AI-based system and method for dynamic compensation of thoracoabdominal tumor target regions

Abstract

This invention provides an AI-based system and method for dynamic compensation of thoracoabdominal tumor target regions. [Solution] The present invention screens some candidate historical treatment samples as historical treatment samples based on the similarity of BMI values ​​between samples, determines a position prediction accuracy index based on the CT image information of each historical treatment sample, determines the similarity of tumor location based on the CT image information of the current treatment sample and each historical treatment sample, obtains the similarity of the radiotherapy process by fusing the position prediction accuracy index and the similarity of tumor location, determines the tumor location at the future irradiation time of each historical treatment sample, and performs a weighted addition process on the tumor location using the similarity of the radiotherapy process to obtain the tumor location at the future irradiation time of the current treatment sample. The present invention can predict the tumor location at the future irradiation time of the current treatment sample and trigger the irradiation device in advance to perform accurate radiotherapy.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to a system and method for dynamic compensation of target regions of thoracic and abdominal tumors based on AI (Artificial Intelligence).

Background Art

[0002] Thoracic and abdominal tumors refer to tumors that occur in the thoracic cavity and abdominal cavity. For the radiotherapy of thoracic and abdominal tumors, respiratory motion is one of the important factors affecting the treatment accuracy. The reason is that respiration periodically causes displacement of tumors and surrounding organs. If the tumor position is not effectively managed, it may cause irradiation leakage of the target region, over-irradiation of normal tissues, and distortion of the dose distribution during the treatment process, and further affect the treatment effect and the safety of patients.

[0003] In order to suppress the influence of respiratory motion, dynamic compensation of the tumor target region is performed using respiratory gating technology. The core principle is to control the tumor irradiation timing during radiotherapy. Specifically, by monitoring the patient's respiratory cycle and restricting the treatment to a specific respiratory phase for irradiation, the displacement of the tumor in other respiratory phases other than the specific respiratory phase is reduced, and the treatment accuracy is ensured.

[0004] However, the conventional respiratory gating technology that restricts the treatment to a specific respiratory phase for irradiation may extend the treatment time. Extending the treatment time means that the patient is exposed to radiation for a longer time during treatment. Long-term or excessive radiation exposure may increase the risk of damage to normal tissues and ultimately increase the probability of secondary cancer occurrence. In addition, in the process of realizing the respiratory gating technology, the correlation between the external respiratory signal and the movement trajectory of the tumor position is low. Specifically, the shape and volume of the tumor may change in different respiratory cycles, especially when the patient coughs or moves the body. The conventional respiratory gating technology is difficult to adapt to these changes and may affect the treatment accuracy.

Summary of the Invention

[0005] To address the problem of excessively long treatment times for the conventional dynamic compensation of thoracoabdominal tumor target regions, the object of the present invention is to provide an AI-based system and method for dynamic compensation of thoracoabdominal tumor target regions. The specific technical proposal to be adopted is as follows.

[0006] One embodiment of the present invention provides an AI-based method for dynamic compensation of thoracoabdominal tumor target regions. The steps include obtaining a time-series of CT images and BMI values ​​for the current treatment sample and several candidate historical treatment samples, A step of screening historical treatment samples from all candidate historical treatment samples based on the similarity of BMI values ​​between the current treatment sample and each candidate's historical treatment sample, The steps include determining a position prediction accuracy index corresponding to each historical treatment sample based on the CT image time series of each historical treatment sample, The steps include determining the similarity of tumor location between the current treatment sample and each historical treatment sample based on the CT image time series of the current treatment sample and the CT image time series of each historical treatment sample, A step of fusing the position prediction accuracy index and the similarity of tumor location to determine the similarity of the radiotherapy process for each historical treatment sample, The steps include obtaining the tumor position at a future irradiation time for each treatment sample based on the CT image time series of each treatment sample, where the future irradiation time is the sum of the current imaging time and the irradiation time interval of the irradiation device, The procedure includes the step of using the similarity of the aforementioned radiotherapy processes to perform a weighted addition process on the tumor position at future irradiation times for each of the aforementioned historical treatment samples to obtain the tumor position at future irradiation times for the current treatment sample, and performing dynamically compensating targeted radiotherapy.

[0007] Furthermore, the step of screening historical treatment samples from all candidate historical treatment samples based on the similarity of BMI values ​​between the current treatment sample and each candidate historical treatment sample is as follows: For each candidate's historical treatment sample, calculate the absolute difference between the BMI value of the current treatment sample and the BMI value of the candidate's historical treatment sample. The absolute value of the difference is subjected to a negative correlation normalization process to obtain a first normalized value, and this first normalized value is determined as the degree to which the respiratory patterns of the candidate historical treatment sample can be compared to the current treatment sample. This includes setting a degree threshold and obtaining a historical treatment sample by comparing the comparability of each respiratory pattern with the degree threshold.

[0008] Furthermore, the step of determining the position prediction accuracy index corresponding to each historical treatment sample based on the CT image time series of each historical treatment sample is as follows: Based on the CT image time series of each historical treatment sample, the central position of the tumor at each historical imaging time is determined and recorded as the actual tumor position, and the central position of the tumor at each historical imaging time is predicted and recorded as the pseudotumor position. Calculate the positional difference between the actual tumor location and the false tumor location at the same historical imaging time, This includes fusing and analyzing the positional difference values ​​of all historical imaging times corresponding to the same historical treatment sample to determine the positional prediction accuracy index for each historical treatment sample, Here, the position difference value shows a negative correlation with the position prediction accuracy index.

