GTVnDice is approximately 0.688. After further cleansing and fine-tuning, GTVpDice reaches approximately 0.731, and GTVnDice is approximately 0.697. In full-case validation in the original image space, 64 validation cases were selected, with 26, 25, and 13 cases from the three source centers, respectively. One validation result shows that the foreground Dice is approximately 0.798, GTVpDice is approximately 0.796, GTVnDice is approximately 0.732, the average GTVp / GTVn Dice is approximately 0.764, and the 95th percentile Hausdorff distance (HD95) of GTVp and GTVn is approximately 7.66 mm and 30.86 mm, respectively. When statistically analyzed by center, the average GTVp / GTVn Dice of the three source centers is approximately 0.738, 0.785, and 0.776.

[0044] The above metrics are used to illustrate that the training process can maintain evaluable segmentation performance across multiple source centers and achieve stable predictions on small target regions such as GTVn. Specific values ​​are related to the dataset, model initialization, training epochs, and evaluation space, and do not constitute a limitation on the scope of this invention.

[0045] S10. Obtain a three-dimensional segmentation model for target region segmentation in magnetic resonance images of nasopharyngeal carcinoma without central focus, and then apply the three-dimensional segmentation model.

[0046] In another implementation, multi-sequence expansion can be performed, whereby the magnetic resonance image may include a T2-weighted image and an enhanced T1-weighted image. The system can use the T2-weighted image as the main sequence and the enhanced T1-weighted image as the auxiliary sequence, calculate the protocol signature for each, and then fuse them. During the training phase, consistent spatial perturbation is applied to the multi-sequence images to ensure that the blur intensity in the layer thickness direction and the in-plane resolution scaling ratio remain consistent across sequences, thereby guaranteeing the spatial correspondence between the multi-sequence images.

[0047] In one specific implementation, the magnetic resonance image is a T2-weighted image. The target voxel spacing, 3D image patch size, foreground sampling ratio, and class weights can be determined based on image resolution, video memory size, target volume, and number of centers. Class labels can be set to background=0, GTVp=1, and GTVn=2.

[0048] 3D segmentation networks can employ encoder-decoder structures, anisotropic perceptual convolutional kernels, and normalization methods suitable for mini-batch training in 3D. Loss functions can include Dice loss and cross-entropy loss, and class weights can be set based on the volume differences between GTVp and GTVn.

[0049] Case-level protocol signatures can be represented using two-dimensional, three-dimensional, or higher-dimensional vectors. In a simplified implementation, the protocol signature is [log_voxel_area, log_anisotropy_ratio, log_fft_ratio]. In another implementation, the protocol signature is: [dx,dy,dz,voxel_area,anisotropy_ratio,FFT_LowFrac,FFT_HighFrac,FFT_HighLowRatio].

[0050] In further implementations, the protocol signature can also incorporate image matrix size, field of view, bias field estimate, noise estimate, and edge sharpness estimate.

[0051] The above parameters are merely examples, and those skilled in the art can adjust them according to image resolution, video memory size, number of centers, sequence type, and target region category.

[0052] Example 2 This embodiment is a system embodiment, providing a three-dimensional target region segmentation system for multicenter nasopharyngeal carcinoma MRI images, such as... Figure 3 As shown, it includes: The case record module is used to acquire nasopharyngeal carcinoma magnetic resonance images and their target region annotations from multiple source centers, and to establish a case record; the magnetic resonance images include at least one of T2-weighted images, T1-weighted images, and enhanced T1-weighted images, and the target region annotations include the primary tumor target region and / or metastatic lymph node target regions; the case record includes at least a case identifier, a center identifier, an image path, a target region annotation path, a voxel spacing, and an image size; The protocol signature calculation module is used to calculate a case-level protocol signature based on the case record. The case-level protocol signature includes at least a spatial resolution signature and a frequency domain energy signature. The spatial resolution signature includes at least one of in-plane voxel area, layer thickness, and anisotropy ratio. The frequency domain energy signature includes at least one of low-frequency energy ratio, high-frequency energy ratio, high-low frequency energy ratio, and their logarithmic values. The source center quota sampling module is used to select source centers during the training phase based on center identifiers and case-level protocol signatures, then select cases within the selected source centers, and sample three-dimensional image patches from the selected cases according to target area foreground constraints. The protocol perturbation enhancement module is used to determine the perturbation type and intensity based on the case-level protocol signature and multi-source center protocol distribution of the selected three-dimensional image block, and to perform protocol perturbation enhancement on the three-dimensional image block according to the protocol perturbation type and protocol perturbation intensity to generate a protocol perturbation image block corresponding to the target area annotation space. The training module is used to input protocol perturbation image patches into a 3D segmentation network and output target region category probabilities; calculate training loss based on the target region category probabilities and the target region annotations, update the parameters of the 3D segmentation network based on the training loss, and repeatedly perform source center quota sampling, protocol perturbation enhancement and network parameter updates according to preset training stages; The inference module is used to perform preprocessing, sliding window prediction, overlapping region fusion, and original space re-attachment on the magnetic resonance images to be segmented. The output module is used to output a three-dimensional segmentation mask of the primary tumor target area and / or metastatic lymph node target area.

