Intraoperative treatment support system for head and neck cancer based on deep transfer reinforcement learning

Abstract

The application relates to the technical field of deep transfer reinforcement learning, and specifically discloses a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning. The system first deeply fuses multi-modal images and clinical texts, uses a nonlinear measurement based on generalized mutual information and an asymmetric orthogonal decoupling mechanism to establish the dominant position of clinical priori, and accurately extracts complementary residual information in the image data that cannot be explained by clinical indicators. Further, to solve the model generalization bottleneck caused by medical small samples, the system uses a transfer reinforcement learning paradigm, trains a general perception strategy using public source domain data, combines general features and fusion representations of specific patients through parameter transfer and target domain fine-tuning. Finally, the system realizes multi-gene mutation probability prediction through a forward inference pipeline, thereby providing objective and quantitative intraoperative diagnosis and treatment support for clinicians.

Description

Technical Field

[0001] This application relates to the field of deep transfer reinforcement learning technology, and more specifically, to a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning. Background Technology

[0002] Head and neck squamous cell carcinoma (HNSCC) is the sixth most common malignant tumor worldwide, with approximately 890,000 new cases and 450,000 deaths annually. It originates from the mucosal epithelium of the oral cavity, oropharynx, hypopharynx, and larynx. Despite significant advancements in multimodal treatments including surgery, radiotherapy, chemotherapy, and immunotherapy, the 5-year survival rate for HNSCC remains around 50%–60%, primarily due to late-stage diagnosis and the high molecular heterogeneity of the tumor.

[0003] In recent years, the rapid development of radiogenomics has demonstrated that CT and MRI imaging features (including tumor morphology, texture, and boundary characteristics) can serve as non-invasive biomarkers for predicting gene mutations. By combining deep learning-based image phenotypic analysis of multimodal CT / MRI images with clinical omics data, the mutation status of key genes related to HNSCC (including 12 genes such as TP53, CDKN2A, NOTCH1, FAT1, PIK3CA, HRAS, CCND1, FBXW7, CASP8, AXIN2, EGFR, and BRAF) can be predicted, thus providing a non-invasive means for molecular subtyping and personalized treatment planning. However, existing gene mutation prediction schemes still have significant technical limitations: First, existing technologies generally suffer from linear traps and modality parity. Commonly used feature selection strategies (such as methods based on Pearson correlation coefficients) can only capture the linear statistical association between image texture and clinical indicators, ignoring the nonlinear coupling relationships that are widespread in tumor heterogeneity (e.g., specific gene mutations only activate image phenotypes within specific texture threshold ranges). Furthermore, they incorrectly place highly semantic clinical gold standard data on an equal footing with noisy image data for variance screening, leading to key pathological indicators being masked or mistakenly rejected by noise. Second, due to the difficulty and high cost of collecting multimodal medical image data with gene mutation detection annotations, and the differences in equipment between different centers, directly trained models often face severe domain shift problems, making it difficult to maintain robustness in new clinical scenarios. Summary of the Invention

[0004] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning.

[0005] According to one aspect of this application, a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning is provided, comprising: The patient data acquisition module is used to acquire multimodal imaging data and clinical text data of patients with head and neck cancer. The data preprocessing module is used to preprocess multimodal image data and clinical text data to obtain a multimodal head and neck cancer dataset; The multimodal fusion feature extraction module is used to extract fused multimodal feature vectors from the multimodal head and neck cancer dataset; The model training and optimization module is used to train and optimize the source domain deep reinforcement learning model based on the acquired public source domain dataset to obtain the pre-trained model parameters. The transfer learning fine-tuning module is used to initialize the target task network based on the pre-trained model parameters, and to perform transfer learning fine-tuning on the target task network using fused multimodal feature vectors as target domain data to obtain a head and neck cancer prediction model. The gene mutation prediction module is used to input the fusion features to be predicted into the head and neck cancer prediction model to predict the probability of multi-gene mutations, so as to obtain the prediction results of the mutation probability of each gene.

[0006] Compared with existing technologies, this application provides a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning. Firstly, it deeply fuses multimodal images and clinical text. Utilizing a nonlinear metric based on generalized mutual information and an asymmetric orthogonal decoupling mechanism, it establishes clinical prior dominance while accurately extracting complementary residual information from image data that cannot be explained by clinical indicators. Furthermore, to address the model generalization bottleneck caused by small medical sample sizes, the system adopts a transfer reinforcement learning paradigm. It trains a general perception strategy using publicly available source domain data and combines general features with patient-specific fusion representations through parameter transfer and target domain fine-tuning. Finally, it achieves multi-gene mutation probability prediction through a forward inference pipeline, thereby providing clinicians with objective and quantitative intraoperative diagnostic and treatment support. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a system block diagram of a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application.

[0009] Figure 2This is a schematic diagram of data flow in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application.

[0010] Figure 3 This is a block diagram of the multimodal fusion feature extraction module in the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application.

[0011] Figure 4 This is a block diagram of a multimodal feature fusion unit in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application.

[0012] Figure 5 This is a schematic diagram of the training set loss and validation set loss curves in the tumor segmentation training process of the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application.

[0013] Figure 6 This is a schematic diagram showing the curves of the training set Dice coefficients and the validation set Dice coefficients in the tumor segmentation training process of the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application.

[0014] Figure 7 This is a schematic diagram showing the intersection-union ratio of the training set and the intersection-union ratio of the validation set in the tumor segmentation training process of the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application.

[0015] Figure 8 This is a schematic diagram showing the training set accuracy versus validation set accuracy curves of the tumor segmentation training process in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application.

[0016] Figure 9 This is a schematic diagram of the training set recall and validation set recall curves in the tumor segmentation training process of the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application.

[0017] Figure 10 This is a schematic diagram comparing the training set Dice coefficients and the validation set Dice coefficients of a tumor segmentation training process in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] To address the technical challenges of existing head and neck cancer treatment support solutions, such as feature mining being limited by linear assumptions, neglecting noise interference caused by semantic hierarchy differences between modalities, and insufficient model generalization ability due to a lack of medical data, this application proposes a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning. This system first performs frequency domain reconstruction and heterogeneous stitching of multimodal images and clinical text, constructing a high-dimensional feature pool using multi-strategy texture analysis. Then, abandoning traditional linear screening strategies, it introduces a generalized mutual information metric based on kernel density estimation to capture complex nonlinear dependencies and performs image-clinical asymmetric orthogonal decoupling. Mathematically, it forcibly removes redundant projections in images interpreted by clinical priors, retaining only image residual features with high complementary value. Based on this, the system adopts a transfer reinforcement learning strategy of "source domain pre-training - target domain fine-tuning". It uses public datasets to train deep Q-networks to obtain general perception parameters, and then combines the high-purity fusion features of the target domain to perform adaptive weight fine-tuning and gradient boosting tree optimization on the network. Finally, it achieves accurate prediction of gene mutation status through forward inference architecture, thereby effectively solving the problem of cross-modal deep feature mining and robust inference in small sample scenarios while ensuring the clinical prior dominance.

[0020] Figure 1 This is a system block diagram of a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application. Figure 1 and Figure 2 As shown, the head and neck tumor intraoperative treatment support system 100 based on deep transfer reinforcement learning according to an embodiment of this application includes: a patient data acquisition module 110 for acquiring multimodal image data and clinical text data of head and neck cancer patients; a data preprocessing module 120 for preprocessing the multimodal image data and clinical text data to obtain a multimodal head and neck cancer dataset; a multimodal fusion feature extraction module 130 for extracting fused multimodal feature vectors from the multimodal head and neck cancer dataset; a model training and optimization module 140 for training and optimizing a source domain deep reinforcement learning model based on the acquired public source domain dataset to obtain pre-trained model parameters; a transfer learning fine-tuning module 150 for initializing a target task network based on the pre-trained model parameters, and using the fused multimodal feature vectors as target domain data to perform transfer learning fine-tuning on the target task network to obtain a head and neck cancer prediction model; and a gene mutation prediction module 160 for inputting the fused features to be predicted into the head and neck cancer prediction model to predict the probability of multiple gene mutations, so as to obtain the prediction results of the probability of each gene mutation.

[0021] In the aforementioned head and neck tumor intraoperative treatment support system 100 based on deep transfer reinforcement learning, the patient data acquisition module 110 is used to acquire multimodal image data and clinical text data of head and neck cancer patients. It should be understood that most existing image processing solutions for head and neck surgery only use single CT data. However, simple CT images cannot fully and clearly display soft tissue structures and lack the auxiliary validation of clinical data. This results in an information bottleneck in the model's ability to capture deep image phenotypic features related to gene mutation states, failing to meet the precise clinical needs for molecular subtyping and individualized treatment planning. Therefore, in the technical solution of this application, a composite database containing multimodal images such as MRI and CT, patient clinical omics indicators, and disease description text is constructed to overcome the information bottleneck of a single modality and achieve complementary fusion of anatomical structural information and functional metabolic information. This provides high-dimensional multimodal data support for subsequent deep reinforcement learning models, covering image morphology, clinical indicators, and gene mutation ground truth annotations, significantly improving the model's ability to perceive the correlation between tumor image phenotypes and gene mutations and its prediction accuracy for multi-gene mutation states.

