7t magnetic resonance head and neck tumor radiotherapy response assessment system

Abstract

The application discloses a 7T magnetic resonance head and neck tumor radiotherapy response evaluation system, belongs to the field of medical imaging and tumor radiotherapy, and comprises an ultrahigh-field multi-modal image acquisition module, a multi-phase accurate registration and radiomics feature extraction module, an early response biomarker identification module and an adaptive treatment decision support module. The high resolution, high signal-to-noise ratio and multi-nuclide imaging capability of the 7T ultrahigh-field magnetic resonance are utilized to acquire multi-modal images before and during treatment, more than 200 features are extracted through accurate registration and radiomics analysis, the treatment response is predicted based on the Delta feature and the deep learning model 2 to 3 weeks after the start of radiotherapy, the prediction accuracy reaches 87.5%, the prediction is 3 to 4 weeks earlier than the conventional evaluation, decision support is provided for timely adjustment of the treatment strategy, a closed-loop feedback mechanism is established, the clinical verification results are used to optimize the parameters of each module reversely, and continuous improvement is realized.

Description

Technical Field

[0001] This invention belongs to the fields of medical imaging technology and tumor radiotherapy technology, specifically involving a 7T magnetic resonance head and neck tumor radiotherapy response assessment system, and particularly involving the use of multimodal radiomics and deep learning technology to achieve early and accurate prediction of tumor treatment response and adaptive treatment decision support. Background Technology

[0002] Head and neck malignancies are the seventh most common cancer worldwide, with squamous cell carcinomas of the head and neck, such as nasopharyngeal carcinoma, oropharyngeal carcinoma, laryngeal carcinoma, and hypopharyngeal carcinoma, accounting for the majority. Radiotherapy is one of the main treatment methods for head and neck tumors, with approximately 70% of patients requiring radical radiotherapy or postoperative adjuvant radiotherapy. However, due to tumor heterogeneity, the sensitivity to radiotherapy varies greatly among patients. Some patients respond well to standard-dose radiotherapy and achieve complete remission, while others exhibit radioresistance, leading to failure of local control.

[0003] Current radiotherapy response assessments primarily rely on imaging examinations three months after treatment, measuring morphological changes in the tumor to determine treatment effectiveness. This assessment method has a significant lag, failing to identify ineffective treatment in the early stages and missing the optimal opportunity to adjust treatment strategies. Numerous studies have shown that functional and metabolic responses to radiotherapy often appear 2 to 4 weeks earlier than morphological changes; therefore, developing imaging assessment techniques capable of capturing these early changes has significant clinical value.

[0004] Magnetic resonance imaging (MRI) has been widely used in the diagnosis and treatment planning of head and neck tumors due to its superior soft tissue contrast. In recent years, 3T MRI has become a standard clinical practice and has demonstrated its advantages in tumor assessment. However, 3T MRI still has limitations in spatial resolution, signal-to-noise ratio, and metabolic imaging capabilities, making it difficult to capture subtle changes in the microstructure of tumors. 7T ultra-high field MRI represents the cutting edge of MRI technology, with a magnetic field strength more than twice that of 3T, theoretically providing nearly four times the signal-to-noise ratio and higher spatial resolution. The second-generation 7T MRI system, approved by the FDA in 2024, has achieved ultra-high resolution imaging at the 0.4mm level, showing great potential in neuroimaging and tumor imaging.

[0005] Radiomics is an emerging image analysis technique that extracts a wealth of quantitative features from medical images to reveal tumor heterogeneity information invisible to the naked eye. Radiomic features include hundreds of quantitative parameters describing tumor shape, texture, and intensity distribution, which are closely related to tumor biological behavior and treatment response. Delta radiomics further focuses on the magnitude of changes in features before and after treatment, offering stronger predictive power compared to feature analysis at a single time point. The introduction of deep learning techniques enables radiomics analysis to automatically learn complex feature patterns, significantly improving the performance of predictive models.

[0006] Chinese patent CN114240855A discloses a method for assessing the skeletal muscle index in patients with head and neck tumors using CBCT. This method utilizes CBCT images routinely used during radiotherapy to calculate the skeletal muscle index by measuring the skeletal muscle area in the cross-section of the third cervical vertebra, for dynamic assessment of the patient's nutritional status during radiotherapy. However, this technology has the following limitations: CBCT's soft tissue resolution is far lower than that of magnetic resonance imaging, making it difficult to clearly display subtle changes in tumor tissue; this method only focuses on the nutritional indicator of skeletal muscle area, failing to provide direct information on the tumor's response to radiotherapy; measurements based on a single anatomical level cannot reflect the three-dimensional heterogeneity of the tumor; it lacks functional and metabolic imaging information, failing to capture early biological changes in the tumor; and its assessment purpose is limited to nutritional monitoring, failing to achieve treatment response prediction and adaptive treatment decision support.

[0007] Therefore, developing a radiotherapy response assessment system for head and neck tumors based on 7T ultra-high field magnetic resonance imaging (MRI) is of great clinical value and scientific significance. This system would fully utilize the technological advantages of 7T MRI in soft tissue imaging, functional imaging, and metabolic imaging, and combine it with advanced radiomics analysis and deep learning technology to achieve early and accurate prediction of tumor treatment response, providing timely treatment decision support for clinicians. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a 7T magnetic resonance head and neck tumor radiotherapy response assessment system. By integrating ultra-high field multimodal imaging, advanced image analysis technology and intelligent prediction model, it can achieve early and accurate assessment of the radiotherapy response of head and neck tumors and adaptive treatment decision support.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The 7T MRI head and neck tumor radiotherapy response assessment system includes an ultra-high field multimodal image acquisition module, a multi-temporal precise registration and radiomics feature extraction module, an early response biomarker identification module, and an adaptive treatment decision support module.

[0011] The ultra-high field multimodal image acquisition module utilizes a 7T magnetic resonance imaging (MRI) system to acquire high-quality images of head and neck tumors at multiple time points before and during treatment. The 7T ultra-high field intensity provides nearly four times the signal-to-noise ratio improvement of 3T MRI, enabling the system to clearly display the tumor's microstructure at an isotropic resolution of 0.5 mm. The acquired images include various complementary sequences: morphological sequences provide anatomical information about the tumor, diffusion-weighted imaging sequences reflect tumor cell density and cell membrane integrity, and magnetic resonance spectroscopy sequences provide metabolic information. The integration of multimodal information allows the system to comprehensively assess the tumor status from morphological, functional, and metabolic dimensions.

[0012] The multi-temporal precise registration and radiomics feature extraction module accurately registers images acquired at different time points, ensuring the accuracy of feature extraction. The registration algorithm employs a non-rigid registration method based on mutual information, capable of handling deformations of the tumor and surrounding tissues during radiotherapy, achieving sub-millimeter level registration accuracy. From the registered images, the system extracts over 200 radiomics features, covering multiple dimensions describing tumor shape, intensity distribution, and texture patterns. The ultra-high resolution of 7T MRI enables the system to capture subtle texture features that are difficult to detect with 3T equipment; these features are closely related to tumor heterogeneity and radiotherapy response.