[0009] Furthermore, the step of fusing and analyzing the positional difference values ​​of all historical imaging times corresponding to the same historical treatment sample to determine the positional prediction accuracy index corresponding to each historical treatment sample is as follows: For each historical shooting time, a negative correlation normalization process is performed on the position difference value to obtain a second normalized value, and the second normalized value is determined as the prediction criterion corresponding to the historical shooting time. This includes performing a power calculation on the prediction criterion for all historical imaging times corresponding to the same historical treatment sample, and determining the obtained first power value as a position prediction accuracy index corresponding to the corresponding historical treatment sample.

[0010] Furthermore, the step of determining the similarity of tumor location in each historical treatment sample to the current treatment sample based on the CT image time series of the current treatment sample and the CT image time series of each historical treatment sample is as follows: Using the time-series CT images of the current treatment sample as input data, the predicted tumor location of the CT image at the next time point is obtained, For each historical treatment sample, a CT image from the CT image time series of the historical treatment sample is selected at the target historical acquisition time, the location of the target historical tumor in the selected CT image is determined, and the image number at the target historical acquisition time is the same as the CT image number at the next time. The method includes determining the degree of similarity of the tumor location of each historical treatment sample to the current treatment sample based on the positional similarity between the predicted tumor location and each of the target historical tumor locations, and between the current tumor location and the historical tumor location of the same image number, and confirming that the current tumor location is the tumor location in the CT image of the current treatment sample.

[0011] Furthermore, the step of determining the similarity of the tumor location of each historical treatment sample to the current treatment sample, based on the positional similarity between the predicted tumor location and each of the target historical tumor locations, and between the current tumor location and the historical tumor location of the same image number, is as follows: For any historically treated sample, calculate the distance between the predicted tumor location and the historical tumor location of that historically treated sample, and calculate the distance between the current tumor location and the historical tumor location for the same image number. This includes obtaining the second power of all distance values ​​corresponding to the historical treatment sample, performing negative correlation normalization on all second powers of distance values ​​to obtain a third normalized value, and using the third normalized value as the similarity of tumor location.

[0012] Furthermore, the step of fusing the position prediction accuracy index and the similarity of the tumor location to determine the similarity of the radiotherapy process for each historical treatment sample is: For each historical treatment sample, the method includes multiplying the position prediction accuracy index of the historical treatment sample by the similarity of the tumor position to obtain the product, and using the product as the similarity of the radiotherapy process of the corresponding historical treatment sample.

[0013] Furthermore, the step of obtaining the tumor position at a future irradiation time for each historical treatment sample based on the CT image time series of each historical treatment sample is: For each historically treated sample, the location of the tumor at each historical imaging time is obtained based on the time series of CT images of the historically treated sample. Based on the time of each historical imaging and the historical tumor position at each historical imaging time, a position acquisition equation is constructed, This includes using the aforementioned future irradiation time as the independent variable in the position acquisition equation, and obtaining the tumor position at the future irradiation time of the corresponding historical treatment sample.

[0014] Furthermore, the step of using the similarity of the radiotherapy process to perform a weighted addition process on the tumor position at future irradiation times for each of the historical treatment samples to obtain the tumor position at future irradiation times for the current treatment sample is: For each historical treatment sample, the ratio of the similarity of the radiotherapy process of the historical treatment sample to the cumulative similarity of all radiotherapy processes is used as the position weight for the corresponding historical treatment sample. The product of the positional weight of each historically treated sample and the tumor position at the future irradiation time of the corresponding historically treated sample is calculated as the tumor subposition at the future irradiation time of the current treatment sample. This includes performing a fusion process on all tumor sub-locations to obtain a fused location, and setting the fused location as the tumor location at a future irradiation time for the current treatment sample.

[0015] Another embodiment of the present invention provides an AI-based system for dynamic compensation of thoracoabdominal tumor target regions, comprising a processor and memory, wherein the processor processes instructions stored in the memory to realize an AI-based method for dynamic compensation of thoracoabdominal tumor target regions.

Advantages of the Invention

[0016] The present invention has the following beneficial effects. Compared with the excessive radiation caused by the extended treatment time of the conventional respiratory gating technology and the irradiation deviation caused by the lack of correlation between the external signal and tumor movement, the present invention utilizes the CT image time series of the historical treatment samples obtained during the target treatment, and analyzes the similarity of the treatment process between it and the current treatment sample, so as to predict the tumor position at the next irradiation time of the current treatment sample, trigger the irradiation equipment in advance, maintain the irradiation equipment in the irradiation state with the highest possible frequency, and shorten the treatment time. In addition, the present invention breaks through the limitation that the conventional gating technology depends only on the patient's respiratory signal through historical population data-driven and real-time dynamic fusion, effectively overcomes the tumor displacement problem caused by respiratory periodicity, and realizes the target treatment of dynamic compensation.

[0017] First, based on the similarity of the BMI values between the current treatment sample and each candidate historical treatment sample, the historical treatment samples are screened from all candidate historical treatment samples. Screening the historical treatment samples is because there are differences in the respiratory patterns between treatment samples with significantly different body types. The screened historical treatment samples have the possibility of comparing the respiratory pattern with the current treatment sample, which is beneficial for more accurate subsequent tumor position prediction, reduces the analysis of unnecessary historical treatment samples, and is also beneficial for saving the time of computational analysis.

[0018] Next, the similarity of the radiotherapy process of each historical treatment sample to the current treatment sample is analyzed from two perspectives: the similarity of the position prediction accuracy index and the similarity of the tumor position. The position prediction accuracy index can represent the accuracy of the tumor position prediction of the historical treatment sample itself, while the similarity of the tumor position represents the similarity of the tumor position of the historical treatment sample to the current treatment sample. A high numerical accuracy of the similarity of the radiotherapy process determined by analysis from the two perspectives is beneficial for accurately analyzing the importance of each historical treatment sample during prediction analysis, and improving the numerical accuracy of the tumor position at the future irradiation time of the currently determined current treatment sample.

