Method and system for identifying breast cancer molecular subtypes based on dce-mri habitatomics analysis

Abstract

This invention belongs to the fields of medical image processing and computer-aided diagnostic technology, and discloses a method and system for molecular subtyping of breast cancer based on DCE-MRI habitat omics analysis. By segmenting and preprocessing breast DCE-MRI images, a dual-path collaborative habitat segmentation module is used to analyze tumor heterogeneity from two complementary dimensions: pharmacokinetics and dynamic enhancement patterns. The segmented habitat features undergo deep interaction and adaptive weighting through a multi-level fusion module to generate a highly discriminative joint feature representation, which is finally output as a molecular subtype diagnosis by a classifier. This network enables deep interaction and complementary enhancement of the two feature paths, improving feature representation capabilities. In the testing phase, for complete modality data, a complete diagnostic process from habitat analysis to subtype classification can be completed. This invention achieves high-precision auxiliary identification of breast cancer molecular subtypes, with advantages such as strong tumor heterogeneity analysis capability, high feature discrimination power, and automated process.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and computer-aided diagnosis technology, specifically relating to a molecular subtyping method and system for breast cancer based on DCE-MRI habitat omics analysis. Background Technology

[0002] Breast cancer is one of the most common malignant tumors in women. As a highly heterogeneous disease, it includes several molecular subtypes that differ significantly in biological behavior, treatment response, and prognosis. Clinically, it is usually classified into major subtypes such as luminal A, luminal B, human epidermal growth factor receptor 2 positive (HER2-positive), and triple-negative breast cancer (TNBC). Accurate identification of these molecular subtypes is a key prerequisite for developing personalized precision treatment plans. Therefore, developing a non-invasive method that can comprehensively characterize tumor heterogeneity and accurately predict molecular subtypes has significant clinical value.

[0003] Currently, tissue biopsy-based pathological testing remains the gold standard for molecular subtyping. However, this method is invasive and, due to limited spatial sampling, struggles to fully reflect the overall heterogeneity within the tumor. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) can non-invasively capture the hemodynamic characteristics of the tumor microenvironment in both temporal and spatial dimensions, providing a powerful tool for assessing tumor heterogeneity. Traditional radiomics methods based on DCE-MRI extract artificially designed quantitative features and establish associations with molecular subtypes, but these methods have limited ability to describe complex intratumoral spatial heterogeneity, and feature design is highly dependent on experience. Although deep learning models can automatically learn feature representations from images, most models focus on global tumor characterization and similarly lack explicit modeling of the spatial pattern of intratumoral heterogeneity.

[0004] Habitat radiomics, by dividing tumors into sub-regions with different imaging phenotypes—i.e., habitats—achieves fine-grained analysis of tumor heterogeneity, enabling more precise characterization of spatial differences related to tumor perfusion, metabolism, and other activities. In recent years, studies have introduced unsupervised learning methods such as clustering into habitat radiomics to depict tumor habitats and improve molecular subtype prediction. For example, Professor Fan Ming's team at Hangzhou Dianzi University, in collaboration with research teams from Virginia Tech, used a fully unsupervised convex set mixture analysis method to unmix dynamic sequences of DCE-MRI, decomposing the tumor into three sub-regions representing plasma input, rapid flow dynamics, and slow flow dynamics, respectively. Radiomic features were then extracted and fused from each sub-region for classification. The Chinese Academy of Medical Sciences, in collaboration with Peking Union Medical College, proposed a data-driven method based on voxel-level DCE-MRI time-intensity curve (TIC) feature mapping. This method classifies curves into different dynamic subtypes, uses these to delineate habitat regions, and constructs a dataset for model training.

[0005] However, the aforementioned existing technologies still have significant limitations. Professor Fan Ming's team mainly focused on the spatial functional decomposition of image sequences and the parameters of TIC curves, while Peking Union Medical College emphasized curve morphology analysis in the time dimension. Neither approach fully and synergistically utilizes the deep information jointly contained in the two dimensions of temporal dynamic changes and spatial heterogeneous distribution of DCE-MRI data. This incomplete utilization of information limits the model's ability to comprehensively characterize the complex biological characteristics of tumors, potentially affecting the accuracy and robustness of the final molecular subtyping prediction. Therefore, a new technical solution is urgently needed to more effectively integrate multi-dimensional information and achieve superior non-invasive molecular subtyping of breast cancer.

[0006] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0007] (1) Existing habitat classification methods usually start from a single dimension or perspective and fail to integrate complementary information of image data in the two core dimensions of temporal dynamic evolution and spatial heterogeneous distribution, which affects the characterization ability of the extracted features and thus restricts the accuracy and generalization ability of molecular subtyping models.

[0008] (2) Architectural limitations of existing methods: Most existing models adopt a serial or single-branch architecture and fail to build a collaborative learning mechanism that can process temporal dynamics and spatial heterogeneous information in parallel. This structural limitation makes it difficult for the model to simultaneously capture the functional evolution of lesions during the enhancement process and their morphological heterogeneity in spatial distribution, resulting in bottlenecks in the discrimination and robustness of the constructed molecular subtyping diagnostic system.

[0009] (3) Existing methods often employ simple feature splicing or weighted averaging strategies in the feature fusion stage. Essentially, these are physical stackings of features, failing to delve into the complex nonlinear relationships and interactions between different features, resulting in the ineffective utilization of the deep complementary and redundant information contained in the features. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides a method, system, device, and terminal for molecular subtyping of breast cancer based on DCE-MRI habitat omics analysis.

[0011] This invention is implemented as follows: a molecular subtyping method for breast cancer based on DCE-MRI habitat omics analysis. This molecular subtyping method includes: segmenting and preprocessing the tumor region in the acquired breast DCE-MRI images; dividing the processed image data into training and testing sets; constructing a habitat segmentation module containing two different habitat segmentation methods to segment tumor habitats from multiple temporal and spatial dimensions; constructing a multi-level feature fusion module to extract features from the segmented habitat data using a feature extractor, and then further interacting the two branches of features through a cross-self-attention module, an orthogonal fusion module, and a pooling fusion module to achieve feature enhancement; during training, the entire network is trained using training set data labeled with molecular subtyping tags. The training objective is to minimize the classification loss function (cross-entropy loss), and the parameters of the multi-level feature fusion module and the final classifier are simultaneously optimized through backpropagation, enabling the model to learn fused feature representations that can accurately identify different molecular subtypings. During the testing phase, for the input DCE-MRI data, the habitat segmentation module first generates a dual-path habitat image, which is then processed by the feature extraction and multi-level fusion module mentioned above. The resulting global joint features are then fed into the trained classifier, which directly outputs the diagnostic results of the molecular subtyping of breast cancer.

[0012] Furthermore, the molecular subtyping method for breast cancer based on DCE-MRI habitat omics analysis includes the following steps:

[0013] Step 1: Segmentation and processing of the original image; In the obtained breast dynamic contrast-enhanced resonance imaging, the tumor region in the original image is manually segmented using the Medical Image Interaction Toolkit (MITK) software; DICOM format data is converted to RAW format data; The number of slices containing tumors in all patients is normalized to obtain three-dimensional images of the same size; All preprocessed samples are randomly divided into training and testing sets for model training and evaluation.

[0014] Step two involves a two-branch habitat segmentation module, employing two different methods: pixel-level pharmacokinetics and region-level enhancement patterns. In pharmacokinetics, the Completely Unsupervised Convex Analysis of Mixtures (CAM) algorithm, through unsupervised learning and convex optimization, classifies voxels into subregions of several baseline hemodynamic patterns with clear physiological significance. In the enhancement pattern branch, based on superpixel segmentation and Signal Enhancement Ratio (SER) calculation, the image is first segmented into superpixel regions with similar textures and spatial locations. Then, each region is classified according to its time-intensity curve morphology, thus achieving habitat segmentation.

