Molecular marker and MRI image fusion tumor early detection method and system

Abstract

The application discloses a molecular marker and MRI image fusion tumor early detection method and system, first collects serum molecular marker detection data and multi-sequence MRI image data of a digestive tract tumor screening object, constructs a molecular feature vector and extracts high-dimensional MRI image features respectively; a tumor prior probability is calculated through a molecular pre-screening deep neural network, and an adaptive spatial attention guide weight is generated; the weight is weighted and fused with the MRI image features to realize fine feature enhancement of a high-risk area and accurate lesion detection; based on a Bayesian evidence fusion reasoning framework, molecular detection evidence and image detection evidence are combined to update a posterior probability, and finally, a lesion positioning, detection confidence and visual detection report are output. The application realizes efficient complementation of molecular evidence and image evidence, greatly improves the accuracy, sensitivity and specificity of early detection of digestive tract tumors, and is suitable for non-invasive accurate screening of high-risk groups in clinical practice.

Description

Technical Field

[0001] This invention relates to the field of tumor detection technology, specifically to a method and system for early tumor detection by fusing molecular markers with MRI images. Background Technology

[0002] In the field of early detection of gastrointestinal tumors, current technologies mainly rely on two independent technical approaches. One approach is the detection of serum tumor molecular markers, commonly including carcinoembryonic antigen (CEA), alpha-fetoprotein (AFP), carbohydrate antigen (CA) series markers, and circulating tumor DNA (ctDNA) methylation. These molecular markers can provide certain evidence for the detection of gastrointestinal tumors at the molecular biological level, reflecting information related to the presence, growth, or metastasis of tumor cells.

[0003] On the other hand, there are medical imaging examinations, especially multi-sequence magnetic resonance imaging (MRI) scans, including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and contrast-enhanced scans. Multi-sequence MRI scans can provide doctors with information on the morphology, signal intensity, and function of gastrointestinal tumor lesions, which helps in the detection of tumor lesions and the initial assessment.

[0004] However, existing technologies have significant drawbacks and limitations. MRI imaging alone lacks sensitivity for detecting small lesions in the early stages of gastrointestinal tumors, easily leading to missed diagnoses and causing some early-stage tumors to go undetected. Relying solely on molecular marker detection has poor specificity and a high false-positive rate, easily causing misdiagnosis and potentially leading to overtreatment. More importantly, in practical applications, the results of these two technologies are mostly interpreted independently, lacking a systematic multimodal fusion mechanism, and failing to achieve effective complementarity between molecular and imaging evidence. This limits the overall accuracy, sensitivity, and specificity of the detection, posing a significant challenge to the accurate early screening of gastrointestinal tumors. Summary of the Invention

[0005] This invention provides a method and system for early tumor detection by fusing molecular markers with MRI images, aiming to solve the problem of accurate screening in early tumor detection where the sensitivity of single imaging examinations is insufficient and the false positive rate of molecular detection is high, leading to missed diagnoses and misdiagnoses.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for early tumor detection by fusing molecular markers with MRI images, comprising the following steps: S1. Collect serum molecular marker detection data and standardized multi-sequence MRI image data of subjects undergoing gastrointestinal tumor screening; S2. Perform multidimensional feature construction and anomaly quantification on the serum molecular marker detection data respectively, and perform lesion localization and multidimensional image feature extraction on the multi-sequence MRI image data to obtain standardized molecular feature vectors and high-dimensional MRI image feature sets. S3. Based on the standardized molecular feature vector, calculate the prior probability of gastrointestinal tumors through a molecular pre-screening deep neural network, and generate adaptive spatial attention-guided weights based on the prior probability. S4. The adaptive spatial attention-guided weights are weighted and fused with MRI image features to perform refined feature enhancement and precise lesion detection in high-risk areas, thereby obtaining image detection evidence. S5. Based on the Bayesian evidence fusion reasoning framework, molecular detection evidence and imaging detection evidence are jointly updated to obtain the final posterior probability of gastrointestinal tumor detection. S6. Output lesion localization label, detection confidence and visualization detection report based on the final detection posterior probability.

[0007] Preferably, in step S1, the serum molecular marker detection data includes quantitative data of carcinoembryonic antigen (CEA), alpha-fetoprotein (AFP), carbohydrate antigen (CA) series tumor markers and ctDNA methylation level data; the multiple sequence MRI image data includes T1WI, T2WI, DWI and dynamic contrast-enhanced sequence image data, and the image data is processed by grayscale normalization and registration, with the resolution within the preset medical imaging standard range and the slice thickness within the clinical routine scanning range.