[0022] Specifically, in one example of this application, the collection and screening of multi-source heterogeneous data are first performed. The sources of the multimodal imaging data and clinical text data are divided into two parts: public datasets and hospital clinical datasets. The public datasets include the TCIA head and neck cancer multimodal imaging dataset with industry-standard annotations and related public genomics datasets, which are used to train the model's general perception ability of the association between imaging phenotypes and gene mutations in head and neck squamous cell carcinoma. The hospital clinical datasets are collected through clinical information systems and image archiving systems, and are specifically screened to include head and neck cancer patient cases covering both MRI and CT imaging modalities and with gene mutation detection results, which are used for fine-tuning the target domain of the gene mutation probability prediction model. During the data collection process, strict inclusion criteria for clinical samples were established: ensuring that each sample had clear multi-gene mutation detection results (covering 12 key genes for head and neck squamous cell carcinoma, including TP53, CDKN2A, NOTCH1, FAT1, PIK3CA, HRAS, CCND1, FBXW7, CASP8, AXIN2, EGFR, and BRAF, with each gene marked as 0 for no mutation and 1 for mutation), and confirming that the patient had not received any radiotherapy or chemotherapy intervention before undergoing CT / MRI imaging. Simultaneously, clinical omics indicators (such as TNM stage and tumor size) and unstructured disease description text corresponding to the imaging samples were collected, thus completing the acquisition of raw multimodal head and neck cancer data and laying the foundation for subsequent dataset construction.

[0023] In the aforementioned head and neck tumor intraoperative treatment support system 100 based on deep transfer reinforcement learning, the data preprocessing module 120 is used to preprocess multimodal image data and clinical text data to obtain a multimodal head and neck cancer dataset. It should be understood that, due to the significant differences in spatial resolution, imaging perspective, and contrast range between the original acquired multimodal image data (including MRI and CT), and the fact that clinical text data is unstructured discrete information, both cannot be directly processed simultaneously by a deep learning model. Without unified standardization and fusion, the model will struggle to extract effective associated features. Therefore, in the technical solution of this application, further data preprocessing is performed on the multimodal image data and clinical text data, including frequency domain fusion, soft tissue window correction, region of interest truncation, and text structured encoding, to eliminate heterogeneity between modalities and enhance the saliency of lesion information. This generates a multimodal head and neck cancer dataset with a unified format, aligned features, and containing high-value pathological information, laying a standardized data foundation for subsequent multimodal fusion feature extraction.

[0024] Specifically, in this embodiment, the data preprocessing module is used to: after performing frequency domain weighted reconstruction and spatial domain inverse transformation on the multimodal image data, concatenate it with clinical text data to obtain a fused image and text data stream; perform contrast correction and lesion masking on the fused image and text data stream to obtain a region of interest image and text data stream; perform key-value pair parsing and numerical encoding on the text information in the region of interest image and text data stream to obtain a heterogeneous feature pair data stream, wherein the heterogeneous feature pair data stream includes paired region of interest images and clinical feature vector pairs; perform sample ID allocation and format standardization encapsulation on the heterogeneous feature pair data stream to obtain a multimodal head and neck cancer dataset.

[0025] More specifically, in a concrete example of this application, the original CT and MRI images are first imported using ITK-SNAP software. Fourier transform technology is then used to convert the images of different modalities to the frequency domain for weighted reconstruction. After inverse spatial transform, a fused image containing both anatomical and metabolic information is generated. This fused image is then logically concatenated with the corresponding clinical text data to form a fused image and text data stream. At this point, the fused image is uniformly adjusted to horizontal, coronal, and sagittal perspectives. Subsequently, a 60*400 soft tissue window is set for contrast correction of the fused image and text data stream. Based on a gold standard mask drawn by clinical experts, a lesion region of a specific volume is extracted, generating a region of interest (ROI) image and text data stream focusing on the lesion tissue. Next, for the disease description text and omics indicators in the ROI image and text data stream, key-value pair parsing technology is used to extract structured information, which is then numerically encoded and transformed to generate a heterogeneous feature pair data stream containing paired ROI images (including CT grayscale values ​​and MRI signal intensity features) and clinical feature vectors. Finally, a unique sample ID is assigned to each pair of heterogeneous features, and the samples are encapsulated in a standardized storage format to obtain the final multimodal head and neck cancer dataset.

[0026] In the aforementioned head and neck tumor intraoperative treatment support system 100 based on deep transfer reinforcement learning, the multimodal fusion feature extraction module 130 is used to extract fused multimodal feature vectors from a multimodal head and neck cancer dataset. It should be understood that, since head and neck cancer lesions typically exhibit high microscopic texture heterogeneity, a single feature extraction strategy is insufficient to fully capture their spatial dependence and grayscale variation patterns. Furthermore, simple image features often contain redundant information and lack prior clinical semantic guidance; directly using raw features can lead to model overfitting and poor robustness. Therefore, in the technical solution of this application, a fused multimodal feature vector is further extracted from the multimodal head and neck cancer dataset to significantly improve the subsequent model's ability to discriminate tumor characteristics and gene mutation status.

[0027] Figure 3 This is a block diagram of a multimodal fusion feature extraction module in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application. Figure 3 As shown in the embodiments of this application, the multimodal fusion feature extraction module 130 includes: a multi-strategy texture feature analysis unit 131, used to perform texture space dependency and consistency analysis on the region of interest images in the multimodal head and neck cancer dataset to obtain a multi-strategy texture feature matrix; and a multimodal feature fusion unit 132, used to perform feature fusion based on a multi-head attention mechanism on the multi-strategy texture feature matrix to obtain a fused multimodal feature vector.

[0028] Specifically, the multi-strategy texture feature analysis unit 131 is used to perform texture spatial dependency and consistency analysis on the region of interest images in the multimodal head and neck cancer dataset to obtain a multi-strategy texture feature matrix. It should be understood that, due to the highly non-uniform tissue density and metabolic distribution within head and neck cancer lesions, it is difficult to effectively characterize the arrangement rules and microstructural differences of tumor cells using only a single gray-level statistical measure. Furthermore, the excessively high dynamic range in the original image data easily introduces computational noise, leading to a decrease in the stability of feature extraction. Therefore, in the technical solution of this application, texture spatial dependency and consistency analysis are further performed on the region of interest images in the multimodal head and neck cancer dataset to obtain a multi-strategy texture feature matrix, thereby constructing a comprehensive texture feature description system. This transforms abstract pathological images into statistically significant quantitative indicators, providing a highly discriminative input basis for subsequent feature fusion and disease prediction.

[0029] More specifically, in the embodiments of this application, the multi-strategy texture feature analysis unit is used to: perform gray quantization and discretization processing on the region of interest image to obtain a quantized image data stream; extract spatial dependency features from the quantized image data stream based on the gray-level co-occurrence matrix to obtain an intermediate data stream containing gray-level co-occurrence features; jointly extract gray-level runout and gradient features from the intermediate data stream to obtain an original multidimensional feature vector set; and perform dimensional unification and normal distribution mapping on the original multidimensional feature vector set to obtain a multi-strategy texture feature matrix.

[0030] Accordingly, the region of interest (ROI) image is subjected to grayscale quantization and discretization to obtain a quantized image data stream. It should be understood that since the original ROI image (especially CT and MRI modalities) typically has a high bit depth (e.g., 12-bit or 16-bit), its grayscale range is extremely wide and contains a large amount of subtle noise fluctuations that do not substantially contribute to texture analysis. Directly calculating statistical features such as the co-occurrence matrix based on the original grayscale levels would not only result in an overly sparse computational matrix but also lead to an exponential increase in computational load, severely impacting the efficiency and robustness of feature extraction. Therefore, in the technical solution of this application, grayscale quantization and discretization are further performed on the ROI image to compress the grayscale dynamic range and smooth high-frequency noise. This significantly reduces the data dimensionality and improves the statistical significance of subsequent texture matrix calculations while preserving the core texture structure of the tumor.

[0031] In a specific example of this application, each sample in the multimodal head and neck cancer dataset is first traversed, and its corresponding region of interest image data is read to obtain the pixel grayscale value distribution range of the original image. Then, a target number of discretized grayscale levels is set (e.g., 32 or 64 levels), and an isometric quantization algorithm or a Lloyd-Max quantization algorithm is used to map the original continuous or high dynamic range pixel grayscale values ​​to discrete integer index intervals. During this process, a histogram equalization strategy is applied to the image data to adjust the uniformity of the grayscale distribution, enhance the local contrast of the lesion area, and repackage the quantized pixel matrix according to its original spatial location to generate a standard format quantized image data stream, providing low-noise and compact input data for subsequent construction of the grayscale co-occurrence matrix and other texture statistical matrices.

[0032] Accordingly, based on the gray-level co-occurrence matrix, spatial dependency feature extraction is performed on the quantized image data stream to obtain an intermediate data stream containing gray-level co-occurrence features. It should be understood that although the quantized image data stream reduces noise, its essence remains a discrete pixel array, unable to directly reflect the spatial arrangement and correlation of pixel pairs within the tumor. These spatial dependencies are precisely the key indicators revealing tumor heterogeneity. Therefore, in the technical solution of this application, spatial dependency feature extraction is further performed on the quantized image data stream based on the gray-level co-occurrence matrix to quantify the co-occurrence probability of pixels at different gray levels in specific directions and distances. This allows the spatial structural information of the image to be transformed into computable statistical features such as energy and contrast, generating an intermediate data stream capable of characterizing the microscopic texture patterns of head and neck cancer lesions.

[0033] In a specific example of this application, the quantized image data stream is first received, and spatial dependency analysis parameters are set, including the distance step size between pixel pairs (e.g., d=1) and four main discrete directions (0°, 45°, 90°, 135°). Then, each pixel in the quantized image is traversed, and the frequency of simultaneous occurrence of different gray-level combinations (i,j) under the set distance and direction conditions is counted, constructing a multi-directional gray-level co-occurrence matrix. Based on the generated gray-level co-occurrence matrix, a Haralick texture feature set is calculated, specifically extracting energy features to reflect the uniformity of the texture, contrast features to describe the sharpness and groove depth of the texture, correlation features to measure the linear dependence of the gray-level distribution, and entropy features to characterize the randomness and complexity of the texture information. Finally, the spatial statistical features obtained above are logically associated and encapsulated with the original quantized image data to generate an intermediate data stream containing rich texture spatial dependency information, laying the foundation for subsequent joint extraction of more complex gradient features.

