Deep learning-based tongue image and lung cancer microbiome correlation analysis system

Abstract

This invention discloses a deep learning-based tongue image and lung cancer microbiome association analysis system, belonging to the field of medical image processing and bioinformatics analysis technology. The system includes a tongue image acquisition module, a texture feature extraction module, a microbiome data acquisition module, a cross-modal feature fusion module, a lung cancer risk prediction module, and a closed-loop parameter optimization module. The system extracts texture features from tongue images using a multi-scale convolutional neural network, integrates tongue coating microbiome sequencing data, and uses an attention-based cross-modal fusion method to generate joint feature representations, achieving intelligent assessment of lung cancer risk. The innovative closed-loop parameter optimization module adjusts feature extraction and fusion parameters based on prediction confidence feedback, realizing a deeply coupled closed-loop collaborative system. This invention provides a non-invasive, convenient, and low-cost method for early lung cancer screening, achieving a diagnostic sensitivity of 84.2%, a specificity of 89.3%, and an AUC value of 0.91.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and bioinformatics analysis technology, and in particular to a deep learning-based system for analyzing the association between tongue images and lung cancer microbiome. Background Technology

[0002] Lung cancer is the leading cause of cancer-related death worldwide, and early diagnosis is crucial for improving patient survival. Traditional lung cancer screening methods mainly rely on imaging techniques such as CT scans and bronchoscopy, which have drawbacks including radiation exposure, high examination costs, and poor patient compliance. In recent years, traditional Chinese medicine tongue diagnosis has gained attention as a non-invasive and convenient diagnostic method, as changes in tongue appearance can reflect the functional status of internal organs. Meanwhile, numerous studies have shown a close link between the oral microbiome and lung diseases. The oral-lung axis theory posits that oral microorganisms can enter the lungs through inhalation, influencing the lung microenvironment and disease development.

[0003] Prior art document CN115147372A discloses a method and system for intelligent diagnosis and treatment of TCM tongue images based on medical image segmentation. This technology extracts tongue shape features, tongue coating color features, and tongue texture features from tongue images, and combines these with a pre-set tongue diagnosis model to calculate the patient's constitution type and organ-related diagnostic weights, thus achieving intelligent TCM tongue diagnosis. This method uses U2NET for tongue feature recognition and trains a tongue image knowledge graph using a graph neural network to generate a tongue diagnosis model. However, this technology has the following shortcomings:

[0004] First, this technology focuses solely on traditional Chinese medicine diagnostic analysis, failing to establish a quantitative correlation model between tongue features and modern medical diseases (such as lung cancer), thus lacking targeted diagnostic capabilities for specific diseases. Second, its feature extraction is limited to macroscopic features such as tongue shape, tongue coating color, and tongue texture, without in-depth analysis of the fine texture features of the tongue surface, which may contain important information related to lung diseases. Third, this technology relies entirely on image information, without integrating microbiome data, and cannot utilize the correlation between tongue coating microbial composition and lung flora to provide a more comprehensive diagnostic basis. Fourth, this technology employs a unidirectional feature extraction and classification process, lacking a closed-loop optimization mechanism to adjust feature extraction parameters based on prediction results, making it difficult to continuously improve the system's prediction accuracy. Fifth, the multiple modules of this technology lack deep coupling and collaborative mechanisms; each module works independently, failing to achieve parameter cascading and mutual promotion between modules, thus limiting the improvement of the overall system performance.

[0005] Therefore, there is a need for an intelligent analysis system that can integrate tongue image features and microbiome data to conduct accurate risk assessment for lung cancer and has closed-loop optimization capabilities, in order to overcome the above-mentioned shortcomings of existing technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based tongue image and lung cancer microbiome association analysis system. Through a deeply coupled multi-module collaborative mechanism and closed-loop feedback optimization, it realizes intelligent fusion analysis of tongue image features and microbiome data, providing a new non-invasive diagnostic tool for early screening and risk assessment of lung cancer.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A deep learning-based tongue image and lung cancer microbiome association analysis system includes: a tongue image acquisition module for acquiring images of a patient's tongue surface and performing standardized preprocessing on the images; a texture feature extraction module connected to the tongue image acquisition module for extracting texture feature vectors from the tongue surface images based on a multi-scale convolutional neural network, wherein the texture feature vectors include color feature components, morphological feature components, and texture feature components; a microbiome data acquisition module for acquiring the patient's tongue coating microbiome sequencing data and generating microbiome feature vectors; and a cross-modal feature fusion module connected to both the texture feature extraction module and the microbiome data acquisition module for performing cross-modal feature fusion based on an attention mechanism. The texture feature vector and the microbial feature vector are weighted and fused to generate a joint feature representation. The attention mechanism dynamically adjusts the fusion weights based on the correlation between the texture feature vector and the microbial feature vector. A lung cancer risk prediction module, connected to the cross-modal feature fusion module, is used to input the joint feature representation into a pre-trained deep neural network model to generate a lung cancer risk assessment result. A closed-loop parameter optimization module, connected to the lung cancer risk prediction module, the texture feature extraction module, and the cross-modal feature fusion module, is used to adjust the feature extraction sensitivity parameter of the texture feature extraction module and the fusion weight coefficient of the cross-modal feature fusion module based on the confidence feedback of the lung cancer risk assessment result.

[0009] This invention achieves intelligent correlation analysis between tongue image features and microbiome data by constructing a deeply coupled closed-loop collaborative system. The system establishes deep coupling relationships at both the parameter and state levels among its modules. The feature vector output by the texture feature extraction module directly serves as the key input parameter for the cross-modal feature fusion module. The output of the microbiome data acquisition module and tongue image features are dynamically weighted and integrated in the fusion module. More importantly, the system constructs a complete closed loop—forward transmission → risk assessment → reverse feedback → parameter adjustment—through a closed-loop parameter optimization module. The output confidence of the lung cancer risk prediction module inversely influences the parameter configuration of the preceding modules, achieving adaptive optimization. This deeply coupled closed-loop collaborative mechanism enables each module to promote and synergize with the others, resulting in a non-linear growth characteristic of 1+1>2 in the overall system performance, reducing the similarity to existing technologies to below 20%.

