Transfer learning driven fully automatic segmentation method for intracranial tumor image data

Abstract

The application provides a transfer learning driven intracranial tumor image data full-automatic segmentation method, relates to the technical field of medical image segmentation, and comprises the following steps: pre-processing a brain MRI image of a target user, obtaining a standard brain MRI image, and analyzing to obtain brain image features; generating an image feature vector, setting an adaptive transfer learning scheme, calling a sample intracranial tumor image segmentation model library, and integrally constructing a first intracranial tumor image segmenter; training a convolutional neural network by using a sample ependymoma image dataset to construct a second intracranial tumor image segmenter, and setting an adaptive fusion weight; after the standard brain MRI image is segmented by using the first intracranial tumor image segmenter and the second intracranial tumor image segmenter, the intracranial tumor image segmentation result is fitted according to the adaptive fusion weight. The technical problem of weak generalization ability and easy segmentation deviation of an intracranial tumor image segmentation model in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation, and more particularly to a fully automated segmentation method for intracranial tumor image data driven by transfer learning. Background Technology

[0002] Precise segmentation of intracranial tumor images is a core technical support for clinical tumor diagnosis and classification, surgical planning, and postoperative efficacy monitoring. The segmentation accuracy directly determines the accuracy of doctors' judgment on the tumor boundary range, infiltration depth, and relationship with surrounding nerves and blood vessels, thereby affecting the formulation of personalized treatment strategies and the prognostic assessment of patients.

[0003] However, existing intracranial tumor image segmentation technology faces insurmountable bottlenecks: on the one hand, intracranial tumor pathological types cover dozens of subtypes, and the imaging characteristics of different subtypes differ significantly, requiring extremely high subtype adaptability of segmentation models; on the other hand, rare subtypes have low clinical incidence and scarce professionally labeled samples, resulting in insufficient training data for dedicated segmentation models for these subtypes. Consequently, intracranial tumor image segmentation models have weak generalization ability and are prone to segmentation bias, making it difficult to meet the clinical demand for high-precision and high-reliability segmentation results.

[0004] Therefore, there is an urgent need for a fully automated segmentation method for intracranial tumor imaging data driven by transfer learning to improve the model's adaptability to different pathological subtypes and segmentation accuracy. Summary of the Invention

[0005] This invention addresses the technical problems of weak generalization ability and easy segmentation deviation in existing intracranial tumor image segmentation models by providing a fully automatic segmentation method for intracranial tumor image data driven by transfer learning.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] This invention provides a fully automated segmentation method for intracranial tumor imaging data driven by transfer learning, including:

[0008] The brain MRI images of the target user are preprocessed according to the preset image processing strategy to obtain standard brain MRI images, and the brain image features are analyzed.

[0009] Based on the clinical condition characteristics of the target user and the brain imaging characteristics, an image feature vector is generated by fusing them. Based on the image feature vector, an adaptation transfer learning scheme is set and the sample intracranial tumor image segmentation model library is called to integrate and construct the first intracranial tumor image segmenter.

[0010] A second intracranial tumor image segmenter was constructed by training a convolutional neural network using a sample ependymoma image dataset, and the adaptation and fusion weights were evaluated and set based on the sample ependymoma image dataset.

[0011] After segmenting the standard brain MRI image using the first intracranial tumor image segmenter and the second intracranial tumor image segmenter, the intracranial tumor image segmentation result is obtained by fitting according to the adaptation fusion weight.

[0012] The beneficial effects of this invention are:

[0013] Compared to existing technologies, this application first preprocesses the target user's brain MRI images according to a preset image processing strategy to obtain standard brain MRI images and analyzes brain image features, providing high-quality standardized data and tumor-specific features for subsequent feature fusion and model matching. Secondly, it generates image feature vectors by fusing the target user's clinical condition characteristics and brain image features. Based on these image feature vectors, it sets an adaptation transfer learning scheme, calls a sample intracranial tumor image segmentation model library, and integrates to construct a first intracranial tumor image segmenter. This effectively solves the problems of single feature dimension, inefficient model calling, and diluted integration accuracy in transfer learning for intracranial tumors (especially rare subtypes such as ependymoma), improving segmentation accuracy, model adaptability, and processing efficiency. Thirdly, it uses a sample ependymoma image dataset to train a convolutional neural network to construct a second intracranial tumor image segmenter, and evaluates and sets adaptation fusion weights based on the sample ependymoma image dataset. This achieves dynamic adjustment of weights according to sample quality, providing an objective and scientific weight basis for the fusion of the two segmenters. Finally, by using the first and second intracranial tumor image segmenters to segment standard brain MRI images, the intracranial tumor image segmentation results were obtained by fitting according to the adaptive fusion weights, which comprehensively improved the accuracy and precision of intracranial tumor image segmentation.

[0014] Through the above technical solution, this application leverages the generalization ability of transfer learning to construct a first intracranial tumor image segmenter, compensating for the insufficient model training caused by a lack of samples of rare subtypes such as ependymoma. Simultaneously, a dedicated second intracranial tumor image segmenter is constructed using a sample ependymoma image dataset to specifically capture the pathological imaging features of ependymoma, and the adaptation and fusion weights are dynamically set based on data quality. In this way, the broad adaptability of the general model is balanced with the targeted nature of the dedicated model, improving the accuracy, reliability, and generalization ability of intracranial tumor (especially ependymoma) image segmentation, and meeting the needs of clinical diagnosis, classification, surgical planning, and efficacy monitoring for high-quality segmentation results. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the fully automated segmentation method for intracranial tumor image data driven by transfer learning provided by this invention.

