Acute radiodermatitis risk prediction method and system

Abstract

The invention discloses an acute radiodermatitis prediction method and system, the method comprises an optical imaging system and a machine learning prediction model, the optical system realizes image acquisition of a treatment area of a radiotherapy patient based on a spatial frequency domain imaging (SFDI) technology, dynamically monitors quantitative optical and physiological indexes of the skin of the patient in a radiotherapy treatment course, and determines whether the patient is in the radiotherapy treatment course or not. Comprise hemoglobin concentration, oxygen saturation, melanin concentration, reduced scattering coefficient and the like. The prediction module comprehensively considers three aspects of individual differences (clinical features) of patients, radiotherapy parameters (dosiomics features) and skin optical and physiological indexes obtained by an optical system, and constructs a prediction model by applying a machine learning algorithm to realize accurate prediction of acute radiodermatitis. The method solves the problems that clinical evaluation of acute radiodermatitis mainly depends on experience judgment of doctors at present, and scientific and systematic prediction methods and tools are lacked.

Description

Technical Field

[0001] The invention relates to the field of medical technology, and in particular to a method and system for predicting the risk of acute radiation dermatitis. Background Art

[0002] Radiotherapy is an important tumor treatment method besides surgery and chemotherapy. Its status in tumor treatment is increasing year by year due to its wide indications, good efficacy and low overall cost. About 70% of tumor patients need radiotherapy during treatment. However, while playing a therapeutic role, radiotherapy can also cause damage to normal tissues. Skin damage is the most common adverse reaction to radiotherapy, with an incidence of about 91.4%. It manifests as varying degrees of erythema, edema, ulcers, etc. In severe cases, it may even lead to infection and necrosis, affecting the patient's quality of life and treatment efficacy. The incidence of forced interruption of treatment due to severe damage is about 58.1%, and the local control rate will decrease by 1%-3% for every day the radiotherapy course is extended. Therefore, screening patients at high risk of acute radiation dermatitis, warning of the occurrence of radiation damage, and timely adjusting radiotherapy plans or interventional intervention measures are effective means to avoid or reduce the risk of damage.

[0003] There are some problems and limitations in the existing evaluation methods of radiation dermatitis. First, the current clinical evaluation of acute dermatitis caused by radiotherapy at home and abroad relies on the subjective observation of clinicians for grading and scoring. The grading results largely rely on the experience and insight of doctors. In clinical grading, it is difficult to distinguish the differences between mild, moderate, severe and other skin reactions, so quantitative evaluation cannot be achieved. It also leads to certain deviations in multicenter studies and cannot be standardized. Secondly, some studies have tried to evaluate radiation skin damage through skin measurement instruments, reflectance spectrophotometry, ultrasound, thermal imaging and other methods. However, these methods have problems such as long examination time, high cost and consumption of a large amount of medical resources, and it is difficult to achieve tracking and monitoring throughout the radiotherapy. The results are easily affected by multiple factors, such as ambient temperature and the patient's physical condition. In addition, traditional prediction methods often cannot fully consider the influence of multiple factors such as individual differences in patients, radiotherapy dose parameters and skin physiological indicators, which can easily lead to inaccurate prediction results and are difficult to meet clinical needs. Therefore, there is a lack of a scientific and systematic method and tool for predicting acute radiation dermatitis in clinical practice. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for predicting the risk of acute radiation dermatitis, which has high prediction accuracy, can accurately manage risks, and helps to improve the effect of radiotherapy.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions: a prediction method and system for acute radiation dermatitis, including a hardware module and a software module, the hardware module including an optical image acquisition device based on spatial frequency domain imaging, the optical image acquisition device is used to obtain the skin surface image of the patient's radiotherapy treatment area, the software module includes an image demodulation module, a parameter inversion module and a risk prediction module, the image demodulation module is used to demodulate the acquired skin surface image to obtain the DC and AC modulation transfer functions (MTF) of the tested skin at the modulation frequency, the parameter inversion module is used to invert the DC and AC modulation transfer functions (MTF) to calculate the corresponding physiological parameters converted by the optical imaging model mapped by the tissue layered structure, and the risk prediction module is used to construct an explainable machine learning risk prediction model and give a personalized risk score for the patient.

