Non-invasive colorectal cancer early screening and diagnosis method and electronic equipment

Abstract

The invention discloses a non-invasive colorectal cancer early screening and diagnosis method and electronic equipment. The method comprises the following steps: collecting excrement, saliva, a blood sample and a standardized tongue image of a subject; species abundance information of intestinal flora and oral flora is obtained through metagenome sequencing, and the tumor marker level is obtained through blood detection; and inputting the multi-modal data into a pre-trained colorectal cancer risk prediction model, and outputting a risk probability. The model is obtained by training based on a machine learning algorithm by adopting a training data set containing healthy persons and patients, the architecture of the model adopts a late fusion strategy, clinical, flora and tongue picture features are respectively processed through three sub-models of logistic regression, a gradient boosting decision tree and a convolutional neural network, dynamic fusion is performed by utilizing a Self-Attention mechanism, and finally a diagnosis result is output. According to the method, through multi-modal information fusion, the screening accuracy, the standardization degree and the patient compliance are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of biological medicine and artificial intelligence, and particularly relates to a non-invasive colorectal cancer early screening and diagnosis method and an electronic device. BACKGROUND

[0002] Colorectal cancer is a common malignant tumor of the digestive tract, and its early detection and diagnosis are crucial for improving patient survival rate. Currently, colonoscopy is the "gold standard" for CRC diagnosis, but it is an invasive examination with poor patient tolerance, high cost and complex operation, and is not suitable for large-scale population screening. Although fecal occult blood test is simple and non-invasive, its sensitivity and specificity are limited, especially for early adenomas and precancerous lesions.

[0003] In the prior art, a multi-target fecal FIT-DNA joint detection technology represented by "Chang Weiqing" has appeared. This technology detects gene mutations (such as KRAS), gene methylation (such as BMP3, NDRG4) and fecal occult blood indicators in feces, and conducts risk assessment combined with a patented algorithm. However, this technical solution has the following obvious defects: 1. Single detection dimension: mainly relying on fecal DNA, without integrating other dimensional biological information, limiting the further improvement of detection accuracy and reliability.

[0004] 2. Lack of standardization: lack of unified standardized operation process, leading to differences in detection results between different laboratories or operators, affecting the comparability and reliability of the results.

[0005] Therefore, there is an urgent need in the art for a non-invasive colorectal cancer early screening and diagnosis method with higher accuracy, standardization and easy popularization. SUMMARY

[0006] The present application aims to solve the above technical problems in the prior art, and provides a non-invasive colorectal cancer early screening and diagnosis method and an electronic device. The method integrates multi-modal biological information and clinical data, and uses advanced artificial intelligence algorithms for fusion analysis, aiming to improve the accuracy, specificity and sensitivity of screening, while reducing the screening cost and improving patient compliance.

[0007] To solve the above technical problems, the present application adopts the following technical solutions: The non-invasive colorectal cancer early screening and diagnosis method comprises the following steps: S1, multi-modal data acquisition: collecting fecal samples, saliva samples, blood samples and tongue images of the subject; S2, laboratory detection: S21, perform metagenomic sequencing on the fecal sample and the saliva sample to obtain the species abundance information of the gut microbiota and the oral microbiota; S22, perform detection on the blood sample to obtain the tumor marker level, and optionally, obtain other auxiliary indexes (such as D-dimer) related to inflammation or thrombosis; S3, artificial intelligence model analysis: input the tongue image, the species abundance information, and the tumor marker level obtained in steps S1 and S2 into a pre-trained colorectal cancer risk prediction model to obtain the risk probability of the subject suffering from colorectal cancer; The colorectal cancer risk prediction model is a multi-modal fusion model trained based on a machine learning algorithm using a training data set containing healthy subjects and colorectal cancer patients.

[0008] Further, in step S1, when collecting the tongue image, standardized shooting conditions are used, including: lighting under a light source with a color temperature of 6500K, a shooting angle of 10° depression angle with the horizontal line, and using a standard color card as a color reference for shooting; the tongue image is subjected to color correction processing based on the standard color card before being input into the model.

[0009] Further, after color correction of the tongue image, the tongue coating and tongue region pixels are further separated in the RGB color space by the K-means clustering algorithm.