[0016] Figure 2This is a flowchart illustrating the process of setting adaptation and fusion weights in the fully automated segmentation method for intracranial tumor image data driven by transfer learning provided by the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0020] Examples, such as Figure 1 As shown, this embodiment of the invention provides a fully automated segmentation method for intracranial tumor image data driven by transfer learning, including:

[0021] S10: Preprocess the target user's brain MRI images according to the preset image processing strategy to obtain standard brain MRI images and analyze the brain image features.

[0022] Raw brain MRI images of intracranial ependymomas are easily affected by differences in equipment and scanning parameters, resulting in inconsistent grayscale ranges, spatial dimension deviations, and the presence of electronic noise, motion artifacts caused by slight head movements, and cerebrospinal fluid flow artifacts. These interferences can obscure tumor boundaries and blur key imaging information, making it difficult for subsequent models to accurately identify the lesion. Furthermore, intracranial ependymomas have unique pathological imaging characteristics, significantly different from other intracranial tumors such as gliomas and metastases. Using general image processing methods may preserve irrelevant areas such as the skull and scalp, increasing computational redundancy.

[0023] To address the aforementioned issues, this application preprocesses the target user's brain MRI images according to a preset image processing strategy to obtain standard brain MRI images and analyzes the brain image features.

[0024] Specifically, step S10 in the method includes:

[0025] Obtain brain MRI images of a target user, wherein the target user is a patient with an intracranial ependymoma;

[0026] The brain MRI images are denoised and standardized according to a preset image processing strategy. The standardized images are then coarsely localized to the lesion region. Based on the coarse localization results, the brain MRI images are cropped to obtain standard brain MRI images.

[0027] Image features are extracted from the standard brain MRI images to obtain brain image features, wherein the brain image features include at least the probability distribution of tumor location, uniformity of enhancement, enhancement ratio, clarity of boundary, and degree and morphology of peritumoral edema.

[0028] In this embodiment, brain MRI images of the target user are first acquired, wherein the target user is a patient with an intracranial ependymoma. Exemplarily, complete brain MRI images of the patient with an intracranial ependymoma are collected, typically including conventional structural sequences, such as T1-weighted images, T2-weighted images, and T1-enhanced images, to observe tumor morphology and enhancement characteristics. If necessary, functional sequences, such as diffusion-weighted imaging (DWI), can also be added to assist in determining tumor cell density.

[0029] Secondly, the brain MRI images are denoised and standardized according to a preset image processing strategy. The standardized images are then used for coarse localization of lesion regions. Based on the coarse localization results, the brain MRI images are cropped to obtain standard brain MRI images. The preset image processing strategy can be determined based on actual needs and the target user's brain MRI image quality.

[0030] For example, image denoising can employ Gaussian filtering to eliminate high-frequency electronic noise, nonlocal mean filtering (NLM) to remove motion artifacts caused by slight head movements during scanning, and wavelet thresholding to preserve tumor boundary details, thus avoiding noise from obscuring the tumor edge or causing artifacts to be misjudged as lesions. Especially considering that intracranial ependymomas are often adjacent to cerebrospinal fluid, denoising can effectively reduce the interference of cerebrospinal fluid flow artifacts on the tumor boundary.

[0031] For example, standardization can eliminate grayscale differences caused by different MRI equipment and scanning parameters in different hospitals through grayscale value standardization (such as z-score standardization), and unify the location and pixel size of brain structures through spatial standardization, so as to ensure that the key features of ependymoma have comparable grayscale thresholds in different images and avoid missing enhancement areas due to equipment differences.

[0032] For example, coarse localization of lesion areas can be achieved using lightweight algorithms, such as gray-scale threshold-based segmentation and region growing algorithms, to initially delineate the suspected tumor range. Considering that intracranial ependymomas are mostly located in the ventricular system, lesions can be searched for in the periventricular region first, reducing the processing of irrelevant areas such as the cerebral cortex.

[0033] For example, the brain MRI image is cropped based on the coarse localization results of the region, extending 5-10 pixels outside the boundary of the suspected lesion to avoid missing the edge of the tumor, and removing irrelevant tissues such as the skull and scalp, so as to finally obtain a standard brain MRI image containing only the lesion and a small amount of normal brain tissue around it, which reduces the amount of subsequent calculations and ensures that the model’s attention is focused on the core region.

[0034] Finally, image features were extracted from standard brain MRI images to obtain brain image features including at least the probability distribution of tumor location, enhancement homogeneity, enhancement ratio, boundary clarity, and degree and morphology of peritumoral edema. These features provide tumor-specific labels for subsequent fusion with clinical features and matching of transfer learning models, avoiding model mismatch caused by fuzzy feature dimensions.

[0035] Among them, the tumor location probability distribution is generated by using sliding window probability prediction or heatmap generation technology to output the probability of each pixel in a standard brain MRI image belonging to a tumor, forming a probability distribution heatmap. The tumor location probability distribution can help subsequent models quickly match tumor models with ventricles as the main lesion area and exclude irrelevant models such as glioma of the cerebral cortex.

[0036] Among them, enhancement homogeneity is based on T1 enhanced image sequence. First, the enhanced area of ​​the tumor is segmented, and then the gray standard deviation of the area is calculated. The smaller the standard deviation, the more homogeneous the enhancement. Ependymomas typically show homogeneous enhancement due to stable blood supply and less necrosis. Enhancement homogeneity can effectively distinguish glioblastomas, which show heterogeneous enhancement due to necrosis, and improve the subtype identification.

