A method and device for predicting radiation dose distribution based on CT images

Abstract

The present invention discloses a method and device for predicting radiation dose distribution based on CT images. A convolutional neural network is used to process a CT image to be predicted to obtain a first feature map and a second feature map of different resolutions corresponding to the CT image to be predicted, and a first initial differential field and a second initial differential field. The feature map and the initial differential field are further predicted through a decoder stream to obtain a first differential field and a second differential field. The shape of the prior image is continuously sampled through a continuous differential field, so that a prior with correct topological features is deformed through a continuous differential field to maintain the topology. Since the first differential field and the second differential field are differential homomorphic fields, which are continuous deformation fields that lead to a one-to-one mapping, their derivatives are reversible and can obtain a positive Jacobian determinant, so that a clear mapping between the prior shape and the predicted image coordinates is given in dose registration, so that 100% of the topology is maintained, and the accuracy of radiation dose distribution prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of CT image processing, and in particular to a method and device for predicting radiation dose distribution based on CT images. Background Art

[0002] CT (Computed Tomography) is widely used in clinical medicine because it can use precisely collimated X-ray beams, gamma rays, ultrasound and other rays for imaging, and has the characteristics of fast scanning time and clear images. Especially in the field of clinical radiotherapy. Since the rays and radiation produced by radiotherapy are not only destructive to tumor tissues, but also to other organ tissues, CT images are needed to determine the target area of ​​radiotherapy and related treatment information such as radiation dose. However, the current CT image analysis technology has certain limitations and cannot accurately obtain accurate radiation dose distribution prediction results based on CT images. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a radiation dose distribution prediction method based on CT images to improve the accuracy of radiation dose distribution prediction.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A method for predicting radiation dose distribution based on CT images, comprising the steps of:

[0006] Acquire a CT image to be predicted;

[0007] Extracting a first feature map and a second feature map of two different resolutions in the CT image to be predicted, as well as a first initial differential field corresponding to the first feature map and a second initial differential field corresponding to the second feature map by using a convolutional neural network;

[0008] Predicting the first initial differential field and the second initial differential field through a decoder stream to obtain a first differential field and a second differential field; the first differential field and the second differential field are homomorphic differential fields;

[0009] Acquire a priori image, and sample the shape of the priori image through the first differential field to obtain first sampling data;

[0010] Performing secondary sampling on the first sampled data through a second differential field to obtain a predicted value;

[0011] The predicted value is downsampled to the same resolution as the CT image to be predicted to obtain a predicted image.

[0012] In order to solve the above technical problems, another technical solution adopted by the present invention is:

[0013] A radiation dose distribution prediction device based on CT images comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, each step in the above-mentioned radiation dose distribution prediction method based on CT images is implemented.

[0014] The beneficial effects of the present invention are as follows: The beneficial effects of the present invention are as follows: The first feature map and the second feature map of different resolutions corresponding to the CT image to be predicted are processed by a convolutional neural network, and the first initial differential field and the second initial differential field are obtained, and the feature map and the initial differential field are further predicted by a decoder stream to obtain the first differential field and the second differential field, and then the prior image shape is continuously sampled by the continuous differential field formed by the first differential field and the second differential field, so as to maintain the topology by deforming a prior with correct topological features through the continuous differential field, and since the first differential field and the second differential field are differential homomorphic fields, which are continuous deformation fields that lead to a one-to-one mapping, their derivatives are reversible, and a positive Jacobian determinant can be obtained, so as to give a clear mapping between the prior shape and the predicted image coordinates in the dose registration, so as to achieve 100% topology maintenance and improve the accuracy of radiation dose distribution prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a method for predicting radiation dose distribution based on CT images according to an embodiment of the present invention;

[0016] Figure 2 The figure is a schematic diagram of the network structure of a radiation dose distribution prediction method based on CT images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to explain the technical content, achieved objectives and effects of the present invention in detail, the following is an explanation in combination with the implementation modes and the accompanying drawings.