[0019] Finally, obtaining the tumor position of each historical treatment sample at the future irradiation time and using this as reference data for determining the tumor position of the current treatment sample at the future irradiation time, and performing weighted summation processing on the tumor positions of each historical treatment sample at the future irradiation time using the similarity of the radiotherapy process contribute to obtaining a tumor position at the future irradiation time of the current treatment sample with higher accuracy and reliability, which facilitates dynamic compensation target radiotherapy.

Brief Description of the Drawings

[0020] To more clearly explain the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for explaining the embodiments or the prior art are briefly described below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those skilled in the art can also obtain other drawings based on these drawings without creative labor.

[0021] [Figure 1] It is a flowchart of the implementation of the dynamic compensation method for the target area of chest and abdominal tumors based on AI according to an embodiment of the present invention. [Figure 2] It is a flowchart of the implementation of step S3 in the embodiment of the present invention. [Figure 3] It is a flowchart of the implementation of step S4 in the embodiment of the present invention. [Modes for carrying out the invention]

[0022] To further illustrate the technical means and effects employed by the present invention to achieve a predetermined inventive objective, specific embodiments, structures, features, and effects of the technical proposal according to the present invention will be described in detail below with reference to the drawings and preferred embodiments. In the following description, different “one embodiment” or “another embodiment” are not necessarily the same embodiment. Furthermore, specific features, structures, or features in one or more embodiments may be combined in any suitable form.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention.

[0024] The application scenarios for the present invention may be as follows: While respiratory gating technology can improve the accuracy of radiation therapy, extending treatment time means that patients are exposed to radiation for a longer period during treatment. If treatment time is too long, especially with high radiation doses, healthy tissues (e.g., lungs and heart) may receive unnecessary radiation, potentially causing side effects such as pulmonary fibrosis and cardiac problems. Furthermore, in the development of respiratory gating technology, the correlation between respiratory signals and the trajectory of tumor location is low, which is likely to lead to irradiation deviations and affect treatment effectiveness.

[0025] To overcome the shortcomings of conventional respiratory gating techniques applied to radiotherapy, one embodiment of the present invention provides an AI-based dynamic compensation method for thoracoabdominal tumor target regions. As shown in Figure 1, the method includes the following steps.

[0026] As S1, time-series CT images and BMI values ​​are obtained for the current treatment sample and several candidate historical treatment samples.

[0027] Here, a CT image time series refers to a set of images obtained in chronological order by scanning the same examination area multiple times consecutively using CT (Computed Tomography), possessing clear time-series attributes, and having a fixed time interval between adjacent CT images. BMI (Body Mass Index) refers to the patient's height and weight. Current treatment samples are patients with thoracic and abdominal tumors currently undergoing radiation therapy, while candidate historical treatment samples are patients with thoracic and abdominal tumors who have completed radiation therapy in the past.

[0028] To analyze the tumor trajectory during targeted therapy in patients with thoracic and abdominal tumors and determine whether pre-irradiation is necessary at future irradiation times, it is necessary to obtain CT image time series. The CT image time series of the current treatment sample can be used as input data to realize prediction, but since historical data related to the current treatment sample may be missing in the prediction process, supplementary analysis should be performed using CT image time series of candidate historical treatment samples.

[0029] First, we obtain time-series CT images of the current treatment sample and several candidate historical treatment samples.

[0030] In this embodiment, the patient acquires a time series of CT images using an imaging device during the radiation therapy process. The time series of CT images of the current treatment sample consists of CT images at N acquisition times, and the time series of CT images of a candidate historical treatment sample consists of CT images at M historical acquisition times, where N and M are both positive integers and M is greater than N. Here, the number of candidate historical treatment samples can be set by the implementer according to the specific actual situation and is not specifically limited here.

[0031] Here, each CT image in the CT image time series has an image number sorted according to the acquisition time. Since the acquisition frequency of CT images for the current treatment sample and different candidate historical treatment samples is the same, the number of CT images corresponding to different candidate historical treatment samples is the same. The CT images collected by the current treatment sample at each acquisition time have corresponding CT images for each of the different candidate historical treatment samples.

[0032] Furthermore, based on the CT image time series, tumor motion trajectory data sequences are obtained for the current treatment sample and multiple candidate historical treatment samples.

[0033] In this embodiment, for the current treatment sample and each candidate's historical treatment sample, the tumor center position in each CT image in the CT image time series is determined as the tumor position. The tumor position determined by the first CT image is set as the coordinate origin, i.e., (0,0), and a position coordinate system is constructed. All tumor positions corresponding to the same patient are associated with this position coordinate system, and all tumor positions are connected to obtain the tumor motion trajectory.

[0034] The tumor motion trajectory data sequences for current treatment samples are shown in Table 1 below. JPEG0007852964000002.jpg19170

[0035] The tumor kinetic trajectory data sequences for each candidate's historical treatment samples are shown in Table 2. JPEG0007852964000003.jpg43170

[0036] Here, the horizontal axis of the position coordinate system represents the horizontal coordinate of the tumor position, and the vertical axis represents the vertical coordinate of the tumor position. The tumor motion trajectory data sequence consists of tumor positions at different imaging times for the same patient. Each CT image has a corresponding tumor position, and the process for determining the tumor position is prior art and falls outside the scope of the present invention; therefore, a detailed explanation is omitted here.

[0037] Furthermore, in the subsequent step of determining the tumor location, it is not necessary to repeat the calculation and analysis of the tumor location; the corresponding data can be obtained directly at this stage.