[0037] The enhancement ratio can be obtained by calculating the ratio of the enhanced area volume to the total tumor volume. The enhancement ratio of intracranial ependymomas is mostly above 70%, which is much higher than that of metastatic tumors with central necrosis. The enhancement ratio can quantify the integrity of tumor blood supply and help the model optimize the segmentation boundary.

[0038] Among them, boundary clarity can be assessed by calculating the gray-level gradient value or edge continuity of the tumor boundary region. The larger the gray-level gradient value, the clearer the boundary. Intracranial ependymomas have a complete capsule and a clear boundary with normal brain tissue. Boundary clarity can help the model accurately locate the tumor edge and avoid misjudging the blurred area of ​​invasive growth as normal tissue.

[0039] The degree of peritumoral edema can be quantified by the ratio of edema volume to tumor volume, and the morphology of peritumoral edema can be described by indicators such as ellipticity and irregularity. The degree and morphology of peritumoral edema can be used to distinguish severe patchy edema in gliomas, avoid misjudging the extent of edema and resulting in an excessively large segmentation range, and ensure that each feature closely matches the pathological imaging characteristics of ependymoma, providing accurate feature support for subsequent procedures.

[0040] In summary, compared to existing technologies, this application preprocesses the target user's brain MRI images according to a preset image processing strategy to obtain standard brain MRI images and analyzes the brain image features. This provides high-quality standardized data and tumor-specific features for subsequent feature fusion and model matching.

[0041] S20: Based on the clinical condition characteristics of the target user and the brain imaging features, an image feature vector is generated by fusing them. Based on the image feature vector, an adaptation transfer learning scheme is set, and a sample intracranial tumor image segmentation model library is called to integrate and construct the first intracranial tumor image segmenter.

[0042] In traditional transfer learning applications of intracranial tumor image segmentation, the model usually relies solely on brain image features, ignoring individual patient differences. This results in insufficient model adaptability to rare subtypes. Furthermore, there are no clear standards for model selection and integration. Either a single model is directly selected, leading to poor adaptability, or multiple models are blindly integrated, causing the weights of highly similar models to be diluted, making the segmentation accuracy and efficiency unable to meet clinical needs.

[0043] To address the aforementioned issues, this application integrates the clinical condition characteristics of the target user with the brain imaging features to generate an image feature vector, sets an adaptation transfer learning scheme based on the image feature vector, calls a sample intracranial tumor image segmentation model library, and integrates and constructs a first intracranial tumor image segmenter.

[0044] Specifically, step S20 in the method includes:

[0045] The clinical condition characteristics of the target user are obtained, and the clinical condition characteristics are encoded and standardized to obtain a standardized clinical feature vector. The clinical condition characteristics include at least the patient's age, tumor location, clinical manifestations and past medical history.

[0046] Principal component analysis was used to reduce the dimensionality of the brain image features to obtain low-dimensional image feature vectors.

[0047] The standardized clinical feature vector and the low-dimensional image feature vector are spliced ​​and fused to generate an image feature vector.

[0048] In this embodiment, the clinical condition characteristics of the target user are first obtained, and then encoded and standardized to obtain a standardized clinical feature vector. These clinical condition characteristics include at least the patient's age, tumor location, clinical manifestations, and past medical history. These characteristics are directly related to the pathological features of intracranial ependymomas; for example, ependymomas in children are more common in the fourth ventricle, while those in adults are more common in the lateral ventricles. A history of radiotherapy may lead to blurred tumor boundaries. For example, non-numerical features, such as clinical manifestations and tumor location, are converted to numerical values ​​using label encoding or one-hot encoding. For instance, 1 represents headache, 2 represents vomiting, ..., and [1, 0] represents the fourth ventricle, [0, 1] represents the lateral ventricle, ..., thus addressing the problem that non-numerical features cannot participate in model calculations. For example, for numerical features, such as patient age, methods such as z-score normalization and Min-Max normalization can be used to scale the numerical features to the range of [0, 1] or [-1, 1]. For example, 5 years old can be scaled down to 0.1 to eliminate the excessive weight amplification caused by the difference in magnitude, and finally a standardized clinical feature vector is obtained, such as [0.1, 1, 0, 0, 0.2], which correspond to age 5 years old, fourth ventricle tumor, no headache, and no past medical history, respectively.

[0049] Secondly, since brain imaging features typically contain multiple sub-dimensions—for example, the probability distribution of tumor location may cover probability values ​​from 10 brain regions such as the frontal lobe, temporal lobe, and ventricles, totaling 10 dimensions—directly using these for fusion would lead to the curse of dimensionality. Therefore, principal component analysis (PCA) is used to reduce the dimensionality of brain imaging features, resulting in low-dimensional image feature vectors. For example, PCA is used for dimensionality reduction of brain imaging features. While preserving the core information of the image features, PCA maps high-dimensional features to a low-dimensional space through linear transformation. For instance, it reduces the 10-dimensional location probability distribution to a 3-dimensional vector, maximizing the variance of the low-dimensional vector. This means that each low-dimensional dimension reflects the key differences in the original features, ultimately yielding a low-dimensional image feature vector that reduces computational burden while preserving key differences in tumor morphology.

[0050] Finally, the standardized clinical feature vector and the low-dimensional image feature vector are concatenated and fused to generate an image feature vector. For example, the standardized clinical feature vector [0.1, 1, 0, 0, 0.2] and the low-dimensional image feature vector [0.8, 0.3, 0.9, 0.1, 0.7] are directly concatenated in dimensional order to obtain the fused image feature vector [0.1, 1, 0, 0, 0.2, 0.8, 0.3, 0.9, 0.1, 0.7]. In this way, a standardized clinical feature vector that can both characterize the objective morphology of the tumor and reflect the individual patient's condition is obtained, providing accurate labels for subsequent matching transfer learning models.