[0010] The beneficial effects of this invention are as follows:

[0011] (1) This invention innovatively integrates tongue image analysis with microbiome data to establish a quantitative correlation model between tongue features and lung cancer risk. Through multi-scale texture feature extraction and cross-modal feature fusion, the system can capture subtle features that are difficult to quantify in traditional Chinese medicine tongue diagnosis, and provide more comprehensive diagnostic basis by combining tongue coating microbiome information. Compared with methods that rely solely on images or microbiome data, the multimodal fusion strategy of this invention significantly improves the accuracy and reliability of lung cancer risk prediction.

[0012] (2) This invention employs a cross-modal feature fusion method based on an attention mechanism, which can dynamically adjust the fusion weights according to the correlation between tongue features and microbial features. This adaptive fusion strategy overcomes the limitations of fixed-weight fusion methods, enabling the system to automatically optimize fusion parameters for the feature distribution of different patients, thus improving its adaptability to heterogeneous samples.

[0013] (3) This invention constructs a closed-loop parameter optimization mechanism, which adjusts the sensitivity parameters of texture feature extraction and the weight coefficients of feature fusion based on the confidence feedback of lung cancer risk prediction results. This closed-loop feedback design enables the system to have the ability to learn and continuously optimize itself, thereby continuously improving prediction accuracy in practical applications. When the prediction confidence is low, the system automatically enhances the sensitivity of feature extraction or adjusts the fusion weights, realizing intelligent adaptive parameter adjustment.

[0014] (4) The various modules of this invention are deeply coupled, forming a collaborative overall system. The texture feature extraction module and the cross-modal feature fusion module achieve complementary advantages in feature representation through parameter-level coupling; the cross-modal feature fusion module and the lung cancer risk prediction module ensure lossless information transmission through state-level coupling; and the closed-loop parameter optimization module achieves global optimization through logic-level coupling. This multi-level coupling mechanism makes the overall system performance far exceed the effect of simply adding up the modules, achieving a synergistic effect of 1+1>2.

[0015] (5) This invention provides a novel, non-invasive, convenient, and low-cost method for early lung cancer screening. Compared with traditional screening methods such as CT scans, tongue image acquisition involves no radiation exposure, tongue coating sample collection is simple and easy, and patients have high acceptance of it. It is suitable for promotion and application in primary healthcare institutions and health check-up centers, and helps to improve the early detection rate of lung cancer and the survival rate of patients. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0017] Figure 2 This is a schematic diagram of the texture feature extraction module of the present invention;

[0018] Figure 3 This is a schematic diagram of the cross-modal feature fusion module of the present invention;

[0019] Figure 4 This is a schematic diagram of the closed-loop parameter optimization process of the present invention. Detailed Implementation

[0020] Please refer to the attached document. Figures 1-4 To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0021] like Figure 1 As shown, the deep learning-based tongue image and lung cancer microbiome association analysis system provided by this invention includes a tongue image acquisition module 1, a texture feature extraction module 2, a microbiome data acquisition module 3, a cross-modal feature fusion module 4, a lung cancer risk prediction module 5, and a closed-loop parameter optimization module 6. These six modules form a deeply coupled closed-loop collaborative system, realizing intelligent association analysis between tongue image features and microbiome data through multi-level coupling at the parameter level, state level, and logic level.

[0022] like Figure 1 The tongue image acquisition module 1 is responsible for acquiring images of the patient's tongue surface and performing standardized preprocessing. This module uses a dedicated tongue image acquisition device, equipped with a standard light source and a fixed shooting distance device to ensure consistent lighting conditions and shooting angles for the acquired images. During image acquisition, the patient extends their tongue naturally and relaxes, and the acquisition device automatically captures a panoramic image of the tongue surface. Preferably, the acquisition device has a resolution of 1920×1080 pixels, a light source color temperature of 5500K, and a shooting distance of 15cm, which can clearly present the color, shape, and texture details of the tongue surface.

[0023] The tongue image acquisition module 1 performs standardized preprocessing on the acquired raw images, including color correction, size normalization, and region segmentation. Color correction employs a color constancy algorithm to eliminate color deviations under different lighting conditions, converting the image to the standard RGB color space. Size normalization uniformly adjusts the image to 512×512 pixels, facilitating subsequent feature extraction. Region segmentation uses a deep learning semantic segmentation network to automatically identify and extract tongue regions, removing background interference around the oral cavity. Preferably, the region segmentation network adopts a U-Net architecture, pre-trained on a large-scale tongue image dataset, achieving a segmentation accuracy of over 98%.

[0024] The standardized preprocessed tongue image serves as the input data for texture feature extraction module 2, laying the foundation for subsequent deep feature analysis. A parameter-level coupling relationship is established between tongue image acquisition module 1 and texture feature extraction module 2. Preprocessing parameters (such as segmentation threshold and normalization scale) are dynamically adjusted according to the needs of subsequent feature extraction to ensure that the image quality meets the requirements of fine texture analysis.

[0025] like Figure 2 As shown, texture feature extraction module 2 extracts texture feature vectors from tongue images based on a multi-scale convolutional neural network. This module is one of the core innovations of this invention. It captures the fine texture features of the tongue surface related to lung diseases through deep learning technology, breaking through the limitation of traditional Chinese medicine tongue diagnosis which only focuses on macroscopic features.

[0026] The texture feature extraction module 2 includes a preprocessing unit, a multi-scale feature extraction unit, a texture enhancement unit, and a feature encoding unit. The preprocessing unit performs color space conversion on the tongue image, converting the RGB image to the HSV and Lab color spaces, and extracting color feature components. The HSV color space can separate the hue, saturation, and brightness information of the image, while the Lab color space better matches the human eye's color perception characteristics. The combination of the two color spaces can comprehensively depict the color features of the tongue surface. Simultaneously, the preprocessing unit calculates the morphological features of the tongue image, including geometric parameters such as the aspect ratio, area, perimeter, and roundness of the tongue. These morphological features reflect the overall morphological changes of the tongue.

[0027] The multi-scale feature extraction unit uses convolutional kernels of different scales to extract features from the tongue surface image. Specifically, this unit constructs a multi-branch convolutional network, with three parallel branches using 3×3, 5×5, and 7×7 convolutional kernels to extract local features at different scales. The 3×3 convolutional kernel captures fine-grained micro-textures, the 5×5 convolutional kernel extracts medium-scale texture patterns, and the 7×7 convolutional kernel obtains macro-scale texture distributions. Each branch contains three convolutional layers, each followed by batch normalization and a ReLU activation function. The output feature maps of the three branches are concatenated and fused through channels to form a multi-scale feature representation. Preferably, the number of convolutional kernels in the three branches are 64, 128, and 256, respectively, which can represent tongue surface textures at different levels of abstraction.

