Methods, related computer-readable media, and computing devices for multi-omic processing of head and neck cancer patient data

Abstract

The present application provides a method for multi-omics processing through radiotherapy data of a head and neck cancer patient, wherein multi-omics is applied to pre-treatment evaluation of radiotherapy re-planning needs of the head and neck cancer patient, comprising the following steps: (1) pre-processing of radiotherapy patient data of a head and neck cancer patient, the radiotherapy patient data comprising radiotherapy data and clinical feature data of the patient; (2) extracting multi-omics features from the pre-processed radiotherapy patient data; (3) estimating or predicting the RTR needs of the head and neck cancer patient by fitting the extracted multi-omics features based on a prediction model. The present application allows for "pre-treatment" identification of patients at risk for RTR, so that doctors can be aware of these patients in advance and better allocate resources for them. The present application does not change the dose and / or coverage of the tumor during radiotherapy; the patient will still experience changes in his anatomy (volume and position of organs) during radiotherapy, as this is a natural reaction of the organs when they are irradiated during radiotherapy.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for multi-omics processing of radiotherapy patient data of head and neck cancer patients. BACKGROUND

[0002] Head-and-Neck Cancer (HNC) is a common tumor that exists in the upper respiratory and digestive tract, and is the fifth most common malignant tumor worldwide. Intensity modulated radiotherapy (IMRT) is currently the standard treatment method for HNC patients in Hong Kong. Before radiotherapy (RT) is performed on cancer patients, the RT treatment plan is pre-customized according to the anatomical structure of each patient or tumor organ. However, during the entire radiotherapy process, the volume and position of the patient organs receiving radiotherapy will change significantly. Therefore, such changes can weaken the radiation dose coverage of the tumor and / or cause excessive dose to the surrounding healthy normal tissue.

[0003] RadioTherapy Re-planning (RTR) is currently commonly used to modify the original radiotherapy plan to compensate for these changes to maintain a satisfactory treatment rate. For example, anatomical changes occurring during intensity modulated radiotherapy can be compensated for by adaptive radiotherapy (ART) using repeated computed tomography (CT) scans to perform re-planning during the treatment process. The clinical and dosimetric benefits of radiotherapy re-planning for HNC patients have been widely reported, and in today's clinical environment, the use of radiotherapy re-planning has become commonplace worldwide. However, the current implementation efficiency of the radiotherapy re-planning treatment procedure is very low, as it is usually completed in a real-time manner. In addition, the monitoring and risk assessment procedure of the entire radiotherapy re-planning is resource-intensive and time-consuming, as it requires efforts from multiple medical departments from image re-scanning (radiation therapists), organ re-segmentation (medical physicists and oncologists), treatment re-planning (medical physicists) to new plan approval (oncologists) for patient-based need assessment, image re-scanning, organ re-segmentation, and treatment re-planning, etc. In particular, the radiotherapy re-planning need screening of each patient can only be performed after the treatment begins, which requires intensive mid-term monitoring for each HNC patient, and there is no universal screening method for oncology departments worldwide. Considering the above factors, the current radiotherapy re-planning practice is very inefficient and places a huge clinical burden on oncology centers.

[0004] For example, Ting-ting Yu et al. in the paper “Pretreatment Prediction of Adaptive Radiation Therapy Eligibility Using MRI-Based Radiomics for Advanced Nasopharyngeal Carcinoma Patients”, Frontiers in oncology, 9, 1050, published on October 16, 2019. See https: / / doi.org / 10.3389 / fonc.2019.01050 demonstrated the promising ability of MRI-based radiomics in the pretreatment identification of adaptive radiation therapy (ART) eligibility for nasopharyngeal carcinoma (NPC) patients. In particular, the joint T1-T2 model with 6 selected radiomics features in the paper appears to be a more desirable prediction system compared to other research models. This would enable radiation oncologists to prescribe ART more effectively and accurately on an individual patient basis to achieve true NPC patient individualized radiation therapy while streamlining resource management in the clinical setting. In future work, a multi-institutional prospective study using a larger patient sample is needed to improve the clinical efficacy of the model in the paper. However, the paper only involves radiomics features from the primary tumor only by using MR images of head and neck cancer patients. No nomogram was generated in the paper, so there was no risk probability assessment for individual patients.

[0005] Therefore, the inventors propose to apply radiomics to the radiotherapy process for head-and-neck cancer (HNC) patients to provide pre-treatment assessment of the radiotherapy re-planning needs for HNC patients. SUMMARY

[0006] The present invention relates to a method for multi-omics processing by radiotherapy data of head-and-neck cancer patients to provide pre-treatment assessment of radiotherapy re-planning (RTR) needs for head-and-neck cancer patients, wherein multi-omics is applied to the pre-treatment assessment of RTR needs for the HNC patients.

[0007] The technology of the present invention mainly includes three aspects: 1) pre-processing of radiotherapy patient data, 2) extraction of multi-omics features, 3) estimation of RTR needs.

[0008] Therefore, the method of multi-omics processing of radiotherapy data of a head and neck cancer patient according to the present application, wherein multi-omics is applied to the pre-treatment evaluation of the radiotherapy re-planning needs of the head and neck cancer patient, comprises the following steps: (1) pre-processing of the radiotherapy patient data of the head and neck cancer patient to ensure that the features of the input image / dose data of the head and neck cancer patient during the radiotherapy process are consistent with the features of the image / dose data used to construct the prediction model for radiotherapy re-planning; (2) extracting multi-omics features from the pre-processed radiotherapy patient data; and (3) estimating or predicting the radiotherapy re-planning needs of the head and neck cancer patient by fitting the extracted multi-omics features based on the prediction model.

