A cfDNA methylation-imaging-omics-clinical-based lung nodule benign and malignant differential diagnosis system and storage medium

Abstract

This invention relates to the field of medical software development technology and discloses a system and storage medium for differentiating benign and malignant pulmonary nodules based on cfDNA methylation-radiomics-clinical studies. The system includes a multimodal data acquisition module, a feature extraction and alignment module, a multimodal fusion diagnosis module, and an interpretable report generation module. After collecting clinical information, low-dose CT images, and peripheral blood cfDNA methylation sequencing data from the subjects, features are extracted from each modality and mapped to a unified semantic space through contrastive learning. A collaborative attention mechanism is then used to dynamically weight and fuse the aligned multimodal features. The fused features are input into a deep neural network classifier to output a nodule malignancy probability score, and finally, an interpretable diagnostic report with risk stratification is generated. This invention solves the problems of insufficient multimodal heterogeneous data fusion, poor model robustness, and weak interpretability, achieving accurate differentiation between benign and malignant pulmonary nodules, and demonstrating good clinical applicability and promotional value.

Description

Technical Field

[0001] This invention relates to the field of medical software development technology, specifically to a system and storage medium for differentiating between benign and malignant pulmonary nodules based on cfDNA methylation-radiomics-clinical studies. Background Technology

[0002] Early qualitative diagnosis of pulmonary nodules is a core aspect of lung cancer prevention and treatment. With the widespread use of low-dose CT screening technology, the detection rate of pulmonary nodules in clinical practice has been greatly improved, providing important support for the early detection and early intervention of lung cancer. However, the differentiation between benign and malignant nodules remains a challenge in clinical diagnosis and treatment. Accurate differential diagnosis can not only avoid unnecessary invasive examinations for patients with benign nodules, but also prevent patients with malignant nodules from missing the best treatment opportunity due to delayed diagnosis.

[0003] In existing technologies, the differentiation between benign and malignant pulmonary nodules largely relies on single-modal diagnostic information or simple multimodal feature splicing methods, both of which have certain technical shortcomings: In single-modal diagnosis, radiomics can provide morphological features of nodules, but due to the problem of "different images for the same disease and the same image for different diseases," the diagnostic specificity is insufficient; cfDNA methylation detection can reflect epigenetic changes in tumors, but the content of ctDNA in the peripheral blood of early-stage lung cancer patients is low, which easily leads to false negative results; clinical variables can only provide the patient's basic risk background, and when used alone, the diagnostic sensitivity is far from meeting clinical needs. Existing multimodal fusion studies mostly use simple feature splicing methods to integrate data from different modalities, without fully considering the heterogeneity between clinical, imaging, and methylation data, lacking an effective cross-modal feature alignment mechanism, and without dynamic adjustment means for the contribution of each modality feature, resulting in poor robustness and weak generalization ability of the fusion model, and the model decision process lacks interpretability. Clinicians cannot intuitively understand the basis for the diagnostic results, making it difficult to widely promote and apply in clinical practice.

[0004] In view of this, we propose a diagnostic system and storage medium for the differential diagnosis of benign and malignant pulmonary nodules based on cfDNA methylation-radiomics-clinical studies. Summary of the Invention

[0005] The purpose of this invention is to provide a system and storage medium for differentiating between benign and malignant pulmonary nodules based on cfDNA methylation-radiomics-clinical studies, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A system for differentiating benign and malignant pulmonary nodules based on cfDNA methylation-radiomics-clinical studies includes a multimodal data acquisition module, a feature extraction and alignment module, a multimodal fusion diagnostic module, and an interpretable report generation module connected in sequence.

[0008] The multimodal data acquisition module is configured to acquire clinical information, low-dose CT image data, and peripheral blood cfDNA methylation sequencing data of the target subject.

[0009] The feature extraction and alignment module is configured to extract clinical feature vectors, radiomics feature vectors, and methylation feature vectors from the clinical information, CT image data, and cfDNA methylation sequencing data, respectively, and map the three types of feature vectors to a unified semantic space through contrastive learning to obtain aligned clinical features, image features, and methylation features.