[0034] Accordingly, grayscale travel and gradient features are jointly extracted from the intermediate data stream to obtain the original multidimensional feature vector set. It should be understood that since a single grayscale co-occurrence feature can only describe second-order statistical information, it is difficult to fully cover the variations in texture coarseness (such as the homogeneity of necrotic areas) and edge gradient mutations (such as the sharpness of infiltration boundaries) present in head and neck cancer lesions, resulting in blind spots in feature description. Therefore, in the technical solution of this application, grayscale travel and gradient features are further jointly extracted from the intermediate data stream, integrating multiple texture analysis strategies to capture the deep structural information of lesions from different dimensions (connectivity, region size, gradient changes). In this way, a highly complementary and comprehensive high-dimensional feature set can be constructed, fully reflecting the microscopic heterogeneity of tumors in morphology and metabolic function.

[0035] In a specific example of this application, the image matrix is ​​first scanned row-by-row and column-by-column according to four preset discrete directions (0°, 45°, 90°, 135°) for the quantized image data in the intermediate data stream. The length and frequency of consecutive adjacent pixel segments with the same gray value are counted to construct a gray-level run-length matrix reflecting the texture coarseness. Based on this matrix, short-run dominance factors and long-run dominance factors are calculated to quantify the texture coarseness. Simultaneously, an eight-connected region algorithm is used to traverse the image matrix to identify connected regions with the same gray value, and the pixel quantity distribution of each connected region is counted to construct a gray-level region size matrix. Subsequently, a neighborhood gray-level difference matrix is ​​introduced. By calculating the difference distribution between the gray-level mean of the central pixel and its surrounding neighboring pixels, the roughness index perceived by human vision is quantified. Further, the Sobel operator is used to perform horizontal and vertical convolution operations on the image matrix to calculate the gradient magnitude and direction of each pixel. Based on the joint distribution of gray-level values ​​and gradient values, a gray-level gradient co-occurrence matrix is ​​constructed, and the gradient mean and gradient entropy features are extracted. Finally, all the scalar feature values ​​obtained through statistical scanning and convolution calculation are numerically concatenated and vectorized with the existing gray-level co-occurrence features in the intermediate data stream to generate a set of original multidimensional feature vectors containing omnidirectional texture descriptors.

[0036] Accordingly, the original multidimensional feature vector set is subjected to dimensional unification and normal distribution mapping to obtain a multi-strategy texture feature matrix. It should be understood that since the original multidimensional feature vector set contains statistical features from different sources (e.g., energy values ​​are typically between 0 and 1, while travel lengths can be as high as hundreds), these features differ significantly in physical dimensions and numerical ranges. If directly used in subsequent calculations, large numerical features will dominate distance metrics and gradient updates, leading to difficulty in model training convergence or insensitivity to small numerical features. Therefore, in the technical solution of this application, the original multidimensional feature vector set is further subjected to dimensional unification and normal distribution mapping to eliminate scale bias between features and normalize data distribution. This ensures that all texture features participate in subsequent fusion operations on the same numerical benchmark, improving the computational stability of the multi-head attention mechanism and the model's convergence speed.

[0037] In a specific example of this application, the process first iterates through each feature dimension in the original multidimensional feature vector set, calculating the statistical mean and standard deviation of that dimension across all samples. Then, the Z-Score normalization algorithm is used to subtract the corresponding mean from each feature value and divide by the standard deviation, transforming it into a standard normal distribution value with a mean of 0 and a variance of 1, thereby eliminating the influence of physical dimensions. Simultaneously, considering the potential for extreme outliers in some features, a nonlinear mapping function (such as the Sigmoid or hyperbolic tangent function) is introduced to fine-tune the standardized data, limiting the range of numerical fluctuations and preventing outliers from skewing the distribution. Finally, all the standardized and distribution-corrected feature vectors are rearranged and stacked according to the sample order to construct a multi-strategy texture feature matrix with a well-structured row and column structure and a regular numerical distribution, serving as the standard input for the subsequent feature fusion module.

[0038] Specifically, the multimodal feature fusion unit 132 is used to perform feature fusion based on a multi-head attention mechanism on the multi-strategy texture feature matrix to obtain a fused multimodal feature vector. It should be understood that although the multi-strategy texture feature matrix contains rich multidimensional information, it inevitably contains a large number of redundant features and low-value noise. Furthermore, simple image features lack direct guidance from clinical pathological gold standards. Without effective screening and fusion, it is highly susceptible to model dimensionality curse and overfitting. Therefore, in the technical solution of this application, feature fusion based on a multi-head attention mechanism is further performed on the multi-strategy texture feature matrix to identify and retain the most discriminative and stable core feature combinations. This generates a fused multimodal feature vector that is dimensionally concise, semantically complementary, and highly robust, providing high-quality feature input for subsequent disease prediction.

[0039] Figure 4This is a block diagram of a multimodal feature fusion unit in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application. Figure 4 As shown, the multimodal feature fusion unit 132 includes: a texture feature weighting subunit 1321, used to perform subspace projection and scaling dot product operations on the multi-strategy texture feature matrix using a multi-head attention mechanism to obtain a weighted texture feature matrix; a dimension alignment and splicing subunit 1322, used to perform dimension alignment and splicing of the weighted texture feature matrix with clinical feature vectors from the multimodal head and neck cancer dataset to obtain a hybrid feature matrix; a high-correlation redundant feature removal subunit 1323, used to remove high-correlation redundant features from the hybrid feature matrix to obtain a dimension-reduced feature matrix; and a robust feature filtering subunit 1324, used to perform ICC-based robust feature filtering on the dimension-reduced feature matrix to obtain a fused multimodal feature vector.

[0040] Accordingly, the texture feature weighting subunit 1321 is used to perform subspace projection and scaled dot product operations on the multi-strategy texture feature matrix using a multi-head attention mechanism to obtain a weighted texture feature matrix. It should be understood that since the multi-strategy texture feature matrix incorporates features from different statistical perspectives (such as spatial dependence, grayscale consistency, and gradient changes), these features contribute differently to identifying the heterogeneity of head and neck cancer lesions. Treating all features equally would dilute key pathological indicators with secondary information. Therefore, in the technical solution of this application, a multi-head attention mechanism is further used to perform parallel subspace projection and scaled dot product operations on the matrix, thereby adaptively learning the correlation and importance between features from multiple independent semantic spaces. This allows texture features that contribute significantly to tumor characterization to be given higher weights, suppressing irrelevant or redundant feature responses, and generating a weighted texture feature matrix that focuses on core lesion information.

[0041] In a specific example of this application, the number of attention heads is first set (e.g., h=8), and a corresponding number of query, key, and value weight matrices are constructed. Then, the multi-strategy texture feature matrix is ​​mapped to multiple high-dimensional feature subspaces, and a scaling dot product attention calculation is performed independently in each subspace: the transpose of the query matrix and the key matrix is ​​multiplied by a scaling factor to maintain gradient stability, and then normalized using the Softmax function to obtain the attention coefficient matrix. Next, the value matrix is ​​weighted and summed using this attention coefficient matrix to capture the global dependencies of different feature dimensions in the current subspace. Finally, the outputs of all subspaces are concatenated and linearly projected to reconstruct a weighted texture feature matrix containing rich contextual feature associations.

[0042] Accordingly, the dimension alignment and splicing subunit 1322 is used to perform dimension alignment and splicing between the weighted texture feature matrix and the clinical feature vectors from the multimodal head and neck cancer dataset to obtain a hybrid feature matrix. It should be understood that since simple radiomics features can only reflect the physical morphology and texture attributes of lesions, lacking high-dimensional clinical pathological gold standard information such as TNM staging, gene mutation status, and patient signs, there is a significant semantic gap between the two, making single-modal analysis difficult to comprehensively assess the disease status of head and neck cancer. Therefore, in the technical solution of this application, the dimension alignment and splicing operation of the weighted texture feature matrix and the clinical feature vector is further performed to break down the data barriers between modalities and construct a hybrid feature space containing both imaging and clinical dimensions. This allows the model to generate a complementary hybrid feature matrix during inference, based on both microscopic texture cues and macroscopic clinical priors, significantly improving the accuracy and interpretability of disease prediction.

[0043] In a specific example of this application, the clinical feature vector corresponding to the current image sample is first retrieved from a multimodal head and neck cancer dataset. This vector contains numerically encoded TNM staging, multi-gene mutation detection status (covering 12 genes including TP53, CDKN2A, NOTCH1, FAT1, PIK3CA, HRAS, CCND1, FBXW7, CASP8, AXIN2, EGFR, and BRAF, each encoded with 0 or 1), and patient physiological indicators. Subsequently, to address the dimensionality mismatch between the clinical feature vector and the weighted texture feature matrix (the former is typically a one-dimensional vector, while the latter is a two-dimensional matrix), a broadcasting or copying expansion strategy is employed to copy the one-dimensional clinical feature vector along the row direction of the weighted texture feature matrix, expanding it into feature blocks with the same number of rows as the texture matrix. Finally, a channel-dimensional cascade operation is performed to seamlessly concatenate the expanded clinical feature blocks and the weighted texture feature matrix in the column dimension, so that each row of feature vectors simultaneously contains the weighted texture information and the corresponding clinical pathology information of the sample, thereby forming a high-dimensional hybrid feature matrix.

[0044] Accordingly, the highly correlated redundant feature removal subunit 1323 is used to remove highly correlated redundant features from the hybrid feature matrix to obtain a dimensionality-reduced feature matrix. It should be understood that since the hybrid feature matrix contains a large number of radiomics features and clinicopathological features generated based on different texture strategies, these features often exhibit extremely high linear collinearity (for example, the energy features of gray-level co-occurrence matrices at different time lengths may carry repetitive texture information). This data redundancy not only significantly increases the computational load of subsequent models but may also lead to the curse of dimensionality, making the model overly sensitive to noise during training and reducing its generalization ability. Therefore, in the technical solution of this application, highly correlated redundant feature removal is further performed on the hybrid feature matrix to quantify the linear dependence between features and remove statistically repetitive components. This allows the feature space to be compressed to a more independent and representative dimension, resulting in a dimensionality-reduced feature matrix with higher information density and free from collinearity interference.