[0009] The present application aims to allow the "pre-treatment evaluation" of the RTR needs of a HNC patient, i.e. whether this particular patient will have more chances of RTR needs if he is treated according to the initial treatment plan.

[0010] Once a patient at risk is identified, it is up to the physician to adjust (and how) the initial treatment plan, or to better allocate hospital resources (e.g. multidisciplinary manpower, time, machine availability, etc.) in advance, or to take other remedial measures. The present application does not change the dose and / or its coverage for the tumor. It is worth noting that the patient will still experience changes in his anatomy (volume and position of organs) during radiotherapy, as this is a natural reaction of the organs when they are irradiated. The present application does not prevent this from happening.

[0011] The present application allows the "pre-treatment" identification of patients at risk for RTR, so that the physician can be aware of them in advance and better allocate resources for them. Otherwise, the traditional periodic close monitoring of every patient throughout the RT process would undoubtedly impose a huge burden on the clinic, and would be extremely inefficient. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is an exemplary block diagram of the prediction model according to the present application.

[0013] Figure 2 is a nomogram of the prediction of the RTR needs of a HNC patient according to the prediction model of the present application.

[0014] Figures 3(a) and 3(b) are a comparison of the high and low dose scaling invariant moments of a HNC patient according to the prediction model of the present application.

[0015] Figure 4 is a distance profile and an angle histogram of the closest distance from an arbitrary position to the surface of the target volume of a HNC patient according to the prediction model of the present application. DETAILED DESCRIPTION

[0016] The technical solutions and specific embodiments of the present application are described in further detail below with reference to the accompanying drawings.

[0017] Multi-omics, also known as integrative omics, "panomics" or "pan-omics", is a method of biological analysis in which the dataset is multiple "omes", such as genome, proteome, transcriptome, epigenome, metabolome and microbiome, i.e., metagenome and / or metatranscriptome, depending on its sequencing method); in other words, multiple omics technologies are used to study life in a coordinated and consistent manner. By combining these "omes", researchers can analyze complex biological big data to discover new associations between biological entities, determine relevant biomarkers and build fine markers of disease and physiology. In doing so, multi-omics integrates different omics data to find consistent matching gene-phenotype-environmental relationships or associations. The OmicTools service lists over 99 software related to multi-omics feature data analysis, and over 99 databases on this topic.

[0018] As described above, the technology of the present application mainly includes three aspects: 1) preprocessing of radiotherapy patient data, 2) extraction of multi-omics features, and 3) estimation of RTR needs.

[0019] Specifically, first, the technology of the present application includes preprocessing of radiotherapy patient data, in which the radiotherapy patient data involved in RTR need prediction includes RT data and clinical feature data of patients. The RT data adopts the format of Digital Imaging and Communications in Medicine (DICOM) files, including multi-modal radio image (i.e., computed tomography (CT) and multi-parameter magnetic resonance (MRI) image) data, RT dose data and organ contour data. The clinical feature data of patients includes patient demographic data, tumor characteristics and the like.

[0020] DICOM stands for "Digital Imaging and Communications in Medicine" and is an international standard for handling, storing, printing, and transferring information in medical imaging - for managing and transferring medical images and related data. It includes both file format definitions and network communication protocols. DICOM was developed jointly by the National Electrical Manufacturers Association (NEMA) and the American College of Radiology (ACR) to allow interoperability between imaging devices and with other devices. The DICOM standard is responsible for managing the image format as well as the various network protocols required to transfer medical image information generated during many healthcare-related imaging "modalities" (e.g., magnetic resonance, nuclear medicine, computed tomography, and ultrasound). DICOM is both a communication protocol and a file format, meaning it can store medical information such as ultrasound and MRI images, as well as patient information in one file. The format ensures that all data stays together and provides the ability to transfer information between devices that support the DICOM format. For example, a CT scan can result in dozens to hundreds of CT images, one for each.dcm file, with the file name being the sequence number. The content stored in the DICOM standard includes not only the patient's CT values but also information about the CT machine, slice thickness, timestamp, patient basic information, and so on, which can all be found in the CT file. DICOM is used worldwide to store, exchange, and transfer medical images, making it possible to integrate medical imaging devices from different manufacturers, exchange and store radiotherapy patient data and related images in a standard format, and facilitate more accurate treatment for patients.

[0021] Preprocessing is performed on the radiotherapy patient data to ensure that the characteristics of the input image / dose data of the patient during radiotherapy (e.g., image resolution, signal intensity variability, and uniformity) are consistent with those of the image / dose data used to construct the RTR prediction model of the present application. Specifically, voxel size resampling is applied to all images, masks, and doses to maintain a consistent resolution of 1 mm x 1 mm x 1 mm. Bias correction is performed using the "N4 Bias Field Correction Image Filter" function of SimpleITK, a software specifically designed to handle medical images, to reduce voxel intensity bias of MR images caused by inhomogeneous magnetic fields. MR images are normalized by moving the average voxel intensity and rescaling its standard deviation. A voxel is a short form of Volume Pixel, and a volume containing voxels can be rendered in 3D or extracted as a polygonal surface of a given threshold contour. A "voxel" is the smallest unit of digital data in a three-dimensional space division, used in the fields of three-dimensional imaging, scientific data, and medical imaging. Conceptually similar to the smallest unit of two-dimensional space, "pixel", used in two-dimensional computer images.

