Tumor prognosis dynamic evaluation method and system combining lesion segmentation and ai

Abstract

The application relates to the technical field of medical image processing, and discloses a tumor prognosis dynamic evaluation method and system combining lesion segmentation and AI. The method comprises the following steps: preprocessing brain image data to obtain standardized brain image data, and segmenting a brain tumor lesion to obtain a lesion segmentation result; extracting imaging features and morphological features based on the lesion segmentation result; performing prognosis risk evaluation on the standardized brain image data, the lesion segmentation result and fused features through a first evaluation model, a second evaluation model and a third evaluation model, to obtain a first evaluation result, a second evaluation result and a third evaluation result; comparing the current evaluation result with a historical evaluation result, calculating a tumor volume change rate and a morphological evolution index, and combining the three evaluation results, the tumor volume change rate, the morphological evolution index and a signal evolution index to generate a final prognosis evaluation conclusion. The application can realize multi-dimensional dynamic precise evaluation of brain tumor prognosis.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a method and system for dynamic assessment of tumor prognosis that combines lesion segmentation with AI. Background Technology

[0002] Brain tumors are among the most common primary diseases of the central nervous system, and their prognostic assessment is crucial for clinical treatment planning and improving patients' quality of life. Traditional brain tumor prognostic assessment relies primarily on the subjective judgment of physicians, analyzing tumor size, location, and morphology in imaging examinations to assess risk. However, this approach has significant limitations: firstly, the human eye struggles to capture subtle textural changes and signal characteristics in images, easily overlooking key prognostic indicators; secondly, the assessment results are greatly influenced by physicians' individual experience and cognitive differences, lacking objectivity and consistency.

[0003] With the rapid development of medical imaging and artificial intelligence technologies, tumor analysis methods based on radiomics and deep learning have gradually become a research hotspot. Existing technologies sometimes employ a single model for tumor segmentation and classification, failing to fully utilize the complementary information from multiple MRI sequences; other methods focus only on local tumor features, neglecting the spatial relationship between the tumor and surrounding brain tissue and key functional areas; still others use static assessment models, failing to incorporate patient historical follow-up data for dynamic prognostic monitoring. Furthermore, existing systems are relatively singular in feature extraction dimensions, often relying on only one of imaging or clinicopathological features, lacking deep fusion of multi-dimensional information.

[0004] More importantly, the prognosis of brain tumors is a dynamic evolutionary process, and the imaging manifestations of the same patient at different times may show drastically different trends. Some patients respond well to treatment, and the tumor continues to shrink; some patients experience pseudo-progression, with the tumor appearing to enlarge on imaging but actually a response to treatment; and some patients experience rapid recurrence after early stabilization. Current technologies lack effective temporal comparison mechanisms and dynamic adjustment strategies, making it difficult to identify these complex evolutionary patterns, resulting in insufficient accuracy and prospectiveness in prognostic assessment. Therefore, there is an urgent need for a novel prognostic assessment method for brain tumors that can integrate multi-dimensional information and achieve dynamic evaluation.

[0005] Therefore, the present invention provides a method and system for dynamic assessment of tumor prognosis that combines lesion segmentation and AI. Summary of the Invention

[0006] The embodiments in this specification provide the following technical solutions: Step S1: Preprocess the brain imaging data to obtain standardized brain imaging data, and segment the brain tumor lesions based on the standardized brain imaging data to obtain the lesion segmentation results; Step S2: Extract imaging and morphological features based on lesion segmentation results; Step S3: Use the first assessment model, the second assessment model, and the third assessment model to perform prognostic risk assessment on the standardized brain imaging data, lesion segmentation results, and fusion features to obtain the first assessment result, the second assessment result, and the third assessment result; Step S4: Compare the current assessment results with the historical assessment results, calculate the tumor volume change rate and morphological evolution indicators, and generate the final prognostic assessment conclusion based on the first assessment results, the second assessment results, and the third assessment results, combined with the tumor volume change rate, morphological evolution indicators, and signal evolution indicators.

[0007] Compared with the prior art, the beneficial effects of the present invention are at least as follows: The technical solution provided in this application effectively eliminates interference from equipment and individual differences through standardized preprocessing and precise lesion segmentation of multi-sequence image data, providing a high-quality data foundation for subsequent feature extraction and model evaluation. By constructing three evaluation models from different perspectives, prognostic risk analysis is conducted from three levels: whole-brain context, radiomics phenotype, and multi-dimensional fusion features. This achieves comprehensive quantification of tumor location, morphology, internal structure, and clinical information, significantly improving the robustness and accuracy of the evaluation results. By introducing time-series comparative analysis of historical evaluation results, tumor volume change rate, morphological evolution indicators, and signal evolution indicators are calculated, enabling dynamic monitoring and pattern recognition of tumor evolution trends. This upgrades prognostic evaluation from static judgment to dynamic tracking, enabling timely detection of key clinical events such as pseudoprogression and early recurrence. Finally, through cross-validation of multi-model results and dynamic correction of time-series information, a comprehensive prognostic evaluation conclusion and treatment recommendations are generated. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of an embodiment of a dynamic prognostic assessment method for tumors that combines lesion segmentation and AI, as described in this application. Figure 2 This is a schematic diagram of an embodiment of a dynamic prognostic assessment system for tumors that combines lesion segmentation and AI, as described in this application. Detailed Implementation

[0010] This application provides a method and system for dynamic assessment of tumor prognosis combining lesion segmentation and AI. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0011] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 An embodiment of a dynamic tumor prognosis assessment method combining lesion segmentation and AI, as described in this application, includes: Step S1: Preprocess the brain imaging data to obtain standardized brain imaging data, and segment the brain tumor lesions based on the standardized brain imaging data to obtain the lesion segmentation results; Specifically, brain imaging data refers to brain MRI image data. Different brain imaging data are affected by multiple factors during the acquisition process. When the same patient is examined at different times, parameters such as scan slice thickness, slice spacing, and flip angle may change. These differences lead to significant variations in brightness, contrast, and spatial resolution of the original images. If standardization is not performed, subsequent evaluation models may mistakenly identify these irrelevant variations as meaningful features, resulting in decreased segmentation accuracy and unstable evaluation results. In order to convert the original brain MRI image data into a standardized format and eliminate interference caused by differences in equipment, scanning parameters, and individual differences, and to provide stable and uniform input data for subsequent lesion segmentation and feature extraction, the above-mentioned interference factors are eliminated through preprocessing steps. Brain imaging data includes T1-weighted images, T2-weighted images, liquid attenuation inversion recovery sequence images, and enhanced T1-weighted images. After preprocessing, the standardized brain imaging data have the same spatial left system, consistent brightness distribution, and uniform numerical range.

[0012] To accurately identify and separate three key subregions of brain tumors from standardized brain MRI imaging data—the enhanced tumor core region, the peritumoral edema region, and the necrotic region—and obtain a spatial distribution map of the lesion with clear anatomical labels, the brain tumor lesion was segmented based on the standardized brain imaging data. The segmentation results included the tumor core region, the peritumoral edema region, and the necrotic region. These segmentation results provide precise data input for subsequent analysis. The high spatial accuracy of the 3D segmentation mask allows for accurate measurement of the volume and morphological parameters of each subregion. Accurate identification of the tumor core region enables physicians to accurately calculate the maximum vertical diameter required by tumor efficacy evaluation criteria. Quantification of the edema region helps assess the compression effect and infiltration extent of the tumor on surrounding brain tissue. Identification of the necrotic region provides a basis for determining the degree of tumor hypoxia and malignancy grade.

