Esophageal fistula risk prediction method based on clinical data and radiomics features

Abstract

The present invention discloses a method for predicting the risk of esophageal fistula based on clinical data and imaging omics features, which relates to the field of medical artificial intelligence technology. The present invention combines deep learning with traditional imaging omics methods, extracts rich imaging omics features through multiple image processing technologies, and ensures the generalization ability and robustness of the model through a multi-stage feature selection strategy. It develops multiple prediction models based on different feature combinations, including clinical feature models, manually produced imaging omics models, deep learning imaging omics models, clinical-deep learning imaging omics models, and clinical-manually produced imaging omics models. Through cross-validation and model optimization, the efficiency of each model in clinical prediction is ensured, the performance of the models is comprehensively compared, and the prediction model most suitable for clinical application is selected to further improve the accuracy and reliability of the prediction. The present invention improves the accuracy of esophageal fistula risk prediction in esophageal cancer patients after radiotherapy or chemoradiotherapy, provides a strong basis for treatment decision-making for esophageal cancer patients, and has broad clinical application prospects.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence technology, and in particular to a method for predicting the risk of esophageal fistula based on clinical data and imaging genomics features. Background Art

[0002] Esophageal cancer is one of the most common and lethal malignancies worldwide, posing a significant challenge to the field of oncology. For patients with unresectable advanced esophageal cancer, aggressive radiotherapy or chemoradiotherapy can significantly improve prognosis. However, esophageal fistulas, pathological connections between the esophagus and adjacent anatomical structures, are common after cancer treatment. They result from an imbalance between tumor cell death and normal tissue repair, leading to rupture or perforation of the esophageal wall and are one of the most serious complications. The development of esophageal fistulas significantly increases the risk of other complications, such as infection and malnutrition. The incidence of fistulas in patients with esophageal cancer undergoing chemoradiotherapy ranges from 5.3% to 24.1%. For patients with advanced esophageal cancer who do not develop fistulas, the average survival is 11 months; however, once fistulas develop, the average survival decreases to 3.63 months, indicating a grim prognosis and low survival rate. Early prediction and proactive intervention are urgently needed. However, based on the clinical symptoms and imaging diagnosis of patients with esophageal fistula, the presence of esophageal fistula can only be confirmed after the symptoms appear, which highlights the urgent need to develop prediction models to identify high-risk patients early.

[0003] Previous studies have shown that multiple clinical factors are associated with esophageal fistula, including demographic characteristics, physiological parameters, medical history, pathological markers, and previous treatments. In recent years, radiomics has been applied as a valuable tool for predicting esophageal fistula in various medical conditions. Deep learning technology has further enriched the prospects of radiomics, providing powerful image analysis capabilities and novel feature extraction methods.

[0004] Although in the prior art,

[0005] Zhu et al. proposed Development and validation of a prognostic nomogram formalignant esophageal fistula based on radiomics and clinical factors. Thoraciccancer 2021, 12(23): 3110-3120. They developed a radiomics map combining radiomics and clinical factors to evaluate survival after esophageal cancer fistula, but it could not predict the occurrence of esophageal cancer fistula before treatment and could not achieve early intervention.

[0006] Shi et al. proposed Quantitative CT analysis to predict esophageal fistula in patients with advanced esophageal cancer treated by chemotherapy or chemoradiotherapy. Cancer imaging: the official publication of the International Cancer Imaging Society 2022, 22(1): 62, which uses clinical factors and manually extracted imaging quantitative indicators to predict the occurrence of esophageal fistula, but requires a lot of manual work.

[0007] Li et al. proposed the Clinical-Radiomics Nomogram for Risk Prediction of Esophageal Fistula in Patients with Esophageal Squamous Cell Carcinoma Treatedby IMRT or VMAT. International journal of radiation oncology, biology, physics 2023, 1172S:e315, which studied the predictive factors of esophageal fistula in patients receiving intensity-modulated radiotherapy (IMRT) or volumetric modulated arc therapy (VMAT) and established a clinical radiomics nomogram, but it was not aimed at the prediction of esophageal fistula based on deep learning-derived features.

[0008] Recently, Zhu et al. proposed Pre-Treatment CT Radiomics and Clinical FactorsPredict Malignant Esophageal Fistula in Patients with Esophageal Cancer.In.:Research Square;2023, which developed a prediction model for esophageal fistula based on logistic regression and pyradiomics radiomics features, but the method was limited to a test cohort and did not consider deep learning aspects. Summary of the Invention

[0009] In order to overcome the above-mentioned defects in the prior art, the present invention provides an esophageal fistula risk prediction method based on clinical data and imaging genomics features to improve the accuracy of esophageal fistula risk prediction in esophageal cancer patients after radiotherapy or chemoradiotherapy.