[0038] Next, obtain the BMI values ​​for the current treatment sample and the historical treatment sample for each candidate.

[0039] In the course of radiation therapy for thoracic and abdominal tumors, obese patients may find abdominal fat compressing the diaphragm, making diaphragmatic breathing difficult. In this case, the patient's abdomen and chest may also be more difficult to breathe effectively due to the effects of anesthesia and pain, resulting in significant changes in breathing patterns. Lean patients may have lower lung capacity, leading to significant changes in breathing patterns, and these changes differ between obese and lean patients. Therefore, to analyze the similarity of breathing patterns between candidate historical treatment samples and current treatment samples, it is necessary to obtain BMI values ​​for both current treatment samples and each candidate historical treatment sample.

[0040] Here, the BMI value can be obtained from height and weight, and the method for obtaining the BMI value is prior art and is not within the scope of protection of the present invention, so a detailed explanation is omitted here.

[0041] Up to this point, in this embodiment, the time-series CT images and BMI values ​​of the current treatment sample and the historical treatment sample of each candidate were obtained during the data acquisition phase.

[0042] As S2, historical treatment samples are screened from all candidate historical treatment samples based on the similarity of BMI values ​​between the current treatment sample and the historical treatment sample for each candidate.

[0043] Here, BMI similarity refers to the confidence level with which a candidate's historical treatment sample, analyzed from the perspective of breathing patterns, can be compared to the current treatment sample. The more similar the breathing patterns, the higher the confidence level with which the candidate's historical treatment sample can be compared.

[0044] If the difference in body shape between a candidate's historical treatment sample and the current treatment sample is too large, even if the movement trajectories of the tumor location in both samples show high agreement, there will still be deviations in the treatment process of the current treatment sample, and the respiratory trajectory at this time cannot be directly compared and corrected. Therefore, it is necessary to either screen for a subset of candidate historical treatment samples with similar respiratory patterns from all candidate historical treatment samples, or to determine whether the similarity of respiratory patterns between each candidate's historical treatment sample and the current treatment sample meets a predetermined requirement, and then screen for candidate historical treatment samples that meet the requirement.

[0045] In an exemplary embodiment, step S2 can be achieved by the following steps.

[0046] Step 1 involves calculating the absolute difference between the BMI value of the current treatment sample and the BMI value of the candidate's historical treatment sample for each candidate's historical treatment sample.

[0047] In step 2, a negative correlation normalization process is performed on the absolute value of the difference to obtain a first normalized value, and this first normalized value is determined as the degree to which the candidate historical treatment sample can be compared to the current treatment sample in terms of respiratory pattern comparability.

[0048] For example, the formula for calculating the degree of comparability of the respiratory patterns of candidate a-th historical treatment sample to the current treatment sample is as follows: JPEG0007852964000004.jpg6170

[0049] JPEG0007852964000005.jpg37170

[0050] JPEG0007852964000006.jpg36170

[0051] By referring to the comparability of the respiratory patterns of candidate a's historical treatment sample to the current treatment sample, we can obtain the comparability of the respiratory patterns of each candidate's historical treatment sample to the current treatment sample.

[0052] Step 3 involves setting a severity threshold and obtaining historical treatment samples by comparing the comparability of each respiratory pattern with the severity threshold.

[0053] In this embodiment, the degree threshold is set to 0.7, and candidate historical treatment samples whose respiratory pattern comparability is equal to or greater than the degree threshold are selected as historical treatment samples for subsequent data analysis, thereby overcoming to some extent the impact of reduced respiratory pattern comparability. Here, the degree threshold can be set by the implementer according to the specific actual situation and is not specifically limited here.

[0054] Furthermore, when analyzing the comparability of respiratory patterns, it is not possible to compare all candidate historical treatment samples with the current treatment sample. Instead, by screening candidate historical treatment samples that have small differences in body type and a degree of comparability in respiratory patterns, the numerical reliability of the similarity of the radiotherapy process can be improved, and the subsequent computational analysis of less relevant candidate historical treatment samples can be reduced, thereby improving computational analysis efficiency.

[0055] To date, this embodiment screens all candidate historical treatment samples for historical treatment samples whose respiratory patterns are similar to those of the current treatment sample.

[0056] As S3, the position prediction accuracy index corresponding to each historical treatment sample is determined based on the CT image time series of each historical treatment sample.

[0057] Here, the position prediction accuracy index refers to the degree of similarity between the predicted tumor position and the actual tumor position at each historical imaging time for a historical treatment sample. The more similar the predicted tumor position and the actual tumor position are, the more accurate the tumor position prediction result for that historical treatment sample is, and the more important it is for analyzing the current treatment sample.

[0058] As an exemplary embodiment, step S3 can be achieved by steps S31 to S33 shown in Figure 2.

[0059] As S31, the central position of the tumor at each historical imaging time is determined based on the CT image time series of each historical treatment sample and recorded as the actual tumor position, and the central position of the tumor at each historical imaging time is predicted and recorded as the pseudotumor position.

[0060] In this embodiment, since only predictable historical imaging times can be analyzed, each historical imaging time has a corresponding actual tumor location and a pseudo-tumor location.

[0061] Predictive analysis can be performed on the motion trajectory data at each historical imaging time during the radiotherapy process of a historical treatment sample using an LSTM (Long Short-Term Memory) neural network. When the deviation between the predicted motion trajectory data and the actual motion trajectory data at the same historical imaging time is small, that is, when the changes in both the x and y axes of the predicted and actual values ​​are small, it indicates that the motion trajectory data corresponding to the historical imaging time is predictable under normal circumstances. Here, the process by which the LSTM neural network achieves prediction is prior art and is not within the scope of protection of the present invention, so a detailed explanation is omitted here.