[0022] In addition to the voxel size resampling, the preprocessing step also includes: all CECT (contrast-enhanced computed tomography) and MR (magnetic resonance) images are resampled to 1x1x1-mm 3 voxel size before radiomic features are extracted to mitigate the effect of image acquisition parameter differences between different patients. Among them, quantization of gray levels is applied to normalize the image signal intensity in both image types. The gray level intensity of the image is discretized into a fixed bin count range, which ranges from 50 to 350 with an incremental interval of 50. In addition, all images are convolved with a Gaussian Laplacian (LoG) filter at three Gaussian radius parameter levels to generate filtered images, which highlight specific texture radiology features at multiple scales from fine (1 mm), medium (3 mm) to coarse (6 mm). Wavelet filters (HHH, HLL, LHL, LLH, LHH, HLH, HHL, LLL, H represents high, L represents low) are also applied to both CECT and MR images to produce radiology features with multiple resolutions. The primary tumor and neck lymph node lesions on the CECT image are re-segmented to limit the Hounsfield Unit (HU) to the range of [-150, 180] to eliminate non-soft tissue components within the VOI (volume of interest) under study, such as air cavities and bone structures. In particular for MR images, the inhomogeneity correction of image pixel values is achieved by using N4B bias correction provided in the "N4 bias field correction image filter" in SimpleITK (v1.2.4).

[0023] The technology of the present application is performed by the RTR prediction model. The RTR prediction model of the present application is developed based on multi-omics feature data for characterizing patient-specific and treatment-specific information. It contains multiple multi-omics-based predictors (e.g., radiomic data, dosimetric data, geometric data) extracted from multiple organ structures (including ipsilateral and contralateral parotid glands, primary tumor, neck lymph node lesions, spinal cord, brain stem, planning-target-volume of lymph node lesions, etc.) by using pre-treatment images (CECT, MR images) and clinical parameters of multi-modal patients.

[0024] During the model development process, the patient cohort was split into training and test datasets in a 7 to 3 ratio via 20 iterations. Supervised feature selection algorithms were applied only to the training dataset of each iteration to preserve the clinical relevance of the residual features. Unsupervised feature selection algorithms were subsequently applied to remove highly redundant features, resulting in a reduced feature set. The development of the RTR prediction model was performed by using a ridge regression algorithm via 10-fold cross-validation within the training set to reduce the risk of model overfitting.

[0025] Figure 1 An exemplary block diagram of the prediction model according to the present application is shown in FIG. 1. In Figure 1 The inputs of the prediction model are specified multi-modality images, multiple ROIs, initial planning dose, and clinical parameters. The final output is the RTR prediction probability. The first three types of inputs are from the radiotherapy planning data, which need to be cleaned, including data screening and format conversion, to meet the following conditions before inputting into the prediction model:

[0026] 1. Whether the selected image data modality is correct, whether it is an enhanced image, whether the time of shooting is before radiotherapy, and whether the delineation of the ROI and the planning dose distribution belong to the initial plan.

[0027] 2. All images and structures are in the same coordinate system, so as to ensure that each ROI can correspond to the correct position of the image and dose distribution.

[0028] 3. All image, ROI, and dose distribution data are converted from DICOM to Simple ITK image data, in which the ROI after conversion is a 3D binary mask image.

[0029] The establishment of the prediction model of the present application includes multiple intermediate steps, including multi-omics feature extraction, feature selection and linear modeling, calibration and score combination.

[0030] Multimodality imaging data is pre-processed and features are extracted. Radiomic features are computed from the multimodality imaging data and multiple regions of interest (ROIs), dosiomic features are computed from the initial plan dose distribution and multiple ROIs, and geometric features are computed from multiple ROIs. The complete set of multimodality features is used for model training, but only the final selected features are used for model deployment. The inventors perform feature selection and linear modeling for each type of features, and obtain radiomic score, dosiomic score, and geometric score. The three linear models can be described by feature coefficients and intercepts, which are used for model deployment. Finally, the inventors calibrate the three scores and clinical parameters and stack them to obtain the final RTR probability. The calibration and stacking process is implemented by a nomogram. Each scale of the nomogram represents the calibration model of each score and clinical parameter. The calibrated scores are added and calibrated again to obtain the RTR probability.

[0031] Key steps in the development of the prediction model of the present invention include patient cohort division and feature selection. First, in the model development process, the patient cohort is divided into training and test datasets in a 7 to 3 ratio by 20 iterations. Then, supervised feature selection algorithms are applied only to the training dataset of each iteration to maintain the clinical relevance of the remaining features. In this feature selection process, statistical univariate analysis is used to assess the clinical relevance of each extracted multimodality feature and the outcome of interest, i.e., whether the patients in the training dataset have RTR risk determined by clinical oncologists. Subsequently, unsupervised feature selection algorithms (e.g., Pearson’s correlation test algorithm) are applied to remove highly redundant features, thereby reducing the feature set. Thus, the main components of the training dataset include multimodality features (radiomic features extracted from multimodality medical images, geometric features from multiple ROIs, and dosiomic features from the initial plan dose) and clinical parameters of patients, all in numerical form.

[0032] Therefore, by the RTR prediction model of the present invention, the final predicted probability of RTR need is: a linear combination of radiomic features and their coefficients (Rad-Score), and / or a linear combination of dosiomic features and their coefficients (Dose-Score), and / or a linear combination of geometric features and their coefficients (Geo-Score), and / or predicted clinical parameters, such as Figure 2 as shown.

[0033] In one embodiment, the development of the RTR prediction model of the present invention is carried out using the Ridge algorithm (a typical statistical method that uses a linear function to resolve the bias-variance tradeoff), wherein residual features are used via 10-fold cross-validation in the training set to mitigate the risk of overfitting of the prediction model. In this process, radiomics models, dosimics models, and geometric models are first developed separately using the training set dataset; this process can be referred to as modeling.