[0010] To achieve the above object, the present invention adopts the following technical solutions, including:

[0011] The esophageal fistula risk prediction method based on clinical data and radiomics features includes the following steps:

[0012] S1, data collection, including the collection of clinical data and imaging data;

[0013] S2, performing image segmentation and preprocessing on the imaging data, i.e., the CT image, segmenting the esophageal cancer tumor area, and standardizing the voxel size of the CT image;

[0014] S3, extracting features from the preprocessed imaging data and clinical data; wherein the feature extraction of the imaging data includes deep learning radiomics features extracted by a deep learning model and manually produced radiomics features extracted by a radiomics feature extractor;

[0015] S4, performing feature selection on clinical features, deep learning radiomics features, and manually crafted radiomics features, and combining the selected clinical features, deep learning radiomics features, and manually crafted radiomics features to obtain a first sample set consisting of clinical features, a third sample set consisting of deep learning radiomics features, a second sample set consisting of manually crafted radiomics features, a fourth sample set consisting of clinical features and deep learning radiomics features, and a fifth sample set consisting of clinical features and manually crafted radiomics features;

[0016] S5, constructing five prediction models respectively, and training the five prediction models using five sample sets respectively for esophageal fistula risk prediction, obtaining a clinical model generated by training with the first sample set, a deep learning radiomics model generated by training with the second sample set, a manually crafted radiomics model generated by training with the third sample set, a clinical-deep learning radiomics model generated by training with the fourth sample set, and a clinical-manually crafted radiomics model generated by training with the fifth sample set;

[0017] S6. Evaluate the performance of the five prediction models, select the prediction model with the best performance as the final esophageal fistula risk prediction model, and use the final esophageal fistula risk prediction model to predict esophageal fistula risk.

[0018] Preferably, in step S2, the CT image is preprocessed based on a resampling technique using trilinear interpolation to normalize the voxel size to a uniform value, as shown below:

[0019] S21, based on the coordinates (x, y, z) of the original voxel center, calculate the coordinates (x', y', z') of the new voxel center using the formula Among them, s x , sy , s z is the original voxel size, s' x , s' y , s' z is the new voxel size;

[0020] S22, calculate the integer coordinates (i, j, k) of the original voxel center by rounding, and calculate the relative distance d x =x'-i,d y =y'-j,d z =z'-k;

[0021] S23, based on the relative distance, the weights of the new voxel center relative to the eight original voxels around the new voxel center are calculated as follows: mnp =(1-m+(-1) m ×d x )×(1-n+(-1) n ×d y )×(1-p+(-1) p ×d z ), where m, n, and p are all 0 or 1;

[0022] S24, passed The weighted sum is used to obtain the CT value of the new voxel center, where V(i+m,j+n,k+p) is the CT value of the original voxel;

[0023] S25, calculating the CT value of each new voxel center according to the above steps S21-S25, and completing the resampling of the CT image;

[0024] S26, after resampling, normalize the CT value of the region of interest of the CT image, that is, the esophageal cancer tumor region, to a range of 0 to 1. The normalization formula is: P = Where H is the original CT value, P is the normalized value, W is the size of the grayscale range in the region of interest, and C is the center value of the grayscale range in the region of interest.

[0025] Preferably, in step S3, a deep learning model based on a three-dimensional convolutional autoencoder is constructed to extract deep learning radiomics features from CT images; the deep learning model includes an encoder and a decoder, is optimized by an Adam optimizer, uses the mean square error between the input image and the reconstructed image as the loss function, and adopts a structural similarity index to evaluate the perceptual quality of the reconstructed image. During the training process, the ModelCheckpoint and EarlyStopping mechanisms are combined, and a four-fold cross-validation strategy is adopted;

[0026] Manually crafted radiomics features were extracted from the esophageal cancer tumor area of ​​the CT image using the radiomics feature extractor pyradiomics.

[0027] Preferably, in step S3, the preprocessed CT image is used as the original image, and the original image is subjected to wavelet transform, LoG transform, square transform, square root transform, logarithmic transform, exponential transform and gradient transform respectively, and seven transformed images after seven kinds of transform processing are obtained; the deep learning model is used to extract the deep learning imaging omics features of the original image and the seven transformed images respectively; the imaging omics feature extractor is used to extract the deep learning imaging omics features of the original image and the seven transformed images respectively.

[0028] Preferably, in step S4, a three-stage feature selection method is used for feature selection, as shown below:

[0029] S41, statistical analysis methods were used to screen out significant features, including Shapiro-Wilk test, Student's t test, Mann-Whitney U test, chi-square test, and Fisher's exact test;

[0030] S42, LASSO method was used to further screen out features that met the standardization requirements;

[0031] S43, gradually remove the features with the least influence through recursive feature elimination method.

[0032] Preferably, in step S5, the performance evaluation indicators include the area under the ROC curve (AUC), Brier score, accuracy (ACC), F1 score, sensitivity, specificity, positive predictive value PPV and negative predictive value NPV.

[0033] Preferably, the clinical-deep learning imaging omics model is selected as the final esophageal fistula risk prediction model. The input features selected by the clinical-deep learning imaging omics model include 6, namely: 1 clinical feature, namely re-radiotherapy, 3 deep learning imaging omics features obtained by feature extraction of wavelet transform images, 1 deep learning imaging omics feature obtained by feature extraction of exponential transform images, and 1 deep learning imaging omics feature obtained by feature extraction of gradient transform images.