[0006] Furthermore, the risk prediction module includes the following steps: Step 1: Collect patient clinical characteristics, skin optical and physiological parameters, and dose parameter data, and conduct skin toxicity assessment on patients, and finally perform data preprocessing and data feature screening; Step 2: constructing a risk prediction model, the risk prediction model includes a stacked multimodal machine learning architecture, the machine learning architecture includes basic modules and meta-modules interconnected in a cascade manner, the basic module includes seven tree-based bagging and boosting basic classifiers, namely decision tree (DT), random forest (RF), AdaBoost, XGBoost, Extratrees, Light GBM and CatBoost, the seven basic classifiers run in parallel to independently predict input features and use a five-fold cross-validation method to prepare the input of the meta-module, the seven basic classifiers aggregate the prediction results and transmit the aggregated results to the meta-module, the meta-module is composed of meta-classifiers, the meta-classifiers process the aggregated results aggregated by the basic classifiers, and finally obtain the prediction results of acute radiation skin toxicity; Step 3: Use SHAP and LIME tools to interpret the tree-based model to enhance the global and local interpretability of the model. The model interpretation step is implemented in the Python environment using libraries, where the SHAP value calculation uses the shap package and the LIME interpretation uses the lime package; Step 4: Evaluate the risk value and obtain the cutoff value. The receiver operating characteristic (ROC) curve analysis was used to evaluate the ability of the risk value to identify the occurrence of acute radiation dermatitis in the training set, and the cutoff value was obtained according to the Youden index. The subjects were divided into low-risk and high-risk groups. Finally, the model prediction results were compared with the actual acute radiation dermatitis.

[0007] Furthermore, the clinical characteristics of the patient in step 1 include age, gender, hypertension, diabetes, smoking history, drinking history, body mass index (BMI), chemotherapy, and immunotherapy; the skin optical and physiological parameters include total hemoglobin concentration, oxygen saturation, melanin concentration, oxygen saturation, reduced scattering coefficient, scattering power, epidermal roughness, and epidermal thickness; and the dose parameters of the tested skin are finally obtained by using a radiotherapy planning system at a position 5 mm below the outer surface of the patient's skin and converting the obtained skin surface dose into DVH parameters. The dose parameters include the maximum dose, minimum dose, average dose, median dose, and skin V 10 、V 20 、V 30 、V 40 、V 50 、V 60 The missing values ​​in the data preprocessing were replaced by the feature median and all continuous variables were standardized, the categorical variables were encoded, and the final data were divided into a training set and a test set in a ratio of 8:2 and the training data were balanced using Borderline SMOTE.

[0008] Furthermore, the optical image acquisition device comprises an imaging light source, a projection device and a detector. The projection device uses a digital micromirror (1) to achieve structured illumination, projects mixed white light with wavelengths of R (623 nm), G (539 nm) and B (465 nm), and uses a collimating objective lens (2) to collimate and correct the mixed white light. The frequencies of the projected mixed white light are 0 and 0.2 mm. -1 By adjusting the frequency, different imaging depths can be achieved. The imaging area is 3*3 cm, and tissue wide-field imaging is achieved. The detector uses a CCD camera (3) to irradiate spatial modulation patterns of different frequencies. to the sample area and capture the reflected image with a CCD camera (3) .

[0009] Furthermore, the image demodulation module uses a standard three-phase shift method (α=0, 2π / 3, 4π / 3) to demodulate the measured skin at modulation frequencies 0 and DC and AC modulation transfer functions (MTF) under the condition, and then according to the propagation theory of light in turbid media, the medium including absorption coefficient ( ) and the reduced scattering coefficient ( ), etc., when the standard three-phase shift method is used for image demodulation, the three phase images are , and It can be expressed as: {I}_{AC}=\frac {\sqrt {2}} {3}\left [ {\left ( {{I}_{1}-{I}_{2}} \right )^{2}+\left ( {{I}_{2}-{I}_{3}} \right )^{2}+\left ( {{I}_{3}-{I}_{1}} \right )^{2}} \right ]^{\frac {1} {2}} .