[0034] The key to the modeling process is generating an equation for each of the three models. This equation represents a linear combination of all selected predictor variables in the model and is referred to as: the Rad-score for the Radiomic model, the Dose-score for the Dosiomics model, and the Geo-score for the Geometric model. The equations are shown below:

[0035] Radiometric score = intercept r + R1(r1) + R2(r2) + R3(r3) + ... + R i (r i )

[0036] Dose score = intercept d + D1(d1) + D2(d2) + D3(d3) + ... + D j (d j )

[0037] Geometric score = intercept g + G1(g1) + G2(g2) + G3(g3) + ... + G k (g k )

[0038] Here, "intercept" is the translation of "Intercept". Intercept r, intercept d, and intercept g are derived from the RTR prediction model developed in this invention for radiomics, dosimics, and geometry models respectively using the Ridge algorithm. (R, D, G) and (r, d, g) are the predictor variables and their coefficients (weights) in the radiomics, dosimics, and geometry models, respectively. (i, j, k) represent the final number of radiological, dosimetric, and geometric features, respectively.

[0039] Examples of the various intercepts are shown in Figure 2 In the nodal diagram.

[0040] As a result, the radiological score, dose score, and geometric score (i.e., as three variables) were calculated for each patient in the training dataset. These three variables, along with clinical predictor variables, were used as input to develop a predictive nomogram for RTR needs using the logistic regression modeling (lrm) function in the rms package of R Studio software (i.e., the R language integrated development environment). Figure 2 As shown. The clinical predictor variables exhibited statistically significant associations, e.g., p-values ​​< 0.05. During nomogram construction, the weights of each predictor variable (dose score, geometric score, radiographic score, gender, and tumor stage) were calibrated based on the training set dataset; these weights contribute to the final RTR requirements. Figure 2 In the exemplary nomogram, the length of each row for each variable is different, representing their respective weights in relation to the RTR requirement. To further illustrate this nomogram from a practical perspective, for example, if a new male patient (reference nomogram ~2) with stage IV tumor (see ~2 in the nomogram) comes to the clinic, then for this particular patient, with a dose score of 0.4 (reference nomogram ~3), a geometry score of 0.62 (reference nomogram ~11), and a radiation score of 0.8 (reference nomogram ~70), the total score would equal 2 + 2 + 3 + 11 + 70 = 88, which would yield an estimated RTR requirement of 0.9 for this particular patient. Figure 2 The last line of the RTR requirement is shown by the dashed line.

[0041] The higher the predicted RTR risk probability value, the higher the RTR requirement for a particular patient. The nomogram model is pre-trained and developed using all multi-omics features and clinical parameters. Once all features specified by the intended patient model are obtained, the nomogram model of this invention can be used to predict the patient's RTR. Work on this invention (and the predicted nomogram) is ongoing, and model development will be further optimized using a larger data cohort.

[0042] In a second aspect, the present invention includes the extraction of multi-omics features: calculating three types of multi-omics features from preprocessed DICOM data. The three types of multi-omics features include radiomics features, dosimics features, and geometric features, which are described below:

[0043] - Radiomic features: Radiomics is able to characterize the intrinsic tissue biology response at treatment perturbation, the inventors have been working on applying radiomics to predict patient-specific treatment response based on tumor shrinkage, which laid a good foundation for applying radiomics to predict patient-specific RTR demand for HNC patients, because the implementation of RTR can be triggered by shrinkage of multiple organs (e.g. primary tumor, metastatic neck lymph nodes and parotid gland). In the present invention, radiomic features are extracted from multi-modal images within organ contours, and the calculation of radiomic features is performed by using a publicly available Python package-PyRadiomics, following the Image Biomarker Standardization Initiative (IBSI) guidelines. Python is a computer programming language, and PyRadiomics is an open-source python package.

[0044] - Dosiomic features: Dosiomics characterizes the aggressiveness of individual radiotherapy plans by providing dose information from regular dose volume histograms and local spatial patterns of radiotherapy dose distribution within organ contours. Dosiomic features are extracted from individual patient's radiotherapy plans.

[0045] In addition, dosiomic information similar to radiomics has also been widely studied for prognosis and toxicity prediction of various cancers, showing great potential in predicting tissue response at treatment perturbation. In the present invention, dosiomic information similar to radiomics is also calculated. The calculation of dosiomic features is performed by an algorithm developed by the inventors themselves.

[0046] - Geometric features: The geometric relationship between the patient's organs / tumors can affect the degree of anatomical change at treatment, leading to the need for RTR. In the present invention, the patient's geometric organs are represented by distance and angle. The calculation of geometric features is performed by an algorithm developed by the inventors themselves.

[0047] The algorithm of the present invention first includes the design of multi-omics features. The inventors have designed a variety of innovative radiotherapy data multi-omics features to comprehensively and effectively describe radiotherapy plan information, so as to be able to mine more and more targeted prediction factors in the modeling process. Specifically, the multi-omics features used in the present invention include: imageomic features (Radiomic features) of multiple imaging modalities and regions before radiotherapy, geometric features (Geometric features) between the target volume and the target organ contour, and dosiomic features (Dosiomic features) based on three-dimensional plan radiotherapy dose distribution. The geometric features include distance features (Distance features) and angle features (Angle features). The combination of distance and angle can most accurately describe the positional relationship between the target volume and the target organ.

[0048] Radiomics features, dosimetrics features and geometric features are extracted from 3D medical images, dose distribution maps, and distance distribution maps, respectively. Since all three kinds of distribution maps are volume distribution, the three kinds of features share the same set of feature definitions, including traditional radiomics feature definitions and other spatial descriptions for volume distribution. The extraction of traditional radiomics features refers to the standard of Image Biomarker Standardization Initiative (IBSI) by [1] Zwanenburg A, Leger S, Vallières M, S. published online in Radiology on March 10, 2020: 191145. doi: 10.1148 / radiol.2020191145. The three kinds of features can be calculated by open-source Python software PyRadiomics, mainly divided into shape features, first-order features, and high-order features (texture features). First-order and high-order features can be calculated from images after filtering, as follows.