[0008] Preferably, step S2 includes the following specific steps: 1) MRI image feature extraction: The ResNeXtAttentionUNet fusion network is used to process multi-sequence MRI images. First, multi-scale deep semantic features are extracted by ResNeXt, then the lesion region features are enhanced by attention gating mechanism, and finally the pixel-level localization of suspicious lesions is achieved by UNet decoding. A high-dimensional MRI image feature set including signal intensity features, apparent diffusion coefficient features, enhancement dynamics features, and time signal curve features is extracted. 2) Molecular biomarker characterization: The raw data of each serum molecular biomarker were normalized using the following formula: ; Where, x i Let x be the normalized value of the i-th molecular marker. i The original quantitative value, μ i This represents the clinical reference mean for this biomarker, σ. i This is the clinical reference standard deviation for this biomarker; The normalized data is then concatenated to obtain a multidimensional molecular feature vector. ; And the molecular anomaly score was calculated using a weighted anomaly scoring model: ; in, The weight of the contribution of the i-th biomarker to the detection of gastrointestinal tumors is determined, satisfying the following conditions. =1, which is learned from the clinical dataset by the L1 regularized logistic regression model.

[0009] Preferably, step S3 includes the following specific steps: The standardized molecular feature vector is input into the molecular pre-screening deep neural network, which outputs the prior probability P of a positive gastrointestinal tumor. prior The network structure includes multiple fully connected layers + batch normalization + LeakyReLU activation; Based on the prior probability P prior Based on the baseline saliency map M of MRI images, an adaptive spatial attention-guided weight W is generated. att The calculation formula is: ; Where σ is the Sigmoid activation function, which maps the weights to the (0,1) interval; M(u,v) is the basic significance value of the image location (u,v); and δ is an adaptive bias term used to balance the detection sensitivity of regions with low prior probability.

[0010] Preferably, step S4 includes the following specific steps: The adaptive spatial attention-guided weight W att MRI deep feature map F extracted by ResNeXt deep Element-wise weighted fusion is performed to obtain the attention-enhanced feature map, as shown in the following formula: ; in, For Hadamard; F att The lightweight, refined feature extraction network is input, and high-order features of lesion edges, textures, morphology, and dynamics are extracted through multi-scale convolution and hollow spatial pyramid pooling structures to obtain refined image feature maps.

[0011] Preferably, a joint detection algorithm combining bounding box regression and mask segmentation is used to locate lesions and determine their attributes in the refined image feature map. Specific steps include: 1) Candidate region generation: Based on the refined image feature map, a number of candidate regions of interest containing suspicious lesions are generated through a region proposal network; 2) Bounding box regression optimization: Bounding box coordinate regression is performed on the candidate region, and the position and size of the candidate region are corrected by the coordinate offset to obtain the accurate lesion boundary box; 3) Masked branch segmentation: Perform pixel-level masked segmentation on the lesion region located by the bounding box, and output the binarized segmentation mask of the lesion region; 4) Lesion feature fusion: The position and size features of the bounding box are combined with the morphological features of the mask segmentation to obtain the comprehensive lesion features; 5) Lesion Attribute Determination: Input the comprehensive features of the lesion into the classification branch, and output the confidence score that the lesion is malignant; use the bounding box coordinates, segmentation mask, and malignancy confidence score together as image detection evidence E. img .

[0012] Preferably, step S5 includes the following specific steps: Construct a Bayesian dual-evidence joint reasoning framework, assuming the tumor-positive event is H and the molecular evidence is E. mol The video evidence is E img And E mol With E img Conditional independence; The final posterior probability P(H|E) mol E img Update according to the following formula: ; Where: P prior The prior probability of tumors obtained from molecular prescreening; The probability of positive molecular evidence under tumor conditions is obtained from the molecular abnormality score S through a nonlinear mapping. The probability of positive imaging evidence under tumor conditions is output by a classifier based on lesion characteristics. , These represent the false positive probabilities of molecular and imaging tests under non-tumor conditions, respectively, which are learned from clinical negative samples.