[0018] Please refer to Figure 1 , a radiation dose distribution prediction method based on CT images, comprising the steps of:

[0019] Acquire a CT image to be predicted;

[0020] Extracting a first feature map and a second feature map of two different resolutions in the CT image to be predicted, as well as a first initial differential field corresponding to the first feature map and a second initial differential field corresponding to the second feature map by using a convolutional neural network;

[0021] Predicting the first initial differential field and the second initial differential field through a decoder stream to obtain a first differential field and a second differential field; the first differential field and the second differential field are homomorphic differential fields;

[0022] Acquire a priori image, and sample the shape of the priori image through the first differential field to obtain first sampling data;

[0023] Performing secondary sampling on the first sampled data through a second differential field to obtain a predicted value;

[0024] The predicted value is downsampled to the same resolution as the CT image to be predicted to obtain a predicted image.

[0025] From the above description, it can be seen that the beneficial effect of the present invention is that: the first feature map and the second feature map of different resolutions corresponding to the CT image to be predicted are processed by a convolutional neural network, and the first initial differential field and the second initial differential field are obtained, and the feature map and the initial differential field are further predicted by the decoder stream to obtain the first differential field and the second differential field, and then the prior image shape is continuously sampled by the continuous differential field formed by the first differential field and the second differential field, so as to maintain the topology by deforming a prior with correct topological features through the continuous differential field. Since the first differential field and the second differential field are differential homomorphic fields, which are continuous deformation fields that lead to a one-to-one mapping, their derivatives are reversible and can obtain positive Jacobian determinants, so as to give a clear mapping between the prior shape and the predicted image coordinates in the dose registration, so as to achieve 100% topology maintenance and improve the accuracy of radiation dose distribution prediction.

[0026] Furthermore, the extracting of a first feature map and a second feature map of two different resolutions in the CT image to be predicted by a convolutional neural network, as well as a first initial differential field corresponding to the first feature map and a second initial differential field corresponding to the second feature map includes:

[0027] The convolutional neural network includes a first convolutional stream, a second convolutional stream and a third convolutional stream;

[0028] The first convolution flow sequentially performs convolution and maximum pooling for a first preset number of times on the CT image to be predicted to obtain a reduced dimension feature map;

[0029] The second convolution flow sequentially performs upsampling and matrix concatenation on the dimension reduction feature map for a second preset number of times to obtain the first feature map and a first initial differential field;

[0030] The third convolution flow performs upsampling and matrix concatenation on the dimension reduction feature map to obtain the second feature map and the second initial differential field.

[0031] From the above description, it can be seen that the first convolution flow performs multiple maximum pooling on the predicted CT image to reduce the image dimension and obtain more refined features. The second and third convolution flows perform matrix splicing and upsampling on the reduced-dimensional feature map to different degrees, respectively, so as to obtain feature maps of different resolutions and corresponding differential fields.

[0032] Further, the predicting the first initial differential field and the second initial differential field through the decoder stream to obtain the first differential field and the second differential field comprises:

[0033] Performing a scaling operation and a squaring operation on the first initial differential field and the second initial differential field;

[0034] The first initial differential field and the second initial differential field after the scaling operation and the square operation are predicted by a decoder stream to obtain a first differential field and a second differential field, and the activation function of the decoder stream is set to v:

[0035]

[0036] Wherein, u(x) represents the first initial differential field or the second initial differential field.

[0037] From the above description, it can be seen that the first initial differential field and the second initial differential field are forced to be scaled and squared through the decoder stream, and the formulated tanh activation function is applied to the output of the decoder stream. The activation function introduces nonlinear factors to the neurons, so that the neural network can arbitrarily approximate any nonlinear function; it is responsible for mapping the input of the neuron to the output end, so that the image is forced to be displaced between -0.5 voxels and +0.5 voxels, so as to maintain the topological structure and improve the accuracy of the prediction, where the voxel represents the unit volume element on the feature image.

[0038] Furthermore, the extracting of the first feature map and the second feature map of two different resolutions in the CT image to be predicted by the convolutional neural network, and the first initial differential field corresponding to the first feature map and the second initial differential field corresponding to the second feature map also includes:

[0039] The output result of each convolution block in the convolutional neural network is processed by Gaussian smoothing method.

[0040] From the above description, it can be seen that since the differential pure field is represented by a finite number of parameters and is sampled using linear interpolation, the inaccuracies and defects in the image acquisition process can be magnified with the number of components. However, by introducing the Gaussian smoothing method, these effects are reduced to ensure the accuracy of the image.

[0041] Further, the predicting the first initial differential field and the second initial differential field through the decoder stream to obtain the first differential field and the second differential field comprises:

[0042] The first differential field and the second differential field outputted are super-upsampled by bilinear interpolation.