[0013] The early response biomarker identification module is the core innovation of the system. By calculating the changes in features before and after treatment, it identifies Delta feature combinations related to treatment response. The system uses a deep learning model to automatically learn complex feature patterns and predict the probability of the tumor achieving complete remission, partial remission, or disease progression at the end of treatment. The key innovation is that this module can predict the final treatment response based on early changes in tumor function and metabolism 2 to 3 weeks after the start of radiotherapy, before the sum of the longest diameters of the target lesion has shrunk to the RECIST 1.1 partial remission threshold (30%). This prediction time point is 3 to 4 weeks earlier than traditional assessment methods.

[0014] The adaptive treatment decision support module generates individualized treatment adjustment suggestions based on the prediction results. For patients predicted to be radiosensitive, the system suggests maintaining or moderately reducing the dose to minimize toxic side effects; for patients predicted to be radioresistant, the system suggests increasing the dose, modifying the target area, or adding sensitizing drugs. Importantly, this module establishes a closed-loop feedback mechanism, feeding back clinically validated actual treatment results to the early response biomarker identification module to optimize the parameters of the prediction model. It also adjusts the feature selection weights of the multi-temporal precise registration and radiomics feature extraction modules, and even guides the scanning scheme of the ultra-high field multimodal image acquisition module, forming a positive feedback loop for continuous improvement of the entire system.

[0015] The beneficial effects of this invention include:

[0016] This system fully leverages the technological advantages of 7T ultra-high field MRI, offering significant advantages over conventional imaging equipment such as 3T MRI and CBCT in terms of soft tissue contrast, spatial resolution, and functional metabolic imaging capabilities. The ultra-high resolution at the 0.5mm level allows the system to capture subtle changes in tumor microstructure that are difficult to detect on conventional imaging equipment. The integration of multimodal imaging provides comprehensive information on tumor status, assessing treatment response from morphological, functional, and metabolic dimensions, resulting in higher diagnostic accuracy compared to single-modal imaging.

[0017] This system enables early and accurate prediction of treatment response. Based on Delta radiomics characteristics and a deep learning model, it can predict the degree of response at the end of treatment 2 to 3 weeks after the start of radiotherapy. This prediction time point is 3 to 4 weeks earlier than imaging assessments based on RECIST 1.1 standard morphological volume changes. Early prediction allows clinicians to identify ineffective treatment early in the course of treatment, adjust treatment strategies in a timely manner, avoid patients receiving ineffective radiotherapy doses and related toxic side effects, and provide the possibility of dose reduction for radiosensitive patients, improving their quality of life. Clinical data show that this system improves the early identification rate of ineffective treatment by 71.3% and achieves a prediction accuracy of 87.5%, significantly outperforming traditional assessment methods.

[0018] This system establishes a complete closed-loop feedback mechanism, enabling adaptive optimization by feeding back clinical validation results to each functional module. The prediction model incrementally learns based on accumulated clinical data, continuously improving prediction accuracy; feature selection weights are dynamically adjusted based on feedback results, prioritizing features most relevant to treatment response; and scanning protocols are optimized according to model requirements, shortening scan time while maintaining diagnostic quality. This closed-loop mechanism ensures that system performance continuously improves over time, exhibiting excellent sustainability.

[0019] The various modules of this system form a deeply coupled and synergistic relationship, achieving an overall prediction accuracy improvement of at least 15 percentage points compared to individual modules. The high-quality images with isotropic 0.5mm resolution and a signal-to-noise ratio of at least 30 from the ultra-high field multimodal image acquisition module provide a solid foundation for subsequent analysis; the accurate features from the multi-temporal precise registration and radiomics feature extraction module enhance the performance of the prediction model; the prediction results from the early response biomarker identification module guide treatment decisions while simultaneously providing feedback to optimize the preceding modules; and the clinical validation results from the adaptive treatment decision support module drive continuous improvement of the entire system. This multi-module synergistic mechanism makes the overall system performance significantly better than the simple sum of the individual modules, demonstrating the advantages of system-level innovation.

[0020] This system is applicable to various head and neck tumor types and different radiotherapy regimens, and has broad clinical application prospects. It can be used for radiotherapy response assessment of head and neck squamous cell carcinomas such as nasopharyngeal carcinoma, oropharyngeal carcinoma, laryngeal carcinoma, and hypopharyngeal carcinoma, supporting multiple treatment modalities including radiotherapy alone, concurrent chemoradiotherapy, and radiotherapy after induction chemotherapy. The system's modular design allows each functional module to be configured and expanded according to specific clinical needs, providing excellent flexibility and scalability. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0022] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0023] Figure 2 This is a schematic diagram of the structure of the ultra-high field multimodal image acquisition module.

[0024] Figure 3 This is a schematic diagram of the workflow of the multi-temporal precise registration and radiomics feature extraction module.

[0025] Figure 4 This is a schematic diagram of the early response biomarker recognition module.

[0026] Figure 5 This is a functional diagram of the adaptive treatment decision support module. Detailed Implementation

[0027] Please refer to the attached document. Figures 1-5 The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0028] like Figure 1 As shown, the 7T MRI head and neck tumor radiotherapy response assessment system of the present invention includes an ultra-high field multimodal image acquisition module 1, a multi-temporal precise registration and radiomics feature extraction module 2, an early response biomarker identification module 3, and an adaptive treatment decision support module 4. The four modules are closely connected through data interfaces to form a complete workflow, and closed-loop optimization is achieved through feedback pathways.

[0029] like Figure 2As shown, the ultra-high field multimodal image acquisition module 1 is the source of system data acquisition, responsible for acquiring high-quality 7T magnetic resonance images at multiple time points before and during treatment. This module is built based on a 7T ultra-high field magnetic resonance imaging system with a magnetic field strength of 7.0T, which has significant technical advantages compared to the clinically commonly used 3T equipment. Theoretically, the signal-to-noise ratio is proportional to the magnetic field strength, and the 7T system can provide nearly twice the signal-to-noise ratio of the 3T system. In actual measurements, considering factors such as dielectric effects and radiofrequency penetration depth, the 7T system can still achieve an approximately 1.8-fold improvement in signal-to-noise ratio in head and neck imaging. The higher signal-to-noise ratio allows the system to use higher spatial resolution without sacrificing image quality. This system achieves an isotropic resolution of 0.5mm, which can clearly display the microstructural features of tumors.

[0030] The ultra-high field multimodal image acquisition module 1 includes a sequence parameter optimization unit and a data quality monitoring unit. The sequence parameter optimization unit determines the scanning parameters for each sequence based on patient characteristics and treatment stage. For morphological imaging, the system employs T1-weighted imaging, T2-weighted imaging, and contrast-enhanced T1-weighted imaging. Typical parameters for T1-weighted imaging are a repetition time of 600 ms, an echo time of 12 ms, a flip angle of 90°, and a slice thickness of 0.5 mm, used to display the anatomical relationship between the tumor and surrounding normal tissue. Typical parameters for T2-weighted imaging are a repetition time of 3500 ms, an echo time of 80 ms, and a slice thickness of 0.5 mm, which can clearly display tumor edema and necrosis areas. Contrast-enhanced T1-weighted imaging is acquired 5 minutes after intravenous injection of gadolinium contrast agent to assess tumor blood supply.