[0053] in: The spatial resolution signature is quantified by in-plane voxel area, layer thickness, and anisotropy ratio, and the central protocol coverage is calculated by combining the image matrix size and field of view. The in-plane voxel area is the product of the voxel spacing in the x and y directions, the layer thickness is the voxel spacing in the z direction, and the anisotropy ratio is a function of the voxel spacing in the z direction and the in-plane voxel spacing. The frequency domain energy signature is obtained by performing a Fourier transform on two-dimensional slices or three-dimensional volume data of magnetic resonance images, calculating the low-frequency energy ratio, high-frequency energy ratio, and high-low frequency energy ratio based on the amplitude energy of the central and peripheral regions of the spectrum, and introducing the 5% to 95% quantile range of the multi-center protocol distribution as the perturbation constraint boundary.

[0054] The sampling weight of the source center quota sampler is determined by dynamically adjusting the weight based on the ratio of center protocol coverage and target volume; the center protocol coverage is the ratio of image matrix size to field of view; the target volume ratio GTVp / GTVn is the ratio of primary tumor target volume GTVp to metastatic lymph node target volume GTVn; when the volume of metastatic lymph node target volume GTVn is <5cm³, the sampling priority of the corresponding case is increased; The target area foreground constraint includes: when the selected case has a primary tumor target area GTVp or a metastatic lymph node target area GTVn, selecting sampling anchor points near the primary tumor target area or metastatic lymph node target area with a preset probability, and cropping three-dimensional image blocks around the sampling anchor points; the preset probability of selecting anchor points near the primary tumor target area GTVp or metastatic lymph node target area GTVn increases linearly as the target area volume decreases.

[0055] The protocol perturbation types may include layer thickness direction perturbation, in-plane resolution perturbation, intensity perturbation, noise perturbation, bias field perturbation, and frequency domain amplitude perturbation, wherein the intensity perturbation includes at least one of brightness shift, contrast scaling, and gamma transformation; the frequency domain amplitude perturbation must simultaneously satisfy the dual constraints of spatial resolution signature and frequency domain energy signature, adjusting low-frequency or high-frequency amplitude components under case-level protocol signature constraints; The protocol perturbation strength is determined by the Euclidean distance between the case protocol signature and the multicenter protocol distribution, so that the protocol signature of the perturbed image block falls within the coverage envelope area composed of all source center protocol signatures.

[0056] The three-dimensional segmentation network is an anisotropic sensing network. The shallow layers of the anisotropic sensing network use in-plane convolutional kernels to learn high-resolution in-plane anatomical details, while the deep layers use three-dimensional convolutional kernels to fuse contextual information in the layer thickness direction. During downsampling, spatial compression in the layer thickness direction is delayed, and a protocol signature feature channel is added to the input layer to assist in learning cross-center protocol invariance features.

[0057] The training loss and optimization strategy include: using Dice-cross-entropy hybrid loss, dynamically adjusting the class weights of different target regions based on the volume ratio of GTVp to GTVn; recording the target region Dice coefficient and 95% Hausdorff distance index of each center, and iteratively optimizing the source center quota sampling weights through a feedback mechanism.

[0058] For the case to be segmented, the inference module first performs the same preprocessing as in the training phase, including resampling and intensity normalization. Due to the large size of 3D MRI images, the system can use a sliding window approach for inference. Each sliding window outputs a class probability or logits, and overlapping regions are fused using a Gaussian importance map or other weighting methods. The system performs a class selection operation on the fused results to obtain a class segmentation map in the resampled space, and then resamples it back to the original MRI image space using nearest neighbor interpolation, outputting GTVp and / or GTVn 3D masks.