[0062] In step S32, the positional difference value between the actual tumor location and the false tumor location at the same historical imaging time is calculated.

[0063] In this embodiment, the positional difference value refers to the magnitude of the distance between two positional points. The greater the distance between the two positional points, the larger the positional difference value between the corresponding actual tumor location and the pseudo-tumor location. Conversely, a small distance indicates a small positional difference value between the corresponding actual tumor location and the pseudo-tumor location.

[0064] In step S33, the positional difference values ​​for each historical imaging time corresponding to the same historical treatment sample are fused and analyzed to determine the positional prediction accuracy index corresponding to each historical treatment sample.

[0065] Here, the position difference value and the position prediction accuracy index show a negative correlation; that is, the larger the position difference value, the smaller the position prediction accuracy index.

[0066] Specifically, for each historical imaging time, a negative correlation normalization process is performed on the position difference value to obtain a second normalized value, and this second normalized value is determined as the prediction criterion level corresponding to the historical imaging time. A power calculation is performed on the prediction criterion levels of all historical imaging times corresponding to the same historical treatment sample to obtain a first power value, and this first power value is determined as the position prediction accuracy index corresponding to the corresponding historical treatment sample.

[0067] Furthermore, since respiratory movement is a continuous process, high position prediction accuracy at a single historical imaging time does not accurately represent the position prediction situation for the entire treatment process. Therefore, in order to determine a position prediction accuracy index corresponding to a historical treatment sample, it is necessary to perform a fusion analysis of the position difference values ​​from all historical imaging times corresponding to the same historical treatment sample.

[0068] As an example, the formula for calculating the position prediction accuracy index corresponding to the j-th historical treatment sample is as follows: JPEG0007852964000007.jpg9170

[0069] JPEG0007852964000008.jpg71170

[0070] JPEG0007852964000009.jpg13170JPEG0007852964000010.jpg18170

[0071] In the formula for calculating the position prediction accuracy index, the degree of prediction criterion for each historical imaging time is first calculated. The degree of prediction criterion is determined by the similarity between the actual tumor position and the pseudo-tumor position in the x and y axes at the same historical imaging time. The greater the degree of position similarity, the higher the degree of prediction criterion for the corresponding historical imaging time.

[0072] Furthermore, since the false tumor location at a given historical imaging time is predicted from the actual tumor location prior to that historical imaging time, there is no false tumor location for the first historical imaging time, and therefore it will not be included in the analysis.

[0073] Refer to the calculation process for the position prediction accuracy index corresponding to the j-th historical treatment sample above to obtain the position prediction accuracy index corresponding to each historical treatment sample.

[0074] To date, this embodiment has obtained a position prediction accuracy index corresponding to each historical treatment sample.

[0075] As S4, the similarity of tumor location between the current treatment sample and each historical treatment sample is determined based on the CT image time series of the current treatment sample and the CT image time series of each historical treatment sample.

[0076] Here, tumor location similarity refers to the degree of similarity between the tumor location movement trajectories of a historically treated sample and a currently treated sample. The greater the tumor location similarity, the smaller the distances on the x and y axes between the two tumor location movement trajectories.

[0077] As an exemplary embodiment, step S4 can be achieved by steps S41 to S43 shown in Figure 3.

[0078] As S41, the time series of CT images of the current treatment sample is used as input data to obtain the predicted tumor position of the CT image at the next time step.

[0079] In this embodiment, the CT image time series of the current treatment sample was collected during the current scanning time period, and the next time is the next scanning time in the current scanning time period, i.e., a future scanning time, obtained based on the scanning frequency.

[0080] Specifically, the current tumor location corresponding to the CT image time series of the current treatment sample is used as input data for an LSTM neural network to predict the tumor location at the next time step for the current treatment sample, and this is recorded as the predicted tumor location. Here, the current tumor location is the tumor location in the CT image of the current treatment sample.

[0081] For example, if the last acquisition time in the CT image time series of the current treatment sample is the Nth acquisition time, the next acquisition time is the N+1th acquisition time, and the time interval between the Nth acquisition time and the N+1th acquisition time is still the time interval when the CT images are acquired in step S1 above. Here, N is a positive integer and also represents the image number.

[0082] As S42, for each historical treatment sample, a CT image at the target historical acquisition time is selected from the CT image time series of the historical treatment sample, and the target historical tumor location of the selected CT image is determined.

[0083] In this embodiment, the target history acquisition time is the same as the image number at the next time; that is, the target history tumor location corresponding to the image number of the CT image at the N+1th history acquisition time is selected from the CT image time series of each history treatment sample, and the target history tumor location is used for comparative analysis with the predicted tumor location.

[0084] Furthermore, the predicted tumor location and the target historical tumor location belong to the same point on the motion trajectory of the corresponding tumor location. To improve the reliability of the similarity analysis of the motion trajectories of tumor locations between the current treatment sample and the historical treatment sample, it is also necessary to analyze the positional similarity between the predicted tumor location and each target historical tumor location.

[0085] As S43, the similarity of the tumor location of each historical treatment sample to the current treatment sample is determined based on the positional similarity between the predicted tumor location and each target historical tumor location, and between the current tumor location and the historical tumor location in the same image number.

[0086] In this embodiment, the method for determining the similarity of tumor location between each historical treatment sample and the current treatment sample is the same. For the sake of explanation, we will use an arbitrary historical treatment sample as an example to determine the similarity of tumor location between a single historical treatment sample and the current treatment sample.