[0034] A readable storage medium stores a computer program, which, when executed, implements the above-mentioned method for predicting the risk of esophageal fistula based on clinical data and imaging omics features.

[0035] An electronic device comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for predicting the risk of esophageal fistula based on clinical data and imaging genomics features is implemented.

[0036] A computer program product comprises a computer program / instruction, which, when executed by a processor, implements the above-mentioned method for predicting the risk of esophageal fistula based on clinical data and imaging omics features.

[0037] The advantages of the present invention are:

[0038] (1) The technical problem that the present invention aims to solve is to improve the accuracy of risk prediction for esophageal fistula in patients with esophageal cancer after radiotherapy or chemoradiotherapy. By integrating manually crafted radiomics features, deep learning radiomics features, and clinical features, a more powerful and comprehensive tool is developed to effectively identify high-risk patients, promote timely early intervention, and thus improve patient prognosis.

[0039] (2) This invention, based on actual clinical needs, combines a patient's clinical data with imaging data and inputs them into a machine learning model to effectively predict whether patients undergoing radiotherapy or chemoradiotherapy for esophageal cancer will develop esophageal fistula. This model provides strong support for the precise treatment of cancer patients, helps develop personalized treatment plans, and reduces treatment risks.

[0040] (3) The present invention ensures the consistency of data processing and analysis at different stages through a series of image data processing, including acquisition, volume segmentation and image preprocessing technologies, laying a solid foundation for subsequent feature extraction and model construction.

[0041] (4) This paper combines deep learning with traditional radiomics methods, extracting rich radiomics features through a variety of image processing techniques (such as wavelet transform and LoG transform), and ensuring the generalization and robustness of the model through a multi-stage feature selection strategy (including statistical analysis, LASSO, and recursive feature elimination). This enables the model to maintain high accuracy and stability in clinical applications.

[0042] (5) The present invention develops multiple prediction models based on different feature combinations, including clinical feature models, manually generated radiomics models, deep learning radiomics models, and hybrid models. Through cross-validation and model optimization, the efficiency of each model in clinical prediction is ensured. A comprehensive comparison of the model performance helps to select the most suitable prediction model for clinical application and further improve the accuracy and reliability of prediction.

[0043] (6) The method of the present invention is not only applicable to the prediction of treatment effects for esophageal cancer, but can also be extended to other types of tumor radiotherapy. This versatility makes the technology highly valuable in actual clinical applications and can provide support for the precise treatment of more patients.

[0044] (7) Esophageal fistula is a devastating and even life-threatening complication of radiotherapy or chemoradiotherapy for locally advanced esophageal cancer, and remains difficult to predict in clinical practice. This invention establishes a prediction model based on clinical and radiomic features, providing an efficient and accurate prediction method that provides a strong basis for treatment decision-making in esophageal cancer patients and has broad clinical application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The flowchart of the esophageal fistula risk prediction method based on clinical data and imaging genomics features of the present invention.

[0046] Figure 2 Schematic diagram of the esophageal fistula risk prediction method based on clinical data and imaging genomics features of the present invention.

[0047] Figure 3 Flow chart for subject enrollment.

[0048] Figure 4 Schematic diagram of CT image segmentation. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] like Figure 1 and Figure 2 As shown, the method for predicting esophageal fistula risk based on clinical data and radiomics features of the present invention comprises the following steps:

[0051] S1, data collection, including the collection of clinical data and imaging data.

[0052] To ensure objectivity, the review process was blinded to the patients' clinical data.

[0053] In order to ensure the acquisition of high-quality image data, the present invention adopts the technical means of performing enhanced CT scanning in the patient's free breathing state, and the image data collected is the CT image (three-dimensional image) obtained by the enhanced CT scan. A specific model of CT equipment (such as Brilliance CT BigBore) is used for scanning, and the parameters include voltage, effective milliampere-seconds, beam collimation, matrix size, pitch and gantry rotation time. The scanning process includes two steps: non-enhanced CT and dynamic enhanced CT. The latter is achieved by intravenous injection of non-ionic contrast agent and capturing images in the arterial phase to enhance the contrast and details of the image. The thickness of each layer scan is fixed to ensure the stability of the image resolution, and the data is imported into the CT simulation workstation for subsequent processing.

[0054] S2, performing image segmentation and preprocessing on the image data, namely the CT image, segmenting the esophageal cancer tumor area, and standardizing the voxel size of the CT image.

[0055] Image segmentation: This method utilizes a manual 3D delineation technique to precisely locate and segment esophageal cancer tumor regions. This layer-by-layer delineation ensures accurate marking of tumor volumes. This process is performed by experienced radiologists, who then review the delineation with radiation oncologists specializing in upper gastrointestinal malignancies to ensure accuracy and consistency.