[0045] Specifically, in one example of this application, the Pearson correlation coefficient matrix between each feature dimension and other features in the mixed feature matrix is ​​first calculated. This matrix is ​​then traversed to identify all highly correlated feature pairs whose absolute values ​​are greater than a preset threshold (e.g., 0.9). Subsequently, for each pair of features that meet the elimination criteria, the feature dimension with lower correlation to other features or higher single classification ability (AUC value) on the validation set is retained by calculating its mean absolute correlation or other statistical indicators. The other redundant feature dimension is then directly removed from the matrix, resulting in a dimensionality-reduced feature matrix with significantly reduced dimensions.

[0046] In a specific example of this application, the Pearson correlation coefficient between each eigenvector in the mixed feature matrix and all other eigenvectors is first calculated, constructing a symmetric correlation coefficient matrix. Then, a high-correlation removal threshold of 0.9 is set, and the upper triangular elements of the correlation coefficient matrix are iterated. When the absolute value of the correlation coefficient between two feature dimensions (denoted as feature X and feature Y) exceeds 0.9, these two features are determined to have high linear redundancy. At this point, the average correlation coefficient between these two features and all other features is calculated, retaining the feature dimension with the smaller average correlation coefficient (i.e., containing relatively more independent information), and permanently removing the other feature from the matrix. This process is repeated until the correlation coefficients of all remaining feature pairs are below the set threshold. The calculation logic for the correlation coefficients between features during this process is as follows:

[0047] in, The Pearson correlation coefficient represents the relationship between any two eigenvectors X and Y in the mixed feature matrix. The total number of samples; and These represent the specific values ​​of the i-th sample in the two feature dimensions, respectively; and These represent the mean values ​​of these two feature dimensions across all samples. This formula is used in scenarios to accurately quantify the linear and synchronous changing trends between image texture and clinical indicators. When the calculation result approaches 1, it indicates that one feature can be almost linearly represented by the other, thus providing a mathematical basis for eliminating redundancy.

[0048] Specifically, the Pearson correlation coefficient and variance-based screening strategy used in the first embodiment above for removing highly correlated redundant features has significant logical flaws and performance bottlenecks in the specific medical scenario of head and neck cancer prediction. First, the limitations of the linear assumption severely restrict the depth of feature mining. The original scheme can only capture linear statistical associations between features. However, in complex biomedical data, there is often a highly nonlinear coupling relationship between image texture features (such as grayscale entropy reflecting tumor heterogeneity) and clinicopathological indicators (such as TNM staging or multi-gene mutation status reflecting gene phenotype (including 12 key genes such as TP53 and CDKN2A)). For example, mutations in certain key genes may only significantly affect the image phenotype within a specific tumor volume or within an interval that can be activated by a specific texture threshold. This non-monotonic, nonlinear accompanying effect is easily ignored by the Pearson coefficient, leading to nonlinear features with important clinical indicative significance being misjudged as low-correlation and retained, or potentially redundant information being lost due to false linear inconsistencies. Secondly, the logical fallacy of modality equality ignores the prior level of medical data. Furthermore, the first embodiment, when processing highly correlated feature pairs, decides whether to retain or discard them solely based on the magnitude of numerical variance, blindly placing radiomics data and clinical text data on an equal footing. In reality, clinical data is typically a high-semantic, low-noise, strong modality validated by pathological gold standards, while imaging data is a low-semantic, weak modality containing significant acquisition noise. If the two are simply linearly correlated, imaging features are often merely digital projections of clinical representations. Mechanically retaining imaging features with larger variance while discarding discrete clinical indicators with smaller variance effectively discards high-confidence true values ​​and retains surrogate values ​​carrying noise; conversely, directly discarding imaging features may result in the loss of fine-grained lesion texture information hidden in the residuals, which is not covered by clinical indicators. Therefore, the first embodiment cannot effectively extract truly incremental and complementary components from imaging data while ensuring the accuracy of prior clinical data.

[0049] To address the aforementioned issues, a cross-modal nonlinear orthogonal decoupling screening process is employed in the second embodiment of this application. Specifically, in the preferred technical solution of this application, a cross-modal nonlinear orthogonal decoupling and complementary information gain screening strategy is used to overcome the constraints of the linear assumption, establish the dominant role of clinical priors, and mathematically force the removal of redundant components in the image that serve only as echoes of clinical indicators. This allows for the precise preservation of unique, fine-grained lesion texture residuals in the image data that cannot be explained by clinical indicators, constructing a dimensionality-reduced feature matrix with maximized semantic complementarity.

[0050] Specifically, in another specific example of this application, the highly correlated redundant feature removal subunit is used to: perform a generalized mutual information metric based on kernel density estimation on the mixed feature matrix to obtain a nonlinear dependency matrix; perform image-clinical asymmetric decoupling on the nonlinear dependency matrix to obtain a decoupled residual feature set; and perform complementary information gain filtering on the decoupled residual feature set and the mixed feature matrix to obtain a dimensionality-reduced feature matrix.

[0051] More specifically, a generalized mutual information measure based on kernel density estimation is applied to the mixed feature matrix to obtain a nonlinear dependency matrix. It should be understood that the traditional Pearson correlation coefficient can only capture linear statistical associations between features, which presents a linear trap when processing complex biomedical data. Specifically, there is often a highly nonlinear coupling relationship between image texture features of head and neck cancer (such as gray entropy reflecting tumor heterogeneity) and clinicopathological indicators (such as TNM staging or polygenic mutation status reflecting gene phenotype). Simple linear measures easily overlook this non-monotonic, nonlinear accompanying effect (for example, mutations in certain key genes may only significantly affect the image phenotype within a specific tumor volume or within a range that can be activated by a specific texture threshold), leading to the retention of nonlinear latent features with important clinical indicative significance as having low correlation. Therefore, in the technical solution of this application, a generalized mutual information measure strategy based on kernel density estimation is further introduced to analyze the mixed feature matrix, thereby overcoming the constraints of linear measures and capturing and quantifying the complex nonlinear dependencies between image and clinical features at the information theory level. In this way, we can accurately identify those latent feature pairs that appear unrelated statistically but are tightly coupled in biological mechanisms, thereby constructing a nonlinear dependency matrix that precisely reflects the nonlinear coupling relationship between imaging and clinical practice. This provides precise topological guidance for subsequent targeted decoupling and avoids missing nonlinear redundant features.

[0052] In a specific example of this application, the conventional linear metric method is abandoned first, and instead, for each pair of eigenvectors (denoted as eigenvectors) in the mixed feature matrix, and The mutual information between features is calculated. In this process, Gaussian kernel density estimation is used to accurately approximate the joint probability distribution and marginal probability distribution of continuous imaging features and discrete / continuous clinical features based on the data samples, without requiring any prior assumptions about the normality of the data distribution. Based on the approximate probability distribution, the generalized mutual information value between features is calculated, and this value is used to fill the matrix, generating a nonlinear dependency matrix describing the strength of the nonlinear association between features. and The formula for calculating the generalized mutual information value between them is as follows:

[0053]

[0054] in, Representation of features and The generalized mutual information value between them For bandwidth-based Kernel density estimation function: and Let be any two feature vectors to be analyzed; This represents the joint probability density estimate of two features obtained based on kernel density estimation; , These represent their respective marginal probability density estimates; The total number of samples; The bandwidth parameter K is used to control the smoothness. The Gaussian kernel function; For the first The feature observations of each sample. By introducing non-parametric kernel density estimation, it is possible to mathematically quantify any complex functional relationship (including nonlinearity, multimodal distribution, etc.) between head and neck cancer imaging features and clinical indicators without losing details of data distribution. This ensures that the generated nonlinear dependency matrix can truly reflect the tight biological coupling mechanism, rather than just the superficial linear correlation.

[0055] More specifically, an image-clinical asymmetric decoupling is performed on the nonlinear dependency matrix to obtain a decoupled residual feature set. It should be understood that traditional feature fusion and selection strategies often fall into the logical fallacy of modality equality, incorrectly placing pathologically validated high-semantic clinical data (strong modality) on an equal footing with low-semantic image data (weak modality) containing device noise. If selection is based solely on numerical variance, it is highly likely that clinical features, which are considered the gold standard, will be discarded to retain image features with larger variance, or image features will be directly eliminated when the two features are highly correlated, resulting in the dual risk of missing the true value or overlooking details. Therefore, in the technical solution of this application, image-clinical asymmetric orthogonal decoupling is further performed based on the generated nonlinear dependency matrix to correct the modality equality defect and implement a fusion paradigm dominated by clinical priors and complementary image residuals. In this way, redundant projections that serve only as echoes of clinical indicators in image features can be forcibly removed at the mathematical level. This not only fully preserves the baseline explanatory power of clinical features as a strong modality, but also successfully extracts unique fine-grained residual information from weak modality image data that cannot be explained by clinical indicators, thus achieving true cross-modal incremental complementarity and decoupling.