[0010] Further, the metagenomic sequencing in step S21 includes the following sub-steps: S211, DNA extraction and enrichment: extract whole genome DNA from the fecal sample and the saliva sample, and specifically enrich microbial DNA; S212, library construction: fragment the enriched DNA, and perform end repair, adapter ligation, purification, and PCR amplification to prepare a sequencing library; S213, sequencing: use a NovaSeq6000 sequencing platform to perform high-throughput sequencing in PE150 mode; S214, data processing: perform quality filtering on the raw sequencing data obtained from the machine, and remove host genome sequences; S215, sequence assembly and annotation: de novo assemble the filtered data, predict genes for the contigs obtained from the assembly to obtain a non-redundant gene set; align the non-redundant gene set with a species annotation database to obtain the species abundance information of each sample.

[0011] Further, the colorectal cancer risk prediction model uses a late fusion architecture, including: The first sub-model is a logistic regression model for processing structured clinical information, including tumor marker levels and thrombus indexes; The second sub-model is a gradient boosting decision tree (GBDT) model, which is used to process the species abundance information of intestinal flora and oral flora. The third sub-model is a convolutional neural network (CNN) based model, which is used to process tongue images and extract visual features. The fusion layer is used to fuse the feature vectors output by the first sub-model, the second sub-model and the third sub-model, and output the final risk probability. The fusion layer adopts a Self-Attention mechanism to dynamically calculate the weights of different modal feature vectors.

[0012] Further, the execution process of the Self-Attention mechanism includes: The input feature sequence is transformed by linear transformation to obtain a query matrix Q, a key matrix K and a value matrix V. The attention weight matrix is calculated as follows: wherein is the dimension of the key vector; The value matrix V is weighted and summed to obtain the fused feature representation: .

[0013] Further, when training the second sub-model, the bacterial flora features used for training are selected from a plurality of bacterial genera by combining bioinformatics differential analysis with feature importance ranking of the GBDT model.

[0014] Further, the number of significantly different bacterial genera is 7.

[0015] Further, the tumor marker is carcinoembryonic antigen (CEA).

[0016] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the non-invasive colorectal cancer early screening and diagnosis method when executing the computer program.

[0017] The present application successfully constructs a high-performance colorectal cancer early screening and diagnosis model by fusing multi-modal data and using advanced machine learning algorithms. The prediction accuracy of the model on the test set reaches 93.1%, and the AUC value is 0.95, showing high accuracy.

[0018] The present application has the following beneficial effects due to the use of the above technical solutions: 1. Screening accuracy and reliability are improved: The present application fuses intestinal flora, oral flora, tongue image and blood tumor markers for multi-modal analysis. By constructing a multi-modal fusion model based on artificial intelligence, the complementarity and cross-validation of different dimensional biological information are realized. Even if the signal of a certain mode is weak (such as tumor markers not elevated), the strong features of other modes (such as specific flora imbalance and typical tongue changes) can ensure accurate judgment of the model. This three-dimensional diagnostic strategy overcomes the limitations of single-dimensional detection. The final model of the method achieves an accuracy of 93.1% and an AUC value of 0.95 on the validation set, showing excellent performance and effectively reducing the misdiagnosis rate and missed diagnosis rate.

[0019] 2. Standardization and repeatability of the whole process are realized: The present application implements strict standardization control on key links. On the one hand, it clearly stipulates the collection of four sample types of feces, saliva, blood and tongue image, ensuring the comprehensiveness and consistency of the data source; on the other hand, especially for the tongue analysis which is traditionally highly subjective, it is converted into objective and quantifiable imaging data through standardized shooting conditions and color correction. Combined with the highly standardized metagenomic sequencing and blood detection technology, the present application ensures the standardization and consistency of the whole process from sample collection to data analysis, improves the repeatability of the diagnostic results, and lays a solid foundation for large-scale clinical application and promotion.