[0013] Preferably, the positive probability of the molecular evidence The calculation is performed using hyperbolic tangent nonlinear mapping: ; Where k is the scaling factor and b is the translation factor, both determined by training on the labeled dataset; the positive probability of the image evidence... The output is derived from a multi-feature fusion random forest model. Input features include lesion size, diffusion coefficient, enhancement peak, curve slope, and texture entropy.

[0014] Preferably, in step S6, when the final posterior probability of the detection is greater than a preset threshold, it is determined to be a high-risk gastrointestinal tumor; the lesion localization is marked by a bounding box or pixel-level segmentation mask; the visualized detection report includes molecular abnormality items, attention weight heatmap, lesion feature parameters, posterior probability and fusion reasoning basis, and the threshold is adaptively adjusted within the standard range according to the clinical sensitivity and specificity requirements.

[0015] This invention also provides an early tumor detection system that integrates molecular markers with MRI images, comprising: The data acquisition module is used to acquire serum molecular marker data and standardized multi-sequence MRI image data; The dual-modal feature extraction module is used to perform MRI image lesion localization and feature extraction, molecular marker normalization and abnormality score calculation; The molecular pre-screening module is used to calculate the prior probability of tumors based on molecular feature vectors and generate adaptive spatial attention-guided weights. The image screening module is used to perform attention-weighted fusion, refined feature extraction, and accurate lesion detection; The Bayesian dual evidence fusion module is used to combine molecular and imaging evidence to calculate the final posterior probability of detection. The visualization output module is used to output reports on lesion location, detection probability, and detection basis.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention generates adaptive spatial attention-guided weights by pre-screening molecular markers, which are then weighted and fused with deep MRI image features and refined to enhance the features. This effectively solves the technical defects of insufficient sensitivity of single MRI images for early small lesions of gastrointestinal tumors and easy to miss diagnose, significantly improves the localization accuracy and detection capability of small lesions, and greatly reduces the risk of missed diagnosis.

[0017] 2. This invention adopts a Bayesian dual-evidence joint reasoning framework, which combines molecular detection evidence and imaging detection evidence for probability updates. This overcomes the problems of poor specificity and high false positive rate of single serum molecular marker detection, effectively eliminates the interference of false positives in molecular detection, improves the specificity of detection, avoids over-treatment due to misdiagnosis, and improves the reliability of detection.

[0018] 3. This invention constructs a systematic fusion detection mechanism of molecular markers and multi-sequence MRI images, breaking through the technical limitations of independent interpretation and non-complementary nature of traditional dual-modal data. By combining multidimensional molecular feature construction and high-dimensional image feature extraction technology, the sensitivity, specificity and accuracy of detection are optimized simultaneously, realizing early non-invasive, accurate and efficient screening of gastrointestinal tumors, and is suitable for large-scale clinical screening of high-risk populations. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0020] In the attached diagram: Figure 1 This is a flowchart of the method for early tumor detection by fusing molecular markers with MRI images according to the present invention; Figure 2 This is a general module framework diagram of the early tumor detection system that integrates molecular markers and MRI images according to the present invention; Figure 3 This is a schematic diagram of the data acquisition module of the molecular marker and MRI image fusion tumor early detection system of the present invention; Figure 4 This is a schematic diagram of the dual-modal feature extraction module of the molecular marker and MRI image fusion tumor early detection system of the present invention; Figure 5 This is a schematic diagram of the molecular pre-screening module of the molecular marker and MRI image fusion tumor early detection system of the present invention; Figure 6 This is a schematic diagram of the image screening module of the early tumor detection system that integrates molecular markers and MRI images according to the present invention; Figure 7 This is a schematic diagram of the Bayesian dual evidence fusion module of the molecular marker and MRI image fusion tumor early detection system of the present invention; Figure 8 This is a schematic diagram of the visualization output module of the molecular marker and MRI image fusion tumor early detection system of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] Example: This invention provides a method for early tumor detection by fusing molecular markers with MRI images, such as... Figure 1 As shown, the specific steps include the following: Step S1: Data Acquisition 1) Serum molecular marker data collection: 5 mL of fasting venous blood was collected from the subjects and centrifuged at 3000 r / min for 10 min to separate the serum. The quantitative data of CEA, AFP, CA19-9 and CA72-4 were detected by a fully automated chemiluminescence immunoassay analyzer. The serum ctDNA methylation level was detected by a digital PCR instrument to obtain complete serum molecular marker detection data.