[0013] Step S2: Extract imaging and morphological features based on lesion segmentation results; Specifically, in order to extract two different types of numerical features from the lesion segmentation results, the first type is morphological features, which are used to describe the aggregate shape and spatial structural characteristics of the tumor, and the second type is imaging features, which are used to describe the signal distribution pattern and micro-texture characteristics inside the tumor. By extracting these two types of features, the macroscopic morphological appearance and microscopic tissue characteristics of the tumor can be quantified, providing multi-dimensional quantitative basis for prognostic assessment.

[0014] Step S3: Use the first assessment model, the second assessment model, and the third assessment model to perform prognostic risk assessment on the standardized brain imaging data, lesion segmentation results, and fusion features to obtain the first assessment result, the second assessment result, and the third assessment result; Specifically, since the prognosis of brain tumors is influenced by multiple factors, a single model cannot fully capture all relevant information. In order to quantify the prognostic risk of brain tumors from different perspectives and obtain complementary assessment perspectives, providing multi-dimensional decision-making basis for subsequent comprehensive judgment, a first assessment model, a second assessment model, and a third assessment model are used to assess the prognostic risk of standardized brain imaging data, lesion segmentation results, and fusion features, respectively, to obtain the first assessment result, the second assessment result, and the third assessment result. The first assessment model receives complete standardized brain imaging data and assesses the prognostic risk from the perspective of tumor location. The second assessment model receives lesion segmentation results, quantifies the malignancy of the tumor phenotype, and reflects the biological invasiveness of the tumor. The third assessment model receives fusion features and outputs individualized survival predictions. The differences in the outputs of the three assessment models reflect the reliability of different information sources and provide more decision-making information for subsequent comprehensive judgment.

[0015] Step S4: Compare the current assessment results with the historical assessment results, calculate the tumor volume change rate and morphological evolution indicators, and generate the final prognostic assessment conclusion based on the first assessment results, the second assessment results, and the third assessment results, combined with the tumor volume change rate, morphological evolution indicators, and signal evolution indicators.

[0016] Specifically, the prognosis of brain tumors is not static but evolves continuously with treatment intervention and disease progression. The imaging manifestations of the same patient at different time points may show drastically different trends. Some patients respond well to treatment, with the tumor continuously shrinking and the prognosis improving. Some patients experience pseudoprogression, where the tumor temporarily enlarges on imaging but is actually a response to treatment, and blind intervention may be harmful. Some patients experience rapid recurrence after early stabilization, requiring timely adjustment of treatment strategies. In order to compare and analyze the current assessment results with the patient's historical assessment data, quantify the tumor evolution trend, identify changes in growth patterns, and provide a time dimension basis for dynamic prognosis adjustment, a time-series comparison mechanism is established by comparing the current assessment results with historical assessment results, calculating the tumor volume change rate and morphological evolution indicators, and capturing these dynamic changes through longitudinal data analysis to achieve dynamic prognosis assessment and early warning. To integrate the outputs of the three assessment models and the evolutionary information from time-series comparisons, cross-validation is performed based on the first, second, and third assessment results. A final prognostic assessment conclusion is generated by combining tumor volume change rate, morphological evolution indicators, and signal evolution indicators. This is achieved through pre-defined judgment rules, balancing AI-assisted decision-making with clinical safety. The comprehensive judgment steps described above enable intelligent integration and clinical translation of diverse information, reducing the risk of errors from a single model. When one assessment model outputs abnormal results due to data quality issues, other assessment models can correct the error. When the three models are inconsistent, a conservative judgment strategy is triggered, classifying the case as uncertain. Then, the risk level is dynamically adjusted based on time-series comparison information.

[0017] In one specific embodiment, preprocessing brain imaging data to obtain standardized brain imaging data specifically includes the following steps: Adaptive filtering algorithms are used to suppress noise in brain imaging data. After noise suppression, intensity non-uniformity correction is performed on the image data. Positional deviations between different slices within the same sequence are corrected. The MRI brain imaging data are registered to the same coordinate system to achieve voxel-level alignment and registration of different sequences. The registered image data is then subjected to intensity standardization, and the gray values ​​of each sequence are mapped to a unified numerical range to obtain standardized brain MRI imaging data.

[0018] Specifically, to remove random noise and artifacts generated during the scanning process while preserving the sharpness of key areas at the tumor margins, an adaptive filtering algorithm was used to process the brain imaging data. To eliminate low-frequency signal deviations caused by magnetic field inhomogeneity and to ensure consistent signal intensity at different locations within the same tissue, intensity non-uniformity correction was performed on the noise-suppressed image data. To detect and correct image blurring and ghosting caused by minute head movements during the scan, image registration techniques were used to correct positional deviations between different slices within the same sequence, eliminating inter-slice misalignment caused by head movement during the scan. The four MRI sequences were obtained from four independent scans. The patient's head position was different during each scan, and the differences in scanning parameters between sequences resulted in different geometric characteristics of the images. To address these issues, the MRI brain imaging data (referring to the four different MRI sequences, including T1-weighted images, T2-weighted images, liquid attenuation inversion recovery sequence images, and enhanced T1-weighted images) were aligned to the same spatial coordinate system, achieving voxel-level correspondence between images of different contrasts, laying the foundation for subsequent multi-sequence fusion analysis. The gray values ​​of different MRI sequences have different physical meanings and numerical ranges. The gray value range of T1-weighted images may be zero to one thousand, T2-weighted images may be zero to two thousand, and enhanced T1-weighted images may be as high as four thousand due to the effect of contrast agents. Even for the same sequence, the gray value distribution of images scanned by different hospitals, different equipment, and different parameters is different. This diversity of numerical scales brings difficulties to automatic algorithms. It is difficult for algorithms to determine a unified threshold or standard to judge high signal or low signal. Therefore, normalization is used to convert images from different sources to the same numerical scale, realize the standardization of data, and obtain standardized brain MRI image data.

[0019] In one specific embodiment, segmenting a brain tumor lesion to obtain the lesion segmentation result includes the following steps: T1-weighted images, T2-weighted images, fluid attenuation inversion recovery sequence images, and enhanced T1-weighted images were fused to construct a four-channel input tensor. A deep learning segmentation network was used to encode and decode the four-channel input tensor to generate a three-dimensional segmentation mask, which included masks for the enhanced tumor region, the peritumoral edema region, and the necrotic core region. Connectivity analysis was performed on the three-dimensional segmentation mask to obtain optimized lesion segmentation results. The surface boundaries of the optimized lesion segmentation results were smoothed to eliminate jagged edges, resulting in the final lesion segmentation result.