[0031] For functional imaging, the system employs a diffusion-weighted imaging sequence. Diffusion-weighted imaging reflects the microstructure of tissues by detecting the random Brownian motion of water molecules, and is of significant value in tumor assessment. When tumor cells are dense and their cell membranes are intact, water molecule diffusion is restricted, resulting in high signal intensity and a low apparent diffusion coefficient. After radiotherapy causes tumor cell death and cell membrane rupture, water molecule diffusion increases, leading to an elevated apparent diffusion coefficient. This change often precedes tumor volume reduction; therefore, diffusion-weighted imaging is an ideal tool for assessing early treatment response. This system uses a single-shot excitation planar echo sequence, with typical parameters including a repetition time of 4000 ms, an echo time of 70 ms, and b-values ​​of 0 and 1000 s / mm. 2 The apparent diffusion coefficient map was calculated through post-processing. The high signal-to-noise ratio of the 7T magnetic resonance imaging system allows for the use of higher b values, such as 1500 s / mm. 2 and 2000s / mm 2 These high b values ​​are more sensitive to changes in tumor microstructure.

[0032] For metabolic imaging, the system employs magnetic resonance spectroscopy (MRS) sequences. MRSS provides metabolic information by detecting resonance peaks of different metabolites in tissues and is an important non-invasive method for assessing the metabolic status of tumors. Typical metabolic features of head and neck tumors include elevated choline peaks, altered creatine peaks, and the appearance of lactate peaks. Choline is an important component of cell membrane synthesis; its concentration increases when tumor cells proliferate actively, and decreases after radiotherapy inhibits cell proliferation. This system uses point-resolved spectral sequences with typical parameters of 2000 ms repetition time, 30 ms echo time, and 8×8×8 mm voxel size. 3 The high magnetic field strength of the 7T magnetic resonance results in greater chemical shift dispersion and higher spectral resolution, enabling the system to resolve more metabolite peaks, including glutamate, glutamine, and glycine, which are difficult to detect with 3T equipment.

[0033] The data quality monitoring unit evaluates the quality indicators of acquired images in real time to ensure the accuracy of subsequent analysis. For morphological images, the main evaluation metrics are signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). SNR is defined as the ratio of the average signal intensity of the region of interest to the standard deviation of the background noise, while CNR is defined as the ratio of the difference in signal intensity between two tissue regions to the standard deviation of the background noise. The system sets an SNR threshold of 30 and a CNR threshold of 5. When the actual measured value is lower than the threshold, the system prompts technicians to check the scanning parameters or re-acquire the image. For diffusion-weighted imaging, in addition to SNR, the geometric distortion of the image is also evaluated. The degree of distortion is quantified by comparing the registration error between the diffusion-weighted image and the anatomical reference image, with a maximum allowable distortion of 1.0 mm. For magnetic resonance spectroscopy, the linewidth and baseline stability of the spectrum are evaluated. Linewidth reflects magnetic field homogeneity, and baseline stability reflects motion artifacts during acquisition. The system requires the linewidth of the water peak to be less than 10 Hz and the baseline fluctuation amplitude to be less than 5% of the signal peak height.

[0034] The ultra-high field multimodal image acquisition module 1 acquires baseline images before treatment to establish the patient's initial tumor status profile. Baseline images include complete acquisition of all the aforementioned sequences, with a total acquisition time of approximately 25 minutes. To ensure patient comfort and image quality, the system employs a dedicated head and neck coil array containing 32 receiving channels, providing a high signal-to-noise ratio and uniform signal reception. The patient is positioned supine, with the head immobilized using a thermoplastic mask to reduce motion artifacts, and the neck and shoulders supported by pillows to maintain a comfortable posture. After the start of radiotherapy, the system acquires follow-up images at weeks 1, 2, 3, and at the end of treatment. The sequence selection for follow-up images can be appropriately simplified according to clinical needs, prioritizing the acquisition of diffusion-weighted imaging and magnetic resonance spectroscopy, which are sensitive to treatment response, thus reducing the total acquisition time to 15 minutes.

[0035] The sequence parameter optimization unit individually adjusts scanning parameters for different patients and different treatment stages based on previous scanning experience and feedback learning results. For example, for patients with large tumors, the coverage area is appropriately increased to include the entire tumor and surrounding lymph nodes; for patients whose tumors have shrunk significantly in the later stages of treatment, the scanning range can be appropriately reduced and the spatial resolution increased for more detailed observation of residual lesions. For patients predicted to be sensitive in the early stages of radiotherapy, the scanning frequency can be appropriately reduced to lessen the patient's burden; for patients predicted to be resistant, the scanning frequency can be increased to closely monitor tumor changes. This individualized scanning protocol optimization reflects the system's intelligence and human-centered design.

[0036] The acquired image data is stored in DICOM format and transmitted to the image processing workstation via the hospital's image archiving and communication system for analysis by subsequent modules. Each patient's complete dataset includes multimodal images from multiple time points, with a data volume of approximately 5GB to 8GB. The system employs compressed storage and high-speed network transmission technology to ensure rapid data access and processing.

[0037] like Figure 3 As shown, the multi-temporal precise registration and radiomics feature extraction module 2 performs precise registration of images acquired at different time points and extracts multi-dimensional radiomics features. This module serves as a bridge connecting image acquisition and intelligent analysis, and its performance directly affects the accuracy of subsequent prediction models.

[0038] Registration is a crucial preprocessing step in image analysis. Its purpose is to align images acquired at different time points to the same spatial coordinate system, ensuring that the same anatomical location corresponds to the same voxel coordinates in different images. During radiotherapy for head and neck tumors, patient positioning may vary slightly, but more importantly, the tumor and surrounding tissues undergo significant deformation. Tumor shrinkage, reduction of surrounding edema, and restoration of normal tissue all lead to changes in the spatial position of tissues. Therefore, rigid registration methods such as translation and rotation alone cannot meet the required accuracy; non-rigid registration algorithms must be used to handle tissue deformation.

[0039] This module employs a non-rigid registration algorithm based on mutual information. Mutual information, a concept in information theory, measures the statistical dependency between two images. For perfectly aligned images, their mutual information reaches its maximum. Non-rigid registration maximizes the mutual information between the registered image pairs by optimizing the deformation field. The deformation field describes the corresponding position of each voxel in the reference image in the image to be registered, and is a three-dimensional vector field. This system uses a B-spline free deformation model to represent the deformation field, defining the deformation through a control point grid. The control point spacing is set to 10mm to balance registration accuracy and computational efficiency.

[0040] The registration algorithm employs a multi-scale optimization strategy, progressively refining the registration results from low to high resolution. First, the image is downsampled to obtain a low-resolution version, where coarse registration is performed to determine the approximate correspondence. Then, the resolution is gradually increased for fine registration. This multi-scale strategy avoids getting trapped in local optima and improves computational speed. The optimization algorithm uses gradient descent, iteratively updating the deformation field parameters until convergence. The convergence criterion is that the change in mutual information between two consecutive iterations is less than 0.01% or the number of iterations reaches 200.

[0041] To evaluate registration accuracy, the system selected several anatomical landmarks in the tumor and surrounding tissues, such as the tumor boundary, vertebral body, and parotid gland, and calculated the positional errors of these landmarks before and after registration. Test results on 50 patients showed that the average positional error of the landmarks after registration was 0.48 mm, and 95% of the landmark errors were less than 0.8 mm, meeting the requirements for sub-millimeter registration. Accurate registration provides a reliable guarantee for subsequent feature extraction and change calculation.

[0042] After registration is complete, the module proceeds to the radiomics feature extraction stage. Radiomics extracts a large number of quantitative features from medical images, transforming visual information into computable digital features to provide input for machine learning models. This system extracts over 200 radiomics features, covering multiple categories to comprehensively describe the imaging characteristics of tumors.