[0056] In a specific example of this application, the nonlinear dependency matrix generated in the preceding steps is first traversed, and a high dependency discrimination threshold (such as mutual information value > 1) is set. When a certain image feature vector is identified (denoted as...) ) and a certain clinical feature vector (denoted as ) When there is a high-intensity nonlinear dependency exceeding a threshold between the two features, it is determined that there is redundancy at the information level. In this case, instead of directly deleting the image features, an asymmetric orthogonal projection operation is performed on the image features. Specifically, the image feature vector... Forced projection to clinical feature vector In the spanned orthogonal complement space, the component of the image feature in the clinical direction is calculated and subtracted from the original image feature, thus ensuring that the remaining residual remains geometrically strictly orthogonal to the clinical feature. The residual feature vector after orthogonal decoupling is represented as follows:

[0057] in, This represents the image residual feature vector after decoupling, from which clinically relevant information has been removed; The original image feature vector; For its paired highly dependent clinical feature vector; This represents the projection vector of the image features onto the clinical feature direction; Represents the dot product of vectors; Represents the vector norm; To prevent the use of tiny constants with a denominator of zero, this step is equivalent to a deep signal purification process on the head and neck cancer imaging data at the mathematical level. By forcibly subtracting projection components that are collinear with clinical indicators, redundant information in the images that only reflects macroscopic pathology (such as TNM staging) is completely eliminated, thus ensuring the quality of the final residual feature vector. It contains only incremental information unique to the imaging modality, reflecting the heterogeneity of lesion microtexture, and which cannot be replaced by existing clinical indicators.

[0058] More specifically, complementary information gain filtering is performed on the decoupled residual feature set and the hybrid feature matrix to obtain a dimensionality-reduced feature matrix. It should be understood that although the residual features after asymmetric orthogonal decoupling theoretically remove clinically relevant information, not all residuals contain valuable pathological signals. If the majority of the variance of a certain image feature can be explained by clinical features, then the remaining small residuals are very likely to originate only from imaging thermal noise or reconstruction artifacts of the CT / MRI equipment, rather than true microscopic texture heterogeneity. Forcibly introducing these low signal-to-noise ratio pure noise residuals into the model not only fails to provide additional predictive gain but also interferes with the model's learning of key features, reducing generalization ability. Therefore, in the technical solution of this application, complementary information gain filtering based on energy ratio is further performed on the decoupled residual feature set and the hybrid feature matrix to quantify the actual information content of the residual features and prevent the introduction of pure noise features. This ensures that each dimension of the features entering the final model has a unique predictive contribution, thereby constructing a concise and efficient dimensionality-reduced feature matrix. While minimizing the feature dimensionality, it maximizes the joint information entropy of multimodal data, significantly improving the robustness and accuracy of the head and neck cancer prediction model.

[0059] In a specific example of this application, for each feature vector in the decoupled residual feature set generated in the previous step, a value assessment based on the variance ratio is performed. Specifically, the original image feature vector corresponding to the residual feature (i.e., the feature before decoupling) is retrieved from the mixed feature matrix, and the statistical variances of the residual feature vector and the original feature vector are calculated respectively. The ratio between the two is then calculated (i.e., the energy ratio). This energy ratio is used as an intuitive quantitative indicator to reflect the proportion of unique information remaining in the image features after removing clinical commonalities. If the energy ratio is lower than a preset complementary information retention threshold (e.g., a set threshold is used), the assessment is performed. If the energy ratio is 0.1, it indicates that over 90% of the information in the image feature is already included in the clinical features, and the remaining part is likely acquisition noise. Therefore, it is judged as a feature with no gain and is directly removed. Conversely, if the energy ratio is higher than the threshold, it is considered that the residual feature carries complementary information that cannot be provided by clinical data (such as micro-texture heterogeneity), and it is added to the feature pool as a valid feature after normalization. The update logic of the feature preservation discriminant function and the dimensionality reduction feature matrix is ​​as follows:

[0060]

[0061] in, Information efficiency ratio representing residual characteristics; This indicates the operation for calculating variance; The residual eigenvectors after decoupling; The original image feature vector before decoupling; A threshold is reserved for complementary information; This is a normalization function; The final output consists of complementary features retained in the dimensionality-reduced feature matrix. This screening mechanism accurately distinguishes between noise residuals and information residuals, ensuring that the retained features are not only mathematically independent but also have substantial biological increments, thus achieving high-precision feature dimensionality reduction.

[0062] This approach aims to overcome the challenges of modality confusion and nonlinear blind spots in multimodal medical data processing using traditional linear screening mechanisms. By introducing a mutual information metric based on kernel density estimation, this invention can accurately capture and quantify the complex nonlinear coupling relationship between radiomics and clinical indicators, effectively avoiding the omission of nonlinear features caused by the Pearson coefficient. Furthermore, utilizing asymmetric orthogonal decoupling and energy ratio screening strategies, this invention creatively establishes a feature fusion paradigm dominated by clinical priors and complementary image residuals. This paradigm can mathematically rigorously remove redundant components in image features explained by clinical indicators, while accurately preserving the unique, irreplaceable fine-grained pathological texture information (residuals) in the image data. Ultimately, this mechanism achieves a qualitative leap from simple deduplication to orthogonal purification, significantly reducing the redundancy and noise interference in the feature space, and ensuring that the fused feature set has maximized complementary increments at the semantic level, thereby significantly improving the generalization ability and diagnostic accuracy of head and neck cancer prediction models when facing complex clinical heterogeneity.

[0063] Accordingly, the robust feature screening subunit 1324 is used to perform robust feature screening based on intraclass correlation coefficient (ICC) on the dimensionality-reduced feature matrix to obtain a fused multimodal feature vector. It should be understood that the acquisition process of medical image data is highly susceptible to interference from external factors such as equipment parameters, scanning time, patient position, and segmentation differences. This means that while some extracted radiomics features may exhibit high discriminative power in a single measurement, they lack stability in repeated observations or cross-center validation. This poor reproducibility directly leads to a significant decrease in the generalization ability of the prediction model in clinical applications. Therefore, in the technical solution of this application, robust feature screening based on intraclass correlation coefficient (ICC) is further performed on the dimensionality-reduced feature matrix to quantitatively evaluate the anti-interference ability and consistency of each feature under different observation conditions. This allows for the strict filtering out of unstable features sensitive to noise from a massive amount of features, retaining only robust features with high reproducibility, thereby generating a stable and reliable fused multimodal feature vector.

[0064] In a specific example of this application, firstly, based on repeated scan data from a multimodal head and neck cancer dataset or retest data generated by geometrically deforming (e.g., translation, rotation) and perturbing the region of interest with grayscale, the intra-group correlation coefficient of each feature in the dimensionality-reduced feature matrix is ​​calculated. This coefficient measures the stability of the feature by analyzing the ratio of between-group variance to within-group variance, and its value ranges from 0 to 1. Subsequently, a robustness screening threshold of 0.75 is set, and the ICC values ​​of all features to be screened are iterated. Features with ICC values ​​below 0.75 are considered unstable and are removed, retaining only features with ICC values ​​greater than 0.75 (i.e., high intra-group consistency). Finally, the remaining highly robust features after screening are integrated and encapsulated as input data for the model, thus obtaining the fused multimodal feature vector.

[0065] In the aforementioned head and neck tumor intraoperative treatment support system 100 based on deep transfer reinforcement learning, the model training and optimization module 140 is used to train and optimize a source domain deep reinforcement learning model based on the acquired public source domain dataset to obtain pre-trained model parameters. It should be understood that due to the difficulty and high cost of collecting clinically labeled data for head and neck cancer (especially image data with gene mutation detection results), directly training a deep neural network from scratch on a small sample of target domain data is prone to getting stuck in local optima or causing severe overfitting, and the model lacks a basic understanding of general anatomical structures. Therefore, in the technical solution of this application, a public source domain dataset is used for source domain pre-training of deep reinforcement learning. The model is trained on public samples to develop a general perception ability of the correlation between image texture features and gene mutation status in head and neck squamous cell carcinoma, enabling the model to learn a general feature extraction strategy for head and neck lesion image phenotypes. In this way, pre-trained model parameters containing general feature extraction capabilities can be generated. Subsequently, the parameters of the pre-trained model can be transferred to the target task network, providing a high starting point for fine-tuning the transfer on image data with gene mutation detection annotations within the hospital, significantly accelerating model convergence and improving generalization performance.

[0066] Specifically, in the embodiments of this application, the model training optimization module is used to: interactively encapsulate the public source domain dataset using the definitions of state space, action space, and reward function to obtain a source domain reinforcement learning environment instance; construct a function approximator using a deep neural network and set the target network update frequency to obtain an initialized deep Q network; perform iterative training of the deep Q network based on a greedy strategy to obtain a set of model parameters to be optimized; and perform Bayesian hyperparameter search and validation set evaluation on the set of model parameters to be optimized to obtain pre-trained model parameters.

[0067] Among them, the parameters of the source domain pre-trained model are transferred through transfer learning. Mapping to target task network parameters The initialization process is described, and the parameter update range is controlled through a hierarchical freezing strategy. Specifically, the parameter migration and fine-tuning process can be formally represented as follows:

[0068] in, Indicates the target task network number 1 Layer parameters; These are the parameters of the corresponding layer in the pre-trained model from the source domain. This is the threshold for the number of frozen layers; To fine-tune the learning rate; The supervised loss function for the target domain (such as multi-label cross-entropy loss); For the loss function relative to the first The gradient of layer parameters. This formula clarifies the mathematical mechanism by which the general perceptual ability of the source domain is transformed into the task-specific ability of the target domain under the transfer reinforcement learning paradigm. It ensures that the shallow general feature extraction ability is fully preserved, while the deep semantic representation can be adaptively adjusted for the head and neck cancer multi-gene mutation prediction task.

[0069] Specifically, the selection of the freezing layer threshold is based on the hierarchical semantic division of labor in deep neural networks. Shallow networks (the first few convolutional blocks) mainly encode general low-level visual features such as edges and textures. These features are highly shared between the source domain (lesion segmentation) and the target domain (gene mutation prediction) and should be frozen and preserved. Deep networks encode high-level semantic features that are strongly related to the task and need to be unlocked to adapt to the distribution of the target domain.

[0070] In a specific example of this application, the threshold for the number of frozen layers can be determined through the following steps: First, count the total number of layers in the pre-trained network and calculate candidate values ​​for the freezing ratio (e.g., 30%, 50%, 70%); evaluate the model convergence speed and predicted AUC under different freezing ratios on the target domain validation set, and select the number of frozen layers with the best validation set performance and stable convergence of training loss as the final threshold for the number of frozen layers; for a typical configuration with ResNet-50 as the backbone network, the first 2 to 3 residual blocks (i.e., The last two residual blocks and the fully connected layer are retained for fine-tuning.