[0087] As an exemplary embodiment, determining the similarity of tumor location between a current treatment sample and a historical treatment sample includes the following steps:

[0088] The distance between the predicted tumor location and the historical tumor location in the historically treated sample is calculated, and the distance between the current tumor location and the historical tumor location in the same image number is calculated.

[0089] The second power of all distance values ​​corresponding to the historical treatment sample is obtained, and a negative correlation normalization process is performed on all second powers of distance values ​​to obtain a third normalized value, which is then used as the similarity of tumor location.

[0090] As an example, the formula for calculating the similarity of tumor location between the j-th historical treatment sample and the current treatment sample is as follows: JPEG0007852964000011.jpg19170

[0091] JPEG0007852964000012.jpg59170

[0092] JPEG0007852964000013.jpg12170JPEG0007852964000014.jpg7170

[0093] JPEG0007852964000015.jpg31170

[0094] JPEG0007852964000016.jpg7170JPEG0007852964000017.jpg31170

[0095] By referring to the calculation process for the similarity of tumor location of the j-th historical treatment sample to the current treatment sample, the similarity of tumor location of each historical treatment sample to the current treatment sample can be obtained.

[0096] To date, this embodiment has obtained the similarity of tumor location between each historical treatment sample and the current treatment sample.

[0097] In S5, the position prediction accuracy index and tumor location similarity are combined to determine the similarity of the radiotherapy process for each historical treatment sample.

[0098] Here, the similarity of the radiotherapy process refers to the degree of similarity between the radiotherapy process of a historical treatment sample and the radiotherapy process of the current treatment sample. A higher degree of similarity of the radiotherapy process indicates that the current treatment sample is more important to the corresponding historical treatment sample in determining the tumor location at future irradiation times.

[0099] In this embodiment, the similarity of the radiotherapy process is analyzed from two perspectives: namely, the position prediction accuracy index and the similarity of tumor location. A higher position prediction accuracy index indicates that the corresponding historical treatment sample is more accurate in predicting the tumor location, meaning that the distance between the actual tumor location and the predicted pseudotumor location is smaller, indicating a higher similarity, and that the corresponding historical treatment sample is more reliable in achieving tumor location prediction. This index is used to correct the similarity of tumor location. The similarity of tumor location indicates the similarity of the motion trajectories of the tumor location between the current treatment sample and the historical treatment sample. A higher similarity of tumor location indicates a higher similarity of the treatment process when the current treatment sample and the corresponding historical treatment sample perform radiotherapy. When both the position prediction accuracy index and the similarity of tumor location are high, it further indicates that the corresponding historical treatment sample is highly important in determining the tumor location at future irradiation times for the subsequent current treatment sample.

[0100] Specifically, for each historical treatment sample, the product is obtained by multiplying the accuracy index of the location prediction of the historical treatment sample by the similarity of the tumor location. This product is taken as the similarity of the radiotherapy process for the corresponding historical treatment sample.

[0101] To date, this embodiment has obtained similarity in the radiotherapy process to indicate the importance of each corresponding historical treatment sample.

[0102] As S6, the tumor position at future irradiation times for each historically treated sample is obtained based on the CT image time series of each historically treated sample.

[0103] Here, the future irradiation time is the sum of the current imaging time and the irradiation time interval of the irradiation device. The future irradiation time is not equal to any of the imaging times, cannot be directly obtained from the CT image time series, and is much smaller than the irradiation interval of the respiratory gating control technology. For example, if the current imaging time is 10:20:06 and the irradiation time interval of the irradiation device is 1 second, the future irradiation time is 10:20:07.

[0104] In an exemplary embodiment, step S6 described above is achieved by the following steps.

[0105] Step 1 involves obtaining the location of the tumor at each historical imaging time for each historical treatment sample, based on the CT image time series of the historical treatment sample.

[0106] In this embodiment, the central position of the tumor in the CT image at each historical acquisition time is determined in the CT image time series of the historical treatment sample and recorded as the historical tumor position.

[0107] Step 2 involves constructing a position acquisition equation based on the historical imaging time and the historical tumor position at each imaging time.

[0108] In this embodiment, a binary linear equation is constructed with the historical imaging time as the independent variable and the historical tumor position corresponding to the historical imaging time as the dependent variable, and this is used as the position acquisition equation for the corresponding historical treatment sample.

[0109] In step 3, the tumor position at the future irradiation time of the corresponding chronologically treated sample is obtained, using the future irradiation time as the independent variable in the position acquisition equation.

[0110] In this embodiment, the future irradiation time is used as the independent variable in the position acquisition equation, and the dependent variable corresponding to the future irradiation time, i.e., the tumor position at the future irradiation time of the corresponding historical treatment sample, can be obtained by calculation.

[0111] Furthermore, the tumor location at future irradiation times for each historical treatment sample serves as reference data for subsequently predicting the tumor location at future irradiation times for the current treatment sample.

[0112] In another exemplary embodiment, the formula for calculating the transverse coordinate of the tumor location at a future irradiation time for the j-th historical treatment sample is as follows: JPEG0007852964000018.jpg12170

[0113] JPEG0007852964000019.jpg37170

[0114] JPEG0007852964000020.jpg41170

[0115] By referring to the x-coordinate of the tumor location at the future irradiation time for each historical treatment sample, the y-coordinate of the tumor location at the future irradiation time for each historical treatment sample can be obtained.

[0116] To date, this embodiment has obtained the tumor location at future irradiation times for each historical treatment sample.

[0117] As S7, the similarity of the radiotherapy process is used to perform a weighted summation process on the tumor position at future irradiation times for each historical treatment sample, to obtain the tumor position at future irradiation times for the current treatment sample, and then dynamically compensated targeted radiotherapy is performed.