[0059] Example 3 Based on the same inventive concept, this application provides an electronic device embodiment corresponding to Embodiment 1, as detailed in Embodiment 3.

[0060] like Figure 4 As shown, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can implement any of the implementation methods in Embodiment 1.

[0061] Since the electronic device described in this embodiment is the device used to implement the method in Embodiment 1 of this application, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in Embodiment 1 of this application. Therefore, how the electronic device implements the method in the embodiment of this application will not be described in detail here. Any device used by those skilled in the art to implement the method in the embodiment of this application falls within the scope of protection of this application.

[0062] Example 4 Based on the same inventive concept, this application provides a storage medium corresponding to Embodiment 1, as detailed in Embodiment 4.

[0063] like Figure 5 As shown, this embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it can implement any of the implementation methods in Embodiment 1.

[0064] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A training method for a three-dimensional target region segmentation model of multicenter nasopharyngeal carcinoma MRI images, characterized in that: Includes the following steps: S1. Obtain nasopharyngeal carcinoma magnetic resonance images and their target region annotations from multiple source centers. The magnetic resonance images include at least one of T2-weighted images, T1-weighted images, and enhanced T1-weighted images. The target region annotations include the primary tumor target region and / or the metastatic lymph node target region. S2. Establish a case record for each case, wherein the case record includes at least the case identifier, center identifier, image path, target area annotation path, voxel spacing and image size; S3. Calculate the case-level protocol signature based on the case record. The case-level protocol signature includes at least a spatial resolution signature and a frequency domain energy signature. The spatial resolution signature includes at least one of in-plane voxel area, layer thickness, and anisotropy ratio. The frequency domain energy signature includes at least one of low-frequency energy ratio, high-frequency energy ratio, high-low frequency energy ratio, and their logarithmic values. S4. Construct a source center quota sampler based on the center identifier and the case-level protocol signature. During the training phase, the source center quota sampler first selects a source center, then selects a case within the selected source center, and samples a three-dimensional image patch from the selected case according to the target area foreground constraint. S5. Determine the protocol disturbance type and protocol disturbance intensity based on the case-level protocol signature of the case corresponding to the three-dimensional image block and its position in the distribution of multiple source center protocol signatures; S6. According to the protocol perturbation type and protocol perturbation intensity, perform protocol perturbation enhancement on the three-dimensional image block to generate a protocol perturbation image block corresponding to the target area annotation space; S7. Input the protocol-perturbed image patch into the 3D segmentation network and output the target region category probability; S8. Calculate the training loss based on the target region category probability and the target region annotation; S9. Update the parameters of the three-dimensional segmentation network according to the training loss, and repeat source center quota sampling, protocol perturbation enhancement and network parameter update according to the preset training stage; S10. Obtain a three-dimensional segmentation model for target region segmentation in magnetic resonance images of nasopharyngeal carcinoma without central focus.

2. The method according to claim 1, characterized in that: The calculation of the case-level protocol signature includes: The spatial resolution signature is quantified by the in-plane voxel area, layer thickness, and anisotropy ratio, and the central protocol coverage is calculated by combining the image matrix size and field of view. The in-plane voxel area is the product of the voxel spacing in the x and y directions, the layer thickness is the voxel spacing in the z direction, and the anisotropy ratio is a function of the voxel spacing in the z direction and the in-plane voxel spacing. The frequency domain energy signature is obtained by performing a Fourier transform on two-dimensional slices or three-dimensional volume data of magnetic resonance images, calculating the low-frequency energy ratio, high-frequency energy ratio, and high-low frequency energy ratio based on the amplitude energy of the central and peripheral regions of the spectrum, and introducing the 5% to 95% quantile range of the multi-center protocol distribution as the perturbation constraint boundary.

3. The method according to claim 1, characterized in that: The sampling weight of the source center quota sampler is determined by dynamically adjusting the weight based on the ratio of center protocol coverage and target volume; the center protocol coverage is the ratio of image matrix size to field of view; the target volume ratio GTVp / GTVn is the ratio of primary tumor target volume GTVp to metastatic lymph node target volume GTVn. When the GTVn volume of the metastatic lymph node target area is <5cm³, the sampling priority of the corresponding case is increased; The target area foreground constraint includes: when the selected case has a primary tumor target area GTVp or a metastatic lymph node target area GTVn, selecting sampling anchor points near the primary tumor target area or metastatic lymph node target area with a preset probability, and cropping three-dimensional image blocks around the sampling anchor points; the preset probability of selecting anchor points near the primary tumor target area GTVp or metastatic lymph node target area GTVn increases linearly as the target area volume decreases.