[0056] Preprocessing: To ensure the spatial resolution and consistency of CT images, this paper uses a resampling technique based on trilinear interpolation to normalize the voxel size to a uniform value to ensure the spatial resolution and consistency of CT images. The specific process is as follows:

[0057] S21, based on the coordinates (x, y, z) of the original voxel center, calculate the coordinates (x', y', z') of the new voxel center using the formula Among them, s x , s y , s z is the original voxel size, s' x , s' y , s' z is the new voxel size;

[0058] S22, calculate the integer coordinates (i, j, k) of the original voxel center by rounding, and calculate the relative distance d x =x'-i,d y =y'-j,d z =z'-k;

[0059] The relative distance refers to the relative distance between the coordinates of the new voxel center and the integer coordinates of the original voxel center;

[0060] S23, based on the relative distance, the weights of the new voxel center relative to the eight original voxels around the new voxel center are calculated as follows: mnp =(1-m+(-1) m ×d x )×(1-n+(-1) n ×d y )×(1-p+(-1) p ×d z ), where m, n, and p are all 0 or 1;

[0061] S24, passed The weighted sum is used to obtain the CT value of the new voxel center, where V(i+m,j+n,k+p) is the CT value of the original voxel;

[0062] S25, calculating the CT value of each new voxel center according to the above steps S21-S25, and completing the resampling of the CT image; the unit of the CT value is HU;

[0063] S26, after resampling, normalize the CT value in the region of interest of the CT image, i.e., the esophageal cancer tumor region, to a range of 0 to 1. The normalization formula is: Where H is the original CT value, P is the normalized value, W is the size of the grayscale range in the region of interest, and C is the center value of the grayscale range in the region of interest.

[0064] Among them, voxel is the basic unit in 3D image; voxel center is the spatial position at the geometric midpoint of voxel. The coordinates (x, y, z) of the original voxel center are continuous coordinates, which may be decimal values. The coordinates of discrete grid points, i.e. integer coordinates, are obtained by rounding off the continuous coordinates.

[0065] S3, extracting features from the pre-processed imaging data and clinical data. The feature extraction from the imaging data includes deep learning radiomics features extracted using a deep learning model and manually created radiomics features extracted using a radiomics feature extractor.

[0066] In the present invention, a preprocessed CT image is used as the original image and subjected to wavelet transform, Logarithmic transform, square transform, square root transform, logarithmic transform, exponential transform, and gradient transform, resulting in seven transformed images. A deep learning model is used to extract deep learning radiomic features from the original image and the seven transformed images. A radiomic feature extractor is used to extract deep learning radiomic features from the original image and the seven transformed images.

[0067] In this paper, a deep learning model based on a 3D convolutional autoencoder was constructed. The model consists of an encoder and a decoder, optimized using the Adam optimizer. The mean squared error (MSE) between the input image and the reconstructed image was used as the loss function, and the structural similarity index (SSIM) was used to evaluate the perceptual quality of the reconstructed image. To improve training effectiveness and reduce the risk of overfitting, the training process incorporated the ModelCheckpoint and EarlyStopping mechanisms, and adopted a four-fold cross-validation strategy.

[0068] In the present invention, manually created radiomics features (manually defined) are extracted using the radiomics feature extractor pyradiomics for the esophageal cancer tumor region in the CT image.

[0069] S4, feature selection is performed on clinical features, deep learning radiomics features, and manually crafted radiomics features, and the selected clinical features, deep learning radiomics features, and manually crafted radiomics features are combined to obtain a first sample set composed of clinical features, a third sample set composed of deep learning radiomics features, a second sample set composed of manually crafted radiomics features, a fourth sample set composed of clinical features and deep learning radiomics features, and a fifth sample set composed of clinical features and manually crafted radiomics features.

[0070] The present invention adopts a three-stage feature selection method during the training process to ensure the statistical significance of the features and the generalization ability of the prediction model. The specific process is as follows:

[0071] S41, screening for significant features through statistical analysis methods, including Shapiro-Wilk test, Student's t test, Mann-Whitney U test, chi-square test, and Fisher's exact test;

[0072] S42, LASSO method was used to further screen out features that met the standardization requirements;

[0073] S43, the features with the least influence are gradually removed through the recursive feature elimination (RFE) method to ensure that the number of features retained by the prediction model is within a reasonable range and avoid overfitting.

[0074] S5, five prediction models were constructed respectively, and the five prediction models were trained using five sample sets respectively for the risk prediction of esophageal fistula, and the clinical model (Clin-model) generated by training the first sample set, the deep learning imaging omics model (DL_ra-model) generated by training the second sample set, the manually made imaging omics model (Hand_ra-model) generated by training the third sample set, the clinical-deep learning imaging omics model (Clin-DL_ra-model) generated by training the fourth sample set, and the clinical-manually made imaging omics model (Clin-Hand_ra-model) generated by training the fifth sample set were obtained.

[0075] RandomForestClassifier was used in the model development process, and parameters were optimized through GridSearchCV to ensure the accuracy and stability of the model.