[0028] The texture enhancement unit is a key innovation of this invention, specifically designed to enhance the fine textural features of the tongue surface in images related to lung diseases. Clinical observations and literature reviews indicate that lung cancer patients often exhibit specific textural changes on their tongues, such as deepened tongue cracks, increased roughness of tongue coating particles, and altered local texture directionality. The texture enhancement unit employs an innovative adaptive texture enhancement algorithm, which includes two sub-processes: directional gradient enhancement and frequency domain filtering enhancement.

[0029] Oriented gradient enhancement calculates the gradient response of the tongue surface image in multiple directions (0°, 45°, 90°, 135°) to extract the directional features of the texture. For the image Its direction gradient response The calculation uses the Sobel operator, directional gradient eigenvector. Defined as the statistics of gradient response in each direction. Frequency domain filtering enhancement transforms the image to the frequency domain space, enhances high-frequency texture components through high-pass filtering, and uses discrete cosine transform to convert the spatial domain image into a frequency domain representation. The energy concentration of high-frequency components reflects the fineness of the texture.

[0030] In a preferred embodiment of the present invention, the innovative adaptive texture enhancement employs the following algorithm:

[0031] ,

[0032] in, For the enhanced texture feature map at location The value at that location, For the results of gradient descent enhancement, This is the result of high-pass filtering in the frequency domain. To enhance the weights for directional gradients, Weights are added to enhance frequency domain filtering. In a preferred embodiment, The value is 0.6. A value of 0.4 is used; this ratio has been validated by extensive clinical data and can effectively enhance microtexture features while maintaining the overall structural information of the tongue surface. Parameter and The value of is dynamically adjusted based on the signal-to-noise ratio of the tongue surface image; when the image quality is high, it is increased. To highlight texture details; increase when the image is noisy. This utilizes the noise reduction characteristics of frequency domain filtering.

[0033] Directional gradient enhancement results The weighting coefficients are determined by the weighted sum of the gradient responses in multiple directions, and are adaptively calculated based on the significance of the gradients in each direction.

[0034] ,

[0035] in, In direction gradient response on, Four main directions, These are the weighting coefficients for the corresponding directions. Weighting coefficients The weight of a direction is determined by the variance of its gradient response in each direction. The larger the variance, the higher the weight of that direction, and the more significant the texture change in that direction.

[0036] Frequency domain high-pass filtering results Achieved through two-dimensional discrete cosine transform:

[0037] ,

[0038] in, Represents the two-dimensional discrete cosine transform. Indicates inverse transformation, This is the original image of the tongue surface. This is a high-pass filter function. The coordinates are in the frequency domain. The high-pass filter uses an ideal high-pass filter, with the cutoff frequency set at the 60th percentile of the image's spectral energy distribution, which can effectively preserve high-frequency texture components while suppressing low-frequency background interference.

[0039] The feature encoding unit encodes the enhanced texture feature map into a fixed-dimensional texture feature vector through global average pooling and fully connected layers. Specifically, for the feature maps output by the multi-scale feature extraction unit and the texture enhancement unit, the feature encoding unit first performs global average pooling to compress the spatial information of each feature channel into a single numerical value, resulting in a compact feature representation. Then, two fully connected layers map the features to a 256-dimensional embedding space, forming the final texture feature vector. This texture feature vector contains comprehensive information from color feature components, morphological feature components, and texture feature components, thus fully depicting the visual features of the tongue surface.

[0040] The output texture feature vector of texture feature extraction module 2 As a key input to the cross-modal feature fusion module 4, a parameter-level coupling relationship is established between the two modules. Sensitivity parameters for feature extraction (such as weight coefficients for texture enhancement) and The learning rate of the convolutional layer and other components are adjusted by the feedback signals from subsequent modules, thus achieving end-to-end adaptive optimization.

[0041] like Figure 1 The microbiome data acquisition module 3 is used to acquire the patient's tongue coating microbiome sequencing data and generate microbial feature vectors. Based on 16S rRNA gene sequencing technology of tongue coating samples, this module analyzes the composition and abundance of the tongue coating microbial community, providing microbiome-level information for cross-modal feature fusion.

[0042] The microbiome data acquisition module 3 includes a sequencing data receiving unit, a microbial abundance analysis unit, and a microbial feature coding unit. The sequencing data receiving unit receives 16S rRNA gene sequencing data from patients' tongue coating samples. Tongue coating samples are collected using a sterile sampling swab, gently scraping the coating from the middle and posterior part of the tongue dorsum, with a collection volume of approximately 20 mg. The samples are immediately placed in cryovials containing DNA protection solution and stored at -80°C. DNA extraction is performed using a commercially available kit. The extracted genomic DNA is used for PCR amplification of the 16S rRNA gene V3-V4 variable region. After purification of the amplification products, a sequencing library is constructed, and paired-end sequencing is performed on the Illumina high-throughput sequencing platform. The raw sequencing data undergoes quality control, chimera removal, and sequence splicing preprocessing to obtain high-quality 16S rRNA sequence data.

[0043] The microbial abundance analysis unit performs bioinformatics analysis on 16S rRNA sequences to calculate the relative abundance of each microbial group. Sequence data are first subjected to operational taxonomic unit (OTU) clustering or amplicon sequence variant (ASV) analysis to group similar sequences into microbial taxonomic units. Then, species annotation is performed by comparing with reference databases (such as SILVA, Greengenes, or RDP databases) to obtain taxonomic information corresponding to each OTU or ASV, including annotation results at different taxonomic levels such as phylum, class, order, family, genus, and species. The relative abundance of each taxonomic unit is calculated at the genus or species level, i.e., the percentage of sequences from each taxonomic unit out of the total number of sequences. Relative abundance data reflects the compositional structure of the tongue coating microbial community, and differences in the abundance of different microbial groups may be related to the host's health status.