[0043] It can be seen from the above description that the amplified differential field is super-upsampled by bilinear interpolation, and the resolution of the input CT image to be predicted is doubled, and the differential field is further smoothed to minimize the abnormality.

[0044] Furthermore, it also includes:

[0045] Set the loss function: L = L Dice (Y,Y label )+L Grad (φ FT )+L Grad (φ Bulk );

[0046] Among them, L Dice is the dice loss function, Y is the comparison sample, Y label is the prediction sample, φ FT is the first differential field, φ Bulk is the second differential field, L Grad for:

[0047]

[0048] Among them, h and w are parameters corresponding to the first feature map and the second feature map.

[0049] From the above description, it can be seen that the loss function is set to supervise the loss of the predicted value, the first differential field and the second differential field during the training process, that is, the loss during supervised training is minimized as much as possible after multiple iterative training, so as to minimize the difference between the predicted value and the true value and improve the accuracy of the predicted value.

[0050] Furthermore, the loss function also includes a weighting parameter β;

[0051] Then L = L Dice (Y,Y label )+β(L Grad (φ FT )+L Grad (φ Bulk )).

[0052] From the above description, it can be seen that by setting the weighting parameter β to balance the weight between the first loss value and the second loss value, the weighting parameter can be adjusted according to the specific situation to make the training result more targeted.

[0053] Furthermore, the model updating data includes, before executing the extraction of the first feature map and the second feature map of two different resolutions in the to-be-predicted CT image by the convolutional neural network, and the first initial differential field and the second initial differential field corresponding to the first feature map and the second feature map:

[0054] Determine whether model update data is received;

[0055] If so, the step of extracting the first feature map and the second feature map of two different resolutions in the CT image to be predicted, as well as the first initial differential field and the second initial differential field corresponding to the first feature map and the second feature map through a convolutional neural network is performed according to the model update data.

[0056] From the above description, it can be seen that by judging whether the prediction operation is performed after receiving the model update data, that is, by saving all the characteristic parameters in the network training process and forming the model data, different model data can be replaced for different CT images to be predicted during the prediction process, thereby improving the accuracy of the prediction.

[0057] Further, the obtaining of the prior image comprises:

[0058] Preprocessing the prior image includes:

[0059] Sampling the prior image shape from three directions to obtain target data;

[0060] establishing a priority list based on the target data;

[0061] Resampling the target data to obtain resampled data;

[0062] Dose value data is obtained according to the priority list and the resampled data.

[0063] From the above description, it can be seen that the target data is obtained by sampling the prior image shape, and a priority list is generated based on the target data. Since different organs and tissues have different sensitivities to radiation, such as some organs and tissues are very sensitive to radiation and can only accept a small radiation dose, while some organs require more radiation to kill cancer cells, there will be a priority between organs, that is, a priority allocation for treatment targets and organs at risk, so that the results of subsequent dose predictions are more accurate.

[0064] Please refer to Figure 3The present invention also provides a radiation dose distribution prediction device based on CT images, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, each step in the above-mentioned method for predicting radiation dose distribution based on CT images is implemented.

[0065] The above-mentioned radiation dose distribution prediction method and device based on CT images of the present invention can be applied to radiation dose distribution prediction, such as partial prediction of radiation dose in tumor radiotherapy, and is described below through specific implementation methods:

[0066] Embodiment 1

[0067] Please refer to Figure 1 , a radiation dose distribution prediction method based on CT images, comprising the steps of:

[0068] S1, obtaining a CT image to be predicted;

[0069] S2. Extracting a first feature map and a second feature map of two different resolutions in the CT image to be predicted, as well as a first initial differential field corresponding to the first feature map and a second initial differential field corresponding to the second feature map through a convolutional neural network;

[0070] Please refer to Figure 2 , the convolutional neural network includes a first convolutional stream, a second convolutional stream and a third convolutional stream, the first convolutional stream is a CNN node from the input end of the convolutional neural network to the branch, the first convolutional stream is a first CNN branch at the branch, and the second convolutional stream is a second CNN branch at the branch; a decoding branch structure similar to U-net is used, each branch is composed of a series of convolutional blocks, and each convolutional block includes three processing steps of convolution, normalization and ReLU activation function; in an optional embodiment, the output result of each convolutional block in the convolutional neural network is processed by Gaussian smoothing; wherein the role of normalization is to map values ​​of different ranges of variation to the same fixed range; the role of the activation function is to map the input of the neuron to the output end, introduce nonlinear factors, and enhance the learning ability of the network;