[0023] 2) Multi-sequence MRI image data acquisition and preprocessing: A 3.0T clinical medical magnetic resonance scanner, in conjunction with an abdominal phased array coil, was used to complete the scan. T1WI: TR=500ms, TE=12ms, layer thickness 3mm; T2WI: TR=4000ms, TE=85ms, layer thickness 3mm; DWI: b-value = 0, 800 s / mm 2 The layer thickness is 3mm; DCE-MRI: Gadolinate dimeglumine was injected at a dose of 0.1 mmol / kg at a rate of 2 mL / s. After scanning, the multi-sequence images were rigidly registered and grayscale normalized to a uniform resolution of 512×512. Motion artifacts were removed to obtain standardized MRI image data.

[0024] Step S2: Dual-modal feature extraction and construction 2.1 Characterization of serum molecular markers The raw test data of each serum molecular marker were normalized according to the generally accepted clinical reference values. The normalization formula is as follows: The normalized data are concatenated to form a 5-dimensional molecular feature vector: X=[x1′,x2′,...,x n ′]; The detection contribution weights of each biomarker are obtained based on an L1 regularized logistic regression model (satisfying...). The molecular anomaly score is calculated using a weighted anomaly scoring model: ; Among them, the molecular abnormality score S ranges from 0 to 3.5.

[0025] 2.2 Multidimensional Feature Extraction from MRI Images The ResNeXtAttentionUNet fusion network was used to process multi-sequence MRI images. ResNeXt-50 was used to extract multi-scale deep semantic features, and the lesion region features were enhanced by attention gating mechanism. Then, UNet decoding was used to achieve pixel-level localization of lesions. Finally, a 256-dimensional high-dimensional MRI image feature set was extracted, including signal intensity, apparent diffusion coefficient, enhancement dynamics, and time-signal curve features.

[0026] Step S3: Molecular pre-screening and adaptive attention weight generation 1) Calculating the prior probability of tumors: Constructing a molecular pre-screening deep neural network with 3 fully connected layers, batch normalization, and LeakyReLU activation. Inputting molecular feature vectors into the network, it outputs the positive prior probability P of gastrointestinal tumors. priorThe value range is 0.01 to 0.98.

[0027] 2) Adaptive spatial attention-guided weight generation Combining prior probabilities with the baseline saliency map of MRI images, adaptive spatial attention-guided weights are generated via Sigmoid activation function mapping. The calculation formula is as follows: ; Specific values ​​in the formula: δ = 0.1 (adaptive bias term), σ is the Sigmoid activation function, W att (u,v) takes values ​​in the range of 0 to 1.

[0028] Step S4 Attention Fusion and Precise Lesion Detection The adaptive spatial attention-guided weights are fused with the deep MRI feature map element-wise using the following formula: ; In the formula: For the Hadamarda product, F deep (u,v) is a 256-dimensional MRI deep feature map.

[0029] F att (u,v) is input to a lightweight, refined feature extraction network, which extracts high-order features of lesions through multi-scale convolution and dilated spatial pyramid pooling. A joint detection algorithm combining bounding box regression and mask segmentation is employed, with an IOU threshold of 0.5, to generate candidate lesion regions, optimize lesion boundaries, and complete pixel-level segmentation. The algorithm outputs a lesion malignancy confidence score, which is then integrated to form the image detection evidence E. img .

[0030] Step S5 Bayesian dual-evidence fusion reasoning Construct a Bayesian dual-evidence joint reasoning framework, assuming the tumor-positive event is H and the molecular evidence is E. mol The video evidence is E img And E mol With E img Conditionally independent, the final posterior probability calculation formula is: ; The specific values ​​in the formula are: k = 1.2 (scaling factor), b = -0.8 (translation factor).

[0031] The positive probability of imaging evidence is output by a multi-feature fusion random forest model with 100 decision trees; the false positive probability is learned from clinical negative samples. =0.08, =0.12, substitute into the formula to complete the joint probability update.

[0032] Step S6 Detection Result Output Based on the clinically optimal balance parameters, the detection threshold was set to 0.75. A posterior probability > 0.75 was considered high-risk for colorectal cancer. A combination of bounding boxes and pixel-level segmentation masks was used to locate and label lesions with an accuracy ≤ 1 mm. A visual detection report was generated, including molecular anomalies, attention weight heatmaps, lesion characteristic parameters, posterior probability, and fusion reasoning basis. The report format is compatible with clinical PACS systems.