[0071] Specifically, the fine-tuning learning rate must be significantly smaller than the initial learning rate during the source domain pre-training phase to avoid destructive updates (i.e., "catastrophic forgetting") to the converged pre-trained weights. At the same time, the learning rate cannot be too small, otherwise the deep parameters will not be able to achieve effective distribution adaptation on the limited target domain samples.

[0072] In a specific example of this application, the following strategy is used to determine the fine-tuning learning rate: First, the source domain pre-training learning rate is used. (usually) Using as a baseline, the fine-tuning learning rate is set to its value. to Then, Bayesian hyperparameter optimization is introduced (consistent with the Bayesian search strategy in the model training optimization module). Log-scale grid search is performed on the fine-tuning learning rate on the validation set in the target domain. The mean of the multi-gene prediction AUC on the validation set is used as the evaluation metric to select the optimal value. A hierarchical learning rate strategy can be further adopted, setting an increasing learning rate for different depth levels of unlocking (smaller values ​​for shallow unlocking parts and larger values ​​for the deepest layers) to achieve more refined parameter adaptation.

[0073] In a specific example of this application, a publicly available source domain dataset containing multimodal CT images of head and neck cancer and corresponding gene mutation detection results is first loaded. This publicly available dataset has standardized gene mutation state annotations, which are used to train the model's general perception ability of the association between CT image phenotypes and gene mutations in head and neck squamous cell carcinoma. The publicly available source domain dataset is encapsulated into a source domain reinforcement learning environment instance conforming to reinforcement learning specifications. The current image feature block is defined as the state space, feature selection or classification decision as the action space, and the increment based on prediction accuracy as the reward function. Subsequently, a Q-value function approximator is constructed using a deep convolutional neural network based on the PyTorch framework. The network weights are initialized, and a target network update frequency with a fixed step size is set to generate an initialized deep Q-network. Next, multi-round interactive training based on an ε-greedy strategy is performed in the source domain environment, allowing the agent to achieve a balance between exploration and exploitation. By maximizing the cumulative reward and continuously updating the network weights, the model learns the general feature extraction ability of the association between texture features of head and neck squamous cell carcinoma CT images and gene mutation states, resulting in a preliminarily converged set of model parameters to be optimized. Finally, for key hyperparameters such as learning rate and discount factor, a Bayesian optimization algorithm is introduced for global optimization search. Through repeated iterations and verification on the Gaussian process surrogate model, the hyperparameter combination that performs best on the source domain segmentation task is determined, and the network weights in this state are solidified and exported to obtain the final pre-trained model parameters.

[0074] Figure 5 This is a schematic diagram of the training set loss and validation set loss curves in the tumor segmentation training process of the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application. Figure 6 This is a schematic diagram showing the curves of the training set Dice coefficients and the validation set Dice coefficients in the tumor segmentation training process of the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application. Figure 7 This is a schematic diagram showing the intersection-union ratio of the training set and the intersection-union ratio of the validation set in the tumor segmentation training process of the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application. Figure 8 This is a schematic diagram showing the training set accuracy versus validation set accuracy curves of the tumor segmentation training process in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application. Figure 9 This is a schematic diagram of the training set recall and validation set recall curves in the tumor segmentation training process of the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to an embodiment of this application. Figure 10 This is a schematic diagram comparing the training set Dice coefficients and the validation set Dice coefficients during the tumor segmentation training process in a head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, according to an embodiment of this application. Figures 5 to 10 As shown, this diagram illustrates the training and validation results obtained during 120 tumor training rounds in the source domain pre-training process. In this source domain pre-training process, tumor region annotations from publicly available source domain data are used to train the source domain deep reinforcement learning model, enabling it to learn the common anatomical structures of head and neck tumors, the spatial location of tumor regions, and lesion boundary representations. The training is then optimized to form the source domain pre-trained model and its corresponding parameters. Figure 5 The changes in training set loss and validation set loss during tumor segmentation training with training rounds are shown. It can be seen that both losses generally decrease continuously and gradually stabilize in the later stages of training, decreasing to about 0.17 by the 120th training round. Figure 6 The results show how the Dice coefficients of the training set and validation set change with the training rounds. Both metrics generally continue to rise and eventually stabilize at about 0.76, which is close to 0.8 overall. Figure 7 The results show the changes in the intersection-union ratio of the training set and the intersection-union ratio of the validation set with the training rounds. Both eventually stabilized above 0.6, indicating that there is a relatively stable spatial overlap relationship between the predicted segmentation region and the corresponding tumor region annotation. Figure 8 The results show how the training set precision and validation set precision change with training rounds. Both metrics stabilize at around 0.76 in the later stages of training, indicating that the model has a certain ability to control missegmented regions. Figure 9 The results show how the training set recall and validation set recall change with training rounds. Both metrics remain stable at around 0.76 in the later stages of training, indicating that the model has a certain control over missed segmentation regions. Figure 10 The training set Dice coefficient and validation set Dice coefficient for each training round are shown in bar chart form. The overall trends of the two sets of indicators are basically consistent, and no significant and persistent deviations were observed in the later stages of training.

[0075] comprehensive Figures 5 to 10It can be seen that as the number of training rounds increases, the loss value in the tumor segmentation training process gradually decreases, while the Dice coefficient, intersection-over-union ratio (IoU), precision, and recall gradually increase, stabilizing around the 120th training round. This indicates that the source domain pre-training process has basically converged, and the source domain pre-trained model exhibits relatively consistent tumor region segmentation performance on both the training and validation sets. In the delineation of gross tumor target areas and clinical target areas for head and neck radiotherapy, the consistency of manually delineated areas by different doctors is typically around 0.75, influenced by factors such as clinical experience, lesion boundary judgment, and image quality. Compared to this reference level, the Dice coefficients of both the training and validation sets in this embodiment are close to 0.8, and the IoU is higher than 0.6, reflecting that the source domain pre-trained model has relatively stable region overlap capability in tumor segmentation tasks. Simultaneously, the precision and recall are both close to 0.8, indicating that it maintains a relatively balanced control over missegmentation and missed segmentation. The training results above demonstrate that the source domain pre-trained model has acquired preliminary representation capabilities of common anatomical information and lesion boundary features of head and neck tumors. This provides a foundation for the auxiliary identification of tumor core regions and invasive boundaries, and helps reduce differences in target area delineation among different physicians. It is worth noting that the tumor segmentation training described above is used to evaluate and enhance the general image representation capabilities of the source domain pre-trained model; its segmentation results are not the final output of this system. In subsequent processing, the obtained pre-trained model parameters are transferred to the target task network, and a head and neck cancer prediction model is obtained through target domain fine-tuning and pipeline integration. The final output of this system remains the mutation probability prediction results for each target gene.

[0076] In the aforementioned head and neck tumor intraoperative treatment support system 100 based on deep transfer reinforcement learning, the transfer learning fine-tuning module 150 is used to initialize the target task network based on pre-trained model parameters, and to fine-tune the target task network using fused multimodal feature vectors as target domain data to obtain a head and neck cancer prediction model. It should be understood that, due to the relative scarcity of target domain data for head and neck cancer (especially clinical cases containing complete gene mutation annotations and multimodal imaging), and the significant heterogeneity among individual patients, directly using randomly initialized networks for training not only results in slow model convergence but also a high risk of overfitting due to insufficient data, making it difficult to capture deep pathological semantics. Furthermore, when processing high-dimensional structured features (such as fused multimodal features), the decision interpretability and feature selection capabilities of simple deep neural networks are often inferior to those of tree-based ensemble models. Therefore, the technical solution of this application further adopts a hybrid modeling strategy of transfer learning and ensemble enhancement. It initializes the target task network using pre-trained parameters from the source domain, adapts the target domain distribution using hierarchical freezing and fine-tuning techniques, and re-optimizes features using a gradient boosting tree. This balances the generality of deep feature extraction with the robustness of specific task-specific decision-making. In this way, it can fully utilize the big data knowledge of the source domain and the multimodal details of the target domain to construct a head and neck cancer prediction model with strong generalization ability and accurate prediction of gene mutations.

[0077] Specifically, in this embodiment, the transfer learning fine-tuning module adopts a paradigm of transferring general perception from the source domain to precise prediction in the target domain. During the source domain pre-training stage, a model is trained using a public dataset to acquire general perception capabilities regarding the association between CT image phenotypes and gene mutations in head and neck squamous cell carcinoma. The parameters of this pre-trained model are then transferred to the target task network, thereby enabling the target task network to inherit a deep understanding of the association between texture features of head and neck CT images and gene mutations. The transfer learning fine-tuning module is used to: load weights onto the target network using the pre-trained model parameters and perform shallow freezing and deep unlocking operations to obtain an initialized target task network; perform supervised fine-tuning and backpropagation updates on the initialized target task network using fused multimodal feature vectors to obtain a deep semantic feature set; optimize the feature weights of the deep semantic feature set based on gradient boosting trees to obtain an optimized tree ensemble classifier; and integrate the fine-tuned target task network and the tree ensemble classifier into a pipeline to obtain a head and neck cancer prediction model.

[0078] Accordingly, the target network is weighted using pre-trained model parameters, and shallow freezing and deep unlocking operations are performed to obtain the initialized target task network. It should be understood that the shallow structure of a deep neural network is mainly responsible for extracting general visual features such as edges, corners, and basic textures. These features are highly shared across different medical imaging tasks (source and target domains), while the deep structure is responsible for extracting high-level semantic features strongly correlated with specific disease categories (such as polygenic mutation patterns in head and neck cancer). If all parameters are fully updated during target task training, it will not only destroy the general visual capabilities already learned by the pre-trained model, but also cause the model parameters to oscillate in the huge search space due to the small amount of data in the target domain, making rapid convergence impossible. Therefore, in the technical solution of this application, pre-trained model parameters are used for initialization, combined with a parameter control strategy of shallow freezing and deep unlocking, to transfer source domain knowledge to the target network without loss, and adjust the high-level semantic representation only for specific tasks. This significantly reduces the training difficulty and time cost of the target task, while effectively preventing overfitting under small sample conditions, resulting in better initial performance.