[0015] Step 3 involves multi-level feature fusion. After habitat segmentation, a feature extraction module is used to map the image to high-dimensional features. A cross-self-attention module is used to achieve deep semantic complementarity and correction of habitat features obtained from the two habitat segmentation methods. An orthogonal fusion module is used to effectively reduce linear redundancy between features, resulting in higher density and stronger complementarity of the fused path features. A pooling fusion module is used to achieve dynamic, sample-related feature importance balancing, ensuring that the final global joint features have high discriminative power, thereby maximizing classification performance.

[0016] Step four, molecular subtyping diagnosis of breast cancer: For the breast DCE-MRI data to be diagnosed, it is input into the constructed two-branch habitat segmentation module for segmentation; the obtained habitat images are sent to the feature extractor to obtain the feature codes of each habitat image; then, the feature codes of different habitats are fused at multiple levels to obtain global joint features, which are sent to a classifier composed of a fully connected network to obtain the diagnostic results.

[0017] Furthermore, step one, in the acquired breast DCE-MRI image, includes the segmentation and processing of the original image, which includes:

[0018] (1) In the acquired breast images, the tumor region in the original image was manually segmented using MITK software;

[0019] (2) Convert the segmented tumor region data in DICOM format to RAW format;

[0020] (3) Normalize the number of slices containing tumors in all patients to obtain three-dimensional images of the same size;

[0021] (4) Randomly divide all processed samples into training and test sets for model training and evaluation.

[0022] Furthermore, the bibranch habitat division in step two is carried out according to the following steps:

[0023] (1) Construct a habitat division module consisting of two habitat division methods, dividing habitats based on pixel-level pharmacokinetics and regional enhancement modes respectively.

[0024] (2) In the pharmacokinetic branch, five consecutive slices of the tumor core region were selected, and a four-phase signal sequence was constructed for each pixel to form a pixel-time matrix X. After standardizing the matrix, the curve of each pixel was optimized through clustering and convexity optimization. Decomposed into a combination of three reference dynamic components:

[0025]

[0026] in, , , The reference dynamic curves represent the three heteroproton regions respectively. , , The corresponding contribution coefficients are used. Spatial distribution maps for each baseline dynamic are reconstructed through non-negative matrix factorization. Three feature maps are output, quantitatively characterizing the microcirculation functions of plasma input, rapid perfusion, and slow perfusion, respectively.

[0027] (3) In the enhancement mode branch, adaptive SLIC superpixel segmentation is performed on the second-phase image (number N = max(10, min(500, total number of pixels / 1000))), and the regions are mapped to the images of each phase. The average intensity (T1, T2, T3) of each superpixel region in each phase is extracted, and its signal enhancement ratio is calculated:

[0028]

[0029] Regions are divided based on the SER threshold: fast emptying type (SER≥1.2), plateau type (0.9≤SER<1.2), and continuous enhancement type (SER<0.9). Three different dynamic enhancement mode regions are output.

[0030] Furthermore, the multi-level feature fusion in step three is performed according to the following steps:

[0031] (1) The habitat image obtained in the previous step is mapped to high-dimensional features through a feature extractor.

[0032] (2) Input the extracted features from the two branches simultaneously, and use a symmetrical cross-attention mechanism to make the two types of features mutually query each other. With key value This allows for deep interaction and contextual enhancement. The core calculation formula is:

[0033]

[0034] in, It is the dimension of the key vector. It is a fusion feature of cross-attention output.

[0035] (3) Fuse multiple features within each path after interaction enhancement. An ordered concatenation strategy is adopted, and an orthogonal loss function is introduced, such as:

[0036]

[0037] Constraints are applied to reduce redundancy between features, resulting in refined path features, where... For pharmacokinetics Orthogonal loss function of branches, , These are the habitat characteristics of the pharmacokinetic branch.

[0038] (4) The refined features are concatenated with the original features to form an enhanced feature stream. Global max pooling and average pooling are performed on each stream. The results are concatenated and then used to generate adaptive weights through a lightweight MLP. Finally, the weighted sum is used to obtain the global joint features.

[0039] Furthermore, step four, molecular subtyping diagnosis of breast cancer, realizes the task of molecular subtyping diagnosis of breast cancer based on DCE-MRI image analysis, and is performed according to the following steps:

[0040] (1) For the breast DCE-MRI data to be diagnosed, input it into the constructed dual-branch habitat division module for division to obtain habitat image data;

[0041] (2) The obtained habitat images are fed into the feature extractor. After obtaining the feature codes of each habitat image, the feature codes of different habitats are fused at multiple levels to obtain global joint features.

[0042] (3) The diagnostic results can be obtained by feeding the global joint features into a classifier composed of a fully connected network.

[0043] Another objective of this invention is to provide a molecular subtype classification and identification system for breast cancer based on DCE-MRI habitat omics analysis, the system comprising:

[0044] The data acquisition module is used to acquire breast DCE-MRI data and include patient data that meets the criteria.

[0045] The data segmentation and preprocessing module is used for segmentation and preprocessing of breast DCE-MRI data;

[0046] The dual-branch habitat segmentation module is used to segment tumor habitat images. It takes the patient's DCE-MRI data as input and uses two complementary technical approaches, namely CAM algorithm and superpixel segmentation and SER calculation, to collaboratively complete the segmentation of the tumor's internal habitat.

[0047] The multi-level fusion module is used to further learn the features of the two branches after the features are extracted by the encoder from the segmented habitat images, so as to enhance the feature representation and enable the network to obtain the joint feature representation with the strongest discriminative power and the richest information.

[0048] The breast cancer molecular subtyping diagnostic module is used to take the fused habitat feature codes as input and feed them into a classifier composed of a fully connected network to obtain the diagnostic results, thus completing the breast cancer molecular subtyping task.

[0049] Another object of the present invention is to provide a computer device comprising a memory and a processor. The memory stores a computer program that, when executed by the processor, implements the previously described method for molecular subtyping of breast cancer based on DCE-MRI habitat omics analysis.

[0050] Another object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the processor performs the previously described method for molecular subtyping of breast cancer based on DCE-MRI habitat omics analysis.

[0051] Another objective of this invention is to provide an information data processing terminal configured to run the aforementioned breast cancer molecular subtype classification and identification system based on DCE-MRI habitat omics analysis, so as to achieve all the functions of the aforementioned method.

[0052] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0053] First, this invention utilizes dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) data of the breast to simultaneously achieve non-invasive automated diagnosis of key molecular subtypes such as Luminal A, Luminal B, HER2-positive, and triple-negative breast cancer, based on a multi-dimensional and detailed analysis of intratumoral heterogeneity. This method significantly improves the model's ability to characterize complex biological characteristics and its diagnostic accuracy by synergistically analyzing temporal dynamics and spatial heterogeneity information. It is also easy to deploy clinically, and its diagnostic results can assist in manual interpretation.

[0054] Secondly, this invention provides a non-invasive and precise molecular subtyping method for breast cancer. Its core lies in its innovative "dual-branch habitat division" and "multi-level feature fusion" architecture, which solves the problems of insufficient utilization of multi-dimensional information and shallow feature fusion in existing technologies. This enables a more comprehensive and in-depth analysis of tumor heterogeneity, thereby significantly improving the performance of subtyping diagnosis. It features a small model size, flexible application scenarios, and high diagnostic accuracy.