[0043] First-order statistical features describe the distribution of pixel intensity values ​​within the tumor region, including mean, standard deviation, skewness, kurtosis, maximum, minimum, median, and quartiles. The mean reflects the average signal intensity of the tumor, the standard deviation reflects the dispersion of signal intensity, i.e., the homogeneity of the tumor, skewness describes the symmetry of the distribution, and kurtosis describes the sharpness of the distribution. These statistical features were calculated for T1-weighted images, T2-weighted images, contrast-enhanced images, and apparent diffusion coefficient maps, yielding approximately 40 first-order features.

[0044] Second-order texture features, based on the gray-level co-occurrence matrix (GLCM) and gray-level run-length matrix (LLM), describe the complex patterns of pixel spatial relationships within a tumor. The GLCM statistically analyzes the frequency of pixel pairs with specific gray-level value combinations at specific directions and distances, reflecting the texture coarseness and directionality of the image. Features such as energy, contrast, correlation, and entropy can be derived from the GLCM. Energy reflects the uniformity of image texture, contrast reflects the drasticness of gray-level changes, correlation reflects the linear relationship between pixels, and entropy reflects the complexity of the texture. The gray-level run-length matrix statistically analyzes the length distribution of consecutive occurrences of specific gray-level values ​​in a specific direction, reflecting the roughness of the image. Features such as short-run emphasis, long-run emphasis, gray-level non-uniformity, and run-length non-uniformity can be derived from the gray-level run-length matrix. The system calculates texture features in four directions: 0°, 45°, 90°, and 135°, with a distance parameter set to 1 pixel, obtaining approximately 80 second-order texture features.

[0045] Higher-order wavelet features decompose an image into sub-bands of different frequencies and directions using wavelet transform, extracting statistical features from each sub-band. Wavelet transform is equivalent to applying filters of different scales to the image, capturing information at different spatial frequencies. This system employs three-level wavelet decomposition using the Coiflet wavelet basis, calculating the mean, standard deviation, energy, and other statistics for each sub-band, obtaining approximately 60 wavelet features. Wavelet features are sensitive to subtle textural changes within tumors, revealing heterogeneous patterns that are difficult to discern with the naked eye.

[0046] Shape-based geometric features describe the three-dimensional morphological characteristics of tumors, including volume, surface area, sphericity, elongation, and flatness. Volume is a fundamental indicator of tumor size; surface area reflects the degree of contact between the tumor and surrounding tissues; sphericity is defined as the ratio of tumor volume to the volume of a sphere with the same surface area, reflecting the regularity of the tumor shape; and elongation and flatness describe the extent of tumor extension in different directions. For head and neck tumors, morphological features have significant prognostic value; tumors with low sphericity and irregular boundaries are often highly invasive and have poor treatment response. The system uses a three-dimensional surface reconstruction algorithm to accurately calculate these geometric features, obtaining approximately 30 shape features.

[0047] Metabolic features were extracted from magnetic resonance spectroscopy imaging to reflect the metabolic state of the tumor. The system performed baseline correction, phase correction, and frequency alignment on the spectral data, and then used a peak fitting algorithm to separate and quantify the peaks of each metabolite. The main metabolites of interest included choline, creatine, N-acetylaspartate, and lactate. The choline peak was located at 3.2 ppm and is a marker of cell membrane metabolism; the creatine peak was located at 3.0 ppm, reflecting the energy metabolism status; the N-acetylaspartate peak was located at 2.0 ppm and is a marker of neurons; and the lactate peak was located at 1.3 ppm, reflecting the degree of anaerobic glycolysis. The system calculated the absolute concentrations of each metabolite and their ratios, such as the choline-to-creatine ratio and the choline-to-N-acetylaspartate ratio, obtaining approximately 10 metabolic features. These metabolic features are sensitive to the biological behavior of tumors and can reflect early metabolic changes induced by treatment.

[0048] All feature extraction processes are performed within the segmented tumor region. Tumor segmentation employs a semi-automatic method, where experienced radiologists delineate the 3D contour of the tumor on high-quality T2-weighted or contrast-enhanced images, and then the segmentation results are automatically refined using region growing and boundary optimization algorithms. To ensure segmentation consistency, the system records segmentation parameters and applies the same segmentation strategy to images at different time points.

[0049] After feature extraction, the system standardizes the features, mapping features with different dimensions and numerical ranges to the same scale. Z-score standardization is used, subtracting the mean from each feature and dividing by its standard deviation, resulting in a standardized feature with a mean of 0 and a standard deviation of 1. Standardization eliminates the influence of differences in feature dimensions, preventing features with large numerical ranges from dominating subsequent analysis and ensuring that all features are treated fairly.

[0050] To reduce feature redundancy and overfitting risks, the system performs feature selection. A minimum redundancy maximum correlation algorithm is employed, which considers both the correlation between features and the target variable and the redundancy among features, selecting a subset of features with high relevance but low redundancy. Specifically, the algorithm first selects the features most correlated with the treatment response, then iteratively selects features with low relevance to the selected features but high relevance to the response, until a preset number of features is reached. Based on experience and cross-validation results, the system retains 40 to 50 of the most predictive features for subsequent modeling.

[0051] The output of the multi-temporal precise registration and radiomics feature extraction module 2 is a structured feature matrix, where each row corresponds to a patient's feature vector at a specific time point, and each column corresponds to a radiomics feature. For patients undergoing multiple scans, the system calculates the feature change between each scan and the baseline scan, i.e., the Delta feature. The Delta feature is defined as the feature value at the current time point minus the baseline feature value, reflecting the dynamic changes of the tumor during treatment. Studies have shown that the Delta feature has stronger predictive power than the absolute feature value at a single time point because it directly quantifies the changes caused by treatment. The module passes the feature matrix and the Delta feature matrix to the early response biomarker identification module 3 as input to the prediction model.

[0052] like Figure 4 As shown, the early response biomarker identification module 3 is the core innovative part of the system. By calculating the changes in features before and after treatment, it identifies Delta feature combinations related to treatment response and predicts the final treatment response of the tumor based on a deep learning model. This module enables the system to predict the degree of response at the end of treatment in the early stages of radiotherapy, before the sum of the longest diameters of the target lesion has shrunk to the RECIST 1.1 partial remission threshold, providing a basis for timely adjustment of treatment strategies.

[0053] The early response biomarker identification module 3 includes a feature change calculation unit, a response prediction model unit, and a confidence assessment unit.

[0054] The feature change calculation unit processes the feature data transmitted by the multi-temporal precise registration and radiomics feature extraction module 2, and calculates the feature changes relative to the baseline at each time point. For the first... The image acquired at the nth time point, its nth The formula for calculating the change in each characteristic is:

[0055] ,

[0056] in, For the first The feature in the first The change at each point in time For the first The feature in the first Feature values ​​at each time point For the first The system calculates the feature values ​​at the baseline. Additionally, it calculates the relative rate of change of each feature:

[0057] ,

[0058] in, For the first The feature in the first The relative rate of change at each point in time A small constant of 0.001 is set to avoid a denominator of zero. The relative rate of change eliminates the influence of baseline feature values ​​and more robustly reflects the degree of feature change. For 40 to 50 selected features, the system calculates their absolute change and relative rate of change at each follow-up time point, forming a complete Delta feature set.