[0020] 3. The universality and patient compliance of screening are improved: All samples used in the present application can be obtained by non-invasive or minimally invasive methods, greatly improving the experience of the subjects and avoiding the pain and risk brought by traditional colonoscopy. This convenient and efficient detection method can effectively improve the willingness and compliance of the target population, and has important positive significance for promoting the popularization of early screening of colorectal cancer, achieving early detection, early diagnosis and early treatment. BRIEF DESCRIPTION OF DRAWINGS

[0021] The present application will be further described below in conjunction with the drawings: Figure 1 is the overall flowchart of the present application; Figure 2 is the feces random forest three-classification chart of the present application; Figure 3 is the saliva random forest three-classification chart of the present application. DETAILED DESCRIPTION

[0022] As shown in Figure 1 , it is the non-invasive colorectal cancer early screening and diagnosis method of the present application. Example 1

[0023] 1. Data preparation and sample collection According to the aforementioned criteria, a total of 300 subjects were recruited, including 100 healthy controls (HC), 100 patients with non-metastatic colorectal cancer (CRC stage I-III), and 100 patients with metastatic colorectal cancer (CRC stage IV). All CRC patients were pathologically diagnosed and met the TNM staging criteria.

[0024] 1.1 Standardized sample collection procedure Fecal and saliva samples: Collected using the NIH Human Microbiome Project (HMP) standard sampling kit. After collection, the samples were immediately frozen in a -80°C ultra-low temperature freezer until DNA extraction.

[0025] Blood samples: Peripheral blood was collected using EDTA anticoagulant tubes, and plasma was separated after centrifugation for the detection of tumor markers such as carcinoembryonic antigen (CEA).

[0026] Tongue image collection: 1) Environment and equipment: In an indoor natural light environment, use HD-130 macro LED ring flash (color temperature 6500K) for uniform lighting. The shooting device is a Canon 77D digital single-lens reflex camera equipped with a fixed-focus macro lens.

[0027] 2) Parameters and posture: Camera parameter settings are aperture f / 8 and ISO 1600. The patient naturally stretches the tongue, and the photographer adjusts the camera so that its lens plane is parallel to the patient's tongue surface, with a shooting angle of 10° depression angle from the horizontal line.

[0028] 3) Color correction: The patient holds the X-Rite Color Checker Classic Mini 24-color standard color card so that it is in the same focal plane as the tongue, ensuring accurate color reproduction accuracy. Each image is saved in RAW+JPEG format.

[0029] 1.2 Data preprocessing Tongue image preprocessing: 1) Color correction: Convert the RAW format image containing the color card to DNG format using Adobe Lightroom Classic, and generate an International Color Consortium Profile (ICC Profile) through the built-in Color Checker function. Then, apply this ICC Profile to batch color correction of tongue images taken in the same environment.

[0030] 2) Tongue body segmentation: Python programming was used to convert the corrected image from the RGB color space to the LAB color space using the OpenCV library. Subsequently, the K-means clustering algorithm (k = 2) was applied to cluster the pixels, automatically separating the tongue fur region from the tongue body region, and extracting the pure tongue body region image for subsequent feature extraction.

[0031] 2. Laboratory testing and analysis 2.1 Metagenomic sequencing DNA extraction and enrichment: Whole genome DNA was extracted from 200 mg of fecal samples and 2 ml of salivary sample precipitate using the QIAamp DNA Mini Kit. Subsequently, the extracted DNA (160 μl of MBD-Fc-bound magnetic beads were used for every 4 μg) was incubated using the NEB Next Microbiome DNA Enrichment Kit to specifically enrich microbial DNA, and purified by ethanol precipitation. The DNA concentration and purity (A260 / A280 ratio between 1.8-2.0) were detected using the NanoDrop 2000.

[0032] Library construction and sequencing: The qualified DNA was randomly broken down to about 350 bp fragments using the Covaris ultrasonic disrupter. After end repair, 3' end "A" addition, ligation of sequencing adapters, and purification screening, PCR amplification was performed to complete the construction of the sequencing library. After library quality control, high-throughput sequencing was performed in PE150 (double-end 150 bp) mode on the Illumina NovaSeq 6000 platform.

[0033] Bioinformatics analysis: 1) Data quality control: Fastp software was used to quality control the raw sequencing data (Raw Data), removing sequences with more than 5% N bases and low-quality sequences (Q≤20 bases accounting for more than 50% of the sequence length).