[0020] Specifically, since a single MRI sequence can only reflect some characteristics of the tissue, T1-weighted imaging shows clear anatomical structures but is not sensitive to edema, while T2-weighted imaging is sensitive to edema but has lower anatomical contrast. Fluid attenuation inversion recovery sequence can suppress cerebrospinal fluid signals, making the boundary between the lesion and the ventricle clearer. Enhanced T1-weighted imaging shows areas of blood-brain barrier disruption but may miss non-enhancing infiltrating components. Each of the four sequences has its advantages and limitations. Only by combining them can we comprehensively capture different aspects of the tumor's characteristics. Therefore, based on the principle of multi-channel data representation, four three-dimensional matrices are stacked into a multi-channel tensor. Each sequence is treated as an independent channel, maintaining the original spatial resolution. The fused tensor contains four values ​​at each spatial location, corresponding to the signal intensity of the corresponding location in the four sequences. This representation method preserves the original information of each sequence while establishing spatial correspondence between them, facilitating subsequent joint analysis.

[0021] Brain tumors exhibit complex spatial structures on MRI, comprising multiple subregions with distinct biological characteristics. The enhanced tumor core represents active tumor tissue with disrupted blood-brain barrier function, serving as a direct therapeutic target. Peritumoral edema represents vasogenic edema and potential tumor infiltration, impacting neurological function and prognosis. The necrotic core represents ischemic necrosis of inactive tissue, indicating rapid tumor growth and hypoxia. Therefore, accurate identification of these three regions is crucial for treatment planning and efficacy evaluation. To identify and label these three key tumor subregions, a semantically labeled 3D segmentation mask is generated. A deep learning network based on an encoder-decoder architecture processes the four-channel input tensor... The encoding and decoding process involves the encoder extracting abstract features from the input data through multi-layer convolution and downsampling operations, from low-level edge textures to high-level tumor morphology. Each encoder layer halves the spatial resolution and doubles the number of feature channels, resulting in a compact feature representation containing global information at the highest layer. The decoder recovers the spatial resolution through upsampling and convolution operations, mapping the abstract features back to their original size. Skip connections pass high-resolution features from the encoder to the corresponding decoder layer, helping to recover fine spatial details. Finally, the output predicts the probability of each voxel for four categories, including the background and three tumor sub-regions, and takes the category with the highest probability as the segmentation label.

[0022] Deep learning network segmentation results may suffer from two common problems. The first is isolated small regions, such as several voxels far from the main lesion being incorrectly labeled as tumors. This could be a false positive caused by noise or artifacts. The second is voids within the mask, such as unlabeled background voxels appearing within enhanced tumor regions. This could be due to atypical signal characteristics of necrotic cores leading to missed segmentation. These defects affect the accuracy of segmentation results and the reliability of subsequent analysis, requiring post-processing correction. Simultaneously, medical image segmentation needs to meet certain topological constraints; for example, tumor regions should be connected or have a finite number of connected components. Therefore, connected component analysis identifies sets of interconnected voxels in 3D space, each set representing a connected component. By setting a volume threshold, connected components smaller than the threshold are removed, eliminating isolated false positive regions. Furthermore, morphological filling operations identify background voxels within the mask. If a voxel is completely surrounded by foreground voxels, it is labeled as foreground, filling the voids. These operations maintain the integrity of the large-scale structure while correcting local segmentation errors.

[0023] Since deep learning networks make predictions on individual voxels, the prediction results for adjacent voxels may be inconsistent. In other words, the segmentation boundaries of deep learning networks may appear stepped or jagged at the voxel level. For example, the tumor boundary should be a smooth curved surface, but the segmentation result may be pixelated and stepped. This irregularity affects the calculation of morphological features, such as surface area and irregularity, which may be overestimated. At the same time, jagged boundaries have poor visual effects in 3D visualization, which is not conducive to doctors' understanding and use. Therefore, a surface smoothing algorithm is used to operate directly on the 3D surface of the segmentation mask. The algorithm first extracts the isosurface of the mask to obtain a surface mesh composed of triangular facets. Then, the vertex positions of the mesh are adjusted to make the distance between adjacent vertices more uniform and the surface curvature change more gradual. The smoothing process is achieved through iterative optimization. In each iteration, the vertices are moved towards the average position of their surrounding neighbors, while the moving distance is limited to avoid over-smoothing and boundary offset. After several iterations, the surface is made smooth while the overall shape and position remain unchanged.

[0024] This step completes the entire process from raw MRI data to precise 3D segmentation. Multi-sequence feature fusion integrates various contrast information, 3D segmentation enables automated tumor identification, segmentation optimization eliminates errors and defects, and boundary smoothing improves visual and computational quality. The obtained lesion segmentation results contain three semantically clear sub-regions with accurate boundary localization and reasonable topological structure, providing a reliable spatial basis for subsequent prognostic assessment.

[0025] In one specific embodiment, obtaining a first evaluation result through a first evaluation model includes the following steps: Standardized brain MRI image data is input into a 3D convolutional neural network encoder to extract multi-scale whole-brain features. Based on the multi-scale whole-brain features, the cerebrospinal fluid region, gray matter region, and white matter region are segmented to construct a brain tissue probability atlas. Key brain functional regions are identified based on the standard brain function atlas and the brain tissue probability atlas. The spatial distance between the tumor region and the key brain functional regions is calculated. The multi-scale whole-brain features are compressed and integrated to extract a global feature vector. Based on the feature vector and spatial distance, the first assessment result is output through a fully connected layer. The first assessment result includes the global prognostic risk level and the first confidence level.

[0026] Specifically, the prognosis of brain tumors depends not only on the characteristics of the tumor itself but also on its location in the brain. When a tumor is located near a functional area, complete surgical removal is difficult, resulting in a high risk of residual tumor. When a tumor compresses the ventricular system, it can easily cause hydrocephalus and increased intracranial pressure. When a tumor grows across the midline, it indicates strong invasiveness. The global spatial relationships mentioned above need to be obtained through whole-brain analysis. Focusing only on the local tumor area will result in the loss of this contextual information, leading to an incomplete prognostic assessment. Therefore, it is necessary to encode brain MRI images and extract multi-scale features containing spatial context. This step is based on... Leveraging the multi-scale feature extraction capabilities of deep convolutional neural networks, this method extracts features from local to global perspectives layer by layer through stacked convolutional and downsampling layers. Shallow networks capture low-level visual features such as edges and textures, mid-level networks capture mid-level semantic features such as organs and lesions, and deep networks capture high-level cognitive features such as spatial relationships and contextual associations. Each layer's feature map has a different spatial resolution, forming a feature pyramid that collectively constitutes a multi-scale whole-brain feature representation. This multi-scale whole-brain feature representation includes feature maps from multiple resolution layers, ranging from fine local details to coarse-grained global context, forming a complete feature pyramid. This step extracts rich, multi-level features from MRI images. Shallow features preserve fine spatial details, aiding in accurate tumor boundary localization, while deep features capture global contextual relationships, helping to assess the tumor's impact on the whole brain. The feature pyramid structure allows subsequent analysis to be performed at different scales, considering both local characteristics and global relationships. Compared to methods using only local tumor features, whole-brain feature encoding improves the accuracy of prognostic assessment, particularly in determining functional area invasion.