[0010] Furthermore, the image demodulation module uses a single snapshot multi-frequency demodulation method (SSMD) to demodulate the measured skin at modulation frequencies 0 and DC and AC modulation transfer functions (MTF) under the condition, and then according to the propagation theory of light in turbid media, the medium including absorption coefficient ( ) and the reduced scattering coefficient ( ), etc., when single snapshot multi-frequency demodulation (SSMD) is used for image demodulation, the structured light pattern incident on the surface of the tissue sample and the light intensity pattern obtained by tissue backscattering can be expressed as: {I}_{0}\left ( {x,y} \right )={I}^\left ( {0} \right )_{DC}+\sum ^{k}_{i=1} {{I}^\left ( {0} \right )_{AC,i}}\cos {\left [ {2π\left ( {{f}_{x,i}x+{f}_{y,i}y} \right )\left ( {} \right )+{\varnothing}_{i}} \right ]} I\left ( {x,y} \right )={I}_{DC}\left ( {x,y} \right )+\sum ^{k}_{i=1} {{I}_{AC,i}}\left ( {x,y} \right )\cos {\left [ {2π\left ( {{f}_{x,i}x+{f}_{y,i}y} \right )+{\varnothing}_{i}} \right ]} where k ≥ 1 is the number of AC components, and The spatial frequencies are and Phase The next The incident amplitude and backscattered amplitude of each AC component of SSMD and the DC component amplitude The calculations are as follows: .

[0011] Furthermore, the absorption coefficient equation of the tested skin tissue is established as , where λ is the wavelength of light, N is the number of chromophores, is the molar extinction coefficient of the chromophore, is the chromophore concentration, described by Mie scattering , where a represents the scattering amplitude, which is related to the number of scatterers, b Represents the scattering ability related to the size of the scatterer, obtained by low-frequency structured light imaging under multiple wavelengths Can be further based on Obtained, measured total hemoglobin concentration in the layered structure and oxygen saturation , equivalent light absorption and scattering are achieved by mapping typical photons between the homogeneous medium and the original two-layer structure, i.e. {μ}_{a}L={μ}_{a,melanin}\left ( {λ} \right )h+\left [ {{μ}_{a,Hb{O}_{2}}\left ( {λ} \right )+{μ}_{a,Hb}\left ( {λ} \right )} \right ]\left ( {Lh} \right ) ,in h is the thickness of the epidermis of the skin, L For the spatial frequency and the average detection depth of structured light at wavelength λ.

[0012] The beneficial effects of the present invention are as follows: the present invention realizes image acquisition of the treatment area of ​​radiotherapy patients through an optical image acquisition device based on spatial frequency domain imaging (SFDI), and monitors the microcirculation and structural characteristics of the patient's skin during radiotherapy in real time. The risk prediction module comprehensively considers multiple factors such as individual differences of patients (clinical characteristics), radiotherapy parameters (dosimetric characteristics), skin optics and physiological indicators, and uses a multimodal machine learning architecture learning algorithm to construct an interpretable risk prediction model to achieve quantitative evaluation and prediction of acute radiation skin toxicity, avoiding the discomfort and risk of infection at the operation site caused by invasive operations in some existing detection methods, and also improving the accuracy and objectivity of skin toxicity evaluation. Compared with the current clinical evaluation of acute radiation dermatitis, which mainly relies on the subjective experience of doctors, the present invention can provide a reference for the development of radiotherapy for patients with radiotherapy. On the other hand, compared with the problems of long inspection time, difficult tracking and monitoring, and being greatly affected by various objective factors in some traditional prediction methods, the detection method of the present invention can efficiently and effectively predict and monitor dermatitis in radiotherapy patients after constructing an explainable risk prediction model, which is convenient for more accurate risk management of radiotherapy patients in clinical practice, and improves the adaptability of overall monitoring and early warning to individual differences of patients and rapid changes in complex diseases. During the radiotherapy treatment process, doctors can track and evaluate the current condition of radiotherapy patients in a timely manner, improve the effect of radiotherapy, reduce the discomfort of radiotherapy patients after radiotherapy, and improve the quality of radiotherapy. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a structural principle diagram of the optical image acquisition device based on spatial frequency domain imaging in Example 1 of the present invention.

[0014] Figure 2 This is a flow chart of SFDI obtaining optical parameters in Example 1 of the present invention.

[0015] Figure 3 It is a flowchart of the overall construction of the risk prediction model in Example 1 of the present invention.

[0016] Figure 4 It is a flowchart of the overall construction of the stacking model in Example 1 of the present invention.

[0017] Figure 5 It is the personalized risk prediction result of a high-risk patient and a low-risk patient in Example 1 of the present invention. DETAILED DESCRIPTION

[0018] In order to make the technical means, innovative features and functions implemented by the present invention easy to understand, the present invention will be further described below.