[0034] 2) Host decontamination: The quality-controlled data was aligned with the human reference genome (hg19) using Bowtie2, and the aligned sequences were removed to obtain pure microbial sequencing data (Clean Data).

[0035] 3) Assembly and annotation: MEGAHIT software was used to de novo assemble the Clean Data of each sample to obtain Contigs (contigs). MetaGeneMark was used to predict genes from Contigs, and CD-HIT software was used to cluster the predicted genes to construct a non-redundant gene set (Unigenes) after redundancy removal.

[0036] 4) Species and functional annotation: Unigenes were aligned against NR database (version: NR meta) to obtain species classification information and abundance table from door to species level using LCA algorithm. At the same time, Unigenes were aligned against KEGG and GO databases for functional annotation and pathway enrichment analysis.

[0037] 2.2 Tumor marker detection The CEA level in plasma was detected using Roche Cobas e 601 electrochemiluminescence immunoassay analyzer and matching reagent kit, and the detection strictly followed the operation procedure of the manufacturer. The critical value was set to 5.0 ng / mL.

[0038] 3, Construction and training of artificial intelligence model 3.1 Feature engineering and dataset division 3.1.1 Feature source: Clinical features: including CEA value (standardized by Z-score), patient age, gender and thrombus index (such as D-dimer, standardized by Z-score).

[0039] Microbiota features: from the species annotation results, the top 400 genera in terms of abundance at the genus level were selected, and their relative abundance was used as the initial feature pool. To select the most relevant genera to colorectal cancer, we used the "bioinformatics analysis + GBDT" double feature screening method: 1) Bioinformatics differential analysis: Wilcoxon rank-sum test was used to compare the relative abundance of each genus between the healthy control group and the colorectal cancer group, and the significantly different genera (p-value <0.05) were selected, a total of 35 candidate genera.

[0040] 2) Machine learning feature importance ranking: the relative abundance data of the above 35 candidate genera were input into the GBDT model for training, and the SHAP value analysis was used to calculate the importance ranking of each genus on the model prediction results.

[0041] 3) Determine the key feature set: based on the ranking results of the above two methods, we finally determined 7 genera with the highest contribution as the final microbial features for model construction. The 7 genera are: Fusobacterium, Bacteroides, Prevotella, Roseburia, Coprococcus, Alistipes and Ruminococcus.

[0042] Subsequently, only the relative abundance data of the 7 key genera were used as the flora feature vector to input into the final GBDT sub-model for training.

[0043] Tongue image feature: The segmented tongue region image after preprocessing was uniformly scaled to 224x224 pixels and input into a pre-trained ResNet-50 convolutional neural network (CNN) to extract a 1024-dimensional feature vector before the last fully connected layer as the tongue image feature.

[0044] 3.1.2 Dataset division: The complete dataset of 300 samples (containing the above-processed multi-modal features and labels) was randomly divided into a training set (210 samples) and a validation set (90 samples) in a ratio of 7:3.

[0045] 3.2 Model architecture and training The multi-modal fusion model of the application adopts a late fusion strategy, and its architecture is as follows: First sub-model (clinical data): a logistic regression model is used, and the input is the standardized clinical feature vector.

[0046] Second sub-model (flora data): a gradient boosting decision tree (GBDT) model is used, and the input is the relative abundance vector of the 7 key genera selected above.

[0047] Third sub-model (tongue image data): a ResNet-50 model pre-trained on ImageNet (with the last classification layer removed) is used, and the input is the standardized tongue image, and the output is a 1024-dimensional feature vector.

[0048] Fusion layer: the feature vectors output by the above three sub-models are spliced to form a fusion feature vector. The fusion vector is input into a fusion layer using the Self-Attention mechanism.

[0049] Self-Attention mechanism implementation: the fusion feature vector is regarded as a sequence. It is linearly transformed into query (Q), key (K), and value (V) matrices through a learnable weight matrix. Through the formula , calculation is performed to enable the model to dynamically learn and assign different importance weights to different modal features.

[0050] Output layer: the output of the Self-Attention layer is passed to a fully connected layer, and finally two nodes are output through the Softmax function, representing the probabilities of the sample being predicted as "healthy" and "colorectal cancer", respectively.