[0010] The multimodal fusion diagnostic module is configured to use a collaborative attention mechanism to perform weighted fusion of the aligned clinical features, imaging features and methylation features to obtain fused features, and input the fused features into a trained deep neural network classifier to output a malignancy probability score for the lung nodules;

[0011] The interpretability report generation module is configured to generate a diagnostic report containing risk stratification and visualization evidence based on the malignancy probability score and the contribution of each modality feature to the diagnostic results.

[0012] Preferably, the clinical information includes age, gender, smoking history, family history of cancer, history of tuberculosis, nodule size, nodule density, and nodule location; the low-dose CT imaging data is DICOM format scan data with a slice thickness ≤1.25mm; the cfDNA methylation sequencing data is the level value of 54 methylation regions obtained by targeted methylation sequencing technology, and the 54 methylation regions cover at least the methylation regions of 10 genes: SHOX2, PTGER4, HOXA7, HOXA9, RASSF1A, TAC1, CCDC181, TBR1, ZNF781, and BDNF.

[0013] Preferably, the feature extraction and alignment module maps features to a unified semantic space through contrastive learning in the following ways: constructing a trimodal positive sample pair composed of clinical feature vectors, radiomics feature vectors, and methylation feature vectors of the same subject; constructing a trimodal negative sample pair composed of random combinations of feature vectors of different subjects; and training the three types of feature vectors using a multimodal contrastive loss function to minimize the feature distance of the positive sample pair and maximize the feature distance of the negative sample pair.

[0014] Preferably, the multimodal contrast loss function is the InfoNCE loss function. During training, the clinical feature vector, radiomics feature vector, and methylation feature vector are respectively input into an independent 2-layer MLP projection head and mapped to a 128-dimensional space. The temperature coefficient of the InfoNCE loss function is set to 0.07.

[0015] Preferably, the collaborative attention mechanism includes an intramodal self-attention unit, a cross-modal cross-attention unit, and a modal gating fusion unit;

[0016] The intramodal self-attention unit is used to perform self-attention calculations on aligned clinical features, imaging features, and methylation features respectively, to capture feature associations within each modality;

[0017] The cross-modal cross-attention unit uses aligned image features as queries, aligned clinical features and methylation features as keys and values, calculates image-clinical and image-methylation cross-attention respectively, and adds the cross-attention results to the self-attention results of image features to obtain enhanced image features;

[0018] The modality-gated fusion unit is used to dynamically calculate the fusion weight based on the data quality of the original feature vectors of each modality, and to perform weighted fusion of the aligned clinical features, the enhanced image features, and the aligned methylation features based on the fusion weight.

[0019] Preferably, the modal gating fusion unit evaluates the quality of each modal data through three independent 2-layer MLP gating networks. The gating network takes the original feature vector of each modality as input, outputs a quality score in the 0-1 interval, and obtains the corresponding fusion weight by normalizing the quality scores of each modality.

[0020] Preferably, the deep neural network classifier is a 3-layer MLP classifier with a network structure of 128→64→32→2. The first two layers use the ReLU activation function and dropout layer with the dropout coefficient set to 0.3. The last layer uses the softmax activation function to output the probability values ​​of benign and malignant lung nodules, and the malignant probability value is taken as the malignant probability score.

[0021] Preferably, the interpretability report generation module uses the Shapley value attribution analysis method to calculate the contribution of each modality feature to the diagnostic result, and uses Grad-CAM technology to generate an attention heatmap of CT images. The diagnostic report is in PDF format and includes basic patient information, malignancy risk score and risk stratification, feature contribution ranking, CT image attention heatmap, list of key abnormal indicators, and clinical diagnostic recommendations.

[0022] A computer-readable storage medium storing a computer program, which, when executed by a processor, performs all the functions of the aforementioned cfDNA methylation-radiomics-clinical differential diagnostic system for benign and malignant pulmonary nodules.

[0023] Preferably, when the computer program is executed by the processor, it also implements the training process of the model in the differential diagnosis system for benign and malignant pulmonary nodules. The training process includes constructing a multicenter pulmonary nodule patient sample set, dividing the training set and validation set in a 7:3 ratio, controlling the number of model training iterations through an early stopping method, and setting the early stopping patience to 10.