[0079] More specifically, in a concrete example of this application, a deep convolutional neural network with the exact same topology as the source domain pre-training stage is first instantiated as the target network architecture. Then, the pre-trained model parameters (including convolutional kernel weights and bias terms) are read from the storage medium and loaded layer by layer into the corresponding layers of the target network using a dictionary mapping mechanism, completing the warm-start initialization of the weights. Next, the hierarchical structure of the target network is traversed, and the parameter gradient properties of the first N convolutional layers (e.g., the first three residual blocks or feature extraction stages) are set to non-trainable, i.e., shallow parameter freezing is performed to keep their values ​​unchanged during subsequent backpropagation; simultaneously, the parameter gradient properties of the deep convolutional layers and fully connected layers at the back end of the network are set to trainable, i.e., deep parameter unlocking is performed. Through this differentiated parameter configuration, an initialized target task network with both general feature extraction capabilities and task adaptability is constructed, preparing for subsequent fine-tuning training.

[0080] Accordingly, the initialized target task network is subjected to supervised fine-tuning and backpropagation updates using fused multimodal feature vectors to obtain a set of deep semantic features. It should be understood that while the initialized target task network inherits the general feature extraction capabilities of the source domain, its deep parameter distribution is primarily adapted to source domain tasks (such as general organ segmentation) and has not yet been modified for the specific pathological features of head and neck cancer (such as image texture patterns associated with multi-gene mutations). Furthermore, the input fused multimodal feature vectors differ significantly in distribution characteristics from the pure image data of the source domain. Therefore, in the technical solution of this application, the fused multimodal feature vectors are further used as input data for the target domain. Under the supervision of real clinical labels (such as gene mutation positive / negative), the network's unfrozen deep parameters are fine-tuned and updated via backpropagation to drive the network to learn the high-dimensional semantic representation specific to head and neck cancer. This corrects the distribution bias between the source and target domains, enabling the network to extract deep semantic features with high discriminative power for tumor evolution and mutation prediction from the multimodal input.

[0081] More specifically, in a concrete example of this application, a clinically collected and constructed multimodal head and neck cancer dataset is first divided into a training set and a validation set. The fused multimodal feature vectors are extracted as input data, along with corresponding multi-gene mutation status labels (covering 12 key genes, each encoded with 0 or 1) as supervision labels. Subsequently, fine-tuning of the training hyperparameters is configured (setting a small learning rate, such as 1e-4, to avoid disrupting pre-trained weights). The feature vectors are then input into the initialized target task network for forward propagation, and the loss function value (e.g., cross-entropy loss) between the predicted output and the true label is calculated. Next, a backpropagation algorithm is executed using stochastic gradient descent or the Adam optimizer to calculate the gradient of the loss function relative to the deep trainable parameters of the network, and the weights of only the unfrozen layers are updated accordingly. After several rounds of iterative training until the validation set loss converges, the trained network is used as a feature extractor, extracting the output vector of the penultimate layer before the fully connected classification layer. This yields a highly compressed deep semantic feature set rich in pathological information, which is then used by the subsequent tree ensemble model.

[0082] Accordingly, the feature weights of the deep semantic feature set are optimized using gradient boosting trees to obtain an optimized tree ensemble classifier. It is understandable that while the deep semantic feature set extracted by deep neural networks possesses high abstraction and non-linear expressive power, its feature dimensionality is often still high and lacks explicit interpretability. Directly using simple fully connected layers for classification may fail to fully utilize the combined interaction advantages between features, and it is also impossible to accurately measure the importance weight of each feature dimension. Gradient boosting trees, as a strong ensemble learner, excel at handling structured high-dimensional features and can automatically perform feature selection and weight optimization through the tree splitting process. Therefore, in the technical solution of this application, the gradient boosting tree algorithm is further used to perform secondary modeling and optimization of the deep semantic feature set to construct a classification decision head with stronger robustness and feature recognition capabilities. This endows the model with adaptive attention to key pathological features, further improving the prediction accuracy for complex tasks such as gene mutations in head and neck cancer.

[0083] More specifically, in a concrete example of this application, the extracted deep semantic feature set is first used as the input feature matrix, and the corresponding gene mutation label is used as the target variable to construct an XGBoost classifier model framework. Then, parameters such as the number of trees, maximum depth, and learning rate are set, and the iterative training process of the ensemble model is initiated. In each iteration, the algorithm constructs a new decision tree based on the prediction residuals of the previous round. During the node splitting process of the decision tree, all deep feature dimensions are traversed, and the splitting contribution of each feature is calculated based on the information gain or Gini index decrease. The weights of features in the decision path are automatically adjusted accordingly, ensuring that key features that contribute significantly to the classification result are used preferentially and frequently. After multiple rounds of accumulation and optimization of the additive model, a strong classifier composed of multiple weighted decision trees is finally generated, i.e., the optimized tree ensemble classifier. This classifier internally has the optimal feature combination and decision logic embedded, enabling high-precision category determination of the input deep features.

[0084] Accordingly, the fine-tuned target task network and the tree ensemble classifier are integrated into a pipeline to obtain a head and neck cancer prediction model. It should be understood that since the fine-tuned target task network and the optimized tree ensemble classifier are trained independently during the training phase—the former generating deep semantic representations as a powerful nonlinear feature extractor, and the latter optimizing feature weights as a robust decision-maker—it is impossible to achieve automated reasoning from raw input to final result if they are separated, and it is difficult to deploy them uniformly in clinical settings. Therefore, in the technical solution of this application, software-level pipeline integration is further performed on these two independently trained modules to construct an end-to-end hybrid architecture inference system. This seamlessly connects the perceptual capabilities of deep learning with the decision-making capabilities of ensemble learning, forming a fully functional head and neck cancer prediction model that directly supports automated output from multimodal input to disease prediction results.

[0085] More specifically, in a concrete example of this application, the fine-tuned model structure and weight parameters of the target task network are first loaded, its original fully connected classification layer is removed, and it is solidified into a feature extraction module capable of outputting a set of deep semantic features. Simultaneously, a trained tree ensemble classifier (such as a saved XGBoost model file) is loaded and used as the backend decision module. Subsequently, an inference pipeline script is written, defining the data flow path: the input data first undergoes forward propagation computation through the feature extraction module, outputting a high-dimensional feature vector; this vector is then automatically passed to the decision module as input features for gradient boosting tree path determination. Finally, the above-mentioned concatenation logic, network parameters, and tree model file are uniformly encapsulated into a callable model object or executable file, thus obtaining the final deployable head and neck cancer prediction model.

[0086] In the aforementioned head and neck tumor intraoperative treatment support system 100 based on deep transfer reinforcement learning, the gene mutation prediction module 160 is used to input the fusion features to be predicted into the head and neck cancer prediction model to predict the probability of multiple gene mutations, thereby obtaining the prediction results of the mutation probabilities of each gene. It should be understood that clinical decision-making regarding the diagnosis and treatment of head and neck cancer not only relies on the current pathological state (covering the mutation probabilities of 12 key genes including TP53, CDKN2A, NOTCH1, FAT1, PIK3CA, HRAS, CCND1, FBXW7, CASP8, AXIN2, EGFR, and BRAF), but also urgently requires predicting the future development trend of the disease (such as changes in tumor volume or recurrence risk). A single static classification result cannot meet this need for full-cycle management encompassing diagnosis and prognosis. Therefore, in the technical solution of this application, the head and neck cancer prediction model is further utilized to perform multi-task parallel inference calculations, simultaneously performing binary classification of gene mutations and temporal extrapolation of disease evolution, thereby maximizing the mining of static molecular phenotypic information and dynamic biological behavior patterns contained in the fusion features. This allows doctors to receive a comprehensive auxiliary report that integrates precise medication guidance and long-term prognosis assessment, enabling a three-dimensional and quantitative analysis of the patient's condition.

[0087] Specifically, in the embodiments of this application, the gene mutation prediction module includes: a forward inference unit, used to perform forward inference on the fusion features to be predicted using a head and neck cancer prediction model to obtain the multi-gene classification log odds; and a mutation state prediction unit, used to perform probability mapping on the multi-gene classification log odds to obtain the mutation probability prediction results of each gene.

[0088] Accordingly, the forward inference unit is used to perform forward inference on the fusion features to be predicted using the head and neck cancer prediction model to obtain the multi-gene classification log odds. It should be understood that since the diagnosis and treatment of head and neck cancer involves the precise determination of the mutation status of multiple genes (covering 12 key genes including TP53, CDKN2A, NOTCH1, FAT1, PIK3CA, HRAS, CCND1, FBXW7, CASP8, AXIN2, EGFR, and BRAF), constructing an independent inference model for each gene would not only multiply the inference time and computational resource consumption, but also prevent the sharing of underlying multimodal feature interaction information. Therefore, in the technical solution of this application, the deep semantic features shared within the head and neck cancer prediction model are further utilized to simultaneously output the classification log odds of 12 target genes in a single forward inference, thereby efficiently obtaining static indicators describing the molecular phenotype of multiple genes in a single inference process. This significantly improves the system's response efficiency and ensures that the prediction tasks of each gene achieve information sharing and mutual enhancement at the feature level.