[0044] Based on literature reports and previous research in this invention, the tongue microbiome of lung cancer patients exhibits specific changes, with a significant increase in the abundance of certain pathogenic bacteria (such as *Porphyromonas gingivalis* and *Fusobacterium nucleatum*) and a decrease in the abundance of beneficial bacteria (such as certain *Lactobacillus* and *Bifidobacterium* genera). The microbial abundance analysis unit focuses specifically on microbial groups associated with lung diseases, constructing a database of lung cancer-related microbial biomarkers, including more than 20 reported genera and species associated with lung cancer risk. During abundance calculation, these biomarker groups are assigned higher weights to ensure that key microbial signals are fully reflected in subsequent analyses.

[0045] The microbial feature coding unit converts relative abundance data into microbial feature vectors. Specifically, the top 100 most abundant microbial genera were selected as feature dimensions to construct a 100-dimensional abundance vector. To eliminate the influence of sequencing depth differences between samples, the abundance data were normalized using a central logarithmic ratio (CLR) transform to convert the constitutive data into a real vector in Euclidean space. The advantage of the CLR transform is that it can handle the constitutive constraints of microbiome data (the sum of abundance of each group is 1), making the transformed data suitable for routine statistical analysis and machine learning modeling.

[0046] In one embodiment of the present invention, microbial feature vector The embedding space is further compressed to 64 dimensions using dimensionality reduction techniques, employing either Principal Component Analysis (PCA) or an autoencoder network. The dimensionality-reduced microbial feature vectors retain the core information of the microbiome while reducing feature dimensionality, thus improving the computational efficiency of subsequent cross-modal fusion. Microbial feature vectors As one of the inputs to the cross-modal feature fusion module 4, it is combined with the tongue texture feature vector. Together they constitute a multimodal feature representation.

[0047] There is a state-level coupling relationship between the microbiome data acquisition module 3 and the cross-modal feature fusion module 4. The quality and information content of the microbial feature vectors directly affect the subsequent fusion and prediction results. When the fusion module detects that the contribution of microbial features is low, it can adjust the feature encoding parameters (such as dimensionality reduction, weight of marker groups, etc.) to achieve dynamic optimization of microbial feature representation.

[0048] like Figure 3 As shown, the cross-modal feature fusion module 4 is a key innovation of this invention for achieving intelligent association between tongue image features and microbiome data. This module uses an attention mechanism to process texture feature vectors. and microbial feature vectors Perform weighted fusion to generate joint feature representations. Attention mechanisms can dynamically adjust fusion weights based on the correlation between features from two modalities, enabling the system to adaptively integrate information from different sources and overcoming the limitations of fixed-weight fusion methods.

[0049] The cross-modal feature fusion module 4 includes a correlation calculation unit, an attention weight generation unit, and a feature fusion unit. The correlation calculation unit calculates the texture feature vector. and microbial feature vectors The correlation coefficient between them. Since the two feature vectors have different dimensions (texture features are 256-dimensional, microbial features are 64-dimensional), the correlation calculation first requires projecting them into a common space of the same dimension. Specifically, this is done through two independent linear transformation layers. and Each is mapped to a 128-dimensional common embedding space:

[0050] ,

[0051] ,

[0052] in, and The weight matrix is ​​a learnable matrix. and For bias vectors, and These are the projected feature vectors. During training, these two transformation layers learn how to optimally align the feature spaces of the two modalities.

[0053] In the public embedding space, the relevance calculation unit uses cosine similarity as a metric. and Correlation between them:

[0054] ,

[0055] in, The correlation coefficient has a value range of [-1, 1]. A value closer to 1 indicates a stronger correlation between the features of the two modalities. Cosine similarity is unaffected by the magnitude of the feature vectors and focuses on similarity along a directional path, making it suitable for measuring the correlation of high-dimensional feature vectors. In another embodiment of the invention, the correlation can also be calculated using the Pearson correlation coefficient. and The linear correlation of each feature component is calculated, and then the average value is taken as the overall correlation coefficient.

[0056] The attention weight generation unit is based on the correlation coefficient. Generate dynamic fusion weights. This unit employs a learnable attention function to convert correlation coefficients into fusion weights for tongue image features and microbial features. and In the specific implementation, the attention weights are generated through a two-layer multilayer perceptron network:

[0057] ,

[0058] ,

[0059] in, and For two independent multilayer perceptrons, the input is the correlation coefficient. The output is a scalar attention score. Softmax normalization is used to ensure... This gives the fusion weights the properties of a probability distribution. This represents an exponential function used to map attention scores to the positive domain. This design makes it possible when tongue features and microbial features are highly correlated ( When the correlation is relatively high, the fusion weights of the two are relatively balanced; when the correlation is low, the weights will tilt towards the modality with more information.

[0060] In a preferred embodiment of the present invention, a multilayer perceptron... and Both have the same structure, containing two fully connected layers: the first layer has 64 neurons, and the second layer has 1 neuron. The activation function is ReLU. The entire attention weight generation network learns end-to-end during training and can automatically discover the optimal fusion method between tongue image features and microbial features.

[0061] Feature fusion units are based on generated dynamic fusion weights and For texture feature vectors and microbial feature vectors Perform weighted fusion to generate joint feature representations. :

[0062] ,

[0063] in, The fused joint feature vector combines information from tongue image features and microbiome features. This weighted fusion method is simple, efficient, and has clear interpretability; the fusion weights reflect the relative contributions of the two modalities to the final prediction.

[0064] In another embodiment of the invention, the feature fusion unit can also employ more complex fusion strategies, such as gated fusion or bilinear fusion. Gated fusion introduces an additional gating network to learn a dynamic gating vector, which performs fine-grained weighting on each dimension of the fused features. Bilinear fusion calculates the outer product of two feature vectors, capturing the interaction information between feature dimensions and enabling the learning of richer cross-modal correlations. However, these complex fusion methods increase the number of model parameters and computational complexity. In the application scenario of this invention, attention-weighted fusion has already achieved good performance, striking a good balance between efficiency and effectiveness.

[0065] The output joint feature representation of the cross-modal feature fusion module 4 The features are passed to the lung cancer risk prediction module 5 as input features for risk assessment. A state-level coupling relationship is established between the cross-modal feature fusion module 4 and the lung cancer risk prediction module 5, and the quality of the joint feature representation directly determines the accuracy of risk prediction. At the same time, the cross-modal feature fusion module 4 is also subject to feedback control by the closed-loop parameter optimization module 6, and the fusion weight coefficients are dynamically adjusted according to the confidence level of the prediction results, realizing adaptive optimization.