[0024] Compared with existing technologies, this invention provides a system and storage medium for differentiating benign and malignant pulmonary nodules based on cfDNA methylation-radiomics-clinical studies, which has the following beneficial effects:

[0025] 1. This cfDNA methylation-radiomics-clinical differential diagnostic system and storage medium for distinguishing between benign and malignant pulmonary nodules aims to improve the accuracy of this differential diagnosis. It achieves feature alignment across three modalities (clinical, radiomics, and cfDNA methylation) through contrastive learning and employs a collaborative attention mechanism for dynamic weighted fusion to fully leverage complementary information among the multimodal data. This results in improved accuracy in differentiating between benign and malignant pulmonary nodules. In multicenter validation, the model achieved an AUC of 0.92, significantly outperforming single-modality and simple feature splicing models.

[0026] 2. This cfDNA methylation-radiomics-clinical differential diagnostic system and storage medium for lung nodules aims to enhance the robustness of the lung nodule diagnostic model by dynamically calculating and adjusting the fusion weights based on the data quality of the original feature vectors of each modality through a modality-gated fusion unit. When the signal of a certain modality is weak, its weight is automatically reduced and the weights of other modalities are increased, thereby achieving the goal of enhancing the robustness of the diagnostic model and ensuring the stability of diagnostic results under different data quality.

[0027] 3. This cfDNA methylation-radiomics-clinical differential diagnostic system and storage medium for distinguishing between benign and malignant pulmonary nodules aims to improve the trust and adoption rate of AI-assisted diagnosis in clinical practice. It utilizes Shapley value attribution analysis to calculate feature contribution, Grad-CAM technology to generate CT image attention heatmaps, and integrates these elements to generate interpretable PDF diagnostic reports with visual evidence. This allows clinicians to intuitively understand the model's decision-making basis, thereby enhancing the clinical trust and adoption rate of AI-assisted diagnosis. Attached Figure Description

[0028] Figure 1This is a diagram illustrating the overall architecture of the pulmonary nodule benign and malignant differential diagnosis system based on cfDNA methylation-radiomics-clinical studies of the present invention.

[0029] Figure 2 This is a block diagram of the feature extraction and alignment module of the present invention;

[0030] Figure 3 This is a block diagram of the collaborative attention mechanism of the present invention;

[0031] Figure 4 This is an example diagram of the interpretability diagnostic report of the present invention;

[0032] Figure 5 This is a comparison chart of the ROC curves of the model of this invention and the single-modal model. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Please see Figures 1-5 The present invention provides a technical solution:

[0035] A diagnostic system for the differential diagnosis of benign and malignant pulmonary nodules based on cfDNA methylation-radiomics-clinical studies includes:

[0036] The multimodal data acquisition module is configured to collect clinical information, low-dose CT imaging data, and peripheral blood cfDNA methylation sequencing data from the target subjects. Clinical information includes age, sex, smoking history, family history of cancer, history of tuberculosis, nodule size, nodule density, and nodule location. The CT images are low-dose scan data in DICOM format with a slice thickness ≤1.25mm. The cfDNA methylation sequencing data are the level values ​​of 54 methylation regions, including genes such as SHOX2, PTGER4, HOXA7, HOXA9, RASSF1A, TAC1, CCDC181, TBR1, ZNF781, and BDNF, obtained through targeted methylation sequencing technology.

[0037] The feature extraction and alignment module contains the following sub-units and workflows:

[0038] Clinical feature extraction subunit: First, one-hot encoding is performed on categorical variables (gender, smoking history, family history, tuberculosis history, nodule density), converting each categorical variable into a binary vector; continuous variables (age, nodule size) are standardized using Z-score, with the formula:

[0039]

[0040] in The mean, The standard deviation is used to encode and standardize the data, and then concatenate them to obtain a 12-dimensional clinical feature vector C.