[0027] The localization of brain functional areas requires an understanding of the brain's anatomical structure. The motor cortex is located in the gray matter region of the precentral gyrus, and the language area is located in the gray matter region surrounding the lateral fissure. The location of these functional areas is closely related to the morphology of the sulci and gyri. If it is unknown whether a voxel belongs to gray matter, white matter, or cerebrospinal fluid, its functional attribute cannot be accurately determined. Furthermore, the presence of tumors can distort normal anatomical structures, causing displacement of functional areas. By segmenting brain tissue, a personalized anatomical atlas can be created for each patient. Based on this atlas, the localization and displacement correction of functional areas can be performed. Therefore, this step utilizes multi-scale whole-brain features through additional segmentation. The head network performs three-class classification prediction for each voxel, determining whether it belongs to cerebrospinal fluid, gray matter, or white matter. The segmentation head network consists of several convolutional layers and upsampling layers, which gradually map deep features back to the original spatial resolution, generating a probability map with the same size as the input image. Each voxel has three probability values, corresponding to three tissue categories. The one with the highest probability is taken as the final classification. This step obtains the patient's personalized brain tissue distribution map, clarifies the tissue attributes of each voxel, and provides an anatomical basis for subsequent functional area localization. Even in the presence of tumor lesions and edema compression, the segmentation of surrounding normal brain tissue remains stable and reliable.

[0028] The cerebral cortex is divided into functional zones, with different areas responsible for different neural functions. The primary motor cortex controls voluntary movement of the contralateral limbs, Broca's area is responsible for language production, Wernicke's area is responsible for language comprehension, and the visual cortex processes visual information. Damage to these key functional areas can lead to severe neurological deficits, affecting the patient's quality of life. The distance between the tumor and the functional area determines the resectability of the surgery and the possibility of preserving neurological function. When a tumor invades a functional area, the prognosis is poor and postoperative recovery is difficult. Therefore, it is essential to locate key areas throughout the entire brain and calculate the spatial relationship between the tumor and these areas as a basis for prognostic assessment. This step is based on the registration and mapping of a standard brain functional atlas, which is established on a large number of healthy subjects using functional magnetic resonance imaging. It marks the typical locations of each functional area. This atlas is registered onto the patient's brain tissue probability atlas, taking into account the patient's individual anatomical variations, to determine the specific location of the functional areas. Then, the spatial distance between the tumor region and each functional area is calculated to determine whether there is any overlap. This step quantifies the spatial distance between the tumor and key brain functional areas, enabling doctors to objectively assess the difficulty of surgery and the risk to neurological function. It provides important spatial contextual information for prognostic assessment, allowing risk assessment to be based not only on the tumor itself, but also on its impact on whole-brain function.

[0029] Multi-scale whole-brain features contain a wealth of information, but also possess numerous dimensions. Shallow feature maps have high spatial resolution and a large number of channels, with the total feature dimension potentially reaching millions. Directly inputting such high-dimensional features into fully connected layers would lead to excessive computation, too many model parameters, and a high risk of overfitting. Furthermore, high-dimensional features contain a large amount of redundant information; for example, features of background regions are not very meaningful for prognostic judgment. Therefore, this paper compresses and integrates multi-scale whole-brain features based on attention mechanisms and global pooling operations, extracting compact global feature vectors to reduce feature dimensionality. The attention mechanism learns the importance weights of different spatial locations and feature channels, giving higher attention to important regions and features. Global pooling operations compress the spatial dimension of the feature map into a single value while retaining channel dimension information. By combining channel attention, spatial attention, and global average pooling, multi-scale features are progressively compressed into fixed-length feature vectors. These vectors contain global information of the whole brain, but with the dimensionality reduced to hundreds of dimensions, facilitating subsequent processing by fully connected layers.

[0030] The aforementioned steps extract rich feature information, including global feature vectors of the whole brain and spatial distance between the tumor and functional areas. This information needs to be integrated into the final prognostic judgment. The global feature vector contains the imaging characteristics of the tumor and the contextual information of the whole brain, while the spatial distance contains functional threat information. The combination of the two can comprehensively assess the patient's prognostic risk. Therefore, based on a multi-layer fully connected neural network, the input layer receives the concatenation of global feature vectors and spatial distance indicators, the hidden layer learns the interaction relationship between features through non-linear activation functions, such as combinations of product and ratio between features, and the output layer generates a probability distribution of risk levels, usually including three levels: low, medium, and high. At the same time, it outputs confidence, reflecting the degree of certainty of the first assessment model for the current prediction. The network learns the correlation pattern between features and prognostic outcomes through training with historical cases.

[0031] Through the execution of the above steps, the first assessment model completes the entire process from MRI image data to risk assessment. The output global prognostic risk level provides doctors with a prognostic assessment based on the whole brain context, which supplements the limitations of focusing only on the tumor itself. The output of confidence helps to identify uncertainties in the model, and low confidence indicates that it is necessary to combine it with other models or make a manual judgment.

[0032] In one specific embodiment, obtaining the second evaluation result through the second evaluation model specifically includes the following steps: The lesion segmentation results and standardized brain MRI image data are masked and fused to obtain local MRI image data containing only the lesion area; First-order statistical features are obtained by extracting pixel intensity distribution statistical indicators from local MRI image data. Texture features are obtained by extracting statistical indicators of pixel spatial relationships from local MRI image data based on the gray-level co-occurrence matrix; Wavelet features are obtained by extracting the energy distribution of different frequency sub-bands from local MRI image data using multi-scale wavelet transform. Select the subset of features most relevant to prognosis from first-order statistical features, texture features, and wavelet features; The feature subset is input into the machine learning model to obtain the second evaluation result. The second evaluation result is based on the prognostic score and the second confidence level of the imaging, and the prognostic score is mapped to the corresponding prognostic risk level based on the preset first mapping rule.

[0033] Specifically, whole-brain MRI images contain a large amount of non-tumor tissue, such as normal brain parenchyma, the ventricular system, and extracranial space. The signal characteristics of these areas are irrelevant to tumor prognosis, and including them in the analysis would introduce noise and interference. At the same time, the data volume of whole-brain images is enormous, and direct processing is computationally inefficient. By determining the spatial extent of the tumor through lesion segmentation results and extracting local image data within that extent, the analysis target can be focused, improving the specificity and computational efficiency of feature extraction. Mask fusion operations can ensure that all features come from the tumor itself or its immediate vicinity. In order to separate the spatial region where the tumor is located from the whole-brain MRI images and remove irrelevant background tissue, providing a focused data range for subsequent refined feature analysis, the lesion segmentation results are three-dimensional binary masks, with voxels within the tumor area labeled as true and non-tumor areas labeled as false. The binary mask is subjected to voxel-by-voxel logical AND operation with the standardized MRI image. The image values ​​of the masked locations are retained, while the masked locations are marked as zero or invalid. The output result is local MRI image data with the same size as the original image but containing only valid data of the tumor region. Through this step, the analysis scope is focused from the whole brain to the tumor region, effectively reducing the data dimensionality and shortening the time for subsequent feature extraction.