[0033] like Figures 2-8 As shown, this invention also provides an early tumor detection system that integrates molecular markers and MRI images. This system is used to perform the above-mentioned detection method and is composed of six functional modules. The specific composition and function of each module are as follows: Data acquisition module: Composed of serum molecular marker detection equipment, 3.0T MRI scanner, and standardized data transmission interface, it realizes standardized data acquisition, transmission and storage.

[0034] The dual-modal feature extraction module includes an image preprocessing unit, an MRI image feature extraction unit, and a molecular feature calculation unit. The molecular feature calculation unit is used to normalize molecular features and calculate anomaly scores, outputting a 5-dimensional standardized molecular feature vector and a 256-dimensional high-dimensional MRI image feature set.

[0035] Molecular pre-screening module: It consists of a molecular feature parsing unit, a prior probability calculation unit, and an attention weight generation unit, and outputs the tumor prior probability and adaptive spatial attention-guided weights in the 0-1 interval.

[0036] Image screening module: includes feature weighted fusion unit, refined feature extraction unit, and lesion joint detection unit, to achieve accurate lesion localization and attribute determination, and output image detection evidence.

[0037] The Bayesian dual evidence fusion module consists of an evidence probability mapping unit and a joint probability update unit, which calculates the final posterior probability of detection.

[0038] Visualization output module: It consists of lesion annotation unit, test report generation unit and clinical display terminal, which completes lesion annotation and report generation.

[0039] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early tumor detection by fusing molecular markers with MRI images, characterized in that, Includes the following steps: S1. Collect serum molecular marker detection data and multi-sequence MRI image data of subjects undergoing gastrointestinal tumor screening; S2. Multidimensional feature construction and abnormal quantification are performed on serum molecular marker detection data, and lesion localization and multidimensional image feature extraction are performed on multi-sequence MRI image data to obtain molecular feature vectors and high-dimensional MRI image feature sets. S3. Based on molecular feature vectors, calculate the prior probability of gastrointestinal tumors through a molecular pre-screening deep neural network and generate adaptive spatial attention-guided weights. S4. The adaptive spatial attention-guided weights are weighted and fused with MRI image features to perform refined feature enhancement and precise lesion detection in high-risk areas, thereby obtaining imaging detection evidence. S5. Based on the Bayesian evidence fusion reasoning framework, molecular detection evidence and imaging detection evidence are jointly updated to obtain the final posterior probability of gastrointestinal tumor detection. S6. Output lesion localization label, detection confidence and visualization detection report based on the final posterior probability of detection.

2. The method for early tumor detection by fusing molecular markers with MRI images according to claim 1, characterized in that: In step S1, serum molecular marker detection data includes quantitative data of tumor markers and ctDNA methylation level data; multiple sequence MRI imaging data includes T1WI, T2WI, DWI and dynamic contrast-enhanced sequence imaging data.

3. The method for early tumor detection by fusing molecular markers with MRI images according to claim 1, characterized in that: Step S2 includes the following specific steps: 1) MRI image feature extraction: The ResNeXtAttentionUNet fusion network is used to process multi-sequence MRI images. First, multi-scale deep semantic features are extracted by ResNeXt, then the lesion region features are enhanced by attention gating mechanism, and finally the pixel-level localization of suspicious lesions is achieved by UNet decoding. A high-dimensional MRI image feature set including signal intensity features, apparent diffusion coefficient features, enhancement dynamics features, and time signal curve features is extracted. 2) Molecular biomarker characterization: The raw data of each serum molecular biomarker were normalized using the following formula: ； Where, x i Let x be the normalized value of the i-th molecular marker. i The original quantitative value, μ i This represents the clinical reference mean for this biomarker, σ. i This is the clinical reference standard deviation for this biomarker; The normalized data is then concatenated to obtain a multidimensional molecular feature vector. ； And the molecular anomaly score was calculated using a weighted anomaly scoring model: ； in, The contribution weight of the i-th biomarker to the detection of gastrointestinal tumors is determined by satisfying the following conditions. =1, which is learned from the clinical dataset by the L1 regularized logistic regression model.

4. The method for early tumor detection by fusing molecular markers and MRI images according to claim 1, characterized in that: Step S3 includes the following specific steps: The molecular feature vector is input into the molecular pre-screening deep neural network, which outputs the prior probability P of gastrointestinal tumors. prior The network structure includes multiple fully connected layers + batch normalization + LeakyReLU activation; Based on prior probability P prior Based on the baseline saliency map M of MRI images, an adaptive spatial attention-guided weight W is generated. att The calculation formula is: ； Where σ is the Sigmoid activation function, which maps the weights to the (0,1) interval; M(u,v) is the basic significance value of the image location (u,v); and δ is an adaptive bias term used to balance the detection sensitivity of regions with low prior probability.

