Cervical cancer brachytherapy dose prediction method based on deep learning

Abstract

The invention discloses a cervical cancer brachytherapy dose prediction method based on deep learning. The method comprises the following steps: preprocessing cervical cancer image data, constructing a data set, establishing and training a dose prediction network model, and inputting a test set to obtain a dose prediction distribution result; according to the method, a network model adopts a double-encoder architecture and is used for extracting deep features of an input image in a layered manner and compressing data dimensions, a decoder performs up-sampling step by step so as to recover spatial details and reconstruct dose distribution, and meanwhile, jump connection is introduced between the encoder and the decoder, so that multi-scale feature fusion is realized; therefore, the feature expression ability of the model to the organ boundary region is enhanced, and the accuracy of dose prediction in the edge region is effectively improved. According to the invention, the problems of low dose prediction precision and the like caused by complex structure, large target shape and size change and fuzzy tumor boundary in a cervical cancer dose prediction task are solved, and the accuracy and clinical practicability of a prediction result are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image processing technology, and in particular to a method for predicting the dose of brachytherapy for cervical cancer based on deep learning. Background Technology

[0002] Cervical cancer is one of the most common malignant tumors that seriously threaten women's health. Radiation therapy is the main treatment for this disease. Brachytherapy, in particular, delivers a very high radiation dose to the tumor by placing the radiation source directly on or near the tumor target area. Simultaneously, it utilizes the characteristic that the dose rapidly decreases with distance, maximizing the protection of surrounding organs at risk. Therefore, this method usually plays a very important role in the treatment plan. Accurate dose prediction is a crucial aspect of brachytherapy in order to achieve good treatment results.

[0003] Currently, in clinical practice, dose prediction during cervical cancer radiotherapy mainly relies on manual adjustment and optimization of the target and dose limits until they meet clinical prescription requirements. During this process, the experience level and working conditions of different physicists can significantly affect the final quality of the radiotherapy plan. Even when dealing with the same patient, differences between operators can lead to significant fluctuations in key indicators such as organ radiation dose, thus affecting the standardization of treatment and the stability of efficacy.

[0004] Because cervical cancer tumors have a multicentric origin and growth pattern, their shapes are highly irregular, and their anatomical relationships with organs at risk, such as the bladder, rectum, and sigmoid colon, are complex and variable. Furthermore, during the intervals between brachytherapy fractions, the tumor itself undergoes significant volume reduction and morphological changes. This means that the shape, size, and spatial relationship of the target area with surrounding organs at risk are constantly evolving with each brachytherapy treatment. Therefore, dose prediction methods must possess extremely strong adaptive capabilities. Traditional computer-aided diagnostic techniques have weak adaptive capabilities, resulting in poor accuracy in dose prediction.

[0005] Therefore, designing a method for predicting the dose of brachytherapy for cervical cancer has become an urgent technical challenge. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based method for predicting the dose of cervical cancer radiotherapy. This method demonstrates good dose prediction performance for cervical cancer tumors of different sizes and shapes.

[0007] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0008] This invention includes the following steps:

[0009] S1: Obtain the original DICOM data of cervical cancer radiotherapy from the hospital, preprocess the dataset, divide the preprocessed image dataset into a training set and a test set with a ratio of 7:3, and randomly distribute them into the two datasets according to the ratio.

[0010] S2: Establish a cervical cancer dose prediction network model, including a dense convolutional coding module, a hybrid downsampling module, a guided attention module, a feature fusion module, a Transformer encoder module, a dual-channel collaborative attention module, and a progressive reconstruction module;

[0011] S3: The cervical cancer dose prediction network constructed in S2 is trained using the training set. Based on the loss function, the loss value between the predicted dose and the clinical dose is obtained. The Adam optimizer is used to update the parameters of the model, and finally the trained cervical cancer dose prediction network model is generated.

[0012] S4: Input the new patient data from the test set into the dose prediction network model trained in S3 to obtain the dose prediction results for cervical cancer tumors.