[0034] The signal intensity within a tumor reflects its tissue composition. Increased signal intensity in tumor regions is correlated with the degree of blood-brain barrier disruption and vascular density; signal intensity in necrotic regions is correlated with the degree of tissue necrosis; and signal intensity in edematous regions is correlated with water content. By statistically analyzing the distribution characteristics of pixel intensity, the uniformity and diversity of these tissue characteristics can be quantified. For example, tumors with concentrated signal distributions have a single tissue composition, while tumors with dispersed signal distributions contain multiple tissue components. These statistical characteristics are related to the biological behavior and prognosis of tumors and are fundamental features of imaging analysis. To quantify the distribution characteristics of pixel intensity within a tumor, based on descriptive statistics principles, without considering the spatial relationships between pixels, only the set of pixel values ​​is statistically analyzed to obtain first-order statistical features. These first-order statistical features include mean (reflecting average signal intensity), standard deviation (reflecting the intensity of signal fluctuations), maximum and minimum values ​​(reflecting signal range), median (reflecting the central tendency of the signal), skewness (reflecting the asymmetry of signal distribution), kurtosis (reflecting the sharpness of signal distribution), energy (reflecting the concentration of signal intensity), and entropy (reflecting the disorder of signal distribution). These statistical features characterize the signal distribution from different perspectives.

[0035] The tissue components within a tumor are not randomly distributed but exhibit certain spatial patterns. For example, enhanced tumor areas may show a ring-like enhancement pattern, necrotic areas may be located in the center, and edema areas may gradually transition outwards. These spatial patterns cannot be captured by first-order statistical features and require analysis of the relationships between pixels. Texture features reveal these microstructural patterns by statistically analyzing the intensity combinations of adjacent pixels. Tumors with high texture complexity often have stronger invasiveness and treatment resistance. Texture features provide a unique information dimension for prognostic assessment. To obtain the spatial distribution patterns of pixel intensity within a tumor and quantify the complexity of its microstructure and tissue heterogeneity, based on... The texture analysis method based on gray-level co-occurrence matrix (GLCM) statistically analyzes the joint probability of a pair of pixels taking a specific intensity value at a specific distance and direction. For example, the matrix element (i, j) represents the number of times a pixel takes intensity value i and its neighboring pixels take intensity value j. Based on this matrix, contrast is calculated to reflect the degree of intensity difference between adjacent pixels, correlation to reflect the degree of linear relationship between adjacent pixels, energy to reflect the uniformity and repeatability of the texture, homogeneity to reflect the local uniformity of the texture, and entropy to reflect the complexity of the texture. By calculating at multiple directions and distances, comprehensive texture features are obtained. After texture feature extraction, the spatial structure regularity inside the tumor is quantified into multiple numerical indicators.

[0036] Tumors on imaging simultaneously exhibit large-scale morphological features and small-scale details. The overall outline and location of the tumor belong to low-frequency information, while the internal texture and edges belong to high-frequency information. Different frequencies reflect different characteristics of the tumor. Low-frequency information is related to the tumor's growth pattern and spatial location, while high-frequency information is related to the tumor's microstructure and boundary characteristics. Wavelet transform can decompose images into sub-bands of different frequencies, allowing for the analysis of the energy distribution of each sub-band, thus comprehensively capturing the multi-scale characteristics of the tumor. Multi-scale analysis is particularly important for identifying the tumor's invasive boundaries and internal heterogeneity. To analyze the signal distribution characteristics of the tumor at different frequency scales and capture its macroscopic morphological trends and microscopic detail changes, multi-resolution analysis based on discrete wavelet transform is used to decompose the image into low-frequency approximate sub-bands. High-frequency detail subbands in three directions—horizontal, vertical, and diagonal—are used to preserve the smoothness and main structure of the image in low-frequency subbands, while high-frequency subbands preserve the edge and texture details. Through multi-level decomposition, subbands at different resolution levels are obtained. The energy proportion of each subband is calculated as a wavelet feature, reflecting the signal distribution characteristics of the tumor at that frequency scale and direction. The distribution pattern of energy proportions characterizes the multi-scale structural features of the tumor. Through wavelet feature extraction, the multi-scale frequency characteristics of the tumor are quantified into energy distribution indicators. These indicators can capture comprehensive information about the tumor from macroscopic morphology to microscopic details, and supplement the deficiencies of first-order statistics and texture features in frequency analysis. Wavelet features have unique value in identifying tumor boundaries and assessing the degree of invasion. The distribution of high-frequency energy can reflect the sharpness and complexity of the boundary.

[0037] The radiomics features extracted in the aforementioned steps are highly dimensional, with approximately ten first-order statistical features, twenty texture features, and thirty wavelet features, totaling hundreds of dimensions. These features are correlated; for example, some texture features and wavelet features may reflect similar image characteristics. High-dimensional feature sets can lead to model overfitting, especially when training samples are limited. Furthermore, not all features are relevant to the prognosis; some features may be noise or irrelevant factors. Feature selection can identify the most robust and relevant features, simplify the model, improve interpretability, and reduce the risk of overfitting. Therefore, in order to select the subset of features most relevant to the prognosis from high-dimensional radiomics features, remove redundant and irrelevant features, and improve the model's predictive performance and generalization ability, the univariate correlation between each feature and the prognostic outcome is first calculated, and significantly relevant candidate features are selected. Then, multivariate methods are used to evaluate the redundancy between features, remove highly correlated duplicate features, and finally use model-based methods, such as feature importance in random forests or sparse selection with L1 regularization, to determine the final feature subset. The selection process is carried out under a cross-validation framework to ensure the stability of the selected features on different data subsets. After the feature selection step, the feature dimensionality is reduced, and the correlation between the selected features and the prognosis is significantly higher than that of the removed features.

[0038] After feature extraction and selection, compact and relevant feature representations are obtained. These features need to be integrated into a comprehensive prognostic judgment. The machine learning model can learn complex nonlinear relationships between features; for example, combinations of certain features may be more predictive than single features. Through training on historical data, the machine learning model can establish a mapping relationship between features and prognostic outcomes to predict new patients. The output score has clear prognostic significance, and the confidence level reflects the reliability of the prediction, facilitating clinical decision-making. This step is implemented based on the gradient boosting tree model in ensemble learning. The model consists of multiple decision trees connected in series. Each tree learns the residual of the previous tree, gradually reducing the prediction error. Tree splitting is based on a threshold judgment of feature values, such as entropy greater than 4.5 or contrast less than 0.2. The prediction results of multiple trees are weighted and averaged to obtain the final score. The model outputs the prognostic risk in probabilistic form. At the same time, a second confidence level is calculated based on the voting consistency of the trees. The more consistent the voting, the higher the second confidence level; the more dispersed the voting, the lower the second confidence level. Through the execution of the above steps, the second assessment model completes the prognostic risk assessment based on radiomics. The output prognostic score is based on radiomics features, reflecting the imaging phenotypic characteristics of the tumor. The output prognostic score is also mapped to the corresponding prognostic risk level based on the preset first mapping rule. The first mapping rule sets different prognostic risk levels for different prognostic score ranges.