[0071] Specifically, the method comprises the following steps: S21, the first convolutional flow sequentially performs convolution and maximum pooling for a first preset number of times on the CT image to be predicted to obtain a reduced-dimensional feature map; and extracts an initial image Y(h×w) with a feature map dimension of h×w from the input CT image to be predicted, where h is the image height and w is the width;

[0072] S22, the second convolution flow sequentially performs upsampling and matrix splicing on the reduced-dimensional feature map for a second preset number of times to obtain the first feature map and the first initial differential field; the first initial differential field is a body field, predicting a low-resolution body velocity u Bulk (2×(h×w) / 8), used to warp the shapes in the prior image to the correct region;

[0073] S23, the third convolution flow performs upsampling and matrix concatenation on the dimension reduction feature map to obtain the second feature map and the second initial differential field; the second initial differential field is a fine-tuning field, and a higher resolution fine-tuning field is predicted for fine-tuning the distorted fragments su FT (2×(h×w) / 2);

[0074] Among them, steps S22 and S23 are parallel steps and can be executed successively or simultaneously; in an optional implementation manner, before S2, the step of: determining whether model update data is received; if so, executing the step of extracting the first feature map and the second feature map of two different resolutions in the CT image to be predicted by the convolutional neural network, and the first initial differential field and the second initial differential field corresponding to the first feature map and the second feature map according to the model update data;

[0075] S3, predicting the first initial differential field and the second initial differential field through a decoder stream to obtain a first differential field and a second differential field; the first differential field and the second differential field are homomorphic differential fields;

[0076] Specifically, the steps include: S31, performing a scaling operation and a squaring operation on the first initial differential field and the second initial differential field;

[0077] S32, predicting the first initial differential field and the second initial differential field through a decoder stream to obtain a first differential field and a second differential field, and setting a constraint condition v:

[0078]

[0079] Wherein, u(x) represents the first initial differential field or the second initial differential field; the activation function v=f(u) is used to force the differential field so that the image is forced to be displaced between -0.5 voxels and +0.5 voxels, thereby maintaining the topological structure and improving the accuracy of the prediction;

[0080] S3 further includes: super upsampling the output first differential field and second differential field by bilinear interpolation;

[0081] S4, obtaining a priori image, sampling the shape of the priori image through the first differential field to obtain first sampling data; specifically, the first differential field Φ Bulk (2×2h×2w) samples the prior image P(h×w) and outputs the third differential field Y Bulk =Φ Bulk (P) and the corresponding image of the completed shape distortion; that is, the voxels of the prior shape P are distorted into the first feature map so that the voxels of the first feature map correspond to the voxels of the prior shape P;

[0082] After obtaining the prior image, the method further includes the following steps: S41, sampling the shape of the prior image from three directions to obtain target data;

[0083] S42, establishing a priority list according to the target data; resampling the target data to obtain resampled data;

[0084] S43, obtaining dose value data according to the priority list and the resampled data;

[0085] S5. Perform secondary sampling on the first sampled data through a second differential field to obtain a predicted value. Specifically, the second differential field Φ TF (2×2h×2w) to obtain the third differential field Y Bulk =Φ Bulk (P) gets Y = Φ TF (Y Bulk ), and sample the image that completes the shape distortion to obtain the predicted value Y labal ;

[0086] S6, downsampling the predicted value to the same resolution as the CT image to be predicted to obtain a predicted image; the predicted value Y labal The CT image to be predicted has different feature sizes and resolutions from the initial one; therefore, it is necessary to adjust the two to the same feature size and resolution by performing maximum pooling again.