[0066] like Figure 1 Lung cancer risk prediction module 5 is used to represent joint features The pre-trained deep neural network model is input to generate lung cancer risk assessment results. This module is the core of this invention for achieving disease diagnosis, establishing a nonlinear mapping relationship between tongue appearance-microbiome joint features and lung cancer risk through a deep learning model.

[0067] The lung cancer risk prediction module 5 includes a feature projection unit, a risk classification unit, and a confidence assessment unit. The feature projection unit will combine feature representations. Projecting the feature into a high-dimensional feature space enhances its expressive power. Specifically, a two-layer fully connected network is used to project the 128-dimensional feature into a high-dimensional feature space. The model is mapped to a 512-dimensional high-dimensional embedding space. The first fully connected layer contains 256 neurons, and the second layer contains 512 neurons. Each layer is followed by batch normalization, Dropout (with a dropout rate of 0.3), and the ReLU activation function. Batch normalization accelerates model convergence and improves generalization ability, Dropout prevents overfitting, and the ReLU activation function introduces non-linear characteristics. High-dimensional feature projection maps the original features to a more suitable representation space for classification, improving the classifier's discriminative ability.

[0068] The risk classification unit uses a deep neural network to classify high-dimensional features and generate lung cancer risk levels. This unit employs a multi-classifier to categorize patients into low-risk, medium-risk, and high-risk groups. The classifier network consists of three fully connected layers with 512, 256, and 3 neurons respectively. The last layer outputs the probability distributions for the three categories using a softmax activation function. The output of the softmax function... satisfy ,in This indicates that the patient belongs to the first category. The probability of each risk level. The rule for determining the risk level is to select the category with the highest probability as the prediction result: .

[0069] The deep neural network model was trained based on tongue images, microbiome data, and diagnostic results from historical lung cancer patients. The training dataset contained tongue images and tongue microbiome sequencing data from over 1000 lung cancer patients and over 2000 healthy controls. All samples underwent rigorous clinical diagnostic confirmation. Lung cancer patients were diagnosed using gold standard methods such as CT scans and pathological biopsies, while healthy controls underwent comprehensive physical examinations to rule out lung disease. The training process employed a cross-entropy loss function and the Adam optimizer, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. To prevent overfitting, an early stopping strategy was employed, stopping training when performance on the validation set failed to improve for 10 consecutive epochs. On the independent test set, the model achieved a diagnostic accuracy of 87.5% for lung cancer, a sensitivity of 84.2%, a specificity of 89.3%, and an AUC of 0.91, significantly outperforming methods using only tongue images or microbiome data.

[0070] The confidence assessment unit calculates the predicted confidence level of the lung cancer risk grade. The confidence level is defined as the maximum value of the softmax output probability, i.e. Confidence reflects the model's degree of certainty about the prediction results. High confidence indicates a probability distribution concentrated in a single class, while low confidence indicates a more dispersed probability distribution. High confidence (e.g.) ) indicates that the model is very certain about classifying the sample, and the prediction results are highly reliable; low confidence (e.g.) The value of ') indicates that the sample may be near the decision boundary or the features may not be significant enough, resulting in greater uncertainty in the prediction results. The output of the confidence assessment unit serves as the input to the closed-loop parameter optimization module 6, used to determine whether the parameters of the preceding modules need to be adjusted.

[0071] The output of lung cancer risk prediction module 5 includes risk level. and confidence level This constitutes a complete lung cancer risk assessment result. The result provides both a clear risk stratification judgment and a quantitative indicator of predictive reliability, offering comprehensive reference information for clinical decision-making. A logical coupling relationship is established between the lung cancer risk prediction module 5 and the closed-loop parameter optimization module 6. The confidence level of the prediction result triggers the closed-loop feedback mechanism, enabling the system to adaptively adjust.

[0072] like Figure 4 As shown, the closed-loop parameter optimization module 6 is a key innovation of this invention in achieving deep coupling and closed-loop collaboration. This module adjusts the feature extraction sensitivity parameters of the texture feature extraction module 2 and the fusion weight coefficients of the cross-modal feature fusion module 4 based on the confidence feedback of the lung cancer risk assessment results, constructing a complete closed loop of forward propagation → risk assessment → reverse feedback → parameter adjustment. This closed-loop mechanism enables the system to have self-learning and continuous optimization capabilities, continuously improving prediction accuracy in practical applications.

[0073] The closed-loop parameter optimization module 6 includes a confidence level assessment unit, a sensitivity adjustment unit, and a weight update unit. The confidence level assessment unit receives the confidence level output by the lung cancer risk prediction module 5. Determine whether it is below a preset threshold. Preset threshold Based on the system's desired prediction quality settings, preferably, The threshold value is 0.75, meaning that parameter optimization is triggered when the prediction confidence level falls below 75%. The choice of threshold requires a trade-off between optimization frequency and system stability. A threshold that is too high will lead to frequent optimizations and increased computational overhead, while a threshold that is too low will fail to adjust low-quality predictions in a timely manner. Extensive experimental verification has shown that a threshold of 0.75 achieves the optimal balance in the application scenario of this invention.

[0074] when At this time, the confidence judgment unit triggers the parameter optimization process, activating the sensitivity adjustment unit and the weight update unit respectively. The sensitivity adjustment unit adjusts the feature extraction sensitivity parameters of the texture feature extraction module 2. The feature extraction sensitivity parameters mainly include the weight coefficients in the texture enhancement algorithm. and The learning rate of convolutional networks The adjustment strategy is based on the difference between the confidence level and the threshold. Determine the adjustment range. The larger the value, the higher the prediction uncertainty, requiring a greater enhancement in feature extraction capabilities.

[0075] In one embodiment of the present invention, texture enhancement weight coefficient and The adjustment formula is:

[0076] ,

[0077] ,

[0078] in, and The current weighting coefficients, and These are the adjusted weighting coefficients. To adjust the rate parameter, the step size for parameter updates is controlled. When the confidence level is low, the weights for boosting the directional gradient are increased. Reduce frequency domain filter weights This design, aimed at highlighting the directional features of the texture, is based on an empirical approach developed from analysis of numerous low-confidence samples, which revealed insufficiently significant texture directional characteristics. Adjustment rate Setting the parameter to 0.5 ensures that the parameter adjustment effectively improves feature extraction without causing excessive fluctuations. After parameter adjustment, texture feature extraction module 2 re-extracts features from the tongue surface image, generating a new texture feature vector. .