[0039] In one specific embodiment, obtaining the third evaluation result through the third evaluation model includes the following steps: Standardize and interact imaging features, morphological features, and clinicopathological features to construct multi-dimensional fusion features; Based on multi-dimensional fusion features, a learning algorithm is used to determine the contribution weights of different feature dimensions to generate weighted fusion features; Based on the weighted fusion features, a third assessment result is output through a multilayer perceptron network. The third assessment result includes a comprehensive prognostic index, a predicted survival value, and a third confidence level. Furthermore, based on a preset second mapping rule, the comprehensive prognostic index is mapped to the corresponding prognostic risk level.

[0040] Specifically, the prognosis of brain tumors is influenced by multiple factors. Imaging features reflect the internal signal characteristics of the tumor, morphological features reflect the spatial geometric characteristics of the tumor, and clinicopathological features reflect the individual biological characteristics of the patient. These three types of features come from different measurement dimensions, have different numerical ranges and physical meanings. If they are simply combined directly, features with large numerical ranges will mask features with small numerical ranges, and the interaction between features cannot be reflected. Therefore, standardization is required to eliminate dimensional differences. At the same time, feature interaction mechanisms should be designed to capture the synergistic effect between features of different dimensions. For example, the combined risk of a large tumor volume and poor molecular markers is higher than the risk of a single factor. Therefore, in order to unify the processing and interaction combination of features from different sources and construct a comprehensive feature representation that includes imaging characteristics, morphological features and clinical information, each type of feature is first standardized to map the numerical values ​​of different ranges to a unified scale and eliminate the influence of dimensions. Then, feature interaction is designed to generate new features that reflect the synergistic effect of features through operations such as multiplication and ratio. For example, the combination of entropy value in imaging and mutation state in molecular markers may produce a stronger predictive signal. The final multidimensional fusion feature includes original features and interaction features, comprehensively characterizing the patient's prognostic risk factors.

[0041] The predictive power of each dimension in the fusion feature is not the same. Some features, such as molecular subtyping status, have a decisive impact on prognosis and should be given high weights. Other features, such as wavelet details in images, may only be relevant in a specific population and should have lower weights. Fixed weights cannot adapt to individual differences among patients, while manually setting weights lacks objectivity. Therefore, a data-driven approach is needed to adaptively learn the optimal weights for each feature, enabling the model to automatically identify important prognostic factors and dynamically adjust them based on input data. The learning algorithm includes an attention mechanism and a gating network. The attention mechanism calculates the importance score of each feature dimension. The higher the score, the more important the feature is for predicting the current patient's prognosis. The gating network dynamically adjusts the weights based on the content of the feature values ​​to achieve adaptive feature selection. Weight calculation is implemented through a small neural network. The input is the feature itself, and the output is the weight value. Multiplying the weight by the feature value generates a weighted fusion feature, where important features are amplified and unimportant features are suppressed.

[0042] After feature fusion and adaptive weighting, optimized feature representations have been obtained. These features need to be converted into clinically understandable prognostic judgments. The comprehensive prognostic index should be concise and clear, so that doctors can quickly understand the risk level. The median survival prediction should be specifically quantified to facilitate the development of treatment plans and communication with patients. The confidence level should reflect the certainty of the model to facilitate the identification of difficult cases that require additional attention. Multilayer perceptron networks (MPBs) possess powerful nonlinear fitting capabilities, enabling them to learn complex mappings between features and prognoses while outputting multiple relevant indicators. Therefore, to output a comprehensive prognostic assessment result based on weighted fusion features using nonlinear transformations of a deep neural network, including a risk index, survival prediction, and confidence score, the deep neural network comprises an input layer, three hidden layers, and an output layer. The input layer receives weighted fusion features, while the hidden layers progressively extract high-order feature combinations through nonlinear activation functions, learning complex interactions between features. The output layer generates three parallel predictions: a comprehensive prognostic index output via regression, ranging from zero to one hundred (higher values ​​indicate greater risk); a survival prediction output via regression, measured in months; and a confidence score calculated based on the entropy of the predicted probability (higher confidence scores indicate a more concentrated probability distribution). Finally, a pre-defined second mapping rule maps the comprehensive prognostic index to the corresponding prognostic risk level, setting different prognostic risk levels for different ranges of the comprehensive prognostic index.

[0043] By executing the above steps, the third assessment model completes the prognostic risk assessment based on multi-dimensional information integration. The output comprehensive prognostic index can provide doctors with a concise risk quantification indicator, the survival prediction can provide specific reference for treatment decisions and patient communication, and the third confidence level can clarify the reliability of the current assessment.

[0044] In one specific embodiment, comparing the current evaluation results with historical evaluation results further includes: By querying the patient's historical data, rigid registration and deformation registration are performed between the currently acquired standardized brain MRI images and the historical standardized brain images to obtain the temporal and spatial correspondence. Based on the temporal spatial correspondence, the current total tumor volume and the historical total tumor volume are calculated, and the tumor volume change rate is calculated. When tumor volume data at at least three historical time points are available, tumor growth curves are fitted based on the rate of change of tumor volume to identify linear growth patterns, exponential growth patterns, stable plateau patterns, or treatment response patterns. Based on the temporal-spatial correspondence, morphological evolution indicators and signal evolution indicators are calculated. Morphological evolution indicators include changes in tumor shape irregularity, changes in edema range, and changes in necrosis ratio. Signal evolution indicators include changes in average signal intensity and changes in texture complexity.

[0045] Specifically, patient historical data refers to past MRI images of the same patient. To align MRI images of the same patient at different time points to the same spatial coordinate system and eliminate positional and deformation differences during scanning, a two-level registration strategy is employed. The first level is rigid registration, assuming the brain tissue is a rigid body, correcting only translational and rotational differences while maintaining shape. The second level is deformation registration, allowing non-rigid deformation of the brain tissue to correct for shape changes caused by tumor growth, edema regression, and postoperative alterations. Rigid registration uses mutual information as a similarity metric to find the optimal translational and rotational parameters. Deformation registration uses optical flow or elasticity models to calculate the local displacement vector of each voxel, matching the tumor shape in historical images with the current image. The two levels of registration are performed sequentially, first with coarse alignment and then with fine adjustment. Through this two-level registration technique, historical images are transformed to the spatial coordinate system of the current image, or both are transformed to a standard space, ensuring that the same anatomical location has the same coordinates on images at different time points.

[0046] Tumor volume is the most direct indicator for assessing prognosis and treatment response. The neuro-oncology efficacy evaluation criteria define progression as an increase in volume exceeding 25% or the appearance of new lesions, and partial remission as a reduction in volume exceeding 50%. Accurate volume measurement is the foundation of efficacy evaluation. By comparing volumes at different time points, the rate of change can be calculated to determine whether the tumor is growing, stable, or shrinking. The rate of change in volume reflects dynamic trends better than absolute volume. For example, for tumors of the same volume, those that grow rapidly have a worse prognosis than those that grow slowly. Therefore, for the segmentation results at the current time point, the number of voxels in the tumor region is directly counted and multiplied by the volume of a single voxel to obtain the current total volume. For the segmentation results at historical time points, the inverse transformation of registration is applied to map the historical mask to the current space, and the number of voxels is counted again to obtain the historical total volume. The rate of change is calculated as the difference between the current volume and the historical volume divided by the historical volume. A positive value indicates an increase, and a negative value indicates a decrease.