[0087] Embodiment 2

[0088] The difference between this embodiment and the first embodiment is that a field regularisation mechanism is further provided to monitor the loss of the training process;

[0089] Specific:

[0090] Set the loss function: L = L Dice (Y1,Y label )+L Grad (φ FT )+L Grad (φBulk ), where L Dice is the dice loss function, Y1 is the comparison sample, Y label is the predicted sample, i.e. the predicted result, φ FT is the first differential field, φ Bulk is the second differential field, L Grad for:

[0091]

[0092] h and w are the parameters corresponding to the first feature map and the second feature map, that is, if L is calculated Grad (φ Ft ), then bring in h and w corresponding to the first feature map, if we calculate L Grad (φ Bulk ) then bring in h and w corresponding to the second feature map;

[0093] In an optional embodiment, the loss function further includes a weighting parameter β;

[0094] Then L = L Dice (Y,Y label )+β(L Grad (φ FT )+L Grad (φ Bulk )); The proportion between the antecedent and the latter term is controlled by adjusting the value of the weighting parameter β. If the proportion of the latter term is required to be larger, the weighting parameter β is set to 1; if the latter term is required to account for half of the proportion, 0.5 is taken.

[0095] Embodiment 3

[0096] Please refer to Figure 2 , this embodiment provides a specific dose distribution prediction network framework structure;

[0097] The image input on the left is the CT image to be predicted, which is Y. The image input on the lower right is the prior image, which is P. The image output on the upper right is the predicted value image, which is Y. label ;

[0098] The first convolution flow includes five CNN nodes and four maximum pooling processes in sequence; the bottleneck of the first convolution flow branches into the second convolution flow and the third convolution flow; the second convolution flow is composed of a CNN node (up-sampling block) and a 1×1 convolution, and the third convolution flow is composed of three CNN nodes and a 1×1 convolution; in the up-sampling process, matrix splicing is performed with the corresponding CNN nodes in the first convolution flow respectively, and a two-dimensional field u is generated respectively, namely the first initial differential field and the second initial differential field;

[0099] The decoder stream includes: a differential homeomorphic layer, a Gaussian smoothing layer, and a spatial transformation layer; and forms two branches: a first decoder stream and a second decoder stream, the first decoder stream corresponds to the first convolution stream, and the second decoder stream corresponds to the second convolution stream; meanwhile, the spatial transformation layer of the second decoder stream is used to input the prior image;

[0100] Specific:

[0101] In step S1, the CT image to be predicted Y is obtained through the first convolution stream, and the prior image P is obtained through the spatial transformation layer of the second decoder stream;

[0102] In step S2, after the data reaches the bottleneck of the first convolution flow, the second convolution flow in the first branch obtains the first feature map and the first initial differential field corresponding thereto, and the third convolution flow in the second branch obtains the second feature map and the second initial differential field corresponding thereto;

[0103] In step S3, in the first branch, the first decoder stream predicts the first differential field according to the result of the second convolution stream; in the second branch, the second decoder stream predicts the second differential field according to the result of the third convolution stream;

[0104] In step S4, the a priori image P is sampled by the first differential field in the spatial transformation layer of the first decoder stream to obtain first sampled data;

[0105] In step S5, the first sampled data is subsampled by the second differential field in the spatial transformation layer of the second decoder stream to obtain a predicted value Y label ;

[0106] Step S6, perform maximum pooling again, and convert the predicted value Y output by the second decoder stream into label The predicted image is obtained by sampling to the same resolution as the CT image Y to be predicted.

[0107] Embodiment 4

[0108] Please refer to Figure 3 A radiation dose distribution prediction device based on CT images, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, each step in a radiation dose distribution prediction method based on CT images as described in any one of embodiments one to three is implemented.

[0109] In summary, the present invention provides a method and device for predicting radiation dose distribution based on CT images. The CT image to be predicted is processed by a convolutional neural network to obtain a first feature map and a second feature map of different resolutions corresponding to the CT image to be predicted, as well as a first initial differential field and a second initial differential field. The feature map and the initial differential field are further predicted by a decoder stream to obtain a first differential field and a second differential field. The prior image shape is continuously sampled by a continuous differential field formed by the first differential field and the second differential field, so that a prior with correct topological features is deformed by a continuous differential field to maintain the topology. Since the first differential field and the second differential field are differential homomorphic fields, which are continuous deformation fields that lead to a one-to-one mapping, their derivatives are reversible and can obtain a positive Jacobian determinant, so that a clear mapping between the prior shape and the predicted image coordinates is given in the dose registration, so as to achieve 100% topology maintenance. At the same time, by introducing field supervision mechanisms, Gaussian smoothing, super upsampling and other processing methods, the accuracy of data in the prediction process and the training process is guaranteed, thereby improving the accuracy of radiation dose distribution prediction.