[0051] Furthermore, the construction steps of the "sample intracranial tumor image segmentation model library" include:

[0052] Several intracranial tumor image segmentation models were collected. Each intracranial tumor image segmentation model corresponds to a sample image training dataset. The sample image training dataset includes intracranial tumor images, clinical disease characteristics, and brain image characteristics.

[0053] High-frequency feature clustering was performed on the clinical condition features and brain image features of the samples in several training datasets of sample images, and several sample image feature vector matrices were constructed based on the several high-frequency feature datasets.

[0054] The sample intracranial tumor image segmentation model library is constructed by mapping and combining the several sample intracranial tumor image segmentation models and the several sample image feature vector matrices.

[0055] In this embodiment, several intracranial tumor image segmentation models trained with different subtypes of intracranial tumors are first collected from existing research results or clinical imaging databases. These include intracranial tumor image segmentation models for gliomas, meningiomas, and ependymomas built based on mainstream segmentation networks such as U-Net and V-Net. Each intracranial tumor image segmentation model corresponds to a sample image training dataset, which includes the intracranial tumor image, clinical features, and brain image features. This ensures that the training input feature dimensions of each intracranial tumor image segmentation model are fully matched with the fusion feature vector of the target user, avoiding the inability to effectively reuse the model during transfer learning due to incompatible feature dimensions.

[0056] Secondly, high-frequency feature clustering is performed on the clinical condition features and brain image features of the samples in several training datasets of sample images, and several sample image feature vector matrices are constructed based on the several high-frequency feature datasets. For example, algorithms such as K-Means and hierarchical clustering can be used to perform high-frequency feature clustering on the clinical condition features and brain image features of the samples in several training datasets of sample images. Here, high-frequency features refer to feature combinations that appear frequently and are highly representative in several training datasets of sample images. For example, through high-frequency feature clustering, the high-frequency features of ependymoma are obtained as "child patients, fourth ventricle tumor, mild edema, homogeneous enhancement", and the high-frequency features of glioma are "adult patients, cerebral hemisphere, severe edema, heterogeneous enhancement". The purpose of high-frequency feature clustering is to group samples with similar features into a high-frequency feature dataset, and then construct several sample image feature vector matrices based on several high-frequency feature datasets. Each row of the sample image feature vector matrix is ​​a fused feature vector of a sample, and each column is a feature dimension. In this way, the original tens of thousands of scattered sample features are simplified into several sample image feature vector matrices. During subsequent matching, it is only necessary to compare the similarity between the target vector and the matrix, without having to compare individual samples one by one, thus improving matching efficiency.

[0057] Finally, a sample intracranial tumor image segmentation model library is constructed by mapping and combining several sample intracranial tumor image segmentation models and several sample image feature vector matrices. For example, the sample intracranial tumor image segmentation model trained with ependymoma samples corresponds to the sample image feature vector matrix for "pediatric patients, fourth ventricle tumor, mild edema, homogeneous enhancement," and the sample intracranial tumor image segmentation model trained with glioma samples corresponds to the sample image feature vector matrix for "adult patients, cerebral hemisphere, severe edema, heterogeneous enhancement," thus forming the sample intracranial tumor image segmentation model library. In this way, by mapping and combining to establish a feature index for the sample intracranial tumor image segmentation model library, after the target user inputs the fused feature vector, the most similar feature vector matrix can be found first, and then the corresponding sample intracranial tumor image segmentation model can be directly called, avoiding blindly screening from a massive number of models.

[0058] Specifically, the phrase "based on the image feature vector, setting an adaptation transfer learning scheme, calling the sample intracranial tumor image segmentation model library, and integrating and constructing a first intracranial tumor image segmenter" includes:

[0059] The image feature vector is compared with the feature vector matrices of the sample images to obtain the feature similarity of the sample images.

[0060] If the number of samples with feature similarity greater than the first similarity threshold is not zero, then the intracranial tumor image segmentation model corresponding to the sample with the highest feature similarity is selected as the first intracranial tumor image segmenter.

[0061] If the number of samples with similarity greater than the first similarity threshold is 0 and the number of samples with similarity greater than the second similarity threshold is not 0, then multiple intracranial tumor image segmentation models that meet the conditions are selected to integrate and construct the first intracranial tumor image segmenter, wherein the first similarity threshold is greater than the second similarity threshold.

[0062] If the number of samples with similarity values ​​greater than the second similarity threshold is 0 and the number of samples with similarity values ​​greater than the third similarity threshold is not 0, then all samples are selected to integrate intracranial tumor image segmentation models to construct a first intracranial tumor image segmenter, wherein the second similarity threshold is greater than the third similarity threshold.

[0063] In this embodiment, the image feature vector is first compared with the feature vector matrices of several sample images to obtain the similarity of several sample features. For example, the similarity between the image feature vector and the feature vector matrices of several sample images can be calculated by cosine similarity to obtain several sample feature similarities, and each sample feature similarity corresponds to a sample intracranial tumor image segmentation model.

[0064] Secondly, if the number of samples with feature similarity greater than the first similarity threshold is not zero, the intracranial tumor image segmentation model corresponding to the sample with the highest feature similarity is selected as the first intracranial tumor image segmenter. The first similarity threshold can be set relatively high, such as 0.8-0.9, representing a high degree of fit. For example, if the sample feature similarity between the image feature vector and the feature vector matrix of a certain sample image is 0.85, which is greater than the first similarity threshold of 0.8, then the intracranial tumor image segmentation model corresponding to the sample with the highest feature similarity is directly selected as the first intracranial tumor image segmenter. In this way, when high-frequency features are highly consistent, the accuracy requirements can be met without integration, avoiding increased computation and achieving fast segmentation.