[0047] Tumor growth is not uniform and may exhibit different dynamic patterns. Linear growth indicates that the tumor grows at a constant rate, which is common in low-grade tumors. Exponential growth indicates that the tumor grows rapidly, suggesting an increase in malignancy. A stable plateau pattern indicates that the tumor volume remains unchanged for a long time, which may be benign or well controlled by treatment. Treatment response pattern indicates that the tumor shrinks first and then stabilizes or regrows, reflecting the effectiveness of treatment and the development of drug resistance. Identifying these patterns is crucial for prognostic assessment. For example, the exponential growth pattern requires emergency intervention, while the plateau pattern can be maintained for observation. Only two time points can be used to calculate the rate of change, but not to identify the pattern. At least three time points are needed to fit the curve. Therefore, volume data from three or more time points are collected, and growth curves are plotted with time on the horizontal axis and volume on the vertical axis. The linear, exponential, and plateau models are tried to fit the curves, and the goodness of fit of each model is calculated. The model with the highest goodness of fit is selected as the identification result. For post-treatment data, the pattern of initial decline followed by increase (treatment response pattern) is detected to identify pseudoprogression or recurrence. By analyzing growth trends, we can identify the tumor evolution patterns of patients, providing a dynamic perspective for prognostic prediction. This not only looks at the current state but also at the development trend, thus improving the foresight and accuracy of prognostic assessment.

[0048] Changes in tumors are not only reflected in their volume, but also in their shape, internal structure, and signal characteristics. When the volume remains unchanged, the tumor may become more irregular, indicating increased infiltrative activity. When edema subsides but the enhancement area expands, it indicates tumor progression. The change from homogeneous to heterogeneous internal signals indicates the occurrence of necrosis or hemorrhage. These subtle changes are of great significance for prognostic judgment, but they are easily masked by volume indicators. Therefore, multi-dimensional evolutionary analysis is needed to comprehensively capture the biological changes of the tumor. Thus, morphological evolution indicators and signal evolution indicators are calculated based on the comparison of registered images and feature differences. Morphological evolution indicators are calculated by comparing segmentation masks at different stages, including changes in shape irregularity, changes in the ratio of edema to the tumor core, and changes in the proportion of necrosis. Signal evolution indicators are calculated by comparing the registered original images, including changes in average signal intensity, changes in texture complexity, and changes in contrast enhancement patterns. These indicators characterize the evolutionary characteristics of the tumor from different perspectives. Through multi-dimensional evolutionary analysis, rich evolutionary information beyond volume indicators is obtained. A decrease in shape irregularity suggests that the tumor invasiveness may be reduced, providing a basis for prognostic improvement. An increase in signal homogeneity suggests tissue repair after treatment, providing a supplement to efficacy evaluation. A slight increase in the proportion of edema indicates details that need attention, avoiding blind optimism based solely on volume indicators.

[0049] The above steps complete the entire process from image registration to multidimensional evolutionary analysis. The entire process makes full use of the value of longitudinal follow-up data, upgrading static prognostic assessment to dynamic disease monitoring, and providing key information in the time dimension for clinical decision-making.

[0050] In one specific embodiment, generating the final prognostic assessment conclusion includes the following steps: The prognostic risk level is determined by comparing the results of the first, second, and third assessments to see if the three assessment results are consistent. When the three assessment results are consistent, the common prognostic risk level of the three assessment results is used as the preliminary conclusion; when the three assessment results are inconsistent, manual judgment is triggered to obtain the corresponding preliminary conclusion. Based on the tumor volume change rate, morphological evolution indicators and signal evolution indicators, the tumor evolution trend is judged to be progressive, stable or remission. When the tumor evolution is indeed in a progressive state, the preliminary conclusion is revised upward, and when the tumor evolution trend is remission, the preliminary conclusion is revised downward. Based on the adjusted preliminary conclusions, corresponding treatment recommendations are generated.

[0051] Specifically, three assessment models analyze prognostic risk from different perspectives. The first assessment model is based on whole-brain context, focusing on the relationship between tumor location and functional areas. The second assessment model is based on radiomics, focusing on the tumor's imaging phenotype. The third assessment model is based on multi-dimensional fusion, integrating imaging, morphological, and clinical information. The judgments of the three models may be consistent or differ. When consistent, the judgment is reliable and can be directly adopted. When inconsistent, it indicates uncertainty and requires careful handling. When the three assessment results are consistent, the common prognostic risk level of the three assessment results is used as the preliminary conclusion. Assuming all three are low-risk, the corresponding preliminary conclusion is low-risk, which means a good prognostic assessment, no immediate risk of recurrence, and only regular follow-up is needed. If the three assessment results are inconsistent, it means the judgments of the three models differ, indicating uncertainty. In this case, it is recommended to trigger manual judgment for accurate judgment, that is, to obtain the corresponding preliminary conclusion (which risk level) through manual intervention. Assuming a high-risk level, it means a higher prognostic risk, with a possible risk of recurrence, requiring timely intervention and treatment.

[0052] The preliminary conclusion, based on the assessment model, can only reflect the current state and cannot accurately predict future trends. Furthermore, model-based judgments are susceptible to misjudgment. Therefore, we also used time-series data on tumor volume change rate, morphological evolution indicators, and signal evolution indicators to determine whether the tumor evolution trend is progression, stabilization, or remission. For example, a positive and relatively large tumor volume change rate (indicating tumor enlargement) or changes in morphological evolution indicators such as increased tumor irregularity or a shift from homogeneous to heterogeneous internal signals suggest a progression trend, meaning worsening tumor symptoms. In such cases, the preliminary conclusion needs adjustment; that is, we assume a low-risk outcome. If the risk level is low, the preliminary conclusion needs to be revised upwards to medium risk. If the tumor evolution trend is in remission, such as when the tumor volume is decreasing, the internal signal remains uniform, and the edema area is receding, it indicates that the tumor symptoms are improving. In this case, the preliminary conclusion can be revised downwards. That is, if the preliminary conclusion is medium risk, it should be revised downwards to low risk. The revised preliminary conclusion, which is the final prognostic assessment conclusion, generates corresponding treatment recommendations. If it is low risk, regular follow-up examinations can be recommended; if it is medium risk, a change in treatment strategy can be recommended; and if it is high risk, the treatment intensity can be increased.

[0053] The above describes a method for dynamic assessment of tumor prognosis combining lesion segmentation and AI in embodiments of this application. The following describes a system for dynamic assessment of tumor prognosis combining lesion segmentation and AI in embodiments of this application. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the tumor prognosis dynamic assessment system combining lesion segmentation and AI in this application includes: The image processing unit is used to preprocess brain image data to obtain standardized brain image data, and to segment brain tumor lesions based on the standardized brain image data to obtain lesion segmentation results. The feature extraction unit extracts imaging and morphological features based on the lesion segmentation results; The model analysis unit uses the first assessment model, the second assessment model, and the third assessment model to perform prognostic risk assessments on standardized brain imaging data, lesion segmentation results, and fusion features, respectively, to obtain the first assessment result, the second assessment result, and the third assessment result. The comprehensive assessment unit compares the current assessment results with historical assessment results, calculates the tumor volume change rate and morphological evolution indicators, and generates the final prognostic assessment conclusion based on the first, second and third assessment results, combined with the tumor volume change rate, morphological evolution indicators and signal evolution indicators.