[0041] Image feature extraction subunit: A 3D dynamic convolutional neural network (3DResNet-18 variant) is used to perform end-to-end feature extraction on the original CT images (input size 64×64×64 voxels). The network structure contains 4 dynamic convolutional blocks. Each dynamic convolutional block introduces a scale factor to adaptively adjust the receptive field of the convolutional kernel. For small nodules with a diameter <10mm, the convolutional kernel automatically shrinks the receptive field to 3×3×3; for large nodules with a diameter >20mm, the receptive field is expanded to 7×7×7. At the same time, 1316 handcrafted radiomics features are extracted using PyRadiomics. After LASSO regression screening, 45 of the most relevant features are retained. The depth features (512-dimensional) are concatenated with the handcrafted features (45-dimensional) to obtain a 557-dimensional image feature vector R.

[0042] Methylation feature extraction subunit: Sequencing data from 54 methylated regions were processed, and the methylation level β value for each region was calculated using the following formula:

[0043]

[0044] in The number of methylated reads. The number of unmethylated reads is used to obtain a 54-dimensional methylation feature vector M.

[0045] Contrastive learning aligns subunits: C, R, and M are input into three independent projection heads (2-layer MLP, 256-dimensional hidden layer, 128-dimensional output layer, ReLU activation function), mapped to a 128-dimensional space, and positive sample pairs are constructed: trimodal features of the same patient. Constructing negative sample pairs: randomly combining features from different patients. Training is performed using the InfoNCE loss function:

[0046]

[0047] in For cosine similarity, The temperature coefficient is set to 0.07. The training objective is to minimize the feature distance between positive sample pairs and maximize the feature distance between negative sample pairs. After training, the aligned clinical features are obtained. Image features and methylation characteristics (128 dimensions each);

[0048] The multimodal fusion diagnostic module contains the following sub-units and workflows:

[0049] Intramodal self-attention units: respectively for , , Perform self-attention calculations, in order to For example, the calculation process is as follows:

[0050]

[0051]

[0052] in , , For learnable parameter matrix, =128, through self-attention, the features within each modality are interactively enhanced;

[0053] Cross-modal attention unit: based on image features For query, clinical features and methylation characteristics For keys and values, calculate the cross-attention between image-clinical and image-methylation, respectively:

[0054]

[0055]

[0056] Combine the two cross-attention results with The self-attention results are summed to obtain enhanced image features that integrate clinical and methylation information. .

[0057] Modality Gated Fusion Unit: Three gated networks (2-layer MLP, 64-dimensional hidden layer, 1-dimensional output layer) are designed to evaluate the data quality of clinical, imaging, and methylation modalities respectively. The input is the original feature vector (C, R, M) of each modality, and the output is the quality score q∈[0,1]. The gated network structure is as follows:

[0058]

[0059] in Using the sigmoid activation function, the fusion weights are calculated based on the quality score:

[0060]

[0061] The final fusion features are:

[0062]

[0063] Classifier: Input the fused feature F into a 3-layer MLP classifier (128→64→32→2). The first two layers use ReLU activation function and Dropout (p=0.3), and the last layer uses softmax activation function to output the probabilities of benign and malignant. The probability of malignancy is taken as the final score P (0-100).

[0064] The interpretability report generation module contains the following sub-units and workflows:

[0065] Shapley value attribution analysis subunit: The SHAP framework is used to calculate the marginal contribution of each modality feature to the diagnostic results. For each sample, the fusion feature F is used as input, and the Shapley value of each feature is calculated through 100 Monte Carlo samplings to generate the feature importance ranking.

[0066] Attention Heatmap Visualization Subunit: Grad-CAM technology is used to backpropagate the gradient of the last convolutional layer of the classifier to the input CT image to generate an attention heatmap. The specific steps are: calculate the gradient of the target category (malignant), perform weighted summation on the feature map, and obtain the heatmap after ReLU activation, which is then overlaid on the original image for display.

[0067] Report generation sub-unit: Integrates the above information to generate a PDF diagnostic report that includes basic patient information, malignancy risk score and risk stratification, feature contribution ranking bar chart, CT image attention heatmap, list of key abnormal indicators, and diagnostic recommendations.

[0068] The present invention will be further described in detail below with reference to the embodiments.

[0069] Example 1: Multi-center queue construction and data acquisition

[0070] Patients with pulmonary nodules were prospectively enrolled at the Xinjiang Production and Construction Corps Hospital (main center) and five branch centers, with a total sample size of 3,000. They were randomly divided into a training set (2,100 cases) and a validation set (900 cases) at a ratio of 7:3.