[0076] S6. Evaluate the performance of the five prediction models, select the prediction model with the best performance as the final esophageal fistula risk prediction model, and use the final esophageal fistula risk prediction model to predict esophageal fistula risk.

[0077] In order to evaluate the performance of the model in predicting the treatment effect of esophageal cancer patients, the present invention adopts a variety of performance evaluation indicators, including the area under the ROC curve (AUC), Brier score, accuracy (ACC), F1 score, sensitivity, specificity, positive predictive value PPV and negative predictive value NPV. The 95% confidence intervals of the above indicators were calculated by 1000 bootstrap resamplings. The Friedman rank test and Nemenyi test were used to perform statistical significance tests on the performance differences between the five prediction models, or to compare the differences between the model of interest and other models. The ROC curve and calibration curve are used to intuitively display the discriminatory ability of the model and the accuracy of the predicted probability. Among them, ACC and ROC curve are used to evaluate the discriminatory ability of the model, while Brier score and calibration curve are used to measure the consistency between the actual and predicted probabilities.

[0078] Accuracy (ACC) = (TP + TN) / (TP + FN + FP + TN);

[0079] Sensitivity = TP / (TP+FN); specificity = TN / (TN+FP);

[0080] Positive predictive value (PPV) = TP / (TP+FP); negative predictive value (NPV) = TN / (TN+FN);

[0081] Among them, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives.

[0082] The present invention ultimately obtained a clinical-deep learning radiomics model (Clin-DL_ra-model) that showed excellent performance in predicting the occurrence of esophageal fistula in patients with advanced esophageal cancer receiving radiotherapy or chemoradiotherapy. The model combines clinical features and deep learning radiomics features, which not only improves the accuracy of prediction but also enhances the interpretability of the model. In addition, by integrating deep learning radiomics features, the model is able to capture more subtle and complex patterns in medical images, which may be difficult to capture with traditional methods. This provides clinicians with a powerful tool to help identify high-risk patients in a timely manner, promote early intervention, and thus improve patient prognosis.

[0083] Example 1

[0084] In this embodiment, the patient data collection is specifically as follows:

[0085] Data from 271 patients with esophageal cancer who underwent radiotherapy or chemoradiotherapy between April 2018 and June 2022 were retrospectively collected, and data from 58 patients were prospectively collected between June 2022 and December 2023. Follow-up after radiotherapy to confirm the development of esophageal fistula was at least 1 year. Based on inclusion and exclusion criteria, data from 175 patients were included from the retrospectively collected data set. Data from 122 patients (70%) were used as training data (22 patients with esophageal fistula and 100 patients without esophageal fistula), and data from 55 patients (30%) were used as testing data (10 patients with esophageal fistula and 43 patients without esophageal fistula). Data from 27 patients (5 patients with esophageal fistula and 22 patients without esophageal fistula) were included from the prospectively collected data set as prospective testing data or validation data. Inclusion criteria included the following patients: (1) patients diagnosed with esophageal cancer stage II to IVA based on pathological biopsy; (2) patients with esophageal cancer receiving radical radiotherapy / chemoradiotherapy, palliative radiotherapy / chemoradiotherapy, or secondary radiotherapy; (3) patients diagnosed with esophageal fistula accompanied by atopic acetaminophen by endoscopy, CT, or X-ray examination. Exclusion criteria included: (1) individuals undergoing esophageal surgery; (2) patients with esophageal fistula due to reasons other than medical injury or trauma; (3) patients with fistula before treatment or during disease progression, or patients with malignant tumors; (4) patients older than 80 years old. The subject enrollment process is as follows. Figure 3 As shown, Figure 3 a represents a retrospective cohort, and b represents a prospective testing cohort.

[0086] Baseline clinical data included age, gender, hemoglobin level, total cholesterol, triglyceride concentration, body mass index (BMI), body surface area (BSA), hypertension, diabetes, smoking history, alcohol history, tumor length, maximum lesion diameter, and re-irradiation status. In this example, the clinical characteristics of each group of patients are shown in Table 1 below:

[0087] Table 1

[0088]

[0089]

[0090]

[0091] Figure 3 and Table 1 , n represents the number of patients, i.e., the sample size.

[0092] All patients had received radiotherapy, and patients who underwent re-radiation had received their first radiotherapy within the past five years. All treatments followed the NCCN (National Comprehensive Cancer Network) guidelines. The radiotherapy dose ranged from 50.4 to 60 Gy, administered in 28 to 30 fractions, and chemotherapy included paclitaxel, platinum, and 5-FU. Before radiotherapy or chemoradiotherapy, patients with esophageal cancer underwent enhanced CT scanning using two scanners of the same model and parameters in the free-breathing state. Scanning was performed using a Brilliance CT BigBore (Philips Medical Systems, Cleveland, USA) in the CT simulation position. The scanning parameters were 120 kV, 325 effective mA, beam collimation 161.5 mm, matrix 512 × 512, pitch 0.938, and gantry rotation time 0.75 s. A non-enhanced CT scan was performed first, followed by a dynamic contrast-enhanced CT scan after an intravenous injection of 1.5–2 ml / kg of nonionic contrast agent (iodine hexol injection, 350 mg I / ml, Beijing Beilu Pharmaceutical Co., Ltd., China) at 3 ml / s. The patient was then flushed with 20 ml of normal saline. Arterial phase images were acquired 26 seconds after injection for image feature extraction. Each slice was scanned at a 5 mm thickness, combining chest contrast-enhanced CT and arterial phase CT. Subsequently, the CT images were imported into a CT simulation workstation for precise delineation of regions of interest (ROIs).