5. The method for early tumor detection by fusing molecular markers with MRI images according to claim 1, characterized in that: Step S4 The specific steps include: Guide the adaptive spatial attention weight W att MRI deep feature map F extracted by ResNeXt deep Element-wise weighted fusion is performed to obtain the attention-enhanced feature map, as shown in the following formula: ； in, For Hadamard; F att The lightweight, refined feature extraction network is input, and high-order features of lesion edges, textures, morphology, and dynamics are extracted through multi-scale convolution and hollow spatial pyramid pooling structures to obtain refined image feature maps.

6. The method for early tumor detection by fusing molecular markers with MRI images according to claim 5, characterized in that: A joint detection algorithm combining bounding box regression and mask segmentation is used to locate lesions and determine their attributes in refined image feature maps. The specific steps include: 1) Candidate region generation: Based on the refined image feature map, a number of candidate regions of interest containing suspicious lesions are generated through a region proposal network; 2) Bounding box regression optimization: Bounding box coordinate regression is performed on the candidate region, and the position and size of the candidate region are corrected by the coordinate offset to obtain the accurate lesion boundary box; 3) Masked branch segmentation: Perform pixel-level masked segmentation on the lesion region located by the bounding box, and output the binarized segmentation mask of the lesion region; 4) Lesion feature fusion: The position and size features of the bounding box are combined with the morphological features of the mask segmentation to obtain the comprehensive lesion features; 5) Lesion Attribute Determination: Input the comprehensive features of the lesion into the classification branch, and output the confidence score that the lesion is malignant; use the bounding box coordinates, segmentation mask, and malignancy confidence score together as image detection evidence E. img .

7. The method for early tumor detection by fusing molecular markers with MRI images according to claim 1, characterized in that: Step S5 includes the following specific steps: Construct a Bayesian dual-evidence joint reasoning framework, assuming the tumor-positive event is H and the molecular evidence is E. mol The video evidence is E img And E mol With E img Conditional independence; Final posterior probability Update using the following formula: ； Where: P prior The prior probability of tumors obtained from molecular prescreening; The probability of positive molecular evidence under tumor conditions is obtained from the molecular abnormality score S through a nonlinear mapping. The probability of positive imaging evidence under tumor conditions is output by a classifier based on lesion characteristics. , These represent the false positive probabilities of molecular and imaging tests under non-tumor conditions, respectively, which are learned from clinical negative samples.

8. The method for early tumor detection by fusing molecular markers with MRI images according to claim 7, characterized in that: Molecular evidence positive probability The calculation is performed using hyperbolic tangent nonlinear mapping: ； Where k is the scaling factor and b is the translation factor, both determined by training on the labeled dataset; positive probability of image evidence. The output is derived from a multi-feature fusion random forest model. Input features include lesion size, diffusion coefficient, enhancement peak, curve slope, and texture entropy.

9. The method for early tumor detection by fusing molecular markers with MRI images according to claim 1, characterized in that: In step S6, when the final posterior probability of the detection is greater than the preset threshold, it is determined to be a high-risk gastrointestinal tumor; the lesion is located by using bounding boxes or pixel-level segmentation masks; the visual detection report includes molecular abnormalities, attention weight heatmap, lesion feature parameters, posterior probability and fusion reasoning basis, and the threshold is adaptively adjusted within the standard range according to the clinical sensitivity and specificity requirements.

10. A molecular marker-MRI image fusion system for early tumor detection, used to perform the method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire serum molecular marker data and standardized multi-sequence MRI image data; The dual-modal feature extraction module is used to perform MRI image lesion localization and feature extraction, molecular marker normalization and abnormality score calculation; The molecular pre-screening module is used to calculate the prior probability of tumors based on molecular feature vectors and generate adaptive spatial attention-guided weights. The image screening module is used to perform attention-weighted fusion, refined feature extraction, and accurate lesion detection; The Bayesian dual evidence fusion module is used to combine molecular and imaging evidence to calculate the final posterior probability of detection. The visualization output module is used to output reports on lesion location, detection probability, and detection basis.

Images