[0071] Inclusion criteria: ① Subjects with pulmonary nodules within the Xinjiang Production and Construction Corps region; ② Nodule diameter 5-30 mm; ③ Age > 18 years; ④ Complete clinical medical records and clear diagnosis; ⑤ Pathologically confirmed surgical tissue specimens (not required for patients with benign nodules). Exclusion criteria: History of malignant tumors; inability to undergo resectable surgery; preoperative neoadjuvant therapy; HIV, HBV infection; no preoperative blood specimens; history of allogeneic blood transfusion within six months prior to surgery; pregnancy or lactation; history of bone marrow transplantation; severe heart, liver, or kidney dysfunction.

[0072] Clinical data collection: Patient age, gender, smoking history, family history of cancer, history of tuberculosis, nodule size, density, and location are collected from the hospital's electronic medical record system;

[0073] Image data acquisition: Low-dose CT scan, slice thickness ≤1.25mm, DICOM format images exported from PACS system, and ROI of lung nodules delineated by two senior radiologists using 3DSlicer software;

[0074] cfDNA methylation detection: 10 mL of peripheral blood was collected from the patient into a Streck cfDNA-specific blood collection tube. The plasma was centrifuged and separated within 6 hours, stored at -80℃, and sequenced using a targeted methylation sequencing panel (covering 54 methylation regions, including genes such as SHOX2, PTGER4, HOXA7, HOXA9, RASSF1A, TAC1, CCDC181, TBR1, ZNF781, and BDNF). The average sequencing depth was >1000×, and the methylation level β value of each target region was calculated.

[0075] This embodiment provides a large-sample, high-quality, and multi-dimensional effective dataset for subsequent model training and validation through a standardized and regulated multi-center data acquisition process. This ensures the effectiveness of subsequent trimodal feature extraction, alignment, and fusion diagnosis from the data source, laying a data foundation for improving the accuracy of differential diagnosis of benign and malignant pulmonary nodules.

[0076] Example 2: Workflow of the Feature Extraction and Alignment Module

[0077] Clinical feature extraction: Taking a 58-year-old male patient as an example, with a smoking history of 30 pack-years, nodule diameter of 15mm, partially solid, categorical variable coding: gender male → [1,0], smoking history → [1,0,0] (three-class coding), nodule density partially solid → [0,1,0]. Continuous variable standardization: age 58 years, training set age mean 55 years, standard deviation 10 years, standardized to 0.3; nodule diameter 15mm, mean 12mm, standard deviation 4mm, standardized to 0.75, concatenated to obtain a 12-dimensional vector C;

[0078] Image feature extraction: CT images were resampled to 64×64×64 voxels and input into a 3D dynamic convolutional network. The network first estimated the nodule diameter (approximately 15 mm) through a nodule detection module. The dynamic convolutional layer adjusted the receptive field to 5×5×5 accordingly. After passing through 4 dynamic convolutional blocks (each with 32, 64, 128, and 256 output channels respectively), 512-dimensional depth features were obtained through global average pooling. At the same time, 1316 features were extracted from the original ROI using PyRadiomics. After LASSO regression screening, 45 features were retained and concatenated to obtain a 557-dimensional vector R.

[0079] Methylation feature extraction: Taking the SHOX2 gene as an example, sequencing yielded 780 methylated reads and 220 non-methylated reads. β=780 / (780+220)=0.78. The β value was calculated for all 54 regions, resulting in a 54-dimensional vector M.

[0080] Contrastive learning alignment: C, R, and M are input into the projection head. The projection head structure is: linear layer (input dimension → 256) + BatchNorm + ReLU + Dropout (0.2) + linear layer (256 → 128). The outputs C', R', and M' are all 128-dimensional vectors. During training, 32 patients are randomly selected from each batch to construct 32 positive sample pairs and randomly combined negative sample pairs. When calculating the InfoNCE loss, the temperature coefficient τ = 0.07, the optimizer is Adam, the learning rate is 0.001, the training is conducted for 100 epochs, and the early stop patience is 10.