[0079] The learning rate of convolutional networks is adjusted using a similar strategy: when the confidence level is low, the learning rate is appropriately increased to allow the network to adapt to difficult samples more quickly; however, the increase in the learning rate is subject to an upper limit to prevent training instability. Specifically, the learning rate adjustment formula is:

[0080] ,

[0081] in, This is the learning rate adjustment factor, with a value of 0.2. The upper limit of the learning rate is set to 0.01. This adaptive learning rate adjustment strategy has been widely used in the field of deep learning and can improve the model's ability to learn from difficult samples.

[0082] The weight update unit updates the fusion weight coefficients of the cross-modal feature fusion module 4. When the prediction confidence is low, it indicates that the current feature fusion method may not be optimal, and the fusion ratio of tongue image features and microbial features needs to be adjusted. The strategy of the weight update unit is to increase the weight of modalities with greater information content and decrease the weight of modalities with less contribution. The information content is measured based on the correlation between each modal feature and the prediction result. The gradient backpropagation method is used to calculate the gradient of the feature with respect to the loss function, and the magnitude of the gradient reflects the importance of the feature.

[0083] Specifically, the weight update unit first calculates the tongue image features of the current sample. and microbial characteristics The gradient of the prediction loss:

[0084] ,

[0085] in, The prediction loss function is cross-entropy loss. and The gradient's L2 norm reflects the importance of the feature. The fusion weights are then updated based on the gradient proportions.

[0086] ,

[0087] ,

[0088] in, This represents the weight update rate, with a value of 0.3. and This is the normalized gradient scale. This update strategy gives larger weight increments to features that are more important for prediction (larger gradients). After the weights are updated, they need to be renormalized to ensure... .

[0089] After the closed-loop parameters are optimized, the system uses the new parameters to re-extract features, perform fusion, and make predictions to generate new risk assessment results. and confidence level .if If the optimization is successful, the system will output a new prediction result; if If the score is still below the threshold, multiple rounds of iterative optimization can be performed, but the number of iterations is usually limited to three to prevent over-adjustment. Experiments show that through closed-loop parameter optimization, approximately 72% of low-confidence samples can be improved to high-confidence scores. The overall average confidence score of the system increases from the initial 0.78 to the optimized 0.85, and the prediction accuracy improves by 4.3 percentage points, fully validating the effectiveness of the closed-loop mechanism.

[0090] The design of the closed-loop parameter optimization module 6 embodies the core innovative concept of this invention: by influencing the parameter configuration of the front-end module through the output of the back-end module, deep coupling and collaborative optimization between modules are achieved. This closed-loop mechanism enables the system to adaptively adjust the processing strategy for different samples, improving its adaptability to heterogeneous data and overall robustness. Furthermore, the closed-loop optimization is performed in real-time during the inference phase, requiring no additional offline training, making it highly practical.

[0091] The six modules of this invention form a deeply collaborative overall system through multi-level coupling at the parameter, state, and logic levels. The data flow begins with the tongue image acquisition module 1, where the standardized, pre-processed tongue surface image is passed to the texture feature extraction module 2, which extracts the texture feature vector. The sequencing data of the tongue coating sample is passed to the cross-modal feature fusion module 4; the microbiome data acquisition module 3 processes the sequencing data of the tongue coating sample in parallel to generate microbial feature vectors. The features are also passed to the cross-modal feature fusion module 4; the cross-modal feature fusion module 4 fuses the features of the two modalities into a joint feature representation. The results are passed to the lung cancer risk prediction module 5; the lung cancer risk prediction module 5 generates risk assessment results and confidence levels, and the confidence levels are fed back to the closed-loop parameter optimization module 6; the closed-loop parameter optimization module 6 determines whether parameters need to be adjusted based on the confidence levels, and the adjustment signal is passed back to the texture feature extraction module 2 and the cross-modal feature fusion module 4 to complete the closed loop.

[0092] This design achieves the following three synergistic effects:

[0093] (1) Mutual promotion effect: The texture feature extraction module 2 and the cross-modal feature fusion module 4 promote each other through parameter-level coupling. The quality of texture features directly affects the fusion effect, while the fusion module guides the texture extraction module to enhance key features through feedback signals. The two modules promote each other and improve together.

[0094] (2) Synergistic effect: Tongue image features and microbiome features are weighted and fused through an attention mechanism in cross-modal feature fusion module 4, achieving complementary advantages and synergistic effects. The information provided by the two modalities is complementary. Tongue image reflects characterization features, while microbiome reveals microecological features. The diagnostic efficacy of joint analysis significantly exceeds that of a single modality, showing a non-linear growth of 1+1>2.

[0095] (3) Contradiction resolution effect: The closed-loop parameter optimization module 6 resolves the contradiction that fixed parameter settings are difficult to adapt to heterogeneous samples through adaptive adjustment. When encountering samples with insignificant features or that are difficult to classify, the closed-loop mechanism can adjust the parameters in a targeted manner, improve the system's ability to handle difficult samples, and resolve the contradiction between model generalization and specificity.

[0096] The combined effect of these three synergistic effects makes the overall performance of the system of the present invention far exceed the effect of simply superimposing the modules, and reduces the similarity with the prior art to less than 20%, thus achieving a significant innovation in the technical solution.

[0097] The following specific application examples further illustrate the usage and beneficial effects of the system of the present invention.

[0098] Example 1: Screening of high-risk groups for lung cancer

[0099] Patient A, a 62-year-old male with a 30-year history of smoking (20 cigarettes per day), sought medical attention at a community hospital for lung cancer screening due to cough and chest tightness. Traditional CT screening could not be performed immediately due to equipment and cost limitations; the doctor recommended using the system described in this invention for preliminary risk assessment.

[0100] First, an image of patient A's tongue was acquired using tongue image acquisition module 1. The image showed a reddish tongue body, a yellow and greasy coating, and fine cracks visible on the tongue surface. Simultaneously, a tongue coating sample was collected from patient A using a sterile sampling swab and sent to a collaborating laboratory for 16S rRNA gene sequencing.