[0059] The response prediction model unit is the core component of the module, employing deep learning technology to construct the prediction model. The model's input is the Delta feature set, and its output is the probability that the patient will achieve complete remission, partial remission, or disease progression at the end of treatment. The definition of treatment response follows the evaluation criteria for solid tumor efficacy: complete remission refers to the complete disappearance of all measurable lesions; partial remission refers to a reduction of at least 30% in the sum of the longest diameters of measurable lesions; and disease progression refers to an increase of at least 20% in the sum of the longest diameters of measurable lesions or the appearance of new lesions. Cases not meeting these criteria are defined as disease stability. In practical applications, the system merges complete and partial remissions into a valid response, and disease stability and disease progression into an invalid response, simplifying the prediction task into a binary classification problem.

[0060] The prediction model employs a hybrid architecture combining convolutional neural networks (CNNs) and support vector machines (SVMs). CNNs excel at automatically learning complex feature patterns from high-dimensional data, while SVMs demonstrate good generalization ability in small sample sizes. Combining the two leverages their respective strengths. The CNN portion consists of 5 convolutional layers and 3 fully connected layers. The input layer receives feature vectors with dimensions of 40 to 50. The first convolutional layer contains 64 kernels of size 3 with a stride of 1, using the same padding method to maintain the output dimension, and employing ReLU as the activation function. The second convolutional layer contains 128 kernels with the same parameter settings as the first layer. The third convolutional layer contains 256 kernels, and the fourth and fifth convolutional layers each contain 512 kernels. Each convolutional layer is followed by a batch normalization layer and a max-pooling layer. Batch normalization accelerates training and improves model stability, while max-pooling reduces feature dimensionality. The fully connected layer consists of two hidden layers and one output layer. The first hidden layer has 256 neurons, and the second hidden layer has 128 neurons. The activation function for both is ReLU, and the dropout ratio is 0.5 to prevent overfitting. The output layer has two neurons, one for the valid response and one for the invalid response. The softmax activation function is used to convert the output into a probability distribution.

[0061] The Support Vector Machine (SVM) part employs a radial basis function (RBF) kernel. The RDF kernel maps features to a high-dimensional space, making linearly inseparable data in the original space linearly separable in the high-dimensional space. The kernel function is defined as:

[0062] ,

[0063] in, For the sample and The kernel function values ​​between The width of the kernel parameter control function. This represents the Euclidean distance between two samples. Kernel parameters. Cross-validation was used to determine the test range, from 0.001 to 10, and the parameter values ​​that yielded the highest accuracy on the validation set were selected. The decision function of the support vector machine is:

[0064] ,

[0065] in, The decision function value, The number of support vectors, For Lagrange multipliers, For support vectors The category label takes a value of 1 or -1. This is the bias term. The sign of the decision function determines the predicted class of the sample.

[0066] The hybrid model combines the outputs of a convolutional neural network (CNN) and a support vector machine (SVM) using weighted fusion. Specifically, the probability of a valid response from the CNN output is denoted as... The decision function value of a support vector machine is normalized to the interval between 0 and 1 and denoted as . The final predicted probability is:

[0067] ,

[0068] in, The final predicted effective response probability, and To satisfy the fusion weight The fusion weights are not fixed values, but are dynamically adjusted based on the performance of each model on the validation set. Specifically, the area under the curves of the convolutional neural network and the support vector machine on the validation set is calculated, denoted as... and The fusion weight is calculated as follows:

[0069] ,

[0070] This dynamic weighting strategy allows the better-performing model to have a larger weight in the fusion result, thus improving the overall predictive power of the hybrid model.

[0071] The model is trained using backpropagation and a stochastic gradient descent optimizer. The loss function is cross-entropy loss, which, for a binary classification problem, is defined as:

[0072] ,

[0073] in, The value of the loss function. The number of training samples. For the first The true label for each sample is either 0 or 1. For the model to predict the first The cross-entropy loss measures the probability that a sample belongs to the positive class. The smaller the loss value, the more accurate the prediction. The optimizer uses the Adam algorithm with an initial learning rate of 0.001. Cosine annealing is used to dynamically adjust the learning rate to improve model convergence speed and final performance. The training batch size is 16, and the number of training epochs is 100. An early stopping strategy is used; training stops when the validation set loss fails to improve for 10 consecutive epochs to prevent overfitting.

[0074] To ensure the model's generalization ability, the system employs 5-fold cross-validation for model selection and performance evaluation. The dataset is randomly divided into 5 subsets. Four subsets are used as the training set and one subset as the validation set in each iteration, repeated 5 times so that each subset serves as the validation set once. The average performance across the 5 validation iterations is calculated as the model's final performance metric. Evaluation metrics include accuracy, sensitivity, specificity, area under the curve (AUC), and F1 score. Accuracy represents the proportion of correctly predicted samples out of the total samples; sensitivity is the proportion of true positives out of actual positives, reflecting the model's ability to identify valid responses; specificity is the proportion of true negatives out of actual negatives, reflecting the model's ability to identify invalid responses; the AUC measures the model's overall performance at different thresholds; and the F1 score is the harmonic mean of accuracy and recall.

[0075] In a retrospective study of 150 patients with head and neck tumors, this module achieved excellent performance in predicting the degree of response at the end of treatment based on Delta features at week 2 after the start of radiotherapy. The mean accuracy of 5-fold cross-validation was 87.5%, sensitivity was 89.3%, specificity was 84.6%, area under the curve was 0.91, and F1 score was 0.88. These results significantly outperformed prediction models based on single-time-point features, improving accuracy by approximately 12 percentage points, demonstrating the superiority of the Delta radiomics approach. More importantly, the week 2 prediction was approximately 4 weeks earlier than the post-treatment assessment, providing ample time for clinical decision-making.

[0076] The confidence assessment unit quantifies the reliability of the prediction results. Even if the model provides prediction probabilities, the level of these probabilities does not necessarily fully reflect the reliability of the prediction, especially for samples close to the decision boundary. The confidence assessment uses a bootstrap method to estimate the confidence intervals of the prediction probabilities. Specifically, the training set is sampled with replacement to generate several bootstrap sample sets. The model is trained on each bootstrap sample set and predictions are made for the target samples, obtaining a set of prediction probability values. The mean and standard deviation of this set of probability values ​​are calculated, and a 95% confidence interval is constructed based on the normal distribution assumption.

[0077] ,

[0078] in, Let be the confidence interval. This represents the mean of the probability of self-guided prediction. 1.96 represents the standard deviation of the self-predicted probability, and 1.96 is the Z-score corresponding to a 95% confidence level. The width of the confidence interval reflects the uncertainty of the prediction; the narrower the interval, the more reliable the prediction. The system uses the confidence interval width as the reciprocal of the confidence level. When the confidence level is below a set threshold, such as 80%, the system prompts clinicians to be cautious about the prediction results and suggests increasing the scanning frequency to obtain more information.

[0079] The output of the early response biomarker identification module 3 includes the predicted response category (e.g., effective or ineffective response), response probability, confidence interval, and a list of key features. The list of key features is obtained through feature importance analysis, quantifying the contribution of each feature to the prediction result. For deep learning models, gradient-weighted class activation mapping is used to visualize feature importance; for support vector machines, feature importance is determined by the weights of the support vectors. The system identifies 10 to 15 features that contribute most to the treatment response prediction. These features typically include changes in the apparent diffusion coefficient, the rate of tumor volume reduction, the decrease in the choline to creatine ratio, and the homogenization trend of texture features. These key features not only improve the interpretability of the model but also provide clinicians with a basis for understanding the prediction results.