[0019] like Figure 1-5 As shown, the technical solution of the method and system for predicting the risk of acute radiation dermatitis of the present invention includes a hardware module and a software module. The hardware module includes an optical image acquisition device based on spatial frequency domain imaging, and the optical image acquisition device is used to obtain the skin surface image of the patient's radiotherapy treatment area. The software module includes an image demodulation module, a parameter inversion module and a risk prediction module. The image demodulation module is used to demodulate the acquired skin surface image to obtain the DC and AC modulation transfer functions (MTF) of the tested skin at the modulation frequency. The parameter inversion module is used to invert the DC and AC modulation transfer functions (MTF) to calculate the corresponding physiological parameters converted by the optical imaging model mapped by the tissue layered structure. The risk prediction module is used to construct an explainable machine learning risk prediction model and give a personalized risk score for the patient.

[0020] In Example 1 of the present invention, the specific implementation steps are as follows.

[0021] The first step is to build a non-invasive quantitative evaluation device based on a spatial frequency domain imaging optical system. The optical image acquisition device based on spatial frequency domain imaging includes an imaging light source, a projection device and a detector, such as Figure 1 As shown, the projection device uses a digital micromirror 1 to achieve structured illumination, projects a mixed white light with wavelengths of R (623 nm), G (539 nm), and B (465 nm), and is collimated and corrected by a collimating objective lens 2. The frequencies of the projected mixed white light are 0 and 0.2 mm -1 By adjusting the frequency, different imaging depths can be achieved. The imaging area is 3*3cm, and tissue wide-field imaging is achieved. The detector uses a CCD camera 3, which is used to inject spatial modulation patterns of different frequencies. to the sample area and capture the reflected image with CCD camera 3 .

[0022] The above device is aimed at the patient's radiotherapy area to obtain the optical properties of the patient's skin surface and deep tissue, such as absorption coefficient and reduced scattering coefficient. The optical imaging model of tissue layer structure mapping is used to convert the obtained optical properties into corresponding physiological parameters, such as hemoglobin concentration, oxygen saturation, etc. The following table shows the data parameters obtained after SFDI optical image acquisition of the neck skin of nasopharyngeal carcinoma patients with different degrees of skin damage (normal, slight peeling, redness and swelling) during radiotherapy. There are certain differences in the obtained physiological parameters, and the parameters are shown in the following table. It can be seen from the data in the table that the prediction method and system of the present invention can quantify the degree of skin damage in a digital way, no longer relying too much on the subjective observation and experience of the attending physician, and is conducive to the timely warning and monitoring of skin inflammation in radiotherapy patients after radiotherapy.

[0023] Total hemoglobin concentration (mM) Blood oxygen saturation Melanin concentration Healthy subjects 0.03070089 0.89284314 0.212206 #1 Normal 0.05925879 0.82850969 0.2738407 #2 Slight peeling 0.0872186 0.84217828 0.2451925 #3 Redness and swelling 0.09889203 0.84515176 0.2696357 .

[0024] The second step is to collect patient clinical characteristics, skin optical and physiological parameters, and dose parameter data, and conduct skin toxicity assessment on patients, and finally perform data preprocessing and data feature screening.

[0025] In the above second step, the clinical characteristics of the patients include age, gender, hypertension, diabetes, smoking history, drinking history, body mass index (BMI), chemotherapy, and immunotherapy. The skin optical and physiological parameters include total hemoglobin concentration, oxygen saturation, melanin concentration, oxygen saturation, reduced scattering coefficient, scattering power, epidermal roughness, and epidermal thickness. The dose parameters of the tested skin are finally obtained by using a radiotherapy planning system at a position 5 mm below the outer surface of the patient's skin and converting the obtained skin surface dose into DVH parameters. The dose parameters include the maximum dose, minimum dose, average dose, median dose of the tested skin area, and skin V 10 、V 20 、V 30 、V 40 、V 50 、V 60 The missing values ​​in the data preprocessing were replaced with the feature median and all continuous variables were standardized. The categorical variables were encoded (coded as 0 and 1), such as gender (male / female). In the process of skin toxicity assessment, the grade of acute radiation dermatitis was assessed according to the Common Terminology Criteria for Adverse Events (CTCAE) scale of the National Cancer Institute of the United States. The grade was obtained by the patient's attending physician during weekly follow-up during radiotherapy and daily photography and recording by the researchers of this project. Patients with CTCAE ≥ 2 were classified as high risk, and patients with CTCAE < 2 were classified as low risk. The final data were divided into training set and test set in a ratio of 8:2, and the training data were balanced using BorderlineSMOTE.