[0051] The mathematical principle of the Self-Attention mechanism is as follows: For an input sequence where n is the sequence length and d is the feature dimension. First, three sets of learnable linear transformations are applied to obtain query matrix key matrix and value matrix respectively. Where are model parameters.

[0052] In the calculation process, first, the matching degree (i.e. the correlation score) between each position is calculated by the inner product of the query vector and the key vector: .

[0053] In order to alleviate the problem of too large numerical value of high-dimensional vector inner product, the correlation score is divided by to scale (i.e. scaled dot-product attention). Then, the normalized score is obtained by the Softmax function to obtain the attention weight matrix a:

[0054] Finally, the attention weight matrix is used to weight the sum of the value matrix V to obtain the output representation Z that integrates global information:

[0055] This mechanism enables the model to dynamically learn and assign different importance weights to different modal features.

[0056] 3.3 Model training and verification Training: Use the training set data, use cross-entropy as the loss function, and use the Adam optimizer to train the whole model end-to-end.

[0057] Verification: 1) Evaluate the model performance on the validation set, and the results show that the prediction accuracy of the model for healthy people and colorectal cancer patients reaches 92.7% and 93.1% respectively, and the AUC value is 0.95.

[0058] 2) To demonstrate the necessity of the multi-modal fusion strategy adopted by the invention (rather than relying on a single modality), and to evaluate the feature contribution of different modalities, the project team constructed a random forest model based on saliva flora and fecal flora for three classification (healthy / NC, non-metastatic / NM, metastatic / M) tasks. The results (as shown in Figure 2 and Figure 3 ) clearly show that: In terms of distinguishing healthy people (NC) from non-metastatic patients (NM), the performance of the fecal flora model (AUC is 0.927 and 0.902 respectively) is significantly better than that of the saliva flora model (AUC is 0.679 and 0.674 respectively).

[0059] In terms of identifying metastatic patients (M), the fecal microbiota model (AUC=0.966) also showed better performance than the salivary microbiota model (AUC=0.931).

[0060] From the overall performance (Micro-average AUC), the fecal microbiota model (0.934) was much higher than the salivary microbiota model (0.813).

[0061] The above results demonstrate that the fecal microbiota shows better characterization ability than the oral microbiota throughout the whole process of colorectal cancer (from health to advanced stage), especially in the non-metastatic stage which is crucial for early screening. This strongly supports the rationality of the scheme of taking fecal microbiota as one of the core modalities, and also reveals that there is an upper limit to the performance of a single modality (even the better performing fecal microbiota) in complex disease diagnosis, thereby proving the value of the core invention idea of integrating complementary information through multi-modal fusion to obtain more stable and comprehensive diagnostic results.

[0062] 3) Using 6-fold cross-validation repeated 40 times, the model average AUC value is 0.94±0.02, showing good robustness.

[0063] 4) An additional 100 patients with intestinal adenomas were recruited as an independent test set, and the model's detection sensitivity for them was 88.0%, indicating that it also has good early warning ability for precancerous lesions. Example Two

[0064] A 55-year-old asymptomatic male came for early screening because his father had a history of colorectal cancer.

[0065] Step one, medical staff collected his fecal, saliva, blood samples and tongue image according to the standardized process of 1.1.

[0066] Step two, the laboratory obtained the abundance data of the 7 key genera in the intestinal / oral cavity, CEA value (3.1 ng / mL) and related thrombus indicators according to the methods of 2.1 and 2.2.

[0067] Step three, the image processing personnel performed color correction and segmentation on the tongue image of the subject according to the method of 1.2.

[0068] Step four, input all the processed data above into the multi-modal artificial intelligence model trained in Example 1.

[0069] Step five, the model output the probability of the subject being at high risk of colorectal cancer as 86%. Based on this, the clinician recommended that he undergo a colonoscopy. This example demonstrates the complete workflow and application potential of the method of the present invention in identifying high-risk groups of colorectal cancer and precancerous lesions.

[0070] The application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor implements the steps of the non-invasive colorectal cancer early screening and diagnosis method as above when running the program.

[0071] The above merely describes specific embodiments of the present application, but the technical features of the present application are not limited thereto. Any simple change, equivalent replacement, or modification made on the basis of the present application to solve the basically same technical problem and achieve the basically same technical effect shall be encompassed in the protection scope of the present application.