[0081] This embodiment obtains feature vectors for clinical, radiomics, and cfDNA methylation through targeted feature extraction methods. It then uses contrastive learning to construct positive and negative sample pairs and employs the InfoNCE loss function to train and align the three-modal features in a unified semantic space. This solves the problem of insufficient heterogeneous data fusion, allowing different modal features to be directly compared and fused, fully exploring the complementary information between multimodal data, and improving the accuracy of differential diagnosis of benign and malignant pulmonary nodules.

[0082] Example 3: Workflow of the Multimodal Fusion Diagnostic Module

[0083] Intramodal self-attention: For example, Q(128×128), K(128×128), and V(128×128) are generated through three linear layers. The attention score matrix A = QK^T / √128 is calculated, and after softmax, it is multiplied by V to obtain... Calculate the same and ;

[0084] Cross-modal attention: For Q, For K and V, calculate:

[0085]

[0086] Similarly, calculate ,Will , , Adding them together yields enhanced image features. ;

[0087] Modality-gated fusion: The inputs to the gating network are C (12-dimensional), R (557-dimensional), and M (54-dimensional), which are unified to 64 dimensions through adaptive average pooling. Then, the inputs pass through a linear layer (64→32) → ReLU → a linear layer (32→1) → sigmoid to obtain quality scores q_C, q_R, and q_M. For example, if a sample has insufficient sequencing depth, q_M=0.3; good image quality, q_R=0.9; and complete clinical data, q_C=0.8. The fusion weights are then... =0.8 / (0.8+0.9+0.3)=0.4, =0.45, =0.15, fusion feature:

[0088] ;

[0089] Classification: F Input 3-layer MLP: linear layer (128→64)+ReLU+Dropout(0.3)+linear layer (64→32)+ReLU+linear layer (32→2)+softmax, output [0.08,0.92], malignancy probability P=92 points.

[0090] This embodiment achieves deep interactive enhancement of multimodal features through intramodal self-attention and cross-modal cross-attention. At the same time, it utilizes a modal gating fusion unit to dynamically calculate and adjust the fusion weights based on the data quality of the original feature vectors of each modality. When the signal of a certain modality is weak, its weight is automatically reduced and the weights of other modalities are increased, ensuring the stability of diagnostic results under different data quality and enhancing the robustness of the lung nodule diagnostic model. Furthermore, through dynamic weighted fusion using a collaborative attention mechanism, it further integrates complementary information from multiple modalities, helping to improve the accuracy of differential diagnosis of benign and malignant lung nodules.

[0091] Example 4: Example of generating an interpretability report

[0092] For the above patients, the system generated a report: Malignancy risk score of 92 (high risk); feature contribution ranking showed spiculation on imaging (28%), SHOX2 methylation (22%), age (15%), smoking history (12%), nodule size (10%), PTGER4 methylation (8%), and other (5%); attention heatmap highlighted the spiculated area at the nodule edge; key abnormal indicators included SHOX2 methylation 0.78 (↑↑), PTGER4 methylation 0.65 (↑), and nodule diameter 15mm (↑); the diagnosis recommended immediate referral to a specialist outpatient clinic for consideration of biopsy.

[0093] This embodiment calculates and ranks the contribution of each feature to the diagnostic results through Shapley value attribution analysis, generates CT image attention heatmaps using Grad-CAM technology, and integrates patient information, risk scores, key abnormal indicators, and diagnostic suggestions to generate standardized PDF interpretable reports. This allows clinicians to intuitively understand the decision-making basis of the model and clearly know which imaging features, methylated genes, and clinical factors led to the risk assessment, thereby improving the trust and adoption rate of AI-assisted diagnosis in clinical practice.

[0094] The present invention has been described in detail above. However, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, any modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.