[0093] In this example, the Elekta Monaco planning system V3.8.0 was used to manually delineate the 3D-labeled ROI (i.e., the esophageal cancer tumor area) on each CT scan, covering the entire tumor area. This delineation was performed by a qualified radiologist (with five years of experience in esophageal imaging) and subsequently reviewed by a radiation oncologist specializing in upper gastrointestinal malignancies. Figure 4 An example of image segmentation is given, Figure 4Diagram showing the image ROI (ROI). (a) A transverse section, (b) 3D lesion image, (c) coronal section, and (d) sagittal section. The delineation process was reviewed by two radiation oncologists and defined using a window width of 500 and a window level of 40. The imaging area was defined as 5 mm of esophageal wall thickening, excluding intraluminal gas, oral contrast agents, and adjacent organs.

[0094] In this embodiment, before feature extraction, the resize function data_resampled = skimage.transform.resize(data, new_space, order = 1) of the skimage package is used to perform trilinear interpolation on the original Hounsfield Units (HU) values ​​and resample the image to a 3×3×3 voxel size to ensure that the spatial resolution and consistency of the CT image data are maintained.

[0095] In this embodiment, deep learning imaging genomics features are extracted using a three-dimensional (3D) convolutional autoencoder developed using the Keras API in TensorFlow2 (V2.7.0). The deep learning model architecture consists of an encoder and decoder structure of a deep artificial neural network and is optimized using the Adam optimizer. The mean squared error (MSE) between the image input and the reconstructed pixel values ​​is used as a loss function, while the structural similarity index metric (SSIM) evaluates the perceptual quality of the reconstructed image. To ensure effective training and mitigate the risk of overfitting, the process integrates Keras's ModelCheckpoint and earlystop mechanisms in a four-fold cross-validation strategy. During the cross-validation process within the training cohort, the MSE and SSIM metrics are calculated for the training and validation sets of each fold.

[0096] In this example, Pyradiomics (version 3.0.1) was used to extract handcrafted radiomics features from the esophageal cancer tumor region in the above eight images.

[0097] In this example, feature selection is rigorously implemented during the training process using a three-stage approach. First, statistical analysis methods are used to identify statistically significant features. The choice of statistical test method depends on the nature and distribution of the data and includes the Shapiro-Wilk test, Student's t-test, Mann-Whitney U test, chi-square test, and Fisher's exact test. Next, normalized features that pass the statistical test are further selected using the least absolute shrinkage and selection operator (LASSO) with cv=10, a technique that gradually decays feature coefficients to zero, thereby facilitating dimensionality reduction. Then, recursive feature elimination (RFE) with step=1 is used to strategically remove features with minimal impact until the desired number of features is achieved. The basic RFE algorithm involves training a model with all features, ranking their importance, and iteratively removing the least important features until the desired number is reached. To avoid overfitting, each prediction model is capped at 6 features, with an empirical feature number threshold equal to the sample size of the training cohort divided by 20. This constraint ensures the preservation of essential information while enhancing the model's generalization ability. In the present invention, unless otherwise stated, features with a p-value less than 0.05 were considered statistically significant.

[0098] Specifically, deep-learning radiomics features were extracted using an autoencoder trained on the training cohort for each image, with a structural similarity index (SSIM) consistently exceeding 0.8 for both the training and validation sets. The encoder ended with an output shape of (16, 16, 8, 8), from which 16,384 features were extracted. Subsequently, 15 clinical, 1,762 manually crafted radiomics features, and 57,688 deep-learning radiomics features were obtained. Of these, 4 clinical features, 520 manually crafted radiomics features, and 15,136 deep-learning radiomics features were considered statistically significant. Statistically significant features were organized into five distinct sets: a clinical feature set, a manually crafted radiomics feature set, a combined feature set of clinical and deep-learning radiomics features (1), and a combined feature set of clinical and manually crafted radiomics features (2). Each feature set with more than six elements underwent further feature selection using LASSO and 10-fold cross-validation. Then, based on LASSO, the features of each corresponding feature set were reduced to 4 clinical features, 5 handcrafted radiomics features, 28 deep learning radiomics features, 15 combination feature sets 2, and 25 combination feature sets 1. Finally, feature selection was performed using the RFE method, retaining only the 6 most critical features for each set. Reirradiation status (i.e., re-RT) was the only shared clinical feature among the clinical feature set, combination feature set 1, and combination feature set 2.