[0065] Secondly, if the number of samples with similarity greater than the first similarity threshold is 0 and the number of samples with similarity greater than the second similarity threshold is not 0, then multiple intracranial tumor image segmentation models that meet the conditions are selected and integrated to construct the first intracranial tumor image segmenter. The first similarity threshold is greater than the second similarity threshold, and the second similarity threshold can be set slightly lower than the first similarity threshold, such as 0.6-0.7, representing moderate fit. For example, if the sample feature similarity between the image feature vector and the feature vector matrices of two sample images is 0.72 and 0.68 respectively, both less than the first similarity threshold of 0.8, but greater than the second similarity threshold of 0.65, then the corresponding two sample intracranial tumor image segmentation models are selected and integrated to construct the first intracranial tumor image segmenter. The integration of the first intracranial tumor image segmenter can employ integration strategies such as voting or weighted averaging, fusing the outputs of multiple sample intracranial tumor image segmentation models. Thus, when the target user's features exhibit cross-subtype commonality, a single sample intracranial tumor image segmentation model cannot fully cover the features, requiring multiple models to complement each other. Integration can reduce the bias of a single model and improve segmentation reliability.

[0066] Finally, if the number of samples with similarity greater than the second similarity threshold is 0 and the number of samples with similarity greater than the third similarity threshold is not 0, then all intracranial tumor image segmentation models from the sample intracranial tumor image segmentation model library are integrated to construct the first intracranial tumor image segmenter. The second similarity threshold is greater than the third similarity threshold, and the third similarity threshold can be set slightly lower than the first similarity threshold, such as 0.4-0.5, representing a low fit but with some correlation. For example, if the sample feature similarity between the image feature vector and the feature vector matrices of two sample images is 0.55 and 0.48 respectively, both less than the second similarity threshold of 0.65, but greater than the third similarity threshold of 0.45, then all intracranial tumor image segmentation models from the sample intracranial tumor image segmentation model library are integrated to construct the first intracranial tumor image segmenter. Thus, when the target user's features are special, such as rare adult fourth ventricle ependymoma, where there are no perfectly matching samples in the high-frequency features, then all information from the sample intracranial tumor image segmentation model library is utilized to the maximum extent possible, thereby compensating for the deficiency of insufficient intracranial ependymoma samples.

[0067] Thus, by setting three levels of similarity thresholds, the first intracranial tumor image segmenter is integrated into three scenarios based on the similarity between the image feature vector and the sample image feature vector matrix, ensuring that the selected model is most suitable for the target user's features.

[0068] Furthermore, the phrase "selecting all sample intracranial tumor image segmentation models to integrate and construct the first intracranial tumor image segmenter" includes:

[0069] Calculate the ratio of the feature similarity of each sample to the mean of the feature similarities of several samples, round it down, and set it as the sample model call count. Obtain several sample model call counts, where the sample model call count is greater than or equal to 1.

[0070] Based on the number of calls to the sample models, the first intracranial tumor image segmenter is constructed by integrating the sample intracranial tumor image segmentation models.

[0071] In this embodiment, the ratio of each sample feature similarity to the mean of several sample feature similarities is first calculated and rounded down to the number of sample model calls, thus obtaining several sample model call counts. The number of sample model calls is greater than or equal to 1. For example, if the similarities of three sample features are 0.8, 0.5, and 0.5 respectively, then the mean sample feature similarity = (0.8 + 0.5 + 0.5) / 3 = 0.6. Therefore, the number of sample model calls for the three samples are: 0.8 / 0.6 ≈ 2 (rounded up), 0.5 / 0.6 ≈ 1 (rounded up), and 0.5 / 0.6 ≈ 1 (rounded up), meaning the three sample models are called 2 times, 1 time, and 1 time respectively.

[0072] Secondly, a first intracranial tumor image segmenter is constructed by integrating several sample intracranial tumor image segmentation models based on the number of times each sample model is invoked. For example, based on the calculated number of times each sample model is invoked, sample models corresponding to the number of invocations are selected from all sample intracranial tumor image segmentation models. For instance, three sample models might be invoked 2 times, 1 time, and 1 time, respectively. Then, an integration algorithm is used, such as averaging the segmentation results after multiple invocations or weighted summation based on the number of invocations, to construct the first intracranial tumor image segmenter. Thus, in the scenario of integrating all sample intracranial tumor image segmentation models, the differentiation in the number of invocations ensures that the segmentation logic of sample models with high sample feature similarity dominates, thereby improving the reliability of the integration results.

[0073] In summary, compared to existing technologies, this application generates image feature vectors by fusing the clinical characteristics of the target user and the brain imaging features. Based on these image feature vectors, an adaptive transfer learning scheme is set to call a sample intracranial tumor image segmentation model library, and a first intracranial tumor image segmenter is constructed. This effectively solves the problems of single feature dimensions, inefficient model calling, and diluted integration accuracy in transfer learning for intracranial tumors (especially rare subtypes such as ependymoma), improving segmentation accuracy, model adaptability, and processing efficiency.

[0074] S30: A second intracranial tumor image segmenter is constructed by training a convolutional neural network using a sample ependymoma image dataset, and the adaptation and fusion weights are evaluated and set according to the sample ependymoma image dataset.