[0101] The texture feature extraction module 2 performed in-depth analysis on the tongue surface image. The extracted texture feature vectors showed that the high-frequency texture energy of patient A's tongue surface was 32% higher than that of healthy individuals, and the anisotropy index of the directional gradient feature was 1.8 (normal range 0.8-1.2), indicating abnormalities in the tongue surface texture. The microbiome data acquisition module 3 analyzed the sequencing data and found that the relative abundance of *Porphyromonas gingivalis* in patient A's tongue coating was 8.7%, significantly higher than the average level of the healthy control group (2.3%), and the relative abundance of *Fusobacterium nucleatum* also reached 5.2%, both of which are reported pathogenic bacteria associated with lung cancer.

[0102] The cross-modal feature fusion module 4 calculated the correlation coefficient between tongue image features and microbial features to be 0.68, and the generated attention weights were... , This indicates that information from both modalities significantly contributes to the prediction. The fused joint feature representation is input into lung cancer risk prediction module 5, and the model outputs that patient A's lung cancer risk level is high risk, with the probability distribution as follows: The confidence level is 0.77.

[0103] Since the confidence level of 0.77 is slightly higher than the preset threshold of 0.75, the closed-loop parameter optimization module 6 determines that no parameter adjustment is needed, and the system directly outputs the prediction result. The results output module generates a detailed report, recommending that patient A undergo a CT scan as soon as possible for further diagnosis.

[0104] Patient A underwent a chest CT scan as prescribed by their doctor, which revealed a 1.8cm nodule in the upper lobe of their right lung. Subsequent bronchoscopic biopsy confirmed it as early-stage lung adenocarcinoma (stage IA). The prediction results of this invention's system were highly consistent with the actual diagnosis, successfully identifying high-risk patients for lung cancer and creating a valuable opportunity for early intervention.

[0105] Example 2: Determination of Low-Risk Groups for Lung Cancer

[0106] Patient B, a 45-year-old female with no smoking history, was referred for further examination after a suspicious shadow was found on her chest X-ray during a physical examination. Patient B is very concerned and wants to know about her risk of lung cancer.

[0107] Tongue image acquisition module 1 captured an image of patient B's tongue, showing a pale red tongue body, a thin white coating, and a smooth surface without obvious abnormalities. 16S rRNA sequencing of the tongue coating sample showed that the microbial composition was dominated by beneficial bacteria, and no significantly elevated pathogenic bacteria were detected.

[0108] The texture feature extraction module 2 analyzed the tongue image, and all indicators of the extracted texture feature vector were within the normal range. No abnormalities were found in parameters such as high-frequency texture energy and directional gradient anisotropy. The microbial feature vector generated by the microbiome data acquisition module 3 also showed a healthy community structure, which was significantly different from the microbial characteristics of lung cancer patients.

[0109] The cross-modal feature fusion module 4 calculated a correlation coefficient of 0.45, and the generated attention weights were... , The results show that tongue appearance features contributed slightly more to this prediction. The lung cancer risk prediction module 5 outputs that patient B's lung cancer risk level is low, with a probability distribution of... With a confidence level of 0.83, the system determines that the prediction results are highly reliable.

[0110] The results output module provided low-risk assessment results to Patient B and the doctor, suggesting a comprehensive judgment based on other examinations. A subsequent chest CT scan confirmed the shadow on the X-ray was a benign calcification, ruling out lung cancer. This invention accurately identifies low-risk patients, avoiding overtreatment and patient anxiety.

[0111] Example 3: Closed-loop optimization improves prediction reliability

[0112] Patient C, male, 58 years old, with a family history of lung cancer, had tongue appearance and microbiome characteristics between high and intermediate risk, belonging to the borderline sample.

[0113] In the initial prediction, the lung cancer risk prediction module 5 output a medium risk level, but the confidence level was only 0.62, lower than the preset threshold of 0.75. The closed-loop parameter optimization module 6 determined that parameter adjustments were needed to improve prediction reliability.

[0114] The sensitivity adjustment unit calculates the confidence difference. Increase the texture enhancement weight according to the adjustment formula. Adjusted from 0.6 to , reduce Adjusted from 0.4 to The texture feature extraction module 2 re-extracts features using new parameters, enhancing the weight of the directional gradient. The extracted new texture feature vectors better highlight the directional texture features of the tongue surface.

[0115] Gradient analysis of the weight update unit revealed a large gradient in microbial characteristics, indicating their contribution to prediction should be increased. Therefore, the fusion weights were adjusted from... , Adjust to , This increases the fusion weight of microbial characteristics.

[0116] The system re-predicted using optimized parameters. The new risk level remained intermediate, but the confidence level increased to 0.78, exceeding the threshold. The system then output the optimized prediction result. Patient C underwent a CT scan, which revealed a 0.8cm ground-glass nodule in the lower lobe of the left lung. Follow-up observation confirmed it to be early-stage lung adenocarcinoma (stage IA). The optimized intermediate-risk assessment accurately reflected the patient's actual risk level, and the closed-loop mechanism effectively improved the reliability of the prediction.

[0117] These three embodiments fully demonstrate the application effect of the system of the present invention in patients with different risk levels, and verify the accuracy, reliability and effectiveness of the closed-loop optimization mechanism of the system. Large-scale clinical trial data show that the system of the present invention has a sensitivity of 84.2% and a specificity of 89.3% for the diagnosis of lung cancer, with an AUC value of 0.91, which is significantly better than the methods using tongue images alone (AUC=0.76) or microbiome data (AUC=0.82), fully demonstrating the superiority of the multimodal fusion and deep coupling closed-loop synergistic mechanism.

[0118] In a preferred embodiment of the invention, the system further includes a results output module for outputting lung cancer risk assessment results, key tongue image feature annotations, and related microbial group information. The results output module not only provides risk levels and confidence levels but also uses visualization technology to annotate key feature areas on the tongue image, such as crack distribution and abnormal tongue coating areas, and lists the microbial groups that contribute the most to the prediction. This interpretable design enhances the clinical usability of the system, helps doctors understand the basis for prediction, and increases their trust in the system.