[0099] In this embodiment, the clinical features, deep learning radiomics features, and manually generated radiomics features after feature selection are combined to obtain a clinical feature set, a deep learning radiomics feature set, a manually generated radiomics feature set, a clinical-deep learning radiomics feature set, and a clinical-manually generated radiomics feature set. The details are shown in Table 2 below:

[0100]

[0101]

[0102] This example retrospectively developed and tested the model, as well as provided a prospective test cohort to evaluate the model's effectiveness and accuracy in a real-world clinical setting. Patients enrolled were randomized into a training cohort and a test cohort at a 1:4 ratio between those with and without esophageal fistulas. Patients treated from June 26, 2023, onward were included in the prospective validation cohort. The training cohort was used for feature selection and model development, while the test and prospective test cohorts were used to evaluate and validate the model's performance.

[0103] In this embodiment, based on the above five feature sets, five prediction models were constructed using RandomForestClassifier, and the imbalance problem of positive and negative class samples was solved by setting its parameter class_weight = 'balanced'. The five prediction models are clinical model (clin-Model), deep learning radiomics model (DL_ra-model), manually made radiomics model (Hand_ra-model), clinical-deep learning radiomics model (Clin-DL_ra-model) and clinical-manual radiomics model (Clin-Hand_ra-model). The model parameters were fine-tuned based on 10-fold cross-validation, with accuracy as the key indicator.

[0104] In this embodiment, in order to evaluate the predictive performance of the model in predicting esophageal fistula, we used performance evaluation indicators, including the area under the receiver operating characteristic curve (AUC), brier score, accuracy (ACC), f1 score, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). The 95% confidence intervals of the above indicators were estimated using 1000 bootstrap resamplings. The Friedman rank test and Nemenyi test were used to test the statistical significance of the metric differences between the above five prediction models or to compare the model of interest with other models. The receiver operating characteristic (ROC) curve and calibration curve were used to show the performance. ACC and ROC curves are tools to measure the discriminative ability of the model, while brier score and calibration curve analysis are tools to measure the consistency between the true probability and the predicted probability. The performance comparison of the five prediction models is shown in Table 3 below:

[0105]

[0106]

[0107]

[0108] The resulting clinical-deep learning radiomics model (Clin-DL_ra-model) performed well in distinguishing esophageal from non-esophageal fistulas (as shown in Table 3), with AUC scores of 0.89 (95% confidence interval, CI, 0.83-0.95), 0.81 (95% confidence interval, CI, 0.65-0.94), and 0.85 (95% confidence interval, CI, 0.71-0.97) in the training, test, and prospective testing cohorts, respectively. The ACCs were 0.92 and 0.85 (95% confidence interval, CI, 0.87-0.96 and 0.74-0.94, respectively) in the training and test cohorts, respectively, and further confirmed to be 0.85 (95% confidence interval, CI, 0.71-0.97) in the prospective testing cohort. The Clin-DL_ra-model model had higher specificity (0.93-0.99), high NPV (0.87-0.92) and moderate PPV (0.63-0.93) compared with other models, with an F1 score and sensitivity value of approximately 0.6. Figure 4 (as shown), the Clin-DL_ra-model model was closer to the 45-degree line than the other models, indicating better, although not perfect, alignment between its predicted probabilities and actual values. This relative closeness to the ideal calibration angle was further supported by the model's lowest brier score (0.1-0.13), which was the best calibration of the five models. Friedman rank tests and Nemenyi paired tests were performed on the AUCs, ACCs, and brier scores from 1000 bootstrappings across the three cohorts, and the results showed significant differences among the five models in each cohort (p value 0.05, see Table S5-5). This indicates that the Clin-DL_ra-model model significantly outperformed the other models. The Clin-DL_ra-model model contains six features, one of which is a clinical feature, namely re-irradiation (re-radiotherapy), three are deep learning radiomics features based on wavelet transform of CT images, and the remaining two are deep learning radiomics features based on exponential transform and gradient transform of CT images.

[0109] Esophageal fistula, a devastating and even life-threatening complication of radiotherapy or chemoradiotherapy for locally advanced esophageal cancer, remains difficult to predict in clinical practice. This study establishes a predictive model based on clinical and radiomics features to predict the occurrence of esophageal fistula during esophageal cancer treatment, providing support for personalized treatment plans for patients.