[0026] During the data feature screening process, feature selection was performed on the dataset. This process aims to optimize the performance and clinical interpretability of the machine learning (ML) model while reducing its complexity. Regarding clinical characteristics, the chi-square test or ANOVA was used to select features that were statistically significantly correlated with the target variable (high risk versus low risk of acute radiation dermatitis) (p < 0.05). For the optical and physiological parameter characteristics of the skin and the dose characteristic data, LASSO (L1 regularization) was applied to select features. It can reduce the weight of unimportant features through penalty terms, thereby achieving feature selection. The feature importance score of random forest was used to select the most important features.

[0027] The third step is to construct a risk prediction model, which includes a stacked multimodal machine learning architecture, which includes basic modules and meta-modules interconnected in a cascade manner. The basic module includes seven tree-based bagging and boosting basic classifiers, namely decision tree (DT), random forest (RF), AdaBoost, XGBoost, Extratrees, Light GBM and CatBoost. The seven basic classifiers run in parallel to independently predict input features and use a five-fold cross-validation method to prepare the input of the meta-module. The seven basic classifiers aggregate the prediction results and transmit the aggregated results to the meta-module. The meta-module is composed of meta-classifiers, which process the aggregated results completed by the basic classifiers to finally obtain the prediction results of acute radiation skin toxicity.

[0028] In the third step above, the parameters of each base classifier are set as follows: the number of trees is 100, the depth is 5, and the learning rate is 0.1. Similarly, the meta-module classifier is the Logistic Regression classifier, and the parameters are set as follows: the C value is 1.0 and the penalty parameter is 0.0001.

[0029] Throughout the training phase, the base module uses a five-fold cross-validation method to prepare the input of the meta-module. This method helps alleviate overfitting and improve model stability. This involves randomly splitting the training set into five different subsets of equal size for iterative model training (five times in total). In each iteration, four of the five subsets are used to train the seven base classifiers simultaneously, while the remaining subset is used for internal validation. Specifically, the seven base classifiers predict the results of the remaining subsets so that the performance of the base classifiers relative to the ground truth can be evaluated. At the same time, these prediction results are aggregated to form the input features of the meta-classifier. After completing five iterations, a set of meta-input features will be obtained, so that all features can be used to train the meta-classifier to predict acute radiation skin toxicity. Subsequently, the performance of the stacked model is evaluated on the test set to assess its effectiveness. The classifier predictions are divided into true positive (TP), true negative (TN), false positive (FP), and false negative (FN). These parameters define indicators such as accuracy, recall, precision, AUC-ROC score, etc., through which the performance of the model is evaluated.

[0030] The fourth step is to use SHAP and LIME tools to interpret the tree-based model and enhance the global and local interpretability of the model. The model interpretation step is implemented in the Python environment using libraries, where the SHAP value calculation uses the shap package and the LIME interpretation uses the lime package.

[0031] The SHAP is a game theory based model that explains machine learning predictions by helping to generate visualizations of the SHAP feature values. The SHAP values ​​estimate the impact of features on the predictions, while feature importance measures the impact of features on model fit. The LIME explains the predictions in terms of individual patients and provides personalized probabilistic risk assessments for patients.

[0032] The fifth step was to evaluate the risk value and obtain the cutoff value. The receiver operating characteristic (ROC) curve analysis was used to evaluate the ability of the risk value to identify the occurrence of acute radiation dermatitis in the training set, and the cutoff value was obtained according to the Youden index. The subjects were divided into low-risk and high-risk groups, and finally the model prediction results were compared with the actual occurrence of acute radiation dermatitis.

[0033] The risk model generates a probability score for each subject, ranging from 0 (lowest risk probability) to 1 (highest risk probability), which can be used as a quantitative marker of acute radiation skin toxicity to predict the future risk of acute radiation dermatitis. Receiver operating characteristic (ROC) curve analysis was used to evaluate the ability of risk values ​​to discriminate the occurrence of acute radiation dermatitis in the training set. Then, the cutoff value was obtained according to the Youden index, and the subjects were divided into low-risk and high-risk groups. The prediction results of the model were compared with the risk value range.

[0034] like Figure 5 As shown, Figure 5 The personalized risk prediction results for a high-risk patient and a low-risk patient show the contribution of 7 important features selected from skin optical and physiological characteristics, clinical characteristics and dose characteristics to the prediction model, as well as the patient's actual grade label of acute radiation dermatitis and the risk value predicted by the model. Figure 5 It can be seen in the upper right box that the real grade label of acute radiation dermatitis in high-risk patients is 1, and its predicted value is 0.80, while the real grade label of acute radiation dermatitis in low-risk patients is 0, and its predicted value is 0.06. The predicted values ​​of the two are relatively close to the real grade labels, and can truly reflect the risk size of acute radiation dermatitis in the current patients. It can be considered that the risk prediction model of the present invention can achieve the interpretability of the model and personalized risk prediction.