[0013] The beneficial effects of this invention are:

[0014] This invention presents a deep learning-based method for cervical cancer tumor dose prediction. Compared with existing technologies, this invention innovatively designs a dual encoder-decoder type dose prediction network. The encoder progressively extracts geometric information from the feature map and compresses the spatial dimension to capture high-level abstract features, while the decoder progressively repairs the details and spatial dimensions of the feature map. In the dose prediction network, skip connections are used to fuse features from different levels between the encoder and decoder, helping the decoder to better predict dose values ​​at organ boundaries. This invention solves the problems of low dose prediction accuracy caused by complex structures, large variations in target shape and size, and difficulty in identifying tumor edges in cervical cancer dose prediction tasks, significantly improving the dose prediction accuracy for cervical cancer tumors. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of one embodiment of the method of the present invention;

[0016] Figure 2 This is a schematic diagram of a dense convolutional coding module according to an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of a hybrid downsampling module according to an embodiment of the present invention;

[0018] Figure 4 This is a schematic diagram of a guided attention module according to an embodiment of the present invention;

[0019] Figure 5 This is a schematic diagram of a feature fusion module according to an embodiment of the present invention;

[0020] Figure 6 This is a schematic diagram of a Transformer encoder module according to an embodiment of the present invention;

[0021] Figure 7 This is a schematic diagram of a dual-channel collaborative attention module according to an embodiment of the present invention;

[0022] Figure 8 This is a schematic diagram of a progressive reconstruction module according to an embodiment of the present invention;

[0023] Figure 9 This is a schematic diagram of a dual encoder structure according to an embodiment of the present invention;

[0024] Figure 10 This is a schematic diagram of a decoder structure according to an embodiment of the present invention;

[0025] Figure 11 This is an example of a visualization of cervical cancer dose prediction results obtained using the method of the present invention. Detailed Implementation

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0027] The fundamental principles involved in this invention are:

[0028] 1. Dense convolutional coding module; such as Figure 2 The input feature map of size C×H×W is sequentially processed through a 3×3 convolutional unit with a stride of 1, followed by IN (batch normalization), and then a ReLU activation function for channel compression, resulting in an output feature map of size G-rate×H×W, where G-rate is a specific value. This feature map is then sequentially processed through a 3×3 convolutional unit with a stride of 1, followed by IN and then a ReLU activation function to extract shallow information. The output of the first activation function is added to this output. Then, the same process is repeated, passing the feature map through a 3×3 convolutional unit with a stride of 1, followed by IN and then a ReLU activation function to extract deeper information. Finally, this output is concatenated with the original input feature map along the channel dimension to obtain the output of the dense convolutional module. The final output image size is (C + G-rate)×H×W. The dense convolutional module significantly reduces the number of parameters and effectively alleviates the gradient vanishing problem, improving the stability of the network model training.

[0029] 2. Hybrid downsampling module; such as Figure 3The input feature map of size C×H×W is sequentially processed through a 3×3 convolutional unit with a stride of 2, followed by IN and ReLU activation functions to compress the image size and channels, resulting in an output feature map with G-rate channels and half the original size. Simultaneously, the original feature map is processed through a max-pooling unit to obtain an output feature map with C-rate channels and half the original size. Finally, the two feature maps are concatenated along the channel dimension to obtain an output feature map with C+G-rate channels and half the original size. This parallel downsampling method allows the model to more effectively capture salient features, significantly reducing information loss and achieving a more efficient integration of original and abstract features.

[0030] 3. Attention-guided modules; such as... Figure 4 The two input feature maps of size C×H×W are concatenated along the channel dimension to obtain an output feature map of size 2C×H×W. This output feature map is then sequentially passed through a 3×3 convolutional unit with a stride of 1, followed by IN and ReLU activation functions for feature extraction. Finally, the activated feature map is passed through a 1×1 convolutional unit with a stride of 1 and a Sigmoid function for channel compression and normalization, generating a guided attention map G of size 1×H×W. i , where i takes values ​​of 1, 2, 3, and 4, representing the guided attention map of the i-th layer. This guided attention map is used to enhance the spatial attention of multi-scale features of the two backbone coding networks, enabling the network to prioritize the target region boundaries and key structural areas, effectively improving the backbone network's ability to represent features of important anatomical regions.