[0049] a. Shape features: This kind of feature mainly describes the size and shape of the delineated region, which is independent of the image itself.

[0050] Including 16 features such as length, volume, and circularity of three dimensions.

[0051] b. First-order features: This kind of feature summarizes the statistical distribution of pixel values in the delineated region itself, which is independent of the position of the pixel, and does not involve the difference between different pixels. Including 19 features such as pixel mean, maximum, and minimum.

[0052] c. High-order features: This kind of feature describes the heterogeneity of pixel arrangement, which is directly related to the difference of pixel values at different positions. High-order features can be divided into 5 categories, including 75 features based on 5 description matrices of GLCM, GLRLM, GLSZM, GLDM, and NGTDM.

[0053] For medical images, the inventors extracted first and higher order features on the original images and Gaussian-Laplacian filtered images with sigma = 1mm, 3mm, 6mm, and 8 wavelet filtered images (Coill HHH, HHL, HLH, LHH, HLL, LHL, LLH, LLL). Meanwhile, the inventors discretized the image gray levels with 50, 100, 150, 200, 250, 300, 350 levels, respectively. For dose distribution and distance distribution, we only calculated traditional radiomics features on the non-discretized original images.

[0054] For dose radiomics features, based on the above traditional radiomics features definition, the inventors borrowed the description of three-dimensional dose space distribution in historical studies, including the description of scale-invariant moments of weighted barycenter, 3D gradients, and histogram points to reduce the dimension of volume distribution.

[0055] The concept of scale-invariant moments of weighted barycenter can be found in Buettner F, Miah AB, Gulliford SL, et al. Novel approaches to improve the therapeutic index of head and neck radiotherapy: An analysis of data from the PARSPORT randomised phase III trial. Radiother Oncol. 2012;103(1):82-87. doi:10.1016 / j.radonc.2012.02.006, and the concept of 3D gradients can be found in HS, Buettner F, Sterzing F, Hauswald H, Bangert M. Design and Selection of Machine Learning Methods Using Radiomics and Dosiomics for Normal Tissue Complication Probability Modeling of Xerostomia. Front Oncol. 2018;8. doi: 10.3389 / fonc.2018.00035.

[0056] The mathematical definition of the scaling invariant moment of the weighted center of gravity and the three-dimensional gradient is as follows:

[0057] a. Scaling invariant moment: the weighted average value of the three-dimensional volume distribution in x, y, z directions, which describes the center of gravity of the three-dimensional volume distribution in a certain area, and the direction and other key information. The mathematical expression of the translation invariant moment is as follows:

[0058]

[0059] The scaling invariant moment can be obtained by normalizing the translation invariant moment:

[0060]

[0061] The comparison of the scaling invariant moments of high and low doses is shown in FIGS. 3(a) and 3(b). The inventors calculated three three-dimensional gradient values from the x, y, and z directions, and used 0, 1, 2, and 3 orders to calculate the scaling invariant moments. Since the result of the full 0 order is always 1, it is not included in the multi-omics feature group, and finally 63 scaling invariant moment features are obtained.

[0062] FIGS. 3(a) and 3(b) are comparisons of high and low dose scaling invariant moments. FIG. 3(a) shows three images in each row, which are different sections of the dose distribution of a patient in the x-y plane. It can be seen that the dose distribution of the patient with high η 110 is biased to the upper right corner (first row). FIG. 3(b) shows three images in each row, which are different sections of the dose distribution of a patient in the x-z plane. The center of gravity of the dose distribution of the patient with high η 003 is obviously biased to the top.

[0063] b. Three-dimensional gradient: this type of feature calculates the gradient of the three-dimensional volume distribution in x, y, and z directions. The dose gradient calculation formula in x dimension is as follows:

[0064]

[0065] where D(x, y, z) represents the dose value at voxel location (x, y, z), and I(x, y, z) describes whether the voxel is inside the region of interest.

[0066] The geometric features in the prediction model of the present application include distance map and angle features. The distance map describes the nearest distance from any location to the surface of the target volume, which is positive outside the target volume and negative inside the target volume. The distance map is calculated by the SignedMaurerDistanceMapFilter of SimpleITK, whose example is provided by Figure 4 (b) are shown. It can be seen that the points further away from the surface of the primary tumor have larger distance values, and the points inside the primary target volume have negative distance values. In addition to the distance features, the geometric features also include angle features, which are used to describe the angular distribution of the organs at risk to the target volume. In the present application, the angle features include angle histogram points. The angle histogram describes the overlapping volume of the parallel projection of the target volume and the organs at risk at all angles. The angle histogram can be obtained by integrating the sinogram generated by the modified radon transform: the integral of the density projection of the organs at risk in front of the back part of the target volume at different angles. Figure 4 (c) shows the traditional sinogram of the left parotid gland (Parotid_L) and the contour of the primary tumor region sinogram, and the modified radon transform can reflect the front-back position between the target volume and the organs at risk, as shown in Figure 4 (d) shows that the sinogram has a signal only when the left parotid gland is in front of the primary tumor region.

[0067] Figure 4 (a) represents the primary tumor region and the left parotid gland (primary tumor-GTVnp, and Parotid_L), Figure 4 (b) represents the distance map of the primary tumor region and the left parotid gland, Figure 4 (c) VS the traditional sinogram of the left parotid gland (Parotid_L) and the contour of the primary tumor region sinogram, Figure 4 (d) VS the modified radon transform reflects the front-back position between the target volume and the organs at risk, and the sinogram has a signal only when the left parotid gland is in front of the primary tumor region.