[0110] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's specification and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for predicting radiation dose distribution based on CT images, characterized in that: Includes steps: Acquire a CT image to be predicted; Extracting a first feature map and a second feature map of two different resolutions in the CT image to be predicted, as well as a first initial differential field corresponding to the first feature map and a second initial differential field corresponding to the second feature map by using a convolutional neural network; Predicting the first initial differential field and the second initial differential field through a decoder stream to obtain a first differential field and a second differential field; the first differential field and the second differential field are homomorphic differential fields; Acquire a priori image, and sample the shape of the priori image through the first differential field to obtain first sampling data; Performing secondary sampling on the first sampled data through a second differential field to obtain a predicted value; Downsampling the predicted value to the same resolution as the CT image to be predicted to obtain a predicted image; The step of predicting the first initial differential field and the second initial differential field through a decoder stream to obtain the first differential field and the second differential field comprises: Performing a scaling operation and a squaring operation on the first initial differential field and the second initial differential field; The first initial differential field and the second initial differential field after the scaling operation and the square operation are predicted by a decoder stream to obtain a first differential field and a second differential field, and the activation function of the decoder stream is set to v: ； Wherein, u(x) represents the first initial differential field or the second initial differential field.

2. The method for predicting radiation dose distribution based on CT images according to claim 1, characterized in that: The extracting of the first feature map and the second feature map of two different resolutions in the to-be-predicted CT image by a convolutional neural network, as well as the first initial differential field corresponding to the first feature map and the second initial differential field corresponding to the second feature map comprises: The convolutional neural network includes a first convolutional stream, a second convolutional stream and a third convolutional stream; The first convolution flow sequentially performs convolution and maximum pooling for a first preset number of times on the CT image to be predicted to obtain a reduced dimension feature map; The second convolution flow sequentially performs upsampling and matrix concatenation on the dimension reduction feature map for a second preset number of times to obtain the first feature map and a first initial differential field; The third convolution flow performs upsampling and matrix concatenation on the dimension reduction feature map to obtain the second feature map and the second initial differential field.

3. The method for predicting radiation dose distribution based on CT images according to claim 1, characterized in that: The extracting of the first feature map and the second feature map of two different resolutions in the to-be-predicted CT image by the convolutional neural network, as well as the first initial differential field corresponding to the first feature map and the second initial differential field corresponding to the second feature map further includes: The output result of each convolution block in the convolutional neural network is processed by Gaussian smoothing method.

4. The method for predicting radiation dose distribution based on CT images according to claim 1, characterized in that: The step of predicting the first initial differential field and the second initial differential field through a decoder stream to obtain the first differential field and the second differential field comprises: The first differential field and the second differential field outputted are super-upsampled by bilinear interpolation.

5. The method for predicting radiation dose distribution based on CT images according to claim 1, characterized in that: Also includes: Setting the loss function ; in is the dice loss function, For comparison samples, For the prediction sample, is the first differential field, is the second differential field, for: ； Among them, h and w are parameters corresponding to the first feature map and the second feature map.

6. The method for predicting radiation dose distribution based on CT images according to claim 5, characterized in that: The loss function also includes a weighting parameter β; but .

7. The method for predicting radiation dose distribution based on CT images according to claim 1, characterized in that: The method of extracting the first feature map and the second feature map of two different resolutions in the CT image to be predicted by the convolutional neural network, and the first initial differential field and the second initial differential field corresponding to the first feature map and the second feature map, comprises: Determine whether model update data is received; If so, the step of extracting the first feature map and the second feature map of two different resolutions in the CT image to be predicted, as well as the first initial differential field and the second initial differential field corresponding to the first feature map and the second feature map through a convolutional neural network is performed according to the model update data.

8. The method for predicting radiation dose distribution based on CT images according to claim 1, characterized in that: The step of obtaining the prior image includes: Preprocessing the prior image includes: Sampling the prior image shape from three directions to obtain target data; The target data includes organ data and tissue data; establishing a priority list based on the target data; Resampling the target data to obtain resampled data; Dose value data is obtained according to the priority list and the resampled data.

9. A radiation dose distribution prediction device based on CT images, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, each step of the method for predicting radiation dose distribution based on CT images as described in any one of claims 1 to 8 is implemented.

Images