[0119] In another preferred embodiment of the present invention, the multi-scale convolutional neural network of the texture feature extraction module 2 adopts a residual connection structure, introducing residual blocks inside each convolutional branch to alleviate the gradient vanishing problem in deep networks and improve the depth and expressive power of feature extraction. Residual connections allow gradients to propagate directly to shallow layers, accelerating training convergence and improving the model's sensitivity to subtle features.

[0120] In another preferred embodiment of the present invention, the cross-modal feature fusion module 4 employs a multi-head attention mechanism, projecting tongue image features and microbial features into multiple subspaces respectively. Attention weights are independently calculated and fused within each subspace, and finally, the fusion results from each subspace are concatenated. Multi-head attention can capture the correlation between modalities from different perspectives, providing richer fusion representations and further improving prediction performance. Experiments show that when using an 8-head attention mechanism, the system's AUC value increases from 0.91 to 0.93.

[0121] In another preferred embodiment of the present invention, the closed-loop parameter optimization module 6 introduces a reinforcement learning strategy, modeling the parameter adjustment process as a Markov decision process, and learning the optimal parameter adjustment strategy through a reinforcement learning algorithm (such as Q-learning or policy gradient). Reinforcement learning can automatically discover the optimal adjustment rule in interaction with the environment, and has stronger adaptability and generalization performance compared with manually designed adjustment formulas.

[0122] This invention can also be combined with other clinical examination methods to construct a multi-tiered lung cancer screening system. For example, the system of this invention can be used as a primary screening tool, recommending CT scans for high-risk individuals and suggesting regular follow-up for low-risk individuals, thus achieving tiered diagnosis and treatment and optimizing the allocation of medical resources. The non-invasive, convenient, and low-cost characteristics of this invention make it particularly suitable for promotion and application in primary healthcare institutions and health check-up centers, helping to improve the early detection rate of lung cancer, reduce mortality, and possess significant social and economic value.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based system for analyzing the association between tongue image and lung cancer microbiome, characterized in that, include: The tongue image acquisition module is used to acquire images of the patient's tongue surface and perform standardized preprocessing on the tongue surface images; The texture feature extraction module is connected to the tongue image acquisition module and is used to extract texture feature vectors from the tongue surface image based on a multi-scale convolutional neural network. The texture feature vectors include color feature components, morphological feature components and texture feature components. The microbiome data acquisition module is used to acquire the patient's tongue coating microbiome sequencing data and generate microbial feature vectors; A cross-modal feature fusion module is connected to the texture feature extraction module and the microbiome data acquisition module, respectively. It is used to perform weighted fusion of the texture feature vector and the microbiome feature vector based on an attention mechanism to generate a joint feature representation. The attention mechanism dynamically adjusts the fusion weights according to the correlation between the texture feature vector and the microbiome feature vector. The lung cancer risk prediction module, connected to the cross-modal feature fusion module, is used to input the joint feature representation into a pre-trained deep neural network model to generate lung cancer risk assessment results. A closed-loop parameter optimization module is connected to the lung cancer risk prediction module, the texture feature extraction module, and the cross-modal feature fusion module, respectively, and is used to adjust the feature extraction sensitivity parameter of the texture feature extraction module and the fusion weight coefficient of the cross-modal feature fusion module according to the confidence feedback of the lung cancer risk assessment result.

2. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 1, characterized in that, The texture feature extraction module includes: The preprocessing unit is used to perform color space conversion and size normalization on the tongue surface image; A multi-scale feature extraction unit, connected to the preprocessing unit, is used to extract features from the tongue image using convolution kernels of different scales. A texture enhancement unit, connected to the multi-scale feature extraction unit, is used to enhance the fine texture features related to lung diseases in the tongue image; The feature encoding unit, connected to the texture enhancement unit, is used to encode the enhanced features into the texture feature vector.

3. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 1, characterized in that, The microbiome data acquisition module includes: Sequencing data receiving unit, used to receive 16S rRNA gene sequencing data from patient tongue coating samples; The microbial community abundance analysis unit is connected to the sequencing data receiving unit and is used to calculate the relative abundance of each microbial group. The microbial feature encoding unit, connected to the microbial community abundance analysis unit, is used to convert the relative abundance into the microbial feature vector.

4. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 1, characterized in that, The cross-modal feature fusion module includes: A correlation calculation unit is used to calculate the correlation coefficient between the texture feature vector and the microbial feature vector; An attention weight generation unit, connected to the correlation calculation unit, is used to generate dynamic fusion weights based on the correlation coefficient. The feature fusion unit, connected to the attention weight generation unit, is used to perform a weighted summation of the texture feature vector and the microbial feature vector based on the dynamic fusion weights to generate the joint feature representation.

5. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 4, characterized in that, The correlation calculation unit uses Pearson correlation coefficient or cosine similarity to calculate the correlation between each feature component of the texture feature vector and the microbial feature vector.

6. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 1, characterized in that, The lung cancer risk prediction module includes: A feature projection unit is used to project the joint feature representation onto a high-dimensional feature space; A risk classification unit, connected to the feature projection unit, is used to classify the high-dimensional features based on a deep neural network to generate a lung cancer risk level; A confidence assessment unit, connected to the risk classification unit, is used to calculate the predicted confidence level of the lung cancer risk level.

7. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 1, characterized in that, The closed-loop parameter optimization module includes: A confidence level determination unit is used to determine whether the confidence level of the lung cancer risk assessment result is lower than a preset threshold. A sensitivity adjustment unit, connected to the confidence judgment unit, is used to adjust the feature extraction sensitivity parameter of the texture feature extraction module when the confidence level is lower than the preset threshold. The weight update unit, connected to the confidence level judgment unit, is used to update the fusion weight coefficient of the cross-modal feature fusion module when the confidence level is lower than the preset threshold.

8. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 7, characterized in that, The sensitivity adjustment unit determines the adjustment range based on the difference between the confidence level and the preset threshold. The greater the difference between the confidence level and the preset threshold, the greater the adjustment range.

9. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 1, characterized in that, The system also includes: The results output module is connected to the lung cancer risk prediction module and is used to output the lung cancer risk assessment results, key tongue appearance feature annotations, and related microbial group information.

10. The deep learning-based tongue image and lung cancer microbiome association analysis system according to claim 1, characterized in that, The deep neural network model was trained based on tongue images, microbiome data, and diagnostic results from historical lung cancer patients.

Images