[0055] The significant technological advancements in this molecular subtyping method for breast cancer are reflected in the following aspects:

[0056] 1. Multi-dimensional habitat synergistic analysis of tumor heterogeneity: By constructing parallel pharmacokinetic and enhancement pattern branches, this invention achieves, for the first time, synergistic analysis of temporal dynamic function and spatial morphological patterns at the habitat segmentation level. This overcomes the limitations of existing methods that only consider a single dimension (such as curve morphology or functional decomposition), and can more completely and complementaryly characterize the heterogeneous characteristics within tumors, providing more accurate habitat information for subsequent analysis.

[0057] 2. Deep Feature Interaction and Fusion Mechanism: Unlike simple feature stitching, this invention designs a multi-layered feature fusion module consisting of cross-self-attention, orthogonal fusion, and pooling fusion. This mechanism can deeply explore the nonlinear correlations and semantic complementarity between features of different habitats, while effectively eliminating redundant information, ultimately generating highly discriminative global joint features. This greatly enhances the model's ability to extract key diagnostic information from complex images.

[0058] 3. Enhancing the accuracy and generalization potential of non-invasive subtyping: Through the aforementioned technological innovations, this method maximizes the extraction of pathological information contained in DCE-MRI images without relying on invasive biopsies. Its dual-branch architecture and deep fusion strategy enhance the model's ability to represent and generalize complex cases, providing a powerful new tool for achieving highly accurate non-invasive prediction of breast cancer molecular subtypes in clinical practice, and contributing to the precise formulation of personalized treatment plans.

[0059] 4. Providing interpretable clues for clinical decision-making: The visualizations generated by this method have clear physiological or radiological significance. This not only outputs a diagnostic label but also provides physicians with spatial visualization information about biological characteristics within the tumor, such as angiogenesis and perfusion abnormalities. It offers potential radiological biomarkers for understanding tumor heterogeneity and assessing prognosis, thus supporting clinical decision-making.

[0060] This molecular subtyping method for breast cancer based on DCE-MRI habitat omics analysis significantly improves the ability to analyze the spatiotemporal heterogeneity of tumors and the accuracy and robustness of molecular subtyping diagnosis through an innovative dual-branch habitat synergistic segmentation and multi-level deep feature fusion strategy, demonstrating the technological advancement of radiomics in achieving non-invasive precision diagnosis and treatment.

[0061] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0062] (1) This invention provides a precise and non-invasive molecular subtyping tool for breast cancer, which can directly assist clinicians in developing individualized treatment plans and reduce reliance on invasive tissue biopsies. Its core value lies in the fact that, through innovative dual-branch habitat analysis and fusion technology, it can more comprehensively characterize tumor heterogeneity, thereby potentially improving the accuracy and robustness of molecular subtype prediction, and has clear clinical translation prospects and commercial application potential.

[0063] (2) Existing technologies have obvious limitations: traditional radiomics-based methods rely on artificial features and have limited ability to describe complex heterogeneity; while most deep learning models focus on global representation and lack explicit modeling of intratumoral spatial heterogeneity patterns. This invention creatively constructs a parallel dual-branch architecture, which synergistically integrates deep information extracted from two complementary dimensions: pixel-level pharmacokinetics (CAM algorithm) and region-level enhancement patterns (superpixels and SER), filling the technical gap in existing methods in synergistically utilizing temporal dynamics and spatial heterogeneity information.

[0064] (3) In clinical practice, accurately and non-invasively assessing the highly heterogeneous internal structure of tumors has always been a challenge. Tissue biopsy is invasive and subject to sampling bias, while existing imaging analysis methods struggle to simultaneously capture the functional evolution of lesions during enhancement and their morphological heterogeneity in spatial distribution. This invention, through a technical approach combining "habitat segmentation" and "multi-level feature fusion," achieves a more refined and comprehensive analysis of the tumor microenvironment, providing a new and effective solution for overcoming the clinical challenge of non-invasive and comprehensive characterization of tumor heterogeneity to achieve accurate subtyping.

[0065] (4) Existing methods mostly adopt a serial or single-branch design in terms of architecture, and often use shallow strategies such as simple splicing in the feature fusion stage, failing to deeply explore the complex nonlinear relationships between features of different dimensions. This invention breaks through this technical bias, constructs a parallel dual-branch collaborative learning mechanism, and designs a series of deep interactive modules such as cross-self-attention, orthogonal fusion, and pooling fusion, realizing semantic complementarity and redundancy elimination between features from different sources, enabling the model to learn joint feature representations with stronger discriminative power and richer information.

[0066] Fourth, significant technological advancements in breast cancer molecular subtyping methods based on DCE-MRI habitat omics analysis mainly include:

[0067] 1. A Multi-Dimensional Collaborative Habitat Analysis Framework: By constructing a parallel dual-branch architecture, this framework innovatively integrates two complementary habitat partitioning strategies: pixel-level pharmacokinetics (CAM algorithm) and region-level enhancement patterns (superpixel and SER calculation). This framework achieves, for the first time, simultaneous and explicit modeling of the temporal dynamic enhancement patterns and spatial heterogeneity of tumors at a fine-grained level, overcoming the limitations of existing methods that only analyze from a single dimension, and providing a new technological pathway for comprehensively characterizing tumor heterogeneity.

[0068] 2. A multi-layered feature fusion mechanism for deep interaction: A sequential fusion process consisting of cross-self-attention, orthogonal fusion, and pooling fusion modules was designed. This mechanism not only achieves deep semantic interaction and context enhancement between features from different habitats, but also effectively reduces feature redundancy through orthogonal constraints and achieves dynamic feature trade-offs through adaptive weights. This goes beyond simple feature concatenation and can discover and strengthen the most effective discriminative features for classification.

[0069] 3. Provides new imaging biomarkers for quantifying functional heterogeneity within tumors: Through two habitat segmentation methods, objective imaging biomarkers for quantifying functional heterogeneity within tumors can be obtained, providing in-depth functional information beyond morphology for clinical assessment.

[0070] 4. A scalable and general analytical framework was constructed: The proposed core architecture of "two-branch habitat segmentation + multi-level feature fusion" is modular and flexible. This framework is not limited to the two currently used habitat segmentation methods; its ideas can be extended to integrate more dimensional segmentation strategies or adapt to different feature extraction and fusion modules, providing an innovative methodological foundation for broader tumor radiomics research.

[0071] Overall, this molecular subtyping method for breast cancer based on DCE-MRI habitat omics analysis represents a significant technological advancement in the field of tumor radiomics, particularly for the study of highly heterogeneous breast cancer. Its innovation lies in constructing a dual-branch architecture that synergistically integrates temporal dynamics and spatial heterogeneity information, and achieving a comprehensive and detailed analysis of the tumor microenvironment through a multi-level deep feature fusion mechanism. This method provides a novel technical framework for addressing the clinical challenge of non-invasive and accurate subtyping, and sets a new standard for image-based habitat omics research. Attached Figure Description

[0072] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings will be briefly described below. It should be noted that the drawings are only used to exemplify some specific embodiments of the present invention. Based on the disclosed content, those skilled in the art can conceive of or derive other related implementation methods without creative effort.

[0073] Figure 1 This is a flowchart of a method for classifying and identifying molecular subtypes of breast cancer based on DCE-MRI habitat omics analysis, provided in an embodiment of the present invention.

[0074] Figure 2 This is a schematic diagram of the structure of the breast cancer molecular subtype classification and identification system based on DCE-MRI habitat omics analysis provided in an embodiment of the present invention.

[0075] Figure 3 This is a diagram of the molecular subtyping network architecture for breast cancer based on DCE-MRI habitat omics analysis provided in this embodiment of the invention.

[0076] Figure 4 It is a cross-attention fusion module (taking pharmacokinetics as an example).

[0077] Figure 5 It is an orthogonal fusion module (taking the enhanced mode branch as an example).