[0080] like Figure 5 As shown, the adaptive treatment decision support module 4 generates individualized treatment adjustment suggestions based on the prediction results of the early response biomarker identification module 3, and establishes a closed-loop feedback mechanism to continuously optimize system performance. This module is a key interface for the clinical application of the system, transforming prediction results into actionable clinical decisions, while continuously improving the accuracy and practicality of the system through feedback learning.

[0081] The adaptive treatment decision support module 4 includes a response grading unit, a dosage adjustment suggestion unit, and a feedback learning unit.

[0082] The response grading unit categorizes patients into different risk classes based on predicted response probabilities. According to clinical needs and decision thresholds, the system classifies patients into three categories: sensitive, intermediate, and resistant. Specifically, patients with a predicted effective response probability greater than or equal to 75% are classified as sensitive; those with a predicted effective response probability between 25% and 75% are classified as intermediate; and those with a predicted effective response probability less than 25% are classified as resistant. This three-classification system can identify both patients with significant treatment effects and those resistant to standard treatment, providing a foundation for precision medicine. The threshold selection is based on clinical experience and decision analysis, balancing the risks of false positives and false negatives, and can be adjusted according to specific clinical scenarios.

[0083] The dose adjustment recommendation unit generates individualized treatment adjustment suggestions for patients with different response categories. For sensitive patients, the system suggests maintaining the current treatment regimen or considering a moderate dose reduction to reduce adverse reactions. Major adverse reactions to head and neck radiotherapy include oral mucositis, dysphagia, and salivary gland hypofunction, which severely impact patients' quality of life. For patients highly sensitive to radiotherapy, a moderate dose reduction, such as from the standard 70 Gy to 66 Gy, may significantly reduce adverse reactions without affecting tumor control. The system calculates the recommended dose reduction based on the predicted response probability and confidence level, combined with the patient's clinical characteristics such as age, general condition score, and comorbidities. Dose reduction recommendations require multidisciplinary team discussion and informed consent from the patient; the system provides decision support rather than automatic execution.

[0084] For intermediate-sized patients, the system recommends maintaining standard treatment and close monitoring. The treatment response in these patients is intermediate; it cannot be definitively confirmed that the desired therapeutic effect will be achieved, nor can it be concluded that treatment failure will occur. For these patients, while continuing standard treatment, the monitoring frequency should be increased, such as repeat imaging assessments at weeks 3 and 4, and subsequent strategies should be adjusted promptly based on tumor trends. The system also recommends more detailed biomarker testing for intermediate-sized patients, such as circulating tumor DNA and inflammatory markers, to integrate multidimensional information and improve predictive accuracy.

[0085] For resistant patients, the system recommends considering treatment intensification strategies. These strategies include dose escalation, target volume modification, and the addition of sensitizing agents. Dose escalation involves increasing the total radiotherapy dose or using concurrent boosting techniques to deliver a higher dose to the tumor, such as increasing the total dose to 74-76 Gy or increasing the dose to resistant subregions of the tumor by 20-30%. Target volume modification expands the clinical target volume based on high-risk areas indicated by imaging, ensuring adequate irradiation of potentially infiltrated areas. Adding sensitizing agents involves adding concurrent chemotherapy to radiotherapy alone, or selecting novel targeted therapies and immunotherapies to enhance treatment efficacy through the synergistic effect of drugs and radiotherapy. The system recommends the most appropriate intensification regimen from the above strategies based on the specific characteristics of the tumor and the patient's tolerance.

[0086] The dose adjustment recommendation unit also considers the feasibility and safety of the treatment plan. Adjustments to the radiotherapy dose must be made while ensuring the safety of surrounding normal tissues. The system calls upon the dose calculation module of the treatment planning system to simulate and calculate the recommended dose regimen, assessing the dose coverage of the tumor target area and the dose distribution to organs at risk. Organs at risk in the head and neck region include the spinal cord, brainstem, parotid gland, mandible, and larynx, each with strict dose limitations. For example, the maximum dose to the spinal cord should not exceed 45 Gy, and the average dose to the parotid gland should be as low as possible below 26 Gy to preserve salivary secretion function. The system ensures that the recommended dose regimen increases the tumor dose without exceeding the dose limits of organs at risk, or optimizes the dose by adjusting the radiation field direction or using more precise radiotherapy techniques such as intensity-modulated radiotherapy (IMRT) or proton therapy.

[0087] The feedback learning unit, a crucial component of the adaptive treatment decision support module 4, is responsible for collecting clinical validation data and feeding it back to various system modules to achieve closed-loop optimization. After the patient completes the entire treatment course, the system tracks and records the actual response at the end of treatment and during follow-up. The evaluation of the actual response, based on imaging examinations, pathological examinations, and clinical observations, is the gold standard for model prediction accuracy. The system compares the predicted response with the actual response, calculating performance indicators such as prediction accuracy, sensitivity, and specificity.

[0088] When the accumulated validation data reaches a certain amount, such as 50 cases, or when performance indicators change significantly, such as an accuracy change exceeding 5%, the feedback learning unit triggers model retraining and parameter optimization. For the early response biomarker identification module 3, the system incorporates new clinical data into the training set and uses incremental learning to update the parameters of the deep learning model and support vector machine. Incremental learning is more efficient than training from scratch and can adapt to changes in the distribution of new data while retaining learned knowledge. For feature weights, the system adjusts the parameters of the minimum redundancy maximum correlation algorithm based on feedback results, prioritizing features with strong predictive ability in practical applications and discarding features with weak predictive ability. For scanning protocols, the system analyzes the contribution of different sequences and parameters to predictive performance, optimizes sequence selection and parameter settings, and minimizes scanning time and reduces patient burden while ensuring diagnostic accuracy.

[0089] Feedback learning not only improves the system's predictive performance but also enhances its adaptability to different patient groups and treatment regimens. Head and neck tumors are highly heterogeneous, with varying response patterns to radiotherapy across different anatomical locations, pathological types, and stages. Through continuous feedback learning, the system gradually accumulates empirical data covering various clinical scenarios, continuously improving the model's generalization ability. The system also supports multi-center collaborative learning, integrating data from different medical institutions to build more robust and universal predictive models.

[0090] The Adaptive Treatment Decision Support Module 4 also provides a user-friendly interface that graphically presents predicted results and treatment recommendations. The interface displays the patient's baseline and follow-up images, 3D reconstructed images of the tumor, curves showing changes in key features, predicted response probabilities and confidence intervals, recommended treatment regimens, and their dose distribution maps. Clinicians can interactively view this information and make final decisions based on their clinical experience and the patient's specific circumstances. The system also generates structured reports summarizing predicted results and recommendations, facilitating clinical discussions and medical record documentation.

[0091] A major innovative feature of this invention is the establishment of a complete closed-loop feedback mechanism, which enables not only positive data flow between modules, but also reverse optimization flow, forming a positive feedback loop of continuous improvement.

[0092] The forward data flow begins with the ultra-high field multimodal image acquisition module 1. The acquired high-quality image data is then passed to the multi-temporal precise registration and radiomics feature extraction module 2. The structured data after registration and feature extraction is then passed to the early response biomarker identification module 3. The prediction results from module 3 are then passed to the adaptive treatment decision support module 4 to generate clinical recommendations. This is the basic workflow of the system, realizing the transformation from image data to clinical decisions.