[0068] In a third aspect, the technology of the present application includes the estimation or prediction of RTR need. Based on the prediction model of the present application, the RTR need of an individual patient is estimated by fitting the extracted features described above, the model of the present application is pre-trained and developed by using the multi-omic features and clinical parameters of HNC patients in the training dataset of the inventors.

[0069] In one embodiment, the training data for developing the RTR prediction model comes from a group of nasopharyngeal carcinoma patients. The nasopharyngeal carcinoma patients include: (1) patients diagnosed with biopsy-confirmed primary NPC, who do not have distant metastasis and other types of coexisting tumors at the time of consultation; (2) patients who received curative concurrent chemoradiotherapy (CCRT) or CCRT plus adjuvant chemotherapy (AC); and (3) patients who received helical tomotherapy. The patient nasopharyngeal carcinoma does not include: (1) patients who received induction chemotherapy before CCRT treatment; (2) patients who received radiotherapy alone without concurrent chemotherapy; (3) patients who did not receive contrast agent injection to obtain planned contrast-enhanced CT (CECT) images or planned contrast-enhanced T1-w (CET1-w) MR images; or (4) patients who do not have a complete clinical / image dataset.

[0070] As described above, the present application relates to the application of multi-omics in the stratification of high-risk HNC patients for radiotherapy re-planning and the multi-omic feature calculation method. Among them, the development and verification of the internal algorithm for dose-omic and geometric feature extraction is the most time-consuming part of the present application in the research and development process.

[0071] In summary, the present application includes at least the following contribution features: (1) a pre-treatment artificial intelligence (AI) evaluation method for HNC RTR need, which is first used for the same purpose in the art. (2) Based on multi-omics, the present application first integrates the patient-specific radiomics, dose-omics, geometry and clinical data of HNC patients for comprehensive evaluation of RTR need. (3) The present application solves the problem of high efficiency of current clinical RTR practice. The implementation efficiency of current RTR is extremely low, and patient demand screening can only be carried out after treatment begins, and intensive mid-treatment monitoring needs to be carried out for each individual patient. The whole monitoring and risk assessment procedure is very time-consuming and requires the efforts of multiple medical departments. Once a patient is determined to be an RTR candidate, immediate RTR needs to be implemented for that particular patient, which requires the efforts of multiple medical departments from image re-scanning (radiation therapists), organ re-segmentation (medical physicists and oncologists), treatment re-planning (medical physicists) to new plan approval (oncologists). The present application allows pre-treatment evaluation of RTR need for each patient, greatly reducing the clinical burden brought by intensive mid-treatment evaluation, simplifying the allocation of medical resources and manpower for high-risk patients, and possibly achieving higher procedural efficiency in clinical RTR delivery.

[0072] In addition, the present invention addresses the problem of lack of personalized strategy for clinical RTR need screening. Most of the existing studies' potential RTR triggers are related to general criteria, such as tumor stage, patient's body mass index (BMI) before treatment, or threshold of mid-treatment radiation dose received by parotid glands. However, there is a difference in treatment response between patients under the same treatment; patients with the same tumor stage or pre-treatment BMI or whose parotid glands receive more than the threshold of mid-treatment radiation dose can exhibit heterogeneous treatment response, leading to different degrees of anatomical changes in patients throughout the RT process. Therefore, such suggested RTR triggers are not patient-specific nor effective, especially in the era of personalized medicine. The present invention's multi-omic based prediction allows integration of patient-specific predictors from individual radio-imaging, RT treatment planning, geometric relationships between specific organs, and relevant clinical attributes for personalized assessment of RTR need in HNC.

[0073] The present invention can achieve at least the following advantages: allowing pre-treatment risk assessment, greatly reducing clinical burden: the model is developed based on pre-treatment patient information; enabling personalized risk assessment, the present invention's model is developed based on multi-omic feature data for characterizing patient and treatment specific information; cost-effective, the present invention is a computer-based method that does not require a well-designed workstation or other hardware, while it can provide multi-omic based patient-specific RTR need prediction before treatment starts.

[0074] The present invention's test data is available, the present invention's current model is developed based on a small sample size, patient's clinical parameters can be added or included in it, that is, future model training with a larger cohort and including relevant clinical attributes (e.g. pre-treatment weight, tumor stage, etc.) can be completed. By retraining the model by increasing the sample size and including patient's clinical data, this will enhance the predictability and statistical inference of the present invention's model. The present invention can be used for pre-treatment HNC patient risk assessment for RTR need, and optimization of initial treatment plan according to predicted RTR need.

[0075] The present invention only includes calculating the radiotherapy dose and / or coverage and / or 2D / 3D distribution map or contour curve of radiotherapy dose of HNC patients based on "initial pre-treatment radiotherapy plan" before radiotherapy starts, rather than calculating real-time changes of the radiotherapy dose and / or coverage during radiotherapy. The inputs of the present invention are all provided before RT starts, such as pre-treatment medical images and pre-treatment RT plan, based on which the RTR need is estimated. This allows "pre-treatment assessment". Therefore, by using the present invention, there is no in-treatment assessment / calculating dose / coverage / distribution map / contour curve of the tumor.

[0076] The present invention utilizes multi-omic features from multiple organs (bilateral parotid glands, primary tumor, lymph node lesions, etc.) by using multi-modal images (computed tomography (CT) and multi-parametric magnetic resonance (MR) images) for patients with advanced nasopharyngeal carcinoma. The present invention constructs a nomogram integrating various predictors to assess the risk probability for each patient.