[0078] Figure 6 It is a pooling fusion module. Detailed Implementation

[0079] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be noted that the following embodiments are for illustrative purposes only and do not constitute any limitation on the scope of protection of this invention.

[0080] Example 1: Precision Treatment Decisions for Breast Tumors

[0081] A newly diagnosed breast cancer patient was seen during a multidisciplinary team (MDT) consultation for breast cancer at a top-tier hospital. The patient's biopsy pathology indicated invasive carcinoma, but the sample size was limited and heterogeneous, making it difficult to determine a complete molecular subtype (e.g., differentiating between Luminal A and Luminal B types) through routine immunohistochemistry. Accurate molecular subtyping is crucial for selecting targeted therapy, endocrine therapy, or chemotherapy regimens. Using the method of this invention, the treatment team can:

[0082] 1. Utilize routine examination data: Use the patient's preoperative breast DCE-MRI images as input, eliminating the need for additional, invasive, or expensive examinations.

[0083] 2. Automated habitat analysis and classification: The system of this invention automatically divides the tumor region in MRI images into two branches, and analyzes the heterogeneity of the tumor microenvironment from two dimensions: pharmacokinetics and enhancement patterns.

[0084] 3. Generate and classify highly discriminative features: The system integrates bi-branch information using a multi-level feature fusion module to generate fusion features with strong representational capabilities, and the trained classifier directly outputs the predicted molecular typing results.

[0085] 4. Assisting in the formulation of treatment strategies: Based on the objective and quantitative classification prediction provided by this invention, combined with pathological results, the MDT team can more confidently formulate individualized neoadjuvant or adjuvant treatment plans for patients (such as whether anti-HER2 targeted drugs are needed), realizing a precise bridge from imaging to treatment decisions.

[0086] Example 2: Promoting Multicenter Clinical Research and Standardized Assessment

[0087] A nationwide breast cancer imaging research project required the collection of DCE-MRI data from dozens of different hospitals using various brands and models of MRI equipment to develop and validate a robust molecular subtyping prediction model. However, significant differences in scanning parameters, resolution, and image quality between centers presented a serious challenge of the "center effect" to analyses based on traditional radiomics features. Using the method of this invention, the research team was able to:

[0088] 1. Handling heterogeneous data: The system first performs tumor segmentation and data normalization preprocessing by directly inputting the raw DCE-MRI data provided by each center.

[0089] 2. Focus on the biological essence: By dividing the habitat based on physiological significance (such as pharmacokinetic branches), the "noise" caused by differences in equipment and protocols is eliminated to a certain extent, and more attention is paid to extracting stable patterns related to biological characteristics such as tumor angiogenesis and permeability.

[0090] 3. Enhanced robustness through fusion: Through cross-self-attention and orthogonal fusion modules, the system can adaptively learn and correct consistency information between features from different sources and suppress non-specific variations, thereby generating more robust fused feature representations.

[0091] 4. Establishing a generalizable model: The classification model trained using this framework demonstrated higher generalization performance and stable diagnostic efficacy when tested on unseen data from new centers, providing a feasible technical solution for establishing a cross-institutional, standardized non-invasive breast cancer subtyping assessment process.

[0092] As can be seen from the two examples above, the breast cancer molecular subtyping method based on DCE-MRI habitat omics analysis provided by this invention demonstrates significant application value in practical medical scenarios. Whether in precision diagnosis and treatment decision-making within hospitals or in multi-center clinical research across institutions, this method can deeply mine biological heterogeneity information from routine images, providing a non-invasive, objective, and quantifiable auxiliary tool for breast cancer molecular subtyping. It is expected to significantly improve the objectivity and consistency of diagnosis, the accuracy of treatment decisions, and the generalization ability of research models.

[0093] To address the problems existing in the prior art, this invention provides a method, system, device, and terminal for molecular subtyping and identification of breast cancer based on DCE-MRI habitat omics analysis. The invention will be described in detail below with reference to the accompanying drawings.

[0094] like Figure 1 As shown, the molecular subtyping method for breast cancer based on DCE-MRI habitat omics analysis provided in this embodiment of the invention includes the following steps:

[0095] S101: Acquisition of breast DCE-MRI data, including data from eligible patients;

[0096] S102: Manually segment and preprocess the tumor region in the original image;

[0097] S103: Achieve tumor habitat division through a dual-branch habitat division module;

[0098] S104: Extract habitat features and obtain global joint features through multi-level fusion;

[0099] S105: Input the global joint features into the classifier and output the molecular subtype classification result.

[0100] Processing procedure for molecular subtyping identification of breast cancer based on DCE-MRI habitat omics analysis:

[0101] First, dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) sequences of the patient's breast are acquired from the medical imaging system. Experienced physicians or professionals then manually delineate the tumor boundaries on the raw images using the Medical Imaging Interaction Toolkit (MITK) software, precisely segmenting the tumor region and eliminating interference from surrounding normal tissue and background.

[0102] Subsequently, the segmented tumor region data underwent normalization processing. First, the DICOM format image data was converted to the more easily processed RAW format. Then, the 3D tumor volume of all samples was normalized by the number of slices to ensure consistent input data size. Finally, the preprocessed full sample dataset was randomly divided into training and test sets for model development and independent evaluation.

[0103] During the training phase, the model performs the following core operations sequentially: First, the dual-path habitat segmentation module analyzes each training sample, generating multiple images representing different functional sub-regions (habitats) within the tumor through two complementary paths: pharmacokinetic analysis (based on the CAM algorithm) and dynamic enhancement pattern analysis (based on superpixels and SER calculation). Next, these images are input into a multi-level feature fusion network, undergoing feature extraction, cross-attention mechanism fusion, orthogonal fusion, and adaptive weighted pooling, integrating the dual-path information into a highly discriminative global feature representation. Finally, this feature is used to train the classifier, iteratively optimizing and progressively improving its classification performance for breast cancer molecular subtypes.

[0104] Finally, in the testing phase, the trained model is used to process the test set samples. The dual-path habitat segmentation module automatically analyzes the input breast DCE-MRI images, generating two sets of habitat features: pharmacokinetic and dynamic enhancement patterns. Subsequently, a multi-level feature fusion network encodes and fuses the features to form highly discriminative global features, which are then input into the classifier. Through forward propagation, the model quickly outputs the molecular subtyping results for breast cancer. This fully automated workflow improves diagnostic efficiency and objectivity while providing reliable non-invasive subtyping evidence for clinical practice.

[0105] As a preferred embodiment, the method for classifying and identifying molecular subtypes of breast cancer provided in this invention specifically includes the following steps:

[0106] (1) Segmentation and processing of the original image: In the obtained breast DCE-MRI, the tumor region in the original image was manually segmented using MITK software; the DICOM format data was converted to RAW format data; the number of slices containing tumors in all patients was normalized to obtain three-dimensional images of the same size; the processed dynamic enhanced multi-phase images were used as sample datasets to divide the training set and test set.

[0107] In the acquired breast DCE-MRI, the segmentation and processing of the original image includes:

[0108] (1.1) In the acquired dynamic contrast-enhanced magnetic resonance images of the breast, the tumor region in the original image was manually segmented using MITK software;

[0109] (1.2) Convert the segmented tumor region data from DICOM format to RAW format;

[0110] (1.3) Normalize the number of slices containing tumors in all patients to obtain three-dimensional images of the same size;

[0111] (1.4) The processed dynamic enhanced multi-phase images are used as a sample dataset and divided into training set and test set.