Claims

1. A non-invasive method for early screening and diagnosis of colorectal cancer, characterized in that, Comprising the following steps: S1, multi-modal data acquisition: collecting fecal samples, saliva samples, blood samples and tongue images of the subject; S2, laboratory detection: S21, performing metagenomic sequencing on the fecal samples and saliva samples to obtain species abundance information of intestinal flora and oral flora; S22, detecting the blood samples to obtain tumor marker levels; S3, artificial intelligence model analysis: inputting the tongue images, species abundance information and tumor marker levels obtained in steps S1 and S2 into a pre-trained colorectal cancer risk prediction model to obtain the risk probability of the subject suffering from colorectal cancer; Wherein, the colorectal cancer risk prediction model is a multi-modal fusion model trained based on machine learning algorithm using a training data set containing healthy subjects and colorectal cancer patients.

2. The non-invasive early screening and diagnostic method for colorectal cancer according to claim 1, characterized in that, In step S1, when collecting the tongue image, standardized shooting conditions are adopted, including: supplementing light under a light source with a color temperature of 6500K, a shooting angle of 10° depression angle with the horizontal line, and using a standard color card as a color reference for shooting; the tongue image is subjected to color correction processing based on the standard color card before being input into the model.

3. The non-invasive early screening and diagnostic method for colorectal cancer according to claim 1, wherein, After color correction, the tongue image is further subjected to pixel separation of tongue fur and tongue body regions in the RGB color space by K-means clustering algorithm.

4. The non-invasive early screening and diagnostic method for colorectal cancer according to claim 1, wherein, The metagenomic sequencing in step S21 comprises the following sub-steps: S211, DNA extraction and enrichment: extracting whole genome DNA from the fecal samples and saliva samples, and specifically enriching microbial DNA; S212, library construction: fragmenting the enriched DNA, and preparing a sequencing library through end repair, adapter ligation, purification and PCR amplification; S213, sequencing: using NovaSeq6000 sequencing platform to perform high-throughput sequencing in PE150 mode; S214, data processing: performing quality filtering on the raw sequencing data obtained from the machine, and removing host genome sequences; S215, sequence assembly and annotation: de novo assembly of the filtered data, gene prediction of the contigs obtained by assembly to obtain a non-redundant gene set; aligning the non-redundant gene set with a species annotation database to obtain species abundance information of each sample.

5. The non-invasive early screening and diagnostic method for colorectal cancer according to claim 1, wherein, The colorectal cancer risk prediction model adopts a late fusion architecture, comprising: A first sub-model, which is a logistic regression model, is used to process structured clinical information, including tumor marker levels and thrombus indicators; A second sub-model, which is a gradient boosting decision tree (GBDT) model, is used to process the species abundance information of intestinal flora and oral flora; A third sub-model, which is a convolutional neural network (CNN) based model, is used to process the tongue image and extract visual features; A fusion layer is used to fuse the feature vectors output by the first, second and third sub-models and output the final risk probability; Wherein, the fusion layer adopts a Self-Attention mechanism to dynamically calculate the weights of different modal feature vectors.

6. The non-invasive early screening and diagnostic method for colorectal cancer according to claim 5, wherein, The execution process of the Self-Attention mechanism comprises: linearly transforming the input feature sequence to obtain a query matrix Q, a key matrix K and a value matrix V; Computing the attention weight matrix: where is the dimension of the key vector; The weighted sum of the value matrix V is obtained, and the fused feature representation is obtained: .

7. The non-invasive early screening and diagnostic method for colorectal cancer according to claim 5, wherein, In training the second sub-model, the training flora characteristics are significant difference flora selected from multiple genera by combining bioinformatics differential analysis and GBDT model feature importance ranking.

8. The non-invasive early screening and diagnostic method for colorectal cancer according to claim 7, characterized in that, The number of the significant difference flora is 7.

9. The non-invasive early screening and diagnostic method for colorectal cancer according to claim 1, wherein, The tumor marker is carcinoembryonic antigen CEA.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the non-invasive colorectal cancer early screening and diagnosis method according to any one of claims 1 to 9 when executing the computer program.

Images