[0035] In the first embodiment of the present invention, a standard three-phase shift method or a single snapshot multi-frequency demodulation method is generally used to demodulate and obtain the modulation transfer function of the sample at the modulation frequency.

[0036] When the standard three-phase shift method is used, the image demodulation module uses the standard three-phase shift method (α=0, 2π / 3, 4π / 3) to demodulate the measured skin at modulation frequencies 0 and DC and AC modulation transfer functions (MTF) under the condition, and then according to the propagation theory of light in turbid media, the medium including absorption coefficient ( ) and the reduced scattering coefficient ( ), etc., when the standard three-phase shift method is used for image demodulation, the three phase images are , and It can be expressed as: {I}_{AC}=\frac {\sqrt {2}} {3}\left [ {\left ( {{I}_{1}-{I}_{2}} \right )^{2}+\left ( {{I}_{2}-{I}_{3}} \right )^{2}+\left ( {{I}_{3}-{I}_{1}} \right )^{2}} \right ]^{\frac {1} {2}} .

[0037] The above-mentioned traditional three-phase shift demodulation can effectively suppress the influence of non-uniform illumination and ambient light on the measurement accuracy, thereby providing relatively high optical property measurement accuracy, which is suitable for high-resolution imaging of tissues. However, since three images need to be collected to provide information of three phases, motion artifacts will be generated during high-speed shooting, and there are problems such as slow imaging speed and large noise interference, and only optical information at one frequency can be obtained.

[0038] When the single snapshot multi-frequency demodulation method (SSMD) is used, the image demodulation module uses the single snapshot multi-frequency demodulation method (SSMD) to demodulate the measured skin at modulation frequencies 0 and DC and AC modulation transfer functions (MTF) under the condition, and then according to the propagation theory of light in turbid media, the medium including absorption coefficient ( ) and the reduced scattering coefficient ( ), etc., when single snapshot multi-frequency demodulation (SSMD) is used for image demodulation, the structured light pattern incident on the surface of the tissue sample and the light intensity pattern obtained by tissue backscattering can be expressed as: {I}_{0}\left ( {x,y} \right )={I}^\left ( {0} \right )_{DC}+\sum ^{k}_{i=1} {{I}^\left ( {0} \right )_{AC,i}}\cos {\left [ {2π\left ( {{f}_{x,i}x+{f}_{y,i}y} \right )\left ( {} \right )+{\varnothing}_{i}} \right ]} I\left ( {x,y} \right )={I}_{DC}\left ( {x,y} \right )+\sum ^{k}_{i=1} {{I}_{AC,i}}\left ( {x,y} \right )\cos {\left [ {2π\left ( {{f}_{x,i}x+{f}_{y,i}y} \right )+{\varnothing}_{i}} \right ]} where k ≥ 1 is the number of AC components, and The spatial frequencies are and Phase The next The incident amplitude and backscattered amplitude of each AC component of SSMD and the DC component amplitude The calculations are as follows: , In the sliding window σ, the integral is performed to eliminate the phase The unknown frequency is respectively and size is The sine and cosine basis kernels are used for filtering, and the two images obtained are combined to obtain the corresponding AC component. SSMD makes full use of the orthogonality of the harmonic function to quickly extract the AC and DC components of multiple spatial modulation frequencies in a single image at the same time, meeting the demand for obtaining multiple frequency information at one time. It has fast speed and higher demodulation accuracy, and maintains a high signal-to-noise ratio. Compared with the above-mentioned traditional three-phase shift method that requires the acquisition of three images, the single-snapshot multi-frequency demodulation method only needs one picture to demodulate the information of multiple frequencies, effectively suppressing motion artifacts. It is particularly suitable for dynamic and fast imaging scenarios in clinical practice, and effectively overcomes the problems of the three-phase shift method, such as slow imaging speed, great noise influence, and only being able to obtain optical information at one frequency.