[0112] (2) Dual-branch segmentation of tumor habitat images: A dual-branch module consisting of two habitat segmentation algorithms is constructed to generate habitat images with different biological meanings, starting from pixel-level pharmacokinetics and region-level enhancement modes, respectively. In the pharmacokinetic branch, based on the convex set hybrid analysis algorithm, the time-signal curve of each pixel is decomposed into three baseline kinetic modes—input plasma, rapid perfusion, and slow perfusion—through unsupervised convex optimization, and its spatial distribution map is reconstructed, outputting three functional feature maps. In the enhancement mode branch, based on superpixel segmentation and signal enhancement ratio calculation, the tumor region is divided into three dynamic enhancement mode regions: rapid emptying, plateau, and continuous enhancement. The two branches work together to analyze the heterogeneity within the tumor from different dimensions.

[0113] The process of generating tumor habitat images using a dual-branch method is as follows:

[0114] (2.1) Construct a dual-branch module consisting of two habitat segmentation algorithms to generate habitat images from pixel-level pharmacokinetics and region-level enhancement modes, respectively.

[0115] (2.2) In the pharmacokinetic branch, a multi-phase signal sequence is constructed for each pixel, and decomposed by the CAM algorithm to obtain the contribution coefficients and spatial distribution map of the three reference kinetic components.

[0116] (2.3) In the enhancement mode branch, adaptive superpixel segmentation is performed on the critical phase image, and the signal enhancement ratio of each region is calculated.

[0117] (2.4) Based on the preset signal enhancement ratio threshold, the superpixel region is classified into three types: fast emptying type, plateau type, and continuous enhancement type.

[0118] (2.5) Integrate the outputs of the two branches to generate a total of six types (three pharmacokinetic maps and three types of enhancement mode regions) of habitat images describing the heterogeneity within the tumor.

[0119] (3) Habitat image feature representation and interaction: After obtaining the dual-branch habitat images, each habitat image is mapped to a high-dimensional feature space through a feature extractor (such as a convolutional coding network) to encode the image information. A cross-self-attention module is constructed so that features from the pharmacokinetic branch and the enhancement mode branch can be used as queries and keys for each other, enabling deep interaction and contextual semantic complementarity. Orthogonal fusion constraints are introduced to reduce linear redundancy between features from different branches and extract path features with higher information density and stronger complementarity. Multi-scale information is dynamically integrated through a pooling fusion module to generate a globally joint feature representation with strong discriminative power. This feature integrates comprehensive information from different habitats within the tumor.

[0120] The representation and interaction of habitat image features are carried out according to the following steps:

[0121] (3.1) Input the two-branch habitat images into the feature extractor to obtain the high-dimensional feature codes of each habitat image;

[0122] (3.2) Input the two types of feature encodings into the cross-self-attention module to achieve deep interaction and semantic enhancement of features;

[0123] (3.3) Introduce orthogonal loss constraints to fuse the interactive features to reduce feature redundancy and obtain refined path features;

[0124] (3.4) The refined features and the original features are concatenated, and adaptive weights are generated by global pooling and lightweight multilayer perceptron. The weighted sum is then used to obtain the final global joint feature representation.

[0125] (3.5) During the training phase, the classification loss (cross-entropy loss) is backpropagated to simultaneously optimize the feature fusion module and classifier parameters, so that the model learns the feature representation that is most discriminative for the molecular subtyping diagnosis task.

[0126] (4) Molecular subtyping diagnosis of breast cancer: For the breast DCE-MRI data to be diagnosed, the pharmacokinetic and enhancement mode habitat images are first generated by the dual-branch habitat segmentation module. The obtained habitat images are sent to the feature extractor to obtain feature encoding, and then interacted and enhanced by the multi-level feature fusion module to obtain a global joint feature representation. The global joint feature is sent to the classifier composed of a fully connected network to directly output the diagnostic results of the molecular subtype of breast cancer.

[0127] The molecular subtyping diagnosis of breast cancer is performed according to the following steps:

[0128] (4.1) Input the DCE-MRI data of the breast to be diagnosed into the trained dual-branch habitat segmentation module to generate habitat images;

[0129] (4.2) Input the habitat image into the feature extractor to obtain feature encoding, and then process it through the multi-level feature fusion module to obtain global joint features;

[0130] (4.3) Input the global joint features into the fully connected network classifier to obtain the final molecular subtyping diagnosis result.

[0131] This technology relates to a method for molecular subtyping and identification of breast cancer based on DCE-MRI habitat omics analysis. It aims to acquire tumor habitat images through a dual-branch habitat segmentation module and perform effective feature processing and classification to achieve accurate diagnosis of breast cancer molecular subtyping. The detailed working principle is as follows:

[0132] 1. Preprocessing and standardization of breast tumor images

[0133] Step 1: In the acquired breast DCE-MRI images, the tumor region is manually segmented, and the images are preprocessed, including data format conversion and 3D size normalization, to prepare clean and standardized data for subsequent habitat omics analysis.

[0134] 2. Construction of a dual-branch habitat division module

[0135] Step 2: Construct a dual-branch habitat segmentation module, which includes two complementary segmentation methods: one is segmentation based on pixel-level pharmacokinetics (CAM algorithm), generating spatial maps representing different microcirculatory functions; the other is segmentation based on region-level enhancement patterns (superpixel segmentation and SER calculation), generating region maps with different dynamic enhancement patterns. Both work together to comprehensively characterize the heterogeneous microenvironment (habitat) within the tumor.

[0136] 3. Construction and Training of Multi-Level Feature Fusion Networks

[0137] Step 3: Construct a multi-layered feature fusion network. First, the two habitat images are encoded into high-dimensional features using a feature extractor. Then, through a cross-attention module, an orthogonal fusion module, and a pooling fusion module, deep interaction, redundancy removal, and adaptive weighted fusion of the two feature streams are achieved, ultimately generating a highly discriminative global joint feature representation. The entire network is trained using classification loss (cross-entropy loss) to enable the model to learn the most valuable features for molecular typing.

[0138] 4. Achievement of molecular subtyping diagnosis for breast cancer

[0139] Step 4: Input the breast DCE-MRI data to be diagnosed into the trained system. The system first generates a habitat image through a two-branch habitat segmentation module, then obtains global joint features through feature extraction and multi-level fusion, and finally inputs the features into the classifier to output specific molecular subtyping diagnostic results such as Luminal A, Luminal B, HER2 positive, triple negative, etc.

[0140] 5. Model Testing and Application

[0141] Step 5: During the testing phase, unseen patient DCE-MRI data is input into the system. The system can automatically complete the entire process from habitat division and feature fusion to classification diagnosis, providing objective and accurate molecular subtyping auxiliary diagnostic support for breast cancer in clinical practice.

[0142] This method enables the technology to deeply mine heterogeneous information within tumors from DCE-MRI images. Through innovative habitat segmentation and feature fusion strategies, it achieves highly accurate automatic identification of breast cancer molecular subtypes, which helps guide clinical treatment decisions.

[0143] The breast cancer molecular subtype classification and identification system based on DCE-MRI habitat omics analysis provided in this invention includes:

[0144] The data acquisition module is used to acquire breast DCE-MRI data and include patient data that meets the criteria.

[0145] The data segmentation and preprocessing module is used for segmentation and preprocessing of tumor regions in breast DCE-MRI images;

[0146] The dual-branch habitat segmentation module is used to segment heterogeneous sub-regions (habitats) within the tumor. It takes preprocessed DCE-MRI data as input and uses two complementary technical approaches—based on pharmacokinetics (CAM algorithm) and based on region-level enhancement mode (superpixel segmentation and SER calculation)—to collaboratively generate habitat images representing different physiological characteristics.

[0147] The multi-level feature fusion module is used to interact with and enhance the features extracted from the two-branch habitat image. It takes the two-branch feature encoding as input and generates a more discriminative global joint feature representation through cross-self-attention mechanism, orthogonal fusion and pooling fusion method.

[0148] The breast cancer molecular subtyping diagnostic module takes the fused global joint features as input, feeds them into a classifier composed of a fully connected network, and outputs the diagnostic results of breast cancer molecular subtyping to complete the classification task.