Claims

1. A system for differentiating benign and malignant pulmonary nodules based on cfDNA methylation-radiomics-clinical studies, characterized in that, It includes a multimodal data acquisition module, a feature extraction and alignment module, a multimodal fusion diagnosis module, and an interpretability report generation module, which are connected in sequence. The multimodal data acquisition module is configured to acquire clinical information, low-dose CT image data, and peripheral blood cfDNA methylation sequencing data of the target subject. The feature extraction and alignment module is configured to extract clinical feature vectors, radiomics feature vectors, and methylation feature vectors from the clinical information, CT image data, and cfDNA methylation sequencing data, respectively, and map the three types of feature vectors to a unified semantic space through contrastive learning to obtain aligned clinical features, image features, and methylation features. The multimodal fusion diagnostic module is configured to use a collaborative attention mechanism to perform weighted fusion of the aligned clinical features, imaging features and methylation features to obtain fused features, and input the fused features into a trained deep neural network classifier to output a malignancy probability score for the lung nodules; The interpretability report generation module is configured to generate a diagnostic report containing risk stratification and visualization evidence based on the malignancy probability score and the contribution of each modality feature to the diagnostic results.

2. The lung nodule benign and malignant differential diagnosis system according to claim 1, characterized in that: The low-dose CT image data is scan data in DICOM format with a slice thickness ≤1.25mm; the cfDNA methylation sequencing data is the level value of 54 methylation regions obtained by targeted methylation sequencing technology, and the 54 methylation regions cover the methylation regions of at least 10 genes: SHOX2, PTGER4, HOXA7, HOXA9, RASSF1A, TAC1, CCDC181, TBR1, ZNF781, and BDNF.

3. The differential diagnostic system for benign and malignant pulmonary nodules according to claim 2, characterized in that: The feature extraction and alignment module maps features to a unified semantic space through contrastive learning in the following way: it constructs a trimodal positive sample pair consisting of clinical feature vectors, radiomics feature vectors, and methylation feature vectors of the same subject, and constructs a trimodal negative sample pair consisting of random combinations of feature vectors of different subjects; it uses a multimodal contrastive loss function to train the three types of feature vectors, minimizing the feature distance of the positive sample pair and maximizing the feature distance of the negative sample pair.

4. The differential diagnostic system for benign and malignant pulmonary nodules according to claim 3, characterized in that: The multimodal contrast loss function is the InfoNCE loss function. During training, the clinical feature vector, radiomics feature vector, and methylation feature vector are respectively input into an independent 2-layer MLP projector and mapped to a 128-dimensional space. The temperature coefficient of the InfoNCE loss function is set to 0.

07.

5. The differential diagnostic system for benign and malignant pulmonary nodules according to claim 1, characterized in that: The collaborative attention mechanism includes an intramodal self-attention unit, a cross-modal cross-attention unit, and a modal gating fusion unit.

6. The differential diagnostic system for benign and malignant pulmonary nodules according to claim 5, characterized in that: The modal gating fusion unit evaluates the quality of each modal data through three independent 2-layer MLP gating networks. The gating network takes the original feature vector of each modality as input, outputs a quality score in the range of 0-1, and obtains the corresponding fusion weight by normalizing the quality scores of each modality.

7. The differential diagnostic system for benign and malignant pulmonary nodules according to claim 6, characterized in that: The deep neural network classifier is a 3-layer MLP classifier. The first two layers use the ReLU activation function and dropout layer with the dropout coefficient set to 0.

3. The last layer uses the softmax activation function to output the probability values ​​of benign and malignant lung nodules, and the malignant probability value is taken as the malignant probability score.

8. The differential diagnostic system for benign and malignant pulmonary nodules according to claim 1, characterized in that: The interpretability report generation module uses the Shapley value attribution analysis method to calculate the contribution of each modality feature to the diagnostic results, and uses Grad-CAM technology to generate attention heatmaps of CT images. The diagnostic report is in PDF format and includes basic patient information, malignancy risk score and risk stratification, feature contribution ranking, CT image attention heatmap, list of key abnormal indicators, and clinical diagnostic recommendations.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements all the functions of the cfDNA methylation-radiomics-clinical differential diagnostic system for benign and malignant pulmonary nodules as described in any one of claims 1-8.

10. The computer-readable storage medium according to claim 9, characterized in that, When the computer program is executed by the processor, it also implements the training process of the model in the differential diagnosis system for benign and malignant pulmonary nodules. The training process includes constructing a multicenter pulmonary nodule patient sample set, dividing the training set and validation set in a 7:3 ratio, controlling the number of model training iterations through the early stop method, and setting the early stop patience to 10.

Images