[0075] The first intracranial tumor image segmenter relies on a transfer learning strategy and integrates and constructs a sample intracranial tumor image segmentation model library. It is a general segmentation model that is compatible with multiple intracranial tumor subtypes. By leveraging the general image feature transfer capabilities of existing models, it has good generalization performance and can adapt to the segmentation needs of intracranial tumor images of different scanning devices and different common pathological types. However, it is limited by the low proportion of ependymoma samples in the model library and the fact that the general model has not carried out targeted learning for the specific features of ependymoma. It lacks the exclusive adaptation capability for rare subtypes such as ependymoma and it is difficult to accurately capture its unique pathological image manifestations to meet the needs of high-precision segmentation.

[0076] To address the aforementioned issues, this application employs a sample ependymoma image dataset to train a convolutional neural network to construct a second intracranial tumor image segmenter, and evaluates and sets adaptation and fusion weights based on the sample ependymoma image dataset.

[0077] For example, the sample ependymoma image dataset is labeled data specifically for intracranial ependymoma, containing clinical disease characteristics, brain imaging features, and corresponding manually labeled masks of the tumor region. A convolutional neural network (CNN) can be selected to extract hierarchical features of the images through multiple convolutions. During training, sample images are input into the CNN, which outputs a predicted tumor segmentation mask. The loss between the predicted mask and the manually labeled mask is then calculated, and the network parameters are iteratively optimized using backpropagation to obtain a second intracranial tumor image segmenter. The second intracranial tumor image segmenter is a model specifically for ependymoma, and its recognition accuracy for the unique features of ependymoma (such as the tendency of the periventricular location and uniform enhancement pattern) is much higher than that of general tumor segmentation models. It can provide more targeted segmentation results, forming a complementary relationship of generalization and specificity with the first intracranial tumor image segmenter (a general model based on transfer learning).

[0078] It should be clarified that the convolutional neural network (CNN) used to construct the second intracranial tumor image segmenter in this application has a basic network structure (such as mainstream architectures like U-Net adapted for medical image segmentation) and conventional training details that are existing mature technologies in this field and are not the core innovative direction of this application. Therefore, the specific structural design and basic training steps of the CNN will not be described in detail here.

[0079] Furthermore, such as Figure 2 As shown, step S30 in the method includes:

[0080] Feature extraction was performed on multiple sample ependymoma image data in the sample ependymoma image dataset to construct multiple sample data feature vectors;

[0081] The multiple sample data feature vectors are distributed in a three-dimensional space to evaluate the discreteness of the sample data distribution and output the discreteness of the sample data distribution.

[0082] The sample data volume in the sample ependymoma image dataset is obtained, and the sample data fitness coefficient is evaluated based on the sample data volume and the sample data distribution dispersion, wherein the sample data fitness coefficient is positively correlated with the sample data volume and the sample data distribution dispersion.

[0083] The ratio of the fitness coefficient of the sample data to the fitness coefficient of the preset standard data is multiplied by the second initial weight and rounded to obtain the second adaptation weight, wherein the second initial weight is 0.5 and the second adaptation weight is greater than 0.2 and less than 0.8.

[0084] The first adaptation weight is obtained by subtracting the second adaptation weight from 1, and the first adaptation weight and the second adaptation weight are used as the adaptation fusion weight.

[0085] In this embodiment, firstly, feature extraction is performed on multiple sample ependymoma image data in the sample ependymoma image dataset according to the dimensions that are completely consistent with the target user feature extraction in steps S10 and S20. Multiple sample data feature vectors containing sample clinical condition features and sample brain image features are constructed. In this way, it can be ensured that the sample ependymoma image data are evaluated under the same feature dimensions, laying the foundation for objectively judging the quality of the sample dataset.

[0086] Secondly, multiple sample data feature vectors can be distributed into a three-dimensional space through spatial mapping and algorithmic analysis. If the dimension of multiple sample data feature vectors exceeds 3, principal component analysis can be used to reduce the dimension to 3. Then, algorithms such as cluster compactness, feature variance summation, or convex hull volume are used to evaluate the dispersion of the sample data distribution and output the dispersion of the sample data distribution. High dispersion indicates a wide range of features covered by the samples and strong representativeness of the ependymoma subtype; conversely, low dispersion indicates concentrated sample features and incomplete dataset coverage. In this way, the feature richness of multiple sample data feature vectors can be accurately captured.

[0087] Next, the sample data volume in the ependymoma image dataset is obtained, and the sample data fitness coefficient is evaluated based on the sample data volume and the dispersion of the sample data distribution. The sample data fitness coefficient is positively correlated with both the sample data volume and the dispersion of the sample data distribution. For example, the sample data fitness coefficient can be calculated using the following formula: Sample data fitness coefficient = (Sample data volume / Preset standard sample size) × (Sample data distribution dispersion / Preset standard dispersion). The sample data fitness coefficient is positively correlated with both the sample data volume and the dispersion of the sample data distribution; that is, the larger the sample data volume, the more sufficient the model training data; the higher the dispersion of the sample data distribution, the stronger the model's generalization ability. The sample data fitness coefficient is a quantitative reflection of the training potential of a dedicated model and can provide a basis for subsequent weight adjustments.

[0088] Furthermore, the ratio of the sample data fitness coefficient to the preset standard data fitness coefficient is multiplied by the second initial weight and rounded to obtain the second adaptation weight. The preset standard data fitness coefficient can be dynamically set according to actual conditions; for example, it can be set to 1. The second initial weight is 0.5, and the second adaptation weight is greater than 0.2 and less than 0.8. This boundary constraint on the second adaptation weight is because a weight below 0.2 would excessively weaken the role of the specialized model, while a weight above 0.8 would ignore the generalization and gap-filling capabilities of the general model. In this way, the weight ratio of the second intracranial tumor image segmenter is dynamically adjusted based on the sample data quality, achieving dynamic adaptation where higher sample quality results in a higher weight for the second intracranial tumor image segmenter, avoiding the rigidity of fixed weights.