[0149] The detailed working principle of the breast cancer molecular subtype classification and identification system based on DCE-MRI habitat omics analysis provided in this invention is as follows:

[0150] First, the data acquisition module is responsible for collecting patients' breast DCE-MRI sequence data from the medical image archiving and communication system, and screening out case data with clear molecular subtyping labels according to the clinical pathology gold standard. These images, containing multi-phase dynamic enhancement information, provide the foundation for subsequent habitat analysis.

[0151] Next, the data segmentation and preprocessing module processes the collected raw DCE-MRI data. This module uses medical image interaction tools to manually and accurately delineate the three-dimensional region of the tumor to focus on the analysis target. Subsequently, the segmented tumor regions undergo preprocessing such as format standardization and slice number normalization to ensure that all input data are of consistent size, providing standardized input for model training and inference.

[0152] Then, a dual-branch habitat segmentation module performs deep analysis on the preprocessed tumor images. This module works collaboratively through two complementary technical paths: one path, based on a convex set hybrid analysis algorithm, starts from pixel-level pharmacokinetics to unsupervisedly deconstruct habitat images representing different blood flow patterns; the other path, based on superpixel segmentation and signal-to-intensity ratio calculation, starts from regional enhancement patterns to segment the tumor into habitat images with different enhancement curves. The output of this module is a set of habitat images reflecting the heterogeneity within the tumor, characterizing the tumor microenvironment from both functional and morphological dimensions.

[0153] Subsequently, a multi-level feature fusion module performs deep feature learning and integration on the habitat images segmented by the two branches. This module first encodes each habitat image into high-dimensional features using a feature extractor. Then, a cross-self-attention mechanism is used to enable deep semantic interaction and complementarity between the features of the two paths. Next, an orthogonal fusion strategy effectively reduces redundant information between features. Finally, a pooling fusion module adaptively weighs feature importance to generate a powerful global joint feature representation. This process ensures that the system can fully learn and fuse the multi-dimensional heterogeneous information of the tumor.

[0154] Finally, the breast cancer molecular subtyping diagnostic module takes the fused global joint features as input and feeds them into a pre-trained classifier (such as a fully connected network). The classifier performs rapid calculations based on these features rich in discriminative information and directly outputs the corresponding breast cancer molecular subtyping (such as Luminal A, Luminal B, HER2 positive, and triple negative) diagnostic results.

[0155] In summary, this system achieves automated molecular subtyping of breast cancer DCE-MRI images through data standardization, multi-dimensional habitat segmentation, multi-level feature fusion, and intelligent classification. This system not only improves the objectivity and efficiency of subtyping diagnosis, but its habitat visualization results also reveal the heterogeneous characteristics within the tumor, providing valuable imaging evidence for developing personalized treatment plans.

[0156] This invention can be applied to a computer-aided diagnostic system for molecular subtyping of breast cancer. During the training phase, the system receives complete DCE-MRI image data of pathologically confirmed breast cancer as input. After performing tumor region segmentation and standardization preprocessing on the images, the system uses this data to train network models for habitat segmentation, multi-level feature fusion, and classifiers in stages, and then stores the trained models in the system.

[0157] In practical applications, clinicians can input breast DCE-MRI images of patients awaiting diagnosis into the system. The system first automatically performs tumor segmentation and preprocessing, then calls upon stored models to sequentially complete bibranch habitat division, deep feature extraction, and fusion. Finally, based on the learned, highly discriminative global joint features, the system automatically calculates and outputs the predicted molecular subtype of breast cancer (such as Luminal A type, triple-negative type, etc.), providing important imaging evidence for clinicians to develop personalized treatment plans.

[0158] As a preferred embodiment, the method for molecular subtype classification and identification of breast cancer based on DCE-MRI habitat omics analysis provided in this invention specifically includes the following steps:

[0159] (1) Acquisition of breast DCE-MRI image data, segmentation and processing of the original images, the specific process is as follows:

[0160] (1a) The data used in this example are dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) images of the breast, from the Cancer Imaging Archive at Duke University School of Medicine;

[0161] (1b) The inclusion criteria are as follows: 1. The interval between consecutive enhancement phases is 110-120 seconds; 2. Complete four-phase dynamic contrast-enhanced magnetic resonance imaging data are available; 3. Each patient has at least 5 tumor slides;

[0162] (1c) A total of 372 cases were included, including 264 cases of Luminal A, 10 cases of Luminal B, 34 cases of HER2 positive, and 64 cases of triple negative. 335 of these patients were set as the training set and 37 patients were set as the test set.

[0163] (1d) In the obtained breast DCE-MRI images, the tumor region in the original image was manually segmented using MITK software;

[0164] (1e) Convert the segmented tumor region data in DICOM format to RAW format data.

[0165] (2) Manual segmentation and preprocessing of breast tumor images, the specific process is as follows:

[0166] (2a) In the acquired breast DCE-MRI images, the tumor region in the original image was manually segmented using MITK software;

[0167] (2b) Convert the original image and the segmented tumor region DICOM format data into RAW format data;

[0168] (2c) Normalize the number of slices containing tumors in all patients to obtain three-dimensional images of the same size;

[0169] (2d) Use the processed three-dimensional tumor images as the sample dataset.

[0170] (3) Two-branch habitat division: The heterogeneous regions within the tumor are divided in a synergistic manner based on pixel-level pharmacokinetics and regional enhancement patterns. The specific process is as follows:

[0171] (3a) Pharmacokinetic Context Segmentation: Based on the Convex Set Hybrid Analysis (CAM) algorithm, continuous slices of the tumor core region are selected, and a multi-phase (e.g., four-phase) dynamic signal sequence is constructed for each pixel, forming a pixel-time matrix X. After standardizing X, unsupervised convex optimization is used to refine the curve of each pixel. It is decomposed into a linear combination of three baseline kinetic components with clear physiological significance (e.g., plasma infusion, rapid perfusion, and slow perfusion):

[0172]

[0173] in, , , The reference dynamic curves represent the three heteroproton regions respectively. , , The corresponding contribution coefficients are represented. Three feature maps are output through nonnegative matrix factorization (NMF) to quantitatively characterize the three microcirculatory functional habitats.

[0174] (3b) Enhanced mode habitat segmentation: Adaptive SLIC superpixel segmentation is performed on images from key enhancement phases (e.g., arterial phase), with the number of superpixels N = max(10, min(500, total number of pixels / 1000)). The segmented regions are mapped to images from each phase, and the average intensity of each superpixel region in each phase is extracted to calculate its signal enhancement ratio (SER).

[0175] The regions are divided according to the SER threshold: rapid drainage type (SER≥1.2), plateau type (0.9≤SER<1.2), and continuous enhancement type (SER<0.9). Three different habitat maps with dynamic enhancement modes are output.

[0176] (4) Multi-level feature fusion, which performs deep interaction and enhancement on the features of the bi-branch habitat. The specific process is as follows:

[0177] (4a) Feature extraction: Input the two habitat maps (3 each) generated in step (3) into a feature encoder with shared weights (such as ResNet) to obtain high-dimensional feature representations: (pharmacokinetic features) and (enhanced mode features).

[0178] (4b) Cross-self-attention fusion: The features of the two branches mentioned above are simultaneously input into a symmetrical cross-attention module, so that the two types of features can query each other. ) and key value ( This enables deep semantic interaction. The core calculation formula is:

[0179]

[0180] In the formula, It is the dimension of the key vector. It is a fusion feature of cross-attention output.