[0118] The core of respiratory gating technology is to control the timing of tumor irradiation during the radiotherapy process and monitor the respiratory cycle to limit treatment to a specific respiratory phase, i.e., the end of exhalation or inhalation, thereby reducing tumor displacement and improving treatment accuracy. However, this can prolong treatment time, increase the patient's radiation exposure time, and potentially cause damage to normal tissue or secondary cancers, particularly tissues such as the lungs and heart, which may receive unwanted radiation. To mitigate the above risks, the present invention obtains the tumor position at future irradiation times for the current treatment sample so that radiotherapy is performed at the highest frequency of the irradiation device.

[0119] In one exemplary embodiment, step S7 described above is achieved by the following steps.

[0120] Step 1: For each historical treatment sample, the ratio of the similarity of the radiotherapy process of the historical treatment sample to the cumulative similarity of all radiotherapy processes is used as the position weight for the corresponding historical treatment sample.

[0121] In this embodiment, the range of possible values ​​for the position weight is limited to between 0 and 1, and the cumulative value of the position weight for each historical treatment sample is 1.

[0122] In step 2, the product of the position weight of each historical treatment sample and the tumor position at the future irradiation time of the corresponding historical treatment sample is calculated as the tumor subposition at the future irradiation time of the current treatment sample.

[0123] In step 3, a fusion process is performed on all tumor sub-locations to obtain the fused location, and this fused location is defined as the tumor location at future irradiation times for the current treatment sample.

[0124] In this embodiment, the step of adding up all tumor sub-locations specifically involves using the sum of the horizontal coordinates of the tumor sub-locations as the horizontal coordinate of the tumor location at the future irradiation time of the current treatment sample, and using the sum of the vertical coordinates of the tumor sub-locations as the vertical coordinate of the tumor location at the future irradiation time of the current treatment sample.

[0125] For example, the formula for calculating the x-coordinate of the tumor location at future irradiation times for a current treatment sample is as follows: JPEG0007852964000021.jpg13170

[0126] JPEG0007852964000022.jpg38170

[0127] By referring to the calculation process for the x-coordinate of the tumor location at future irradiation times for the current treatment sample, it is possible to obtain the y-coordinate of the tumor location at future irradiation times for the current treatment sample.

[0128] Furthermore, the time intervals between future irradiation times are much shorter than those of respiratory gating technology. By triggering the irradiation equipment in advance and irradiating the predicted tumor location before the actual irradiation time arrives, the target treatment time can be effectively shortened.

[0129] After obtaining the tumor position of the current treatment sample at future irradiation times, the irradiation device can be triggered in advance at the next irradiation time to perform precise radiotherapy. The above process of determining the tumor position of the current treatment sample at future irradiation times is repeated continuously, and after each irradiation, the tumor position at the irradiation time separated by the highest frequency of the next irradiation device is obtained, the irradiation device is triggered in advance, and precise radiotherapy is continued until the treatment is completed, ultimately achieving targeted radiotherapy that dynamically compensates the current thoracic and abdominal tumor patient. Here, the irradiation frequency of the irradiation device for achieving radiotherapy may be obtained by the practitioner.

[0130] Another embodiment of the present invention provides an AI-based system for dynamic compensation of thoracoabdominal tumor target regions, comprising a processor and memory, wherein the processor processes instructions stored in the memory to realize an AI-based method for dynamic compensation of thoracoabdominal tumor target regions.

[0131] The above embodiments are merely for illustrative purposes and not limiting purposes. While the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the inventions described in each of the above embodiments can still be modified or some of their technical features can be replaced with equivalent substitutions. Such modifications or substitutions do not cause the essence of the corresponding invention to deviate from the scope of the inventions described in each embodiment, and should all fall within the scope of protection of the present invention.

Claims

1. A method for dynamic compensation of a thoracoabdominal tumor target region based on AI processed by a processor, The steps include obtaining the time series of CT images and BMI values ​​of the current treatment sample and several candidate historical treatment samples, A step of screening a candidate historical treatment sample for a respiratory pattern similar to the current treatment sample, based on the similarity of BMI values ​​between the current treatment sample and each of the candidate historical treatment samples, wherein the similarity of BMI values ​​refers to the confidence level of comparing the candidate historical treatment sample, as analyzed from the perspective of respiratory pattern, with the current treatment sample; The steps include determining a position prediction accuracy index corresponding to each of the aforementioned historical treatment samples based on the CT image time series of each of the aforementioned historical treatment samples, A step of determining the similarity of the tumor location of each of the historical treatment samples to the current treatment sample based on the CT image time series of the current treatment sample and the CT image time series of each of the historical treatment samples, A step of fusing the position prediction accuracy index and the similarity of the tumor position to determine the similarity of the radiotherapy process for each of the historical treatment samples, The steps include obtaining the tumor position at a future irradiation time for each of the aforementioned treatment samples based on the CT image time series of each aforementioned treatment sample, where the future irradiation time is the sum of the current imaging time and the irradiation time interval of the irradiation device, The step of using the similarity of the radiotherapy process to perform a weighted addition process on the tumor position at future irradiation times for each of the historical treatment samples to obtain the tumor position at future irradiation times for the current treatment sample, The step of screening for historical treatment samples from the multiple candidate historical treatment samples whose respiratory patterns are similar to those of the current treatment sample, based on the similarity of BMI values ​​between the current treatment sample and each of the multiple candidate historical treatment samples, is as follows: For each of the aforementioned candidate historical treatment samples, the absolute value of the difference between the BMI value of the current treatment sample and the BMI value of the candidate historical treatment sample is calculated. The absolute value of the difference is subjected to a negative correlation normalization process to obtain a first normalized value, and this first normalized value is determined as the degree to which the respiratory patterns of the candidate historical treatment sample can be compared to the current treatment sample. This includes setting a degree threshold and obtaining a historical treatment sample by comparing the comparability of each respiratory pattern with the degree threshold, A method for dynamic compensation of thoracoabdominal tumor target regions based on AI, characterized by the following:

2. The step of determining a position prediction accuracy index corresponding to each of the historical treatment samples based on the CT image time series of each of the historical treatment samples is: Based on the CT image time series of each of the aforementioned historical treatment samples, the central position of the tumor at each historical imaging time is determined and recorded as the actual tumor position, and the central position of the tumor at each historical imaging time is predicted and recorded as the pseudotumor position. Calculate the positional difference between the actual tumor location and the false tumor location at the same historical imaging time, This includes fusing and analyzing the positional difference values ​​of all historical imaging times corresponding to the same historical treatment sample, and determining the positional prediction accuracy index corresponding to each of the historical treatment samples, Here, the position difference value shows a negative correlation with the position prediction accuracy index. A method for dynamic compensation of a thoracoabdominal tumor target region based on AI, as described in feature 1.

3. The step of performing a fusion analysis of the positional difference values ​​of all historical imaging times corresponding to the same historical treatment sample and determining the positional prediction accuracy index corresponding to each historical treatment sample is: For each historical shooting time, a negative correlation normalization process is performed on the position difference value to obtain a second normalized value, and the second normalized value is determined as the prediction criterion corresponding to the historical shooting time. This includes performing a power calculation on the prediction criterion for all historical imaging times corresponding to the same historical treatment sample to obtain a first power value, and determining the first power value as a position prediction accuracy index corresponding to the corresponding historical treatment sample. The method for dynamic compensation of a thoracoabdominal tumor target region based on AI as described in feature 2.

4. The step of determining the similarity of the tumor location of each historical treatment sample to the current treatment sample based on the CT image time series of the current treatment sample and the CT image time series of each historical treatment sample is as follows: Using the time series of CT images of the current treatment sample as input data, the predicted tumor position of the CT image at the next time point is obtained. For each of the aforementioned historical treatment samples, a CT image from the CT image time series of the historical treatment sample is selected for the target historical acquisition time, the location of the target historical tumor in the selected CT image is determined, and the image number for the target historical acquisition time is the same as the CT image number for the next time, The similarity of the tumor location of each historical treatment sample to the current treatment sample is determined based on the positional similarity between the predicted tumor location and each of the target historical tumor locations, and between the current tumor location and the historical tumor location of the same image number, and the current tumor location is the tumor location in the CT image of the current treatment sample. A method for dynamic compensation of a thoracoabdominal tumor target region based on AI, as described in feature 1.

5. The step of determining the similarity of the tumor location of each historical treatment sample to the current treatment sample, based on the positional similarity between the predicted tumor location and each of the target historical tumor locations, and between the current tumor location and the historical tumor location of the same image number, is as follows: For any historically treated sample, calculate the distance between the predicted tumor location and the historical tumor location of that historically treated sample, and calculate the distance between the current tumor location and the historical tumor location for the same image number. This includes obtaining the second power of all distance values ​​corresponding to the historical treatment sample, performing negative correlation normalization on all second powers of distance values ​​to obtain a third normalized value, and using the third normalized value as the similarity of tumor location. The method for dynamic compensation of a thoracoabdominal tumor target region based on AI as described in feature 4.

6. The step of fusing the position prediction accuracy index and the similarity of tumor location to determine the similarity of the radiotherapy process for each of the historical treatment samples is: For each of the aforementioned historical treatment samples, the method includes multiplying the position prediction accuracy index of the historical treatment sample by the similarity of the tumor position to obtain the product, and using the product as the similarity of the radiotherapy process of the corresponding historical treatment sample. A method for dynamic compensation of a thoracoabdominal tumor target region based on AI, as described in feature 1.

7. The step of obtaining the tumor position at a future irradiation time for each of the historical treatment samples based on the CT image time series of each of the historical treatment samples is: For each of the aforementioned historical treatment samples, the location of the historical tumor at each historical imaging time is obtained based on the CT image time series of the historical treatment sample. Based on the time of each historical imaging and the historical tumor position at each historical imaging time, a position acquisition equation is constructed, This includes using the aforementioned future irradiation time as the independent variable in the position acquisition equation, and acquiring the tumor position at the future irradiation time of the corresponding historical treatment sample. A method for dynamic compensation of a thoracoabdominal tumor target region based on AI, as described in feature 1.

8. The step of obtaining the tumor position of the current treatment sample at a future irradiation time by performing a weighted addition process on the tumor position at a future irradiation time for each of the historical treatment samples using the similarity of the radiotherapy process is as follows: For each of the aforementioned historical treatment samples, the ratio of the similarity of the radiotherapy process of the historical treatment sample to the cumulative similarity value of all the radiotherapy processes is used as the position weight for the corresponding historical treatment sample. The product of the position weight of each of the aforementioned historical treatment samples and the tumor position at the future irradiation time of the corresponding historical treatment sample is calculated as the tumor subposition at the future irradiation time of the current treatment sample, This includes performing a fusion process on all tumor sub-locations to obtain a fused location, and setting the fused location as the tumor location at a future irradiation time for the current treatment sample. A method for dynamic compensation of a thoracoabdominal tumor target region based on AI, as described in feature 1.

9. The system includes a processor and memory, wherein the processor is configured to process instructions stored in the memory to realize the AI-based method for dynamic compensation of a thoracoabdominal tumor target region according to any one of claims 1 to 8. An AI-based dynamic compensation system for thoracoabdominal tumor target regions, characterized by the following features.

Images