[0039] In biological tissues, the absorption coefficient is determined by the tissue chromophore and its concentration. In the visible light band, the main chromophores in biological tissues are oxygenated hemoglobin (HbO2), deoxyhemoglobin (Hb) and melanin. Assuming that the absorption coefficient of the tissue in each band is the linear superposition of the spectra of each chromophore (that is, the absorption of each chromophore is independent of each other), the absorption coefficient equation of the tested skin tissue is established as follows: , where λ is the wavelength of light, N is the number of chromophores, is the molar extinction coefficient of the chromophore, is the chromophore concentration, described by Mie scattering , where a represents the scattering amplitude, which is related to the number of scatterers, b Represents the scattering ability related to the size of the scatterer, obtained by low-frequency structured light imaging under multiple wavelengths Can be further based on Obtained, measured total hemoglobin concentration in the layered structure and oxygen saturation , equivalent light absorption and scattering are achieved by mapping typical photons between the homogeneous medium and the original two-layer structure, i.e. {μ}_{a}L={μ}_{a,melanin}\left ( {λ} \right )h+\left [ {{μ}_{a,Hb{O}_{2}}\left ( {λ} \right )+{μ}_{a,Hb}\left ( {λ} \right )} \right ]\left ( {Lh} \right ) ,in h is the thickness of the epidermis of the skin, L For the spatial frequency and the average detection depth of structured light at wavelength λ.

[0040] In summary, compared with the prior art, the technical solution of the present invention effectively improves the prediction accuracy of the risk model by using a multi-level classification structure of basic classifiers and meta-classifiers and a five-fold cross-validation method. It also uses tools such as SHAP and then LIME to interpret the tree-based model, so that users can understand the reasons why the model makes and obtains prediction results, thereby enhancing the interpretability of the risk prediction model. The receiver operating characteristic curve analysis is used to evaluate the ability of risk values ​​to identify the occurrence of acute radiation dermatitis in the training set. This quantitative risk assessment method helps clinicians to conduct more accurate risk management of patients, and effectively avoids the problem that current clinicians need to rely on experience and subjective judgment to assess the current status of patients. In addition, the technical solution of the present invention also uses SFDI to characterize the skin microcirculation in the patient's radiotherapy area. After adding clinical characteristic data and dose characteristic data, the adaptability of the risk prediction model to individual differences and complex disease changes and the predictive performance of the risk prediction model are improved. In clinical applications, the risk prediction method and risk prediction model of the present invention can also be used to evaluate skin toxicity in the radiotherapy area of ​​patients with breast cancer and nasopharyngeal carcinoma, and can also be extended to the evaluation of other radiation skin toxicity caused by radiotherapy. In addition, the device and method of the present invention can also be used in other scenarios that require evaluation of skin toxicity, such as burn rehabilitation, drug testing, etc.

[0041] Although the present invention has been disclosed as above in terms of specific embodiments, they are not intended to limit the present invention. Any technician in this field may make slight changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the scope of the attached claims.

Claims

1. A method and system for predicting the risk of acute radiation dermatitis. Features: The system comprises a hardware module and a software module. The hardware module comprises an optical image acquisition device based on spatial frequency domain imaging, and the optical image acquisition device is used to acquire the skin surface image of the patient's radiotherapy treatment area. The software module comprises an image demodulation module, a parameter inversion module and a risk prediction module. The image demodulation module is used to demodulate the acquired skin surface image to obtain the DC and AC modulation transfer functions (MTF) of the tested skin at the modulation frequency. The parameter inversion module is used to invert the DC and AC modulation transfer functions (MTF) to calculate the corresponding physiological parameters converted by the optical imaging model mapped by the tissue layered structure. The risk prediction module is used to construct an explainable machine learning risk prediction model and give a personalized risk score for the patient.

2. A method and system for predicting the risk of acute radiation dermatitis according to claim 1, Features: The risk prediction module comprises the following steps: Step 1: Collect patient clinical characteristics, skin optical and physiological parameters, and dose parameter data, and conduct skin toxicity assessment on patients, and finally perform data preprocessing and data feature screening; Step 2: constructing a risk prediction model, the risk prediction model includes a stacked multimodal machine learning architecture, the machine learning architecture includes basic modules and meta-modules interconnected in a cascade manner, the basic module includes seven tree-based bagging and boosting basic classifiers, namely decision tree (DT), random forest (RF), AdaBoost, XGBoost, Extratrees, Light GBM and CatBoost, the seven basic classifiers run in parallel to independently predict input features and use a five-fold cross-validation method to prepare the input of the meta-module, the seven basic classifiers aggregate the prediction results and transmit the aggregated results to the meta-module, the meta-module is composed of meta-classifiers, the meta-classifiers process the aggregated results aggregated by the basic classifiers, and finally obtain the prediction results of acute radiation skin toxicity; Step 3: Use SHAP and LIME tools to interpret the tree-based model to enhance the global and local interpretability of the model. The model interpretation step is implemented in the Python environment using libraries, where the SHAP value calculation uses the shap package and the LIME interpretation uses the lime package; Step 4: Evaluate the risk value and obtain the cutoff value. The receiver operating characteristic (ROC) curve analysis was used to evaluate the ability of the risk value to identify the occurrence of acute radiation dermatitis in the training set, and the cutoff value was obtained according to the Youden index. The subjects were divided into low-risk and high-risk groups. Finally, the model prediction results were compared with the actual acute radiation dermatitis.