[0181] (4c) Orthogonal Fusion: Features from multiple habitat maps within each branch are fused. After ordered stitching, an orthogonal loss function constraint is introduced to reduce feature redundancy and increase information density. For the pharmacokinetic branch, an example of the loss function is as follows:

[0182]

[0183] In the formula, , These represent feature vectors of different habitats within this branch. By minimizing this loss, refined path features can be obtained;

[0184] (4d) Pooling Fusion and Weight Generation: The refined features are concatenated with the original attention features to form an enhanced feature stream. Global Average Pooling (GAP) and Global Max Pooling (GMP) are performed separately before concatenation. Adaptive weights α and (1-α) are generated through a lightweight Multi-Layer Perceptron (MLP), and finally, the weighted sum is used to obtain the global joint features. .

[0185] (5) Process the global joint features to achieve the classification and diagnosis of breast cancer molecular subtypes. The specific process is as follows:

[0186] (5a) Feed the global joint representation obtained in step (4) into a classifier composed of fully connected layers;

[0187] (5b) The classifier outputs probability distributions corresponding to different molecular subtypes (e.g., LuminalA, LuminalB, HER2+, Triple-Negative). During training, the network parameters are optimized by minimizing the cross-entropy classification loss function.

[0188]

[0189] In the formula, N represents the total number of samples, pi represents the probability output by the classifier, and yi corresponds to the patient's true label.

[0190] (5c) During the testing phase, the DCE-MRI data to be diagnosed is input, and after habitat division, feature extraction and fusion, the classifier directly outputs the final molecular subtyping diagnosis result of breast cancer.

[0191] This invention uses the area under the receiver operating characteristic (AUC) and accuracy (ACC) as evaluation metrics for model performance, aiming to assess its classification effectiveness for key molecular subtypes of breast cancer (including Luminal and triple-negative). The metrics range from [0,1], with higher values ​​indicating better predictive performance. Specific results of the model on the validation and test sets are shown in Table 1.

[0192] This invention compares the proposed breast cancer molecular subtyping network (Ours) with several other methods, including CAM, SLIC+SER, K-means, and ResNet. Specifically, we used these methods to perform breast cancer molecular subtyping (divided into Luminal / Non-Luminal and TN / Non-TN tasks) on the same test set, and the evaluation metrics included AUC (area under the curve), ACC (accuracy), and p-value (statistical significance).

[0193] Table 1. Results of Molecular Subtype Classification of Breast Cancer

[0194]

[0195] (Significance was calculated), and the classification results are shown in Table 2.

[0196] This invention compares the proposed two-branch habitat segmentation and multi-level feature fusion method with several other habitat feature construction strategies, including: end-to-end methods without habitat segmentation (ResNet-Only), habitat segmentation based on K-means clustering (K-means-ResNet), single habitat segmentation based on pharmacokinetics (CAM-ResNet), and single habitat segmentation based on enhancement patterns (SLIC-SER-ResNet). Specifically, we trained the above comparative models using the same training set and network backbone (ResNet), and evaluated their molecular subtyping performance on the same independent test set. The results are shown in Table 2. Experimental results show that the proposed two-branch fusion method achieves the best diagnostic performance. When using the end-to-end baseline method, the model directly learns the original image and fails to explicitly characterize the heterogeneity within the tumor, resulting in limited classification performance. Although the K-means method introduces an unsupervised habitat concept, its segmentation lacks clear medical prior guidance, and the generated regional features have insufficient discriminative power. Two prior-based single-path methods (CAM and SLIC-SER) showed significant performance improvements, providing effective feature representations from both functional and morphological perspectives, thus confirming the value of habitat analysis. However, habitat feature information from a single perspective is limited and cannot fully characterize the complex tumor microenvironment.

[0197] Table 2: Comparison of Breast Cancer Molecular Subtyping Network with Other Methods

[0198]

[0199] In contrast, the method proposed in this paper deeply integrates complementary information from pharmacokinetic and enhancement modes through a multi-level fusion mechanism, generating a more discriminative and robust joint feature representation. This results in the best performance across all key metrics, validating the superiority of the multi-dimensional habitat fusion strategy.

[0200] Example 1

[0201] Breast DCE-MRI data from 50 breast cancer patients were selected, and molecular subtyping was performed using the method described above. During data preprocessing, the tumor region was manually segmented, converted from DICOM to RAW format, and the number of slices in the 3D tumor images was normalized to 64 layers. The images were randomly divided into a training set of 35 cases and a test set of 15 cases. In the dual-path habitat segmentation, the pharmacokinetic path used a convex set hybrid analysis algorithm to decompose the voxel time-intensity curves, while the dynamic enhancement mode path set the number of superpixels to 100, calculating the signal enhancement ratio to segment the regions. Feature fusion employed a cross-attention module to achieve feature interaction between the two paths, reducing redundancy through orthogonal loss function constraints, and using pooling to fuse adaptive feature weights. The final classifier outputs the subtyping results. The diagnostic accuracy on the test set reached 90%, significantly higher than traditional single-path subtyping methods, demonstrating the inventiveness of dual-path collaborative segmentation and multi-level fusion.

[0202] Example 2

[0203] Data from 80 breast cancer patients encompassing four subtypes were selected: 20 cases of luminal type A, 20 cases of luminal type B, 20 cases of human epidermal growth factor receptor 2 positive, and 20 cases of triple-negative breast cancer. After preprocessing, the data were divided into a training set of 56 cases and a test set of 24 cases. For the pharmacokinetic pathway, a pixel-time matrix was constructed using five consecutive slices of the tumor core region. The number of superpixels in the dynamic enhancement mode was adaptively adjusted to 200 based on the total number of pixels. A ResNet50 feature extractor was used, with the key vector dimension of the cross-attention module set to 64. A lightweight multilayer perceptron with two fully connected layers was employed. During training, the classification loss function was minimized, and the parameters of the fusion module and classifier were simultaneously optimized. The subtype classification accuracy on the test set reached 92.5%, with an inter-subtype confusion rate of less than 8%, validating the ability of the proposed technique to distinguish between different subtypes.

[0204] Example 3

[0205] Data from 30 patients with complex breast cancer were selected. These patients exhibit strong tumor heterogeneity, and traditional classification methods have an accuracy rate of less than 70%. The aforementioned system was used for classification. The data acquisition module incorporated DCE-MRI data from eligible patients; the data segmentation and preprocessing module performed format conversion and normalization; the bi-branch habitat division module generated two sets of habitat maps; the multi-level fusion module achieved in-depth feature interaction and redundancy elimination; and the diagnostic module output the classification results. The final diagnostic accuracy reached 88%, with a 91% accuracy rate in differentiating triple-negative and luminal type B breast cancer. This solution addresses the technical challenge of classifying complex cases, demonstrating the innovation and practicality of the technical approach.

[0206] Example 4

[0207] Data from 60 breast cancer patients were selected and divided into a training set (n=42) and a test set (n=18). A comparative experiment was conducted with two groups: the experimental group used the method described above, while the control group used a single pharmacokinetic pathway classification method. The experimental group employed a dual-pathway collaborative segmentation of the tumor habitat and multi-level feature fusion, while the control group only used pharmacokinetic pathway feature extraction and classification. Results showed that the experimental group achieved a 91% accuracy rate on the test set, compared to only 75% in the control group. Furthermore, the experimental group demonstrated a 20% improvement in the accuracy of differentiating between luminal type A and luminal type B compared to the control group. This experiment proves that the combination of dual-pathway collaborative segmentation and multi-level fusion effectively improves classification accuracy, overcomes the technical limitations of traditional single-pathway methods in comprehensively capturing tumor heterogeneity, and highlights the innovativeness of this technology.