[0110] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the risk of esophageal fistula based on clinical data and radiomics features, characterized by: The following steps are involved: S1, data collection, including the collection of clinical data and imaging data; S2, performing image segmentation and preprocessing on the imaging data, i.e., the CT image, segmenting the esophageal cancer tumor area, and standardizing the voxel size of the CT image; S3, extracting features from the preprocessed imaging data and clinical data; wherein the feature extraction of the imaging data includes deep learning radiomics features extracted by a deep learning model and manually produced radiomics features extracted by a radiomics feature extractor; S4, performing feature selection on clinical features, deep learning radiomics features, and manually crafted radiomics features, and combining the selected clinical features, deep learning radiomics features, and manually crafted radiomics features to obtain a first sample set consisting of clinical features, a third sample set consisting of deep learning radiomics features, a second sample set consisting of manually crafted radiomics features, a fourth sample set consisting of clinical features and deep learning radiomics features, and a fifth sample set consisting of clinical features and manually crafted radiomics features; S5, constructing five prediction models respectively, and training the five prediction models using five sample sets respectively for esophageal fistula risk prediction, obtaining a clinical model generated by training with the first sample set, a deep learning radiomics model generated by training with the second sample set, a manually crafted radiomics model generated by training with the third sample set, a clinical-deep learning radiomics model generated by training with the fourth sample set, and a clinical-manually crafted radiomics model generated by training with the fifth sample set; S6, evaluating the performance of the five prediction models, selecting the prediction model with the best performance as the final esophageal fistula risk prediction model, and using the final esophageal fistula risk prediction model to predict esophageal fistula risk; In step S4, a three-stage feature selection method is used for feature selection, as shown below: S41, statistical analysis methods were used to screen out significant features, including Shapiro-Wilk test, Student's t test, Mann-Whitney U test, chi-square test, and Fisher's exact test; S42, LASSO method was used to further screen out features that met the standardization requirements; S43, gradually remove the features with the least influence through recursive feature elimination method.

2. The method for predicting esophageal fistula risk based on clinical data and radiomics features according to claim 1, characterized in that: In step S2, the CT image is preprocessed based on the resampling technique of trilinear interpolation to normalize the voxel size to a uniform value, as shown below: S21, based on the coordinates (x, y, z) of the original voxel center, calculate the coordinates (x′, y′, z′) of the new voxel center using the formula Among them, s x , s y , s z is the original voxel size, s′ x , s′ y , s′ z is the new voxel size; S22, calculate the integer coordinates (i, j, k) of the original voxel center by rounding, and calculate the relative distance d x =x′-i,d y =y′-j,d z =z'-k; S23, based on the relative distance, the weights of the new voxel center relative to the eight original voxels around the new voxel center are calculated as follows: mnp =(1-m+(-1) m ×d x )×(1-n+(-1) n ×d y )×(1-p+(-1) p ×d z ), where m, n, and p are all 0 or 1; S24, passed The weighted sum is used to obtain the CT value of the new voxel center, where V(i+m,j+n,k+p) is the CT value of the original voxel; S25, calculating the CT value of each new voxel center according to the above steps S21-S24, and completing the resampling of the CT image; S26, after resampling, normalize the CT value in the region of interest of the CT image, i.e., the esophageal cancer tumor region, to a range of 0 to 1. The normalization formula is: Where H is the original CT value, P is the normalized value, W is the size of the grayscale range in the region of interest, and C is the center value of the grayscale range in the region of interest.

3. The method for predicting esophageal fistula risk based on clinical data and radiomics features according to claim 1, characterized in that: In step S3, a deep learning model based on a three-dimensional convolutional autoencoder is constructed to extract deep learning radiomics features from CT images. The deep learning model includes an encoder and a decoder, is optimized using an Adam optimizer, uses the mean squared error between the input image and the reconstructed image as a loss function, and uses a structural similarity index to evaluate the perceptual quality of the reconstructed image. The training process combines the ModelCheckpoint and EarlyStopping mechanisms, and adopts a four-fold cross-validation strategy. Manually crafted radiomics features were extracted from the esophageal cancer tumor area of ​​the CT image using the radiomics feature extractor pyradiomics.

4. The method for predicting esophageal fistula risk based on clinical data and radiomics features according to claim 1, characterized in that: In step S3, the preprocessed CT image is used as the original image, and the original image is subjected to wavelet transform, LoG transform, square transform, square root transform, logarithmic transform, exponential transform and gradient transform respectively, and seven transformed images after seven kinds of transformation are obtained; the deep learning model is used to extract the deep learning imaging omics features of the original image and the seven transformed images respectively; the imaging omics feature extractor is used to extract the deep learning imaging omics features of the original image and the seven transformed images respectively.

5. The method for predicting esophageal fistula risk based on clinical data and radiomics features according to claim 1, characterized in that: In step S5, the performance evaluation indicators include the area under the ROC curve (AUC), Brier score, accuracy (ACC), F1 score, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).

6. The method for predicting esophageal fistula risk based on clinical data and radiomics features according to claim 4, characterized in that: The clinical-deep learning radiomics model was selected as the final esophageal fistula risk prediction model. The input features selected by the clinical-deep learning radiomics model included six: one clinical feature, namely re-irradiation, three deep learning radiomics features obtained by feature extraction of wavelet transform images, one deep learning radiomics feature obtained by feature extraction of exponential transform images, and one deep learning radiomics feature obtained by feature extraction of gradient transform images.

7. A readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, the method for predicting the risk of esophageal fistula based on clinical data and imaging genomics features according to any one of claims 1 to 6 is implemented.

8. An electronic device, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, it implements the esophageal fistula risk prediction method based on clinical data and imaging genomics features as described in any one of claims 1 to 6.

Images