4. The method according to claim 3, characterized in that: The protocol perturbation types may include layer thickness direction perturbation, in-plane resolution perturbation, intensity perturbation, noise perturbation, bias field perturbation, and frequency domain amplitude perturbation, wherein the intensity perturbation includes at least one of brightness shift, contrast scaling, and gamma transformation; the frequency domain amplitude perturbation must simultaneously satisfy the dual constraints of spatial resolution signature and frequency domain energy signature, adjusting low-frequency or high-frequency amplitude components under case-level protocol signature constraints; The protocol perturbation strength is determined by the Euclidean distance between the case protocol signature and the multicenter protocol distribution, so that the protocol signature of the perturbed image block falls within the coverage envelope area composed of all source center protocol signatures.

5. The method according to claim 1, characterized in that: The three-dimensional segmentation network is an anisotropic sensing network. The shallow layers of the anisotropic sensing network use in-plane convolutional kernels to learn high-resolution in-plane anatomical details, while the deep layers use three-dimensional convolutional kernels to fuse contextual information in the layer thickness direction. During downsampling, spatial compression in the layer thickness direction is delayed, and a protocol signature feature channel is added to the input layer to assist in learning cross-center protocol invariance features.

6. The method according to claim 1, characterized in that: The magnetic resonance images include T2-weighted images as the main sequence and enhanced T1-weighted images as auxiliary sequences, and the processing method is as follows: Case-level protocol signatures for the main sequence and auxiliary sequences are calculated separately, and multi-sequence features are dynamically fused through an attention mechanism; Synchronization protocol perturbation is performed on multi-sequence images to ensure that the blur intensity in the layer thickness direction and the in-plane resolution scaling ratio remain consistent between sequences.

7. The method according to claim 1, characterized in that: The training loss adopts a Dice-cross-entropy hybrid loss, and the class weights of different target regions are dynamically adjusted according to the volume ratio of GTVp to GTVn; the Dice coefficient and 95% Hausdorff distance index of each center are recorded, and the source center quota sampling weights are iteratively optimized through a feedback mechanism.

8. A three-dimensional target region segmentation system for multicenter nasopharyngeal carcinoma MRI images, characterized in that: The case record module is used to acquire nasopharyngeal carcinoma magnetic resonance images and their target region annotations from multiple source centers, and to establish a case record; the magnetic resonance images include at least one of T2-weighted images, T1-weighted images, and enhanced T1-weighted images, and the target region annotations include the primary tumor target region and / or metastatic lymph node target regions; the case record includes at least a case identifier, a center identifier, an image path, a target region annotation path, a voxel spacing, and an image size; The protocol signature calculation module is used to calculate a case-level protocol signature based on the case record. The case-level protocol signature includes at least a spatial resolution signature and a frequency domain energy signature. The spatial resolution signature includes at least one of in-plane voxel area, layer thickness, and anisotropy ratio. The frequency domain energy signature includes at least one of low-frequency energy ratio, high-frequency energy ratio, high-low frequency energy ratio, and their logarithmic values. The source center quota sampling module is used to select source centers during the training phase based on center identifiers and case-level protocol signatures, then select cases within the selected source centers, and sample three-dimensional image patches from the selected cases according to target area foreground constraints. The protocol perturbation enhancement module is used to determine the perturbation type and intensity based on the case-level protocol signature and multi-source center protocol distribution of the selected three-dimensional image block, and to perform protocol perturbation enhancement on the three-dimensional image block according to the protocol perturbation type and protocol perturbation intensity to generate a protocol perturbation image block corresponding to the target area annotation space. The training module is used to input protocol perturbation image patches into a 3D segmentation network and output target region category probabilities; calculate training loss based on the target region category probabilities and the target region annotations, update the parameters of the 3D segmentation network based on the training loss, and repeatedly perform source center quota sampling, protocol perturbation enhancement and network parameter updates according to preset training stages; The inference module is used to perform preprocessing, sliding window prediction, overlapping region fusion, and original space re-attachment on the magnetic resonance images to be segmented. The output module is used to output a three-dimensional segmentation mask of the primary tumor target area and / or metastatic lymph node target area.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images