[0208] Please note that the technical solution of this invention can be implemented through hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic circuits; the software portion can be stored in memory and executed by an appropriate instruction execution system (such as a microprocessor or dedicated hardware). Those skilled in the art should understand that the methods and systems of this invention can be implemented using computer-executable instructions and / or processor control code. Such code can be carried on physical media including, but not limited to: disks, CDs, or DVD-ROMs; programmable storage media such as read-only memories; or data transmission media such as optical or electronic signal carriers. The devices and modules of this invention can be implemented using hardware circuits such as very large-scale integrated circuits, gate arrays, logic chips, and programmable logic devices, or implemented using software executed by various processors, or implemented in firmware form combining hardware and software.

[0209] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A method for breast cancer molecular subtyping based on DCE-MRI habitatomics analysis, characterized by, Includes the following steps: Step 1: Tumor region segmentation and data preprocessing; Step 2: Dual-path collaborative habitat segmentation. A habitat segmentation module containing two complementary segmentation strategies is constructed, which starts from pixel-level pharmacokinetics and regional enhancement modes respectively, to collaboratively complete the segmentation of heteroprotic regions within the tumor. Step 3, deep feature extraction and multi-level fusion: the segmented habitat data are extracted by feature extractors, and then the features are successively processed by cross-self-attention module, orthogonal fusion module and pooling fusion module to achieve deep interaction, redundancy elimination and adaptive weighted fusion of the two branches of features, so as to obtain a global joint feature representation. Step four, molecular subtype diagnosis of breast cancer: The global joint features are fed into the trained classifier, and the molecular subtype diagnosis results of breast cancer are output.

2. The method of breast cancer molecular typing based on DCE-MRI habitat-omics analysis according to claim 1, wherein, Step one specifically includes: (1) In the acquired dynamic contrast-enhanced magnetic resonance imaging of the breast, the tumor region in the original image was manually segmented using the Medical Imaging Interaction Toolkit software; (2) Convert the segmented tumor region data from DICOM format to RAW format; (3) Normalize the number of three-dimensional tumor images of all patients in the slice direction to obtain three-dimensional image data of the same size; (4) All preprocessed samples are randomly divided into training set and test set for model training and evaluation. 3.The method of breast cancer molecular typing based on DCE-MRI habitat-based oncoimics analysis of claim 1, wherein, Step two, the dual-pathway collaborative habitat delineation, specifically includes: (1) Construct a habitat segmentation module that includes pharmacokinetic pathways and dynamic enhancement mode pathways; (2) In the pharmacokinetic pathway, continuous multi-layer slices of the core tumor region are selected, and a multi-phase dynamic signal sequence is constructed for each voxel to form a pixel-time matrix. After standardizing the matrix, the time-intensity curve of each voxel is decomposed into a combination of three physiologically significant baseline kinetic components by using the convex set hybrid analysis algorithm. The combination relationship of the three baseline kinetic components is expressed by the formula as the sum of the products of each baseline kinetic curve and its corresponding contribution coefficient. The solution is obtained by non-negative matrix decomposition, and three feature maps are output to quantitatively characterize the three microcirculatory functional habitats. (3) In the dynamic enhancement mode path, adaptive simple linear iterative clustering superpixel segmentation is performed on the second phase image. The number of superpixels is set to a maximum of 10, a minimum of 500 and not exceeding the total number of pixels divided by 1000. The segmented regions are mapped to the images of each phase. The average intensity of each superpixel region in each phase is extracted, and its signal enhancement ratio is calculated. The signal enhancement ratio is calculated by dividing the difference between the average intensity of the second phase and the first phase by the difference between the average intensity of the third phase and the first phase. Regions are divided according to the signal enhancement ratio threshold. A signal enhancement ratio greater than or equal to 1.2 is a fast emptying type, a signal enhancement ratio greater than or equal to 0.9 and less than 1.2 is a plateau type, and a signal enhancement ratio less than 0.9 is a continuous enhancement type. Three different dynamic enhancement mode habitat maps are output. 4.The method of breast cancer molecular typing based on DCE-MRI habitat-based oncoimics analysis of claim 1, wherein, Step three, deep feature extraction and multi-level fusion, specifically includes: (1) Input the two sets of habitat images obtained in step two into the feature extractor to realize the mapping from images to high-dimensional feature vectors; (2) Input the extracted two path features into the cross-self attention module at the same time. The two types of features interact deeply through the symmetrical cross attention mechanism. The cross attention is calculated by multiplying the query vector and the transpose of the key vector, dividing by the square root of the key vector dimension, and then multiplying by the value vector after processing by the softmax function. (3) Multiple features in each path after interaction enhancement are fused using an ordered splicing strategy, and an orthogonal loss function is introduced for constraint. The orthogonal loss function is the sum of the absolute value of the dot product of different habitat feature vectors in the same branch and the square of the dot product. The refined path features are obtained by minimizing this loss. (4) The path features refined by orthogonal fusion are concatenated with the original path features to form an enhanced feature flow; global max pooling and global average pooling are performed on each enhanced feature flow, and the pooling results are concatenated and then an adaptive weight is generated through a lightweight multilayer perceptron to perform weighted summation on the features to obtain global joint features. 5.The method of breast cancer molecular typing based on DCE-MRI habitat-based oncoimics analysis of claim 1, wherein, Step four, the molecular subtype diagnosis of breast cancer, specifically includes: (1) For the breast DCE-MRI data to be diagnosed, input it into the trained habitat segmentation module to generate two sets of habitat images: pharmacokinetic and dynamic enhancement mode. (2) The obtained habitat images are fed into the feature extractor to obtain the feature codes of each habitat image; (3) Input the feature codes of the two paths into the multi-level feature fusion module for processing to obtain global joint features; (4) The global joint features are fed into a classifier composed of fully connected layers, and the molecular subtyping diagnosis results of the breast cancer are output, which belong to luminal type A, luminal type B, human epidermal growth factor receptor 2 positive or triple negative breast cancer.

6. The molecular subtyping method for breast cancer based on DCE-MRI habitat omics analysis as described in claim 1, characterized in that, During training, training set data labeled with molecular subtyping tags is used to minimize the classification loss function while optimizing the parameters of the multi-level feature fusion module and the final classifier.

7. A molecular subtype classification and identification system for breast cancer based on DCE-MRI habitat omics analysis, characterized in that, include: The data acquisition module is used to acquire breast DCE-MRI data and include patient data that meets the criteria. The data segmentation and preprocessing module is used for segmentation and preprocessing of breast DCE-MRI data; The dual-branch habitat segmentation module is used to segment tumor habitat images. It takes the patient's DCE-MRI data as input and uses two complementary technical approaches, namely, a convex set hybrid analysis algorithm and a superpixel segmentation and signal enhancement ratio calculation, to collaboratively complete the segmentation of the tumor's internal habitat. The multi-level fusion module is used to further learn the features of the two branches after the features are extracted by the encoder from the segmented habitat images, enhance the feature representation, and obtain the joint feature representation with the strongest discriminative power and the richest information. The breast cancer molecular subtyping diagnostic module is used to take the fused habitat feature codes as input and feed them into a classifier composed of a fully connected network to obtain the diagnostic results, thus completing the breast cancer molecular subtyping task.

8. The breast cancer molecular subtype classification and identification system as described in claim 7, characterized in that, The partitioning logic of the dual-branch habitat partitioning module is consistent with the dual-path collaborative habitat partitioning logic described in claim 3.

9. The breast cancer molecular subtype classification and identification system as described in claim 7, characterized in that, The feature fusion logic of the multi-level fusion module is consistent with the deep feature extraction and multi-level fusion logic described in claim 4.

10. The breast cancer molecular subtype classification and identification system as described in claim 7, characterized in that, The diagnostic logic of the breast cancer molecular subtyping diagnostic module is consistent with the breast cancer molecular subtype diagnostic logic described in claim 5.

Images