[0089] More specifically, in a concrete example of this application, the fused features to be predicted are first received as an input vector and fed into the backbone network of the head and neck cancer prediction model. After deep feature extraction, the data stream enters the classification prediction head, which consists of a fully connected layer and an XGBoost ensemble classifier. This head is used to mine the texture and clinical indicators highly correlated with the expression patterns of each key gene in the features. For 12 target genes, including TP53, CDKN2A, NOTCH1, FAT1, PIK3CA, HRAS, CCND1, FBXW7, CASP8, AXIN2, EGFR, and BRAF, the unnormalized classification log odds are calculated and output. Through this single-inference architecture with shared features, the model completes the parallel analysis of the patient's multi-gene mutation status in milliseconds.

[0090] Accordingly, the mutation state prediction unit is used to perform probability mapping on the multi-gene classification log odds to obtain the mutation probability prediction results for each gene. It should be understood that since the classification log odds output by the model's classification branch for each gene are usually unbounded continuous real numbers, their physical meaning is not intuitive for clinicians and cannot be directly used as a basis for targeted therapy decisions. Clinical practice requires mutation probability values ​​for each gene with clear statistical significance. Therefore, in the technical solution of this application, a nonlinear probability mapping is further performed on the classification log odds of each gene to transform the abstract model score into a mutation probability description that conforms to clinical diagnostic criteria. This provides doctors with an intuitive multi-gene mutation risk assessment, clearly indicating the mutation probability of each molecular target in the patient, thereby assisting in the development of personalized precision medicine plans.

[0091] First, the sigmoid activation function is used to numerically map the classification log odds of each gene, compressing them into a closed interval between 0 and 1 to obtain the posterior probability value representing the likelihood of mutation for each gene. The system outputs corresponding mutation probability values ​​for 12 key genes (TP53, CDKN2A, NOTCH1, FAT1, PIK3CA, HRAS, CCND1, FBXW7, CASP8, AXIN2, EGFR, and BRAF). These probability values ​​directly reflect the risk level of mutation for each gene, providing clinicians with a multi-dimensional molecular subtyping reference. Finally, the mutation probability values ​​and corresponding confidence levels of each gene are combined to output the multi-gene mutation probability prediction result.

[0092] In a specific example of this application, the system can further utilize the evolution prediction parameters output by the model as coefficients of the kinetic equation to perform mathematical extrapolation and simulation of key disease indicators (such as tumor volume) at multiple future time steps, thereby constructing a prediction curve reflecting the temporal trajectory of the disease for clinicians' reference. This function serves as a subsequent clinical application extension of the system and does not constitute the core protected content of this application.

[0093] More specifically, in a concrete example of this application, a standardized data container (such as a JSON object or XML document) is first created, defining key fields such as "gene mutation probability" and "confidence level". Then, the predicted mutation probabilities of each gene are sequentially filled into the corresponding fields according to gene number (covering 12 genes including TP53, CDKN2A, NOTCH1, FAT1, PIK3CA, HRAS, CCND1, FBXW7, CASP8, AXIN2, EGFR, and BRAF, each represented by its corresponding mutation probability value), and the confidence level parameter for each prediction result is attached. Finally, the above structured data, along with the patient's basic metadata, is encapsulated and output for subsequent clinical analysis. Clinicians can further conduct a comprehensive assessment of the disease progression trend based on the predicted mutation probabilities of each gene and the patient's actual condition; this system does not directly output a diagnostic conclusion.

[0094] In summary, the head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning according to the embodiments of this application is elucidated. It first performs deep fusion of multimodal images and clinical text, utilizing a nonlinear metric based on generalized mutual information and an asymmetric orthogonal decoupling mechanism to accurately extract complementary residual information from image data that cannot be explained by clinical indicators while establishing the dominant clinical prior. Furthermore, to address the model generalization bottleneck caused by small medical sample sizes, the system adopts a transfer reinforcement learning paradigm, training a general perception strategy using publicly available source domain data. Through parameter transfer and target domain fine-tuning, it combines general features with the fused representation of specific patients. Finally, it achieves multi-gene mutation probability prediction through a forward inference pipeline, thereby providing clinicians with objective and quantitative intraoperative diagnostic and treatment support.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A head and neck tumor intraoperative treatment support system based on deep transfer reinforcement learning, characterized in that, include: The patient data acquisition module is used to acquire multimodal imaging data and clinical text data of patients with head and neck cancer. The data preprocessing module is used to preprocess multimodal image data and clinical text data to obtain a multimodal head and neck cancer dataset; The multimodal fusion feature extraction module is used to extract fused multimodal feature vectors from the multimodal head and neck cancer dataset; The model training and optimization module is used to train and optimize the source domain deep reinforcement learning model based on the acquired public source domain dataset to obtain the pre-trained model parameters. The transfer learning fine-tuning module is used to initialize the target task network based on the pre-trained model parameters, and to perform transfer learning fine-tuning on the target task network using fused multimodal feature vectors as target domain data to obtain a head and neck cancer prediction model. The gene mutation prediction module is used to input the fusion features to be predicted into the head and neck cancer prediction model to predict the probability of multi-gene mutations, so as to obtain the prediction results of the mutation probability of each gene.

2. The intraoperative treatment support system for head and neck tumors based on deep transfer reinforcement learning according to claim 1, characterized in that, The data preprocessing module is used for: After performing frequency domain weighted reconstruction and spatial domain inverse transformation on multimodal image data, it is concatenated with clinical text data to obtain a fused image and text data stream. Contrast correction and lesion masking are performed on the fused image and text data streams to obtain the region of interest image and text data stream; Key-value pair parsing and numerical encoding are performed on the text information in the region of interest image and text data stream to obtain a heterogeneous feature pair data stream, which includes paired region of interest images and clinical feature vectors. Sample IDs are assigned to heterogeneous feature-based data streams, and the data streams are standardized and encapsulated to obtain a multimodal head and neck cancer dataset.

3. The intraoperative treatment support system for head and neck tumors based on deep transfer reinforcement learning according to claim 1, characterized in that, The multimodal fusion feature extraction module includes: The multi-strategy texture feature analysis unit is used to perform texture spatial dependency and consistency analysis on region-of-interest images in a multimodal head and neck cancer dataset to obtain a multi-strategy texture feature matrix. The multimodal feature fusion unit is used to perform feature fusion on the multi-strategy texture feature matrix based on the multi-head attention mechanism to obtain the fused multimodal feature vector.

4. The intraoperative treatment support system for head and neck tumors based on deep transfer reinforcement learning according to claim 3, characterized in that, The multi-strategy texture feature analysis unit is used for: The region of interest image is subjected to grayscale quantization and discretization to obtain a quantized image data stream; Based on the gray-level co-occurrence matrix, spatial dependency feature extraction is performed on the quantized image data stream to obtain an intermediate data stream containing gray-level co-occurrence features; The intermediate data stream is subjected to joint extraction of grayscale runout and gradient features to obtain the original multidimensional feature vector set; The original multidimensional feature vector set is subjected to dimensional unification and normal distribution mapping to obtain a multi-strategy texture feature matrix.

5. The intraoperative treatment support system for head and neck tumors based on deep transfer reinforcement learning according to claim 3, characterized in that, The multimodal feature fusion unit includes: The texture feature weighted subunit is used to perform subspace projection and scaling dot product operations on the multi-strategy texture feature matrix using a multi-head attention mechanism to obtain a weighted texture feature matrix. The dimension alignment and concatenation subunit is used to dimensionally align and concatenate the weighted texture feature matrix with clinical feature vectors from a multimodal head and neck cancer dataset to obtain a hybrid feature matrix. The highly correlated redundant feature removal subunit is used to remove highly correlated redundant features from the mixed feature matrix to obtain a dimension-reduced feature matrix. The robust feature filtering subunit is used to perform ICC-based robust feature filtering on the dimensionality-reduced feature matrix to obtain a fused multimodal feature vector.

6. The intraoperative treatment support system for head and neck tumors based on deep transfer reinforcement learning according to claim 1, characterized in that, The model training optimization module is used for: By defining state space, action space, and reward function, an interactive environment is encapsulated for a public source domain dataset to obtain a source domain reinforcement learning environment instance. A function approximator is constructed using a deep neural network, and the update frequency of the target network is set to obtain an initialized deep Q-network. The deep Q-network is iteratively trained using a greedy strategy to obtain the set of model parameters to be optimized; Bayesian hyperparameter search and validation set evaluation are performed on the set of model parameters to be optimized to obtain the pre-trained model parameters.

7. The intraoperative treatment support system for head and neck tumors based on deep transfer reinforcement learning according to claim 1, characterized in that, The transfer learning fine-tuning module is used for: The target network is weighted using the pre-trained model parameters, and shallow freeze and deep unlock operations are performed to obtain the initialized target task network. Supervised fine-tuning and backpropagation updates are performed on the initialized target task network using fused multimodal feature vectors to obtain a deep semantic feature set; The feature weights of the deep semantic feature set are optimized based on gradient boosting tree to obtain an optimized tree ensemble classifier; The fine-tuned target task network is integrated with a tree ensemble classifier in a pipeline to obtain a head and neck cancer prediction model.

8. The intraoperative treatment support system for head and neck tumors based on deep transfer reinforcement learning according to claim 1, characterized in that, The gene mutation prediction module includes: The forward inference unit is used to perform forward inference on the fusion features to be predicted using the head and neck cancer prediction model in order to obtain the multi-gene classification log odds. The mutation state prediction unit is used to perform probability mapping on the log odds of multi-gene classification to obtain the mutation probability prediction results of each gene.

9. The intraoperative treatment support system for head and neck tumors based on deep transfer reinforcement learning according to claim 5, characterized in that, The highly correlated redundant feature removal subunit is used for: A generalized mutual information measure based on kernel density estimation is applied to the mixed feature matrix to obtain the nonlinear dependency matrix; Image-clinical asymmetric decoupling is performed on the nonlinear dependency matrix to obtain the decoupling residual feature set; Complementary information gain filtering is performed on the decoupled residual feature set and the hybrid feature matrix to obtain the dimensionality-reduced feature matrix.

Images