[0089] Finally, the first adaptation weight is obtained by subtracting the second adaptation weight from 1, and the first and second adaptation weights are used as the adaptation fusion weights. Since the weights of the first intracranial tumor image segmenter (a transfer learning general model) and the second intracranial tumor image segmenter (a specialized model) need to jointly cover the entire contribution of the segmentation result, the first adaptation weight = 1 - the second adaptation weight. For example, if the second adaptation weight is 0.7, then the first adaptation weight = 1 - 0.7 = 0.3, meaning that 70% of the segmentation result depends on the specificity of the specialized model, and 30% depends on the generalization of the general model. This complementary weight design can leverage the advantages of the specialized model when the sample quality is good, and can rely on the general model to compensate when the sample quality is poor, ensuring that the segmentation result after weighted fitting is stable and reliable.

[0090] In summary, compared to existing technologies, this application uses a sample ependymoma image dataset to train a convolutional neural network to construct a second intracranial tumor image segmenter, and evaluates and sets adaptation and fusion weights based on the sample ependymoma image dataset. This achieves dynamic adjustment of weights according to sample quality, providing an objective and scientific basis for the fusion of the two segmenters.

[0091] S40: After segmenting the standard brain MRI image using the first intracranial tumor image segmenter and the second intracranial tumor image segmenter, the intracranial tumor image segmentation result is obtained by fitting according to the adaptation fusion weight.

[0092] This application utilizes a first intracranial tumor image segmenter and a second intracranial tumor image segmenter to segment the standard brain MRI image, and then obtains the intracranial tumor image segmentation result by fitting according to the adaptation and fusion weights. For example, the first and second intracranial tumor image segmenters respectively segment the preprocessed standard brain MRI image. The first intracranial tumor image segmenter (a general model integrated through transfer learning) has good generalization ability and outputs basic segmentation results that can adapt to various intracranial tumor features, avoiding the risk of missing rare subtypes by a single dedicated model. The second intracranial tumor image segmenter (an ependymoma-specific model) outputs accurate segmentation results for ependymoma-specific features based on dedicated training data, compensating for the general model's insufficient adaptation to sub-subtypes.

[0093] Secondly, based on the adaptation and fusion weights (e.g., the second adaptation weight is 0.7 and the first adaptation weight is 0.3), the two segmentation results are subjected to pixel-level or region-level weighted fitting, and finally integrated to form an intracranial tumor image segmentation result that takes into account both generalization reliability and subtype specificity. In this way, the segmentation deviation of the general model for ependymoma is avoided, and the accuracy risk of the dedicated model due to sample quality fluctuations is reduced, thus comprehensively improving the accuracy and precision of intracranial tumor image segmentation.

[0094] In summary, the embodiments of this application have at least the following technical effects:

[0095] Compared to existing technologies, this application first preprocesses the target user's brain MRI images according to a preset image processing strategy to obtain standard brain MRI images and analyzes the brain image features. This provides high-quality standardized data and tumor-specific features for subsequent feature fusion and model matching.

[0096] Secondly, this application generates image feature vectors by fusing the clinical characteristics of the target user's condition with the brain imaging features. Based on these image feature vectors, an adaptive transfer learning scheme is set up to call up a sample intracranial tumor image segmentation model library, and an integrated first intracranial tumor image segmenter is constructed. This effectively solves the problems of single feature dimension, inefficient model calling, and diluted integration accuracy in transfer learning for intracranial tumors (especially rare subtypes such as ependymoma), improving segmentation accuracy, model adaptability, and processing efficiency.

[0097] Furthermore, this application uses a sample ependymoma image dataset to train a convolutional neural network to construct a second intracranial tumor image segmenter, and evaluates and sets adaptation and fusion weights based on the sample ependymoma image dataset. In this way, the weights are dynamically adjusted according to the sample quality, providing an objective and scientific basis for the fusion of the two segmenters.

[0098] Finally, this application utilizes the first and second intracranial tumor image segmenters to segment the standard brain MRI images, and then fits the segmentation results of the intracranial tumor images according to the adaptive fusion weights. This avoids the segmentation bias of general models for ependymomas and reduces the accuracy risk of dedicated models due to sample quality fluctuations, thus comprehensively improving the accuracy and precision of intracranial tumor image segmentation.

[0099] Through the above technical solution, this application leverages the generalization ability of transfer learning to construct a first intracranial tumor image segmenter, compensating for the insufficient model training caused by a lack of samples of rare subtypes such as ependymoma. Simultaneously, a dedicated second intracranial tumor image segmenter is constructed using a sample ependymoma image dataset to specifically capture the pathological imaging features of ependymoma, and the adaptation and fusion weights are dynamically set based on data quality. In this way, the broad adaptability of the general model is balanced with the targeted nature of the dedicated model, improving the accuracy, reliability, and generalization ability of intracranial tumor (especially ependymoma) image segmentation, and meeting the needs of clinical diagnosis, classification, surgical planning, and efficacy monitoring for high-quality segmentation results.