[0077] While the present invention has been illustrated in detail in connection with limited embodiments, it is to be understood that the invention is not limited to the disclosed embodiments. Other substitutions, alterations, and changes in the number of parts, and arrangements of parts, which will occur to one with ordinary skill in the art, are also intended to fall within the scope of the present invention.

Claims

1. A method for multi-omics processing of radiotherapy patient data from head and neck cancer patients, wherein multi-omics is applied to a pre-treatment assessment of radiotherapy replanning needs in said head and neck cancer patients, comprising the following steps: (1) Preprocessing of radiotherapy patient data of head and neck cancer patients to form input image / dose data of the head and neck cancer patients, ensuring that the characteristics of the input image / dose data of the head and neck cancer patients during radiotherapy are consistent with the characteristics of the image / dose data used to construct a predictive model for radiotherapy replanning; (2) Extracting multi-omics features from preprocessed radiotherapy patient data; (3) Based on the prediction model, the radiotherapy replanning needs of the head and neck cancer patients are estimated or predicted by fitting the extracted multi-omics features; The multi-omics features are extracted through calculation, and the data of the multi-omics features include radiomics feature data, dosimics feature data, and geometric feature data. The geometric feature data represents the geometric or positional relationship between the patient's organs and tumors using distance and angle.

2. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the radiotherapy patient data includes radiotherapy data and the patient's clinical characteristic data, and the characteristics of the patient's input image / dose data include image resolution, signal intensity variability, and uniformity; in, The radiotherapy data is in the format of medical digital imaging and communication files, and includes multimodal radiographic image data, radiotherapy dose data, and organ contour data; as well as, The patient's clinical characteristics data include patient demographic data or tumor characteristic data.

3. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 2, wherein the multimodal radiographic data includes data from computed tomography (CT) scans and multiparametric magnetic resonance imaging (MRI).

4. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the preprocessing step of the radiotherapy patient data for head and neck cancer patients includes: Voxel-scale resampling was applied to all images, masks, and doses of the radiotherapy patient data to maintain a consistent resolution of 1mm x 1mm x 1mm.

5. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 4, wherein the preprocessing step of the radiotherapy patient data for head and neck cancer patients includes: Deviation correction is performed using the N4 deviation field correction image filter function of SimpleITK software to reduce voxel intensity deviation in the MRI images of the radiotherapy patient data caused by inhomogeneous magnetic fields; and the MRI images are normalized by performing a moving average of the voxel intensities and rescaling the standard deviation of the voxel intensities.

6. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 4, wherein the preprocessing step of the radiotherapy patient data for head and neck cancer patients further includes: Before extracting the radiomics features, all contrast-enhanced computed tomography (CT) images and MRI images were resampled to 1 x 1 x 1 mm. 3 The voxel size was adjusted to mitigate the impact of differences in image acquisition parameters between different patients. The grayscale intensity of the contrast-enhanced computed tomography (CT) image and MRI image is discretized into a fixed bucket count range, which is from 50 to 350, with an increment interval of 50. All contrast-enhanced computed tomography and MRI images were convolved with a Laplacian Gaussian filter at three Gaussian radius parameter levels to generate filtered images, which highlight specific texture radiometric features at multiple scales, ranging from 1 mm fine, 3 mm medium to 6 mm coarse. Wavelet filters are also applied to both the contrast-enhanced computed tomography (CT) images and MRI images to produce radiographic features with multiple resolutions. as well as The primary tumor and cervical lymph node lesions on the contrast-enhanced computed tomography images were resegmented to limit the Henle units to a range of -150 to 180 in order to eliminate non-soft tissue components within the volume of interest of the head and neck cancer patients.

7. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the prediction model is developed based on multi-omics feature data, the multi-omics feature data being used to characterize information about a specific patient and the specific treatment they receive, the multi-omics feature data including multiple multi-omics-based predictive factors extracted from multiple organ structures using pre-treatment images and clinical parameters of multimodal patients.

8. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 7, wherein the predictive model is developed through the following steps: dividing the patient cohort into training and test datasets in a 7:3 ratio via 20 iterations; applying a supervised feature selection algorithm only to the training dataset in each iteration to maintain the clinical relevance of residual features; and then applying an unsupervised feature selection algorithm to remove highly redundant features, thereby obtaining a reduced feature set; wherein, The prediction model was developed using a ridge regression algorithm with 10x cross-validation on the training set.

9. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 7, wherein the final predicted probability of the pre-treatment assessment of the radiotherapy replanning needs of the head and neck cancer patients using the prediction model includes: Radiometric score, dose score, geometric score, and / or predicted clinical parameters, wherein the radiometric score is a linear combination of radiomics features and their coefficients, the dose score is a linear combination of dosemics features and their coefficients, and the geometric score is a linear combination of geometric features and their coefficients.

10. The method of multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the radiomics feature data is used to characterize the intrinsic tissue biological response to treatment perturbation, and the radiomics feature data is extracted from multimodal images within the patient's organ contours.

11. The method of multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the dosimics feature data characterizes the aggressiveness of an individual radiotherapy plan by providing dose information from local spatial patterns of radiotherapy dose distribution within conventional dose-volume histograms and organ contours, the dosimics features being extracted from the individual patient's radiotherapy plan.

12. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the radiomics feature data, dosemics feature data, and geometric feature data are extracted from three-dimensional medical images, dose distribution maps, and distance distribution maps, respectively.

13. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 12, wherein the radiomics features include shape features, first-order features, and higher-order features. in, The shape features describe the size and shape of the outlined area, including 16 features comprising three dimensions: length, volume, and roundness. The first-order features represent the statistical distribution of pixel values ​​within the delineated area, including 19 features encompassing the pixel average, maximum, and minimum values; and The higher-order features represent the heterogeneity of pixel arrangement and are directly related to the difference in pixel values ​​at different positions. The higher-order features are divided into 5 categories, including a total of 75 features based on 5 description matrices: GLCM, GLRLM, GLSZM, GLDM, and NGTDM.

14. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 13, wherein for medical images, first-order and higher-order features are extracted from the original image and Gaussian-Laplace filtered images with Sigma = 1mm, 3mm, and 6mm, as well as wavelet filtered images containing Coil1HHH, HHL, HLH, LHH, HLL, LHL, LLH, and LLL; and the image grayscale is discretized using 50, 100, 150, 200, 250, 300, and 350 levels, respectively; and For dose distribution and distance distribution, the radiomics features are calculated only on the undiscrete original image.

15. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the dosimics features include scaling invariant moments, three-dimensional gradients, and histogram points for describing weighted centering to reduce the dimensionality of the volume distribution; The mathematical definitions of the scaling invariant moment of the weighted center of gravity and the three-dimensional gradient are as follows: (1) Scaling invariant moment: The weighted average of the three-dimensional volume distribution in the x, y, and z directions, used to describe key information including the centroid and orientation of the three-dimensional volume distribution in a specific region; The scaling invariant moments are obtained by normalizing the translation invariant moments: Where p, q, and r are the user-specified exponents in the x, y, and z directions, respectively. Translation invariant moment μ pqr The mathematical expression is: (2) Three-dimensional gradient: This type of feature calculates the gradient of the three-dimensional volume distribution in the x, y, and z directions. The formula for calculating the dose gradient in the x direction is: gradient Where D(x,y,z) represents the dose value at voxel position (x,y,z), and I(x,y,z) describes whether the voxel is located within the region of interest.

16. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the distance distribution map of the geometric feature data describes the nearest distance from any location to the surface of the target area, with distances outside the target area being positive and distances inside the target area being negative; the calculation process of the distance distribution map is implemented by the SignedMaurerDistanceMapFilter of SimpleITK software.

17. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 16, wherein the geometric feature data further includes angular features, the angular features being used to describe the angular distribution of the tumor organ to the target area, wherein, The angular features include the points of the angular histogram and describe the overlap volume between the parallel projection of the target area and the tumor organ at all angles.

18. The method of multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 17, wherein the angle histogram is obtained by integrating a sine graph generated by a modified Radon transform: the integral of the density projection of the tumor organ at different angles in the anterior part of the back of the target area, the modified Radon transform reflecting the anteroposterior orientation between the target area and the tumor organ.

19. The method of multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 1, wherein the predictive model is pre-trained and developed using multi-omics features and clinical parameters of the head and neck cancer patients, wherein the need for radiotherapy replanning of the head and neck cancer patients is comprehensively predicted or pre-treatment assessed by integrating patient-specific radiomics features, dosimics features, geometric features, and clinical data of the head and neck cancer patients.

20. The method of multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 19, wherein the predictive model performs multi-omics-based predictions by integrating patient-specific predictors from individual radiographic images, radiotherapy treatment plans, geometric relationships between target organs, and relevant clinical attributes of the head and neck cancer patients to provide personalized predictions or pre-treatment assessments of the need for radiotherapy replanning in the head and neck cancer patients.

21. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 19, wherein the inputs to the prediction model include multimodal images, multiple regions of interest, and initial planning dose distributions from radiotherapy planning data, the inputs to the prediction model further include clinical prediction parameters, and the output of the prediction model is the RTR prediction probability.

22. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 21, wherein the intermediate steps of the prediction model include multi-omics feature extraction, feature selection and linear modeling, and calibration and score overlay, wherein, Multi-omics feature extraction includes preprocessing and feature computation. Radiomic features are calculated from multi-omics images and multiple regions of interest. Dosimic features are calculated from the initial planned dose distribution and multiple regions of interest. Geometric features are calculated only from multiple regions of interest. The prediction model performs feature screening and linear modeling on radiomic features, dosimic features, and geometric features to obtain radiation scores, dose scores, and geometric scores, respectively. The radiation scores, dose scores, geometric scores, and clinical parameters are calibrated and superimposed to obtain the final radiotherapy replanning probability. The calibration and superimposition process is implemented through a nomogram. Each bar in the nomogram represents a calibration model for each score and clinical parameter in the radiation score, dose score, and geometric score. The calibrated scores are summed and calibrated again to obtain the final radiotherapy replanning probability.

23. The method for multi-omics processing of radiotherapy data from head and neck cancer patients according to claim 22, wherein the feature selection and linear modeling include generating an equation for each of the radiomics model, the dosimics model, and the geometric model, wherein the equations represent linear combinations of all selected predictor variables in the models: the radiometric score of the radiomics model, the dose score of the dosimics model, and the geometric score of the geometric model, respectively including: Radiometric score = intercept r + R1(r1) + R2(r2) + R3(r3) + ... + R i (r i ) Dose score = intercept d + D1(d1) + D2(d2) + D3(d3) + ... + D j (d j ) Geometric score = intercept g + G1(g1) + G2(g2) + G3(g3) + ... + G k (g k ) Wherein, intercept r, intercept d, and intercept g are derived from the prediction model using the Ridge algorithm for radiomics, dosimics, and geometry, respectively. R, D, G and r, d, g are the prediction variables and their coefficients in the radiomics, dosimics, and geometry models, respectively, and i, j, k represent the final number of radiological, dosimetric, and geometric features, respectively.

24. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the steps of the method as described in any one of claims 1 to 23.

25. A computing device comprising a processor and a readable storage medium on which a program is stored, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 23.

Images