[0093] The reverse optimization flow begins at the feedback learning unit of the adaptive treatment decision support module 4. Collected clinical validation data is passed to the front-end modules in the opposite direction to the data flow, driving parameter optimization in each module. Specific feedback paths include: feedback data is passed to the early response biomarker identification module 3 to update the parameters of the deep learning model and support vector machine, improving prediction accuracy; feedback data is passed to the multi-temporal precise registration and radiomics feature extraction module 2 to adjust the weights of feature selection, prioritizing features with strong predictive power and eliminating redundant features; and feedback data is passed to the ultra-high field multimodal image acquisition module 1 to optimize scanning sequences and parameters, adjusting the acquisition scheme according to the image characteristics of patients in different response categories.

[0094] This closed-loop feedback mechanism brings about multifaceted synergistic effects. First, high-quality image data improves the accuracy of feature extraction; accurate features enhance the performance of the predictive model; reliable prediction results guide rational clinical decision-making; and each step of the positive data flow lays the foundation for the next, demonstrating the mutually reinforcing relationship between modules. Second, clinical feedback data drives continuous model optimization; the optimized model improves predictive accuracy; accurate predictions reduce decision-making errors; and correct decisions lead to better clinical outcomes; the reverse optimization flow forms a positive feedback loop of performance improvement. Third, the four stages of image acquisition, feature analysis, predictive modeling, and clinical decision-making are tightly coupled; improvements in each stage are amplified to the entire system through a cascading effect, achieving an overall prediction accuracy of 87.5%, a synergistic effect that is at least 15 percentage points higher than any single module operating alone.

[0095] In practical applications, the closed-loop feedback mechanism enables the system to adapt to different clinical environments and patient populations. Different medical institutions have varying MRI equipment models, scanning parameters, radiotherapy techniques, and patient characteristics; directly applying predictive models developed in other institutions may result in performance degradation. Through closed-loop feedback learning, the system can fine-tune the model based on its own institution's data, achieving a transformation from a general-purpose model to a specialized model, thus improving the system's clinical applicability. The system also supports a federated learning framework, allowing multiple institutions to collaboratively train the model while protecting data privacy. This leverages large sample data to improve model performance while avoiding the privacy and security risks associated with centralized data.

[0096] The system of this invention has been clinically validated in multiple medical institutions and has achieved good application results. The following examples illustrate the application process of the system in actual clinical practice.

[0097] The patient, a 58-year-old male, presented with hoarseness for 3 months and was diagnosed with stage T3N2M0 squamous cell carcinoma of the larynx. He was scheduled for radical chemoradiotherapy. Prior to treatment, a baseline scan was performed on a 7T MRI, followed by T1-weighted, T2-weighted, contrast-enhanced T1-weighted, diffusion-weighted, and magnetic resonance spectroscopy (MRS) imaging. The primary tumor was located in the supraglottic region, invading the epiglottis and arytenoid cartilage, with a volume of 12.8 cubic centimeters. Multiple metastatic lymph nodes, the largest with a diameter of 2.3 cm, were observed on the left side of the neck. Radiomic analysis showed high tumor heterogeneity and a low apparent diffusion coefficient of 0.87 × 10⁻³ mm / s. MRS spectroscopy showed a significantly elevated choline peak and a choline-to-creatine ratio of 2.8.

[0098] The patient received intensity-modulated radiotherapy (IMRT), with a prescribed dose of 70 Gy for the primary tumor and metastatic lymph nodes, and 54 Gy for the selective lymphatic drainage area, completed in 33 fractions. Concurrent cisplatin chemotherapy was administered at 40 mg / m² body surface area weekly. The first follow-up scan was performed at week 1 after radiotherapy initiation. At this time, the tumor volume was 11.9 cm³, a reduction of 7%, the apparent diffusion coefficient increased to 1.02 × 10⁻³ mm² / s, and the choline / creatine ratio decreased to 2.4. Delta characteristics were systematically calculated, but due to the small magnitude of change, the prediction made by Module 3 at this time had a low confidence level of 65%. Continued treatment was recommended, with a reassessment at week 2.

[0099] The follow-up scan at week 2 showed a tumor volume of 10.1 cubic centimeters, a 21% reduction from baseline. The apparent diffusion coefficient further increased to 1.28 × 10⁻³ mm² / s, the choline to creatine ratio decreased to 1.9, and texture features showed increased tumor homogeneity. Based on the Delta features at week 2, the system predicted an effective response with a 92% probability and a confidence interval of 87% to 96% for the early response biomarker identification module 3, indicating high confidence. The response grading unit classified the patient as sensitive. The dose adjustment recommendation unit, after analysis, concluded that the patient responded well to the standard dose and recommended maintaining the current treatment regimen without dose adjustment, but suggested moderately reducing the dose to the selective lymphatic drainage area to reduce salivary gland damage. After discussion by the multidisciplinary team, it was decided to continue the original treatment plan.

[0100] The patient successfully completed the full course of chemoradiotherapy. During treatment, the patient experienced grade 2 acute radiation mucositis and grade 2 dysphagia, which were relieved through symptomatic and supportive treatment. One month after treatment, a follow-up examination showed complete disappearance of the primary tumor and metastatic lymph nodes, achieving complete remission. Follow-ups at 3 and 6 months after treatment showed maintenance of efficacy with no signs of recurrence. The patient's actual response was completely consistent with the system's prediction at week 2, validating the system's accuracy. This patient's data was incorporated into a feedback learning database for further optimization of model performance.

[0101] Another patient, a 63-year-old female with nasopharyngeal carcinoma stage T4N3M0, also received radical chemoradiotherapy. At the second week follow-up, the system predicted a 28% effective response probability with a confidence interval of 21% to 36%, classifying her as resistant. The dose adjustment recommendation unit suggested considering treatment intensification, including increasing the primary tumor dose to 74 Gy or adding induction chemotherapy. After discussion, it was decided to add the immune checkpoint inhibitor pembrolizumab to the maintenance radiotherapy dose. After completing treatment, the primary tumor shrank significantly but did not completely disappear, and the cervical lymph nodes partially regressed, achieving partial remission. Although complete remission was not achieved, the efficacy of adding immunotherapy was improved compared to chemoradiotherapy alone. This case suggests that the system's early prediction has clinical value, enabling the identification of high-risk patients and guiding treatment intensification; however, the specific strategies and timing of treatment intensification still need further optimization.

[0102] This invention, through the aforementioned technical solution, achieves early and accurate assessment of the radiotherapy response to head and neck tumors based on 7T ultra-high field magnetic resonance imaging, providing timely and effective treatment decision support for clinicians. The system integrates advanced imaging technology, intelligent analysis methods, and clinical decision-making tools, demonstrating significant innovation and practicality.

[0103] Although the present invention has been described in detail with reference to the accompanying drawings and embodiments, those skilled in the art should understand that various modifications and substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the invention, and all such modifications and substitutions should fall within the scope defined by the claims of the present invention. For example, the system can be extended to the assessment of radiotherapy response in tumors in other locations, the deep learning model architecture used can be adjusted according to specific needs, and the triggering conditions and update strategies for feedback learning can be optimized based on clinical practice; these changes do not deviate from the core ideas of the present invention.