[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0055] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0056] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for dynamic assessment of tumor prognosis combining lesion segmentation and AI, characterized in that, The method includes: Step S1: Preprocess the brain imaging data to obtain standardized brain imaging data, and segment the brain tumor lesions based on the standardized brain imaging data to obtain the lesion segmentation results; Step S2: Extract imaging and morphological features based on lesion segmentation results; Step S3: Use the first assessment model, the second assessment model, and the third assessment model to perform prognostic risk assessment on the standardized brain imaging data, lesion segmentation results, and fusion features to obtain the first assessment result, the second assessment result, and the third assessment result; Step S4: Compare the current assessment results with the historical assessment results, calculate the tumor volume change rate and morphological evolution indicators, and generate the final prognostic assessment conclusion based on the first assessment results, the second assessment results, and the third assessment results, combined with the tumor volume change rate, morphological evolution indicators, and signal evolution indicators.

2. The method according to claim 1, characterized in that, Preprocessing brain imaging data to obtain standardized brain imaging data includes: Adaptive filtering algorithms are used to suppress noise in brain imaging data. After noise suppression, intensity non-uniformity correction is performed on the image data. Positional deviations between different slices within the same sequence are corrected. The MRI brain imaging data are registered to the same coordinate system to achieve voxel-level alignment and registration of different sequences. The registered image data is then subjected to intensity standardization, and the gray values ​​of each sequence are mapped to a unified numerical range to obtain standardized brain MRI imaging data.

3. The method according to claim 1, characterized in that, The segmentation of brain tumor lesions is performed to obtain the segmentation results, including: T1-weighted images, T2-weighted images, fluid attenuation inversion recovery sequence images, and enhanced T1-weighted images were fused to construct a four-channel input tensor. A deep learning segmentation network was used to encode and decode the four-channel input tensor to generate a three-dimensional segmentation mask, which included masks for the enhanced tumor region, the peritumoral edema region, and the necrotic core region. Connectivity analysis was performed on the three-dimensional segmentation mask to obtain optimized lesion segmentation results. The surface boundaries of the optimized lesion segmentation results were smoothed to eliminate jagged edges, resulting in the final lesion segmentation result.

4. The method according to claim 1, characterized in that, The first evaluation result is obtained through the first evaluation model, including: Standardized brain MRI image data is input into a 3D convolutional neural network encoder to extract multi-scale whole-brain features. Based on the multi-scale whole-brain features, the cerebrospinal fluid region, gray matter region, and white matter region are segmented to construct a brain tissue probability atlas. Key brain functional regions are identified based on the standard brain function atlas and the brain tissue probability atlas. The spatial distance between the tumor region and the key brain functional regions is calculated. The multi-scale whole-brain features are compressed and integrated to extract a global feature vector. Based on the feature vector and spatial distance, the first assessment result is output through a fully connected layer. The first assessment result includes the global prognostic risk level and the first confidence level.

5. The method according to claim 1, characterized in that, The second evaluation results are obtained through the second evaluation model, including: The lesion segmentation results and standardized brain MRI image data are masked and fused to obtain local MRI image data containing only the lesion area; First-order statistical features are obtained by extracting pixel intensity distribution statistical indicators from local MRI image data. Texture features are obtained by extracting statistical indicators of pixel spatial relationships from local MRI image data based on the gray-level co-occurrence matrix; Wavelet features are obtained by extracting the energy distribution of different frequency sub-bands from local MRI image data using multi-scale wavelet transform. Select the subset of features most relevant to prognosis from first-order statistical features, texture features, and wavelet features; The feature subset is input into the machine learning model to obtain the second evaluation result. The second evaluation result is based on the prognostic score and the second confidence level of the imaging, and the prognostic score is mapped to the corresponding prognostic risk level based on the preset first mapping rule.

6. The method according to claim 1, characterized in that, The third evaluation results are obtained through the third evaluation model, including: Standardize and interact imaging features, morphological features, and clinicopathological features to construct multi-dimensional fusion features; Based on multi-dimensional fusion features, a learning algorithm is used to determine the contribution weights of different feature dimensions to generate weighted fusion features; Based on the weighted fusion features, a third assessment result is output through a multilayer perceptron network. The third assessment result includes a comprehensive prognostic index, a predicted survival value, and a third confidence level. Furthermore, based on a preset second mapping rule, the comprehensive prognostic index is mapped to the corresponding prognostic risk level.

7. The method according to claim 1, characterized in that, Compare the current assessment results with historical assessment results, including: By querying the patient's historical data, rigid registration and deformation registration are performed between the currently acquired standardized brain MRI images and the historical standardized brain images to obtain the temporal and spatial correspondence. Based on the temporal spatial correspondence, the current total tumor volume and the historical total tumor volume are calculated, and the tumor volume change rate is calculated. When tumor volume data at at least three historical time points are available, tumor growth curves are fitted based on the rate of change of tumor volume to identify linear growth patterns, exponential growth patterns, stable plateau patterns, or treatment response patterns. Based on the temporal-spatial correspondence, morphological evolution indicators and signal evolution indicators are calculated. Morphological evolution indicators include changes in tumor shape irregularity, changes in edema range, and changes in necrosis ratio. Signal evolution indicators include changes in average signal intensity and changes in texture complexity.

8. The method according to claim 1, characterized in that, Generate the final prognostic assessment conclusion, including: The prognostic risk level is determined by comparing the results of the first, second, and third assessments to see if the three assessment results are consistent. When the three assessment results are consistent, the common prognostic risk level of the three assessment results is used as the preliminary conclusion; when the three assessment results are inconsistent, manual judgment is triggered to obtain the corresponding preliminary conclusion. Based on the tumor volume change rate, morphological evolution indicators and signal evolution indicators, the tumor evolution trend is judged to be progressive, stable or remission. When the tumor evolution is indeed in a progressive state, the preliminary conclusion is revised upward, and when the tumor evolution trend is remission, the preliminary conclusion is revised downward. Based on the adjusted preliminary conclusions, corresponding treatment recommendations are generated.

9. A dynamic prognostic assessment system for tumors combining lesion segmentation and AI, used to implement the dynamic prognostic assessment method for tumors combining lesion segmentation and AI as described in any one of claims 1-8, characterized in that, The system includes: The image processing unit is used to preprocess brain image data to obtain standardized brain image data, and to segment brain tumor lesions based on the standardized brain image data to obtain lesion segmentation results. The feature extraction unit extracts imaging and morphological features based on the lesion segmentation results; The model analysis unit uses the first assessment model, the second assessment model, and the third assessment model to perform prognostic risk assessments on standardized brain imaging data, lesion segmentation results, and fusion features, respectively, to obtain the first assessment result, the second assessment result, and the third assessment result. The comprehensive assessment unit compares the current assessment results with historical assessment results, calculates the tumor volume change rate and morphological evolution indicators, and generates the final prognostic assessment conclusion based on the first, second and third assessment results, combined with the tumor volume change rate, morphological evolution indicators and signal evolution indicators.

Images