[0100] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0106] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A transfer learning driven fully automatic segmentation method for intracranial tumor image data, characterized in that, The method comprises: According to the preset image processing strategy, the brain MRI image of the target user is preprocessed, the standard brain MRI image is obtained, and the brain image features are analyzed and obtained; According to the clinical condition characteristics of the target user and the brain image features, an image feature vector is generated, and a sample intracranial tumor image segmentation model library is called based on the image feature vector to integrate and construct a first intracranial tumor image segmentation device; A second intracranial tumor image segmentation device is constructed by training a convolutional neural network using a sample ependymoma image dataset, and an adaptive fusion weight is set according to the sample ependymoma image dataset; After the standard brain MRI image is segmented by the first intracranial tumor image segmentation device and the second intracranial tumor image segmentation device, an intracranial tumor image segmentation result is fitted according to the adaptive fusion weight; According to the sample ependymoma image dataset, the adaptive fusion weight is set, comprising: Feature extraction is performed on multiple sample ependymoma image data in the sample ependymoma image dataset to construct multiple sample data feature vectors; The multiple sample data feature vectors are distributed in a three-dimensional space for sample data distribution discrete evaluation, and sample data distribution discrete degree is output; The sample data amount in the sample ependymoma image dataset is obtained, and a sample data adaptation coefficient is obtained according to the sample data amount and the sample data distribution discrete degree evaluation, wherein the sample data adaptation coefficient is positively correlated with the sample data amount and the sample data distribution discrete degree; The ratio of the sample data adaptation coefficient to a preset standard data adaptation coefficient is multiplied by a second initial weight to obtain a second adaptive weight, wherein the second initial weight is 0.5, and the second adaptive weight is greater than 0.2 and less than 0.8; The first adaptive weight is obtained by subtracting the second adaptive weight from 1, and the first adaptive weight and the second adaptive weight are used as the adaptive fusion weight. 2.The transfer learning driven fully automated segmentation of intracranial tumor image data method of claim 1, wherein, According to the preset image processing strategy, the brain MRI image of the target user is preprocessed, the standard brain MRI image is obtained, and the brain image features are analyzed and obtained, comprising: Obtain the brain MRI image of the target user, wherein the target user is an intracranial ependymoma patient; According to the preset image processing strategy, the brain MRI image is denoised and standardized, and the lesion region is roughly positioned after standardization. According to the region rough positioning result, the brain MRI image is cropped to obtain the standard brain MRI image; Image feature extraction is performed on the standard brain MRI image to obtain brain image features, wherein the brain image features at least include tumor position probability distribution, enhancement uniformity, enhancement ratio, boundary definition and peritumoral edema degree and morphology. 3.The transfer learning driven fully automated segmentation of intracranial tumor image data method of claim 1, wherein, According to the clinical condition characteristics of the target user and the brain image features, an image feature vector is generated, comprising: Obtain the clinical condition characteristics of the target user, encode and standardize the clinical condition characteristics to obtain a standardized clinical feature vector, wherein the clinical condition characteristics at least include patient age, tumor location, clinical manifestations and past medical history; The brain image features are processed by dimension reduction using principal component analysis to obtain low-dimensional image feature vectors; The standardized clinical feature vectors and the low-dimensional image feature vectors are spliced and fused to generate image feature vectors. 4.The transfer learning driven fully automated segmentation of intracranial tumor image data method of claim 1, wherein, The steps of constructing the sample intracranial tumor image segmentation model library include: Collecting a plurality of sample intracranial tumor image segmentation models, wherein each sample intracranial tumor image segmentation model corresponds to a sample image training data set, and the sample image training data includes sample intracranial tumor images, sample clinical condition features, and sample brain image features; Respectively clustering the sample clinical condition features and the sample brain image features in the plurality of sample image training data sets, and constructing a plurality of sample image feature vector matrices based on a plurality of high-frequency feature data sets; Mapping and combining the plurality of sample intracranial tumor image segmentation models and the plurality of sample image feature vector matrices to construct the sample intracranial tumor image segmentation model library.

5. The transfer learning driven intracranial tumor image data fully automatic segmentation method according to claim 4, characterized in that, Based on the image feature vectors, an adaptive transfer learning scheme is set to call the sample intracranial tumor image segmentation model library to integrate and construct a first intracranial tumor image segmentation device, including: Respectively comparing the image feature vectors with the plurality of sample image feature vector matrices to obtain a plurality of sample feature similarities; If the number of sample feature similarities greater than a first similarity threshold is not 0, selecting a sample intracranial tumor image segmentation model corresponding to the maximum sample feature similarity as the first intracranial tumor image segmentation device; If the number of sample feature similarities greater than a first similarity threshold is 0 and the number of sample feature similarities greater than a second similarity threshold is not 0, selecting a plurality of sample intracranial tumor image segmentation models satisfying the condition to integrate and construct the first intracranial tumor image segmentation device, wherein the first similarity threshold is greater than the second similarity threshold; If the number of sample feature similarities greater than a second similarity threshold is 0 and the number of sample feature similarities greater than a third similarity threshold is not 0, selecting all sample intracranial tumor image segmentation models to integrate and construct the first intracranial tumor image segmentation device, wherein the second similarity threshold is greater than the third similarity threshold.

6. The transfer learning driven intracranial tumor image data fully automatic segmentation method according to claim 5, characterized in that, Selecting all sample intracranial tumor image segmentation models to integrate and construct the first intracranial tumor image segmentation device, including: Respectively calculating the ratio of each sample feature similarity to the average of the plurality of sample feature similarities and rounding to an integer to obtain a sample model calling number, wherein the sample model calling number is greater than or equal to 1; Based on the plurality of sample model calling numbers, the plurality of sample intracranial tumor image segmentation models are mapped and called to integrate and construct the first intracranial tumor image segmentation device.

Images