Claims

A 1.7T magnetic resonance imaging (MRI) system for evaluating the response to radiotherapy for head and neck tumors, characterized in that... include: The ultra-high field multimodal image acquisition module is used to acquire 7T magnetic resonance images of head and neck tumors at multiple time points before and during treatment. The images include morphological sequences, diffusion-weighted imaging sequences, and magnetic resonance spectroscopy imaging sequences. The morphological sequences include T1-weighted imaging, T2-weighted imaging, and contrast-enhanced T1-weighted imaging. A multi-temporal precise registration and radiomics feature extraction module, connected to the ultra-high field multimodal image acquisition module, is used to perform sub-millimeter-level precise registration of images acquired at different time points and extract radiomics features of the tumor region to form a multi-dimensional feature set. The radiomics features include first-order statistical features, second-order texture features, higher-order wavelet features, shape-based geometric features, and metabolic features. The first-order statistical features include the mean, standard deviation, skewness, kurtosis, maximum, minimum, median, and quartiles of pixel intensity values ​​within the tumor region. The second-order texture features are obtained based on the gray-level co-occurrence matrix and gray-level run matrix. The higher-order wavelet features are extracted after the image is decomposed into sub-bands of different frequencies and directions by wavelet transform. The shape-based geometric features include the tumor volume, surface area, sphericity, elongation, and flatness. The metabolic features are the concentration characteristics of choline, creatine, N-acetylaspartate, and lactate extracted from the magnetic resonance spectroscopy imaging sequence. The early response biomarker identification module is connected to the multi-temporal precise registration and radiomics feature extraction module. It is used to calculate the changes in features before and after treatment. Through a deep learning model, it identifies feature combinations related to treatment response. In the second to third week after the start of radiotherapy, when the sum of the longest diameters of the target tumor lesions has not yet reached the RECIST 1.1 partial remission threshold of 30%, the module predicts the degree of tumor response to radiotherapy in advance based on the changes in features of functional imaging and metabolic imaging. An adaptive treatment decision support module, connected to the early response biomarker identification module, is used to generate treatment adjustment suggestions based on the predicted response level and feed back the clinical validation results to the early response biomarker identification module to form a closed-loop optimization mechanism. The clinical validation results are used to optimize model parameters, adjust feature weights, and guide subsequent scanning protocols.

2. The system according to claim 1, characterized in that: The ultra-high field multimodal image acquisition module includes a sequence parameter optimization unit and a data quality monitoring unit; The sequence parameter optimization unit is used to determine the scanning parameters of each sequence based on patient characteristics and treatment stage. The scanning parameters include repetition time, echo time, flip angle, and spatial resolution. The data quality monitoring unit is used to evaluate the signal-to-noise ratio and contrast-to-noise ratio of the acquired images in real time, and triggers re-acquisition when the signal-to-noise ratio or contrast-to-noise ratio is lower than a preset threshold.

3. The system according to claim 1, characterized in that: The multi-temporal precise registration and radiomics feature extraction module adopts a non-rigid registration algorithm based on mutual information, achieving a registration accuracy of less than 0.5 mm in average position error and less than 0.8 mm in position error for 95% of the marker points. The radiomics features include more than 200 quantitative features, covering first-order statistical features, second-order texture features, higher-order wavelet features, and shape-based geometric features; The metabolic characteristics include the choline to creatine ratio, N-acetylaspartate concentration, and lactate concentration extracted from magnetic resonance spectroscopy imaging.

4. The system according to claim 1, characterized in that: The early response biomarker identification module includes a feature change calculation unit, a response prediction model unit, and a confidence assessment unit. The feature change calculation unit is used to calculate the difference between the image features before treatment and the image features at each time point during treatment, forming a Delta feature set; The response prediction model unit adopts a hybrid model combining convolutional neural networks and support vector machines. Based on the Delta feature set, it predicts the response category of the tumor at the end of treatment and its corresponding probability. The response categories include complete remission, partial remission and disease progression. Partial remission is defined as a decrease of at least 30% in the sum of the longest diameters of the target lesions compared to the baseline. Disease progression is defined as an increase of at least 20% in the sum of the longest diameters of the target lesions compared to the minimum value or the appearance of new lesions. The confidence assessment unit is used to calculate the confidence interval of the prediction result, and triggers an increase in the scanning frequency when the confidence level is lower than 80%.

5. The system according to claim 4, characterized in that: The convolutional neural network of the response prediction model unit includes 5 convolutional layers and 3 fully connected layers, and adopts ReLU activation function and Dropout regularization; The support vector machine uses a radial basis function kernel, and the kernel parameters are determined through cross-validation. The hybrid model combines the probability output of the convolutional neural network and the decision function value of the support vector machine in a weighted manner, with the weights dynamically adjusted based on the performance of each model on the validation set.

6. The system according to claim 1, characterized in that: The adaptive treatment decision support module includes a response grading unit, a dose adjustment suggestion unit, and a feedback learning unit; The response grading unit classifies patients into three categories—sensitive, moderate, and resistant—based on the predicted response level. The dose adjustment suggestion unit generates dose reduction suggestions for sensitive patients and dose increase or target area modification suggestions for resistant patients. The feedback learning unit collects actual response data after treatment and calculates prediction accuracy. When the accuracy changes by more than 5%, the model is retrained.

7. The system according to claim 1, characterized in that: The system acquires baseline images before treatment and follow-up images at the 1st, 2nd, and 3rd weeks after the start of radiotherapy and at the end of treatment, with each acquisition time not exceeding 25 minutes. The early response biomarker identification module predicts the degree of response at the end of treatment based on Delta features at week 2, with the prediction time point being 3 to 4 weeks earlier than imaging assessment based on RECIST 1.1 standard morphological volume changes.

8. The system according to claim 1, characterized in that: The closed-loop optimization mechanism includes a parameter reverse adjustment process and a model adaptive update process. The parameter reverse adjustment process adjusts the feature selection weights of the multi-temporal precise registration and radiomics feature extraction modules based on clinical validation results. The model adaptive update process incrementally learns the deep learning model of the early response biomarker recognition module based on accumulated clinical data, with an update frequency of once every 50 patient data.

9. The system according to claim 1, characterized in that: The system is suitable for evaluating the response to radiotherapy in nasopharyngeal carcinoma, oropharyngeal carcinoma, laryngeal carcinoma, and hypopharyngeal carcinoma. The radiotherapy methods include radiotherapy alone, concurrent chemoradiotherapy, and radiotherapy after induction chemotherapy.

10. The system according to claim 1, characterized in that: The ultra-high field multimodal image acquisition module, with an isotropic resolution of 0.5 mm and a signal-to-noise ratio of not less than 30, ensures that the average positional error of the registration between the multi-temporal precise registration and the radiomics feature extraction module does not exceed 0.5 mm, and supports the precise extraction of more than 200 radiomics features. The multi-dimensional features of the multi-temporal precise registration and radiomics feature extraction module enhance the predictive ability of the early response biomarker identification module; The prediction results of the early response biomarker identification module guide the adaptive treatment decision support module to optimize the treatment plan, and at the same time provide feedback information to optimize the parameter settings of the front-end module. The clinical validation results of the adaptive treatment decision support module are used to inversely optimize the operating parameters of each module in the entire system, forming a positive feedback loop of continuous improvement.

Images