3. A method and system for predicting the risk of acute radiation dermatitis according to claim 2, Features: The clinical characteristics of the patient in step 1 include age, gender, hypertension, diabetes, smoking history, drinking history, body mass index (BMI), chemotherapy, and immunotherapy. The skin optical and physiological parameters include total hemoglobin concentration, oxygen saturation, melanin concentration, oxygen saturation, reduced scattering coefficient, scattering power, epidermal roughness, and epidermal thickness. The dose parameters of the tested skin are finally obtained by using a radiotherapy planning system at a position 5 mm below the outer surface of the patient's skin and converting the obtained skin surface dose into DVH parameters. The dose parameters include the maximum dose, minimum dose, average dose, median dose, and skin V 10 、V 20 、V 30 、V 40 、V 50 、V 60 The missing values ​​in the data preprocessing were replaced by the feature median and all continuous variables were standardized, the categorical variables were encoded, and the final data were divided into a training set and a test set in a ratio of 8:2 and the training data were balanced using Borderline SMOTE.

4. A method and system for predicting the risk of acute radiation dermatitis according to claim 1, 2 or 3, Features: The optical image acquisition device comprises an imaging light source, a projection device and a detector. The projection device uses a digital micromirror (1) to achieve structured illumination, projects mixed white light with wavelengths of R (623 nm), G (539 nm) and B (465 nm), and uses a collimating objective lens (2) to collimate and correct the mixed white light. The frequencies of the projected mixed white light are 0 and 0.2 mm. -1 By adjusting the frequency, different imaging depths can be achieved. The imaging area is 3*3cm, and tissue wide-field imaging is achieved. The detector uses a CCD camera (3) to irradiate spatial modulation patterns of different frequencies. to the sample area and capture the reflected image with a CCD camera (3) .

5. A method and system for predicting the risk of acute radiation dermatitis according to claim 4, Features: The image demodulation module uses a standard three-phase shift method (α=0, 2π / 3, 4π / 3) to demodulate the measured skin at modulation frequencies 0 and DC and AC modulation transfer functions (MTF) under the condition, and then according to the propagation theory of light in turbid media, the medium including absorption coefficient ( ) and the reduced scattering coefficient ( ), etc., when the standard three-phase shift method is used for image demodulation, the three phase images are , and It can be expressed as: 。 6. A method and system for predicting the risk of acute radiation dermatitis according to claim 4, Features: The image demodulation module uses a single snapshot multi-frequency demodulation method (SSMD) to demodulate the measured skin at modulation frequencies 0 and 1. DC and AC modulation transfer functions (MTF) under the condition, and then according to the propagation theory of light in turbid media, the medium including absorption coefficient ( ) and the reduced scattering coefficient ( ), etc., when single snapshot multi-frequency demodulation (SSMD) is used for image demodulation, the structured light pattern incident on the surface of the tissue sample and the light intensity pattern obtained by tissue backscattering can be expressed as: where k ≥ 1 is the number of AC components, and The spatial frequencies are and Phase The next The incident amplitude and backscattered amplitude of each AC component of SSMD and the DC component amplitude The calculations are as follows: 。 7. A method and system for predicting the risk of acute radiation dermatitis according to claim 5 or 6, Features: The absorption coefficient equation of the tested skin tissue is established as , where λ is the wavelength of light, N is the number of chromophores, is the molar extinction coefficient of the chromophore, is the chromophore concentration, described by Mie scattering , where a represents the scattering amplitude, which is related to the number of scatterers, b Represents the scattering ability related to the size of the scatterer, obtained by low-frequency structured light imaging under multiple wavelengths Can be further based on Obtained, measured total hemoglobin concentration in the layered structure and oxygen saturation , equivalent light absorption and scattering are achieved by mapping typical photons between the homogeneous medium and the original two-layer structure, i.e. ,in h is the thickness of the epidermis of the skin, L For the spatial frequency and the average detection depth of structured light at wavelength λ.

Images