[0031] 4. Feature fusion module; such as Figure 5 The process involves concatenating two input feature maps of size C×H×W along the channel dimension to obtain an output feature map of size 2C×H×W. This output feature map is then compressed through a 1×1 convolutional unit with a stride of 1, resulting in a feature map of size C×H×W. Global average pooling and global max pooling are then applied to this feature map simultaneously, yielding two average pooling and max pooling vectors of size C×1×1. These two vectors are then concatenated along the channel dimension to obtain a vector of size 2C×1×1. This vector is then compressed to C / r through a 1×1 convolutional unit. ReLU activation is applied to the compressed vector, and the activated vector is then increased to C through a 1×1 convolutional unit. Finally, this vector is processed by the Sigmoid function to generate an attention weight vector of size C×1×1. This attention weight vector is then element-wise multiplied with the feature map generated in the first 1×1 convolutional unit, weighting the feature map and ultimately generating a fused feature map Z with an output size of C×H×W. i, where i takes the value 1, 2, 3, 4, representing the fused feature map of the i-th layer. The generated fused feature map is used for subsequent residual connections. The feature fusion module can enhance the feature expression ability while maintaining the integrity of the original feature information, avoid feature shift during gradient propagation, and make the overall network more stable during training.

[0032] 5. Transformer encoder module; such as Figure 6 The input feature map is passed through a 3×3 convolutional unit, expanding the number of channels to 1024. This feature map is then reshaped into a series of two-dimensional image patches, each with a size of p×p, and the number of patches is n = h×w / p. 2 Each image patch is mapped into a D-dimensional space through a fully connected layer. To preserve the original spatial location information of the image patches, a learnable location-specific code is embedded into these image patches. This image patch is then input into a Transformer layer, which consists of 12 Transformer layers. Each Transformer layer contains two layer normalization (LN) layers, a multi-head self-attention (MSA) module, and a multilayer perceptron (MLP) module. Its output can be represented as:

[0033]

[0034] Where LN represents layer normalization, MSA represents multi-head self-attention module, and MLP represents multilayer perceptron module, y y represents the input to the second layer of normalization. i-1 and y i These represent the outputs of the (i-1)th and i-th Transformer layers, respectively.

[0035] 6. Dual-channel collaborative attention module; such as Figure 7 This module uses the fused feature map Z i Upsampled feature map U during the decoding stage j As input, where j represents the number of upsampling layers in the decoder, the two feature maps are first passed sequentially through 3×3 convolutional units, IN, and ReLU activation functions to locally enhance the input features. Then, they are compressed through 1×1 convolutional layers. The dimensionality-reduced feature maps are then processed by the Sigmoid function to generate channel attention weight vectors. Finally, Z... i The weights generated by the branches are applied to U j Branch road, will U j The weights generated by the branch are applied to Z. iBranches are used to achieve bidirectional collaboration between encoder structural information and high-level semantic information at the decoder. Finally, the features enhanced by bidirectional interaction are added element-wise and fused through 3×3 convolution to generate residual feature map A. n , where n takes the values ​​1, 2, 3, and 4, representing the residual feature map generated by the nth layer. Compared with the traditional single convolutional interactive attention structure, this structure can effectively improve the network's sensitivity to key dose regions, tumor boundaries, and tissue structure changes.

[0036] 7. Progressive reconstruction module; such as Figure 8 This module comprises four reconstruction branches, corresponding to different resolution levels of the decoder output. Specifically, each reconstruction branch contains a feature extraction unit, which includes two consecutive 3×3 convolutions, IN, and ReLU activation functions. This unit performs depth-refinement extraction on the feature map of the current level. The output of the feature extraction unit is then passed through a 1×1 convolution, compressing the number of channels in the feature map to 1. The compressed feature map is then passed through a Sigmoid function to generate a reconstructed dose map. To achieve progressive refinement across levels, this module also uses sub-pixel convolution to upsample the output dose map of the current branch, using it as low-level detail information input to higher-level reconstruction branches. This progressive, multi-scale fusion mechanism iteratively adjusts the prediction results and gradually generates the final dose target. It fully utilizes multi-scale and contextual information, significantly improving the sensitivity to local dose constraints and dose gradient information, thereby generating a final dose distribution map with richer texture details and higher accuracy.

[0037] 8. Weighted Loss Function; The weighted loss function can enhance the training efficiency of the model, help the model converge faster, and increase the model's stability and accuracy. Its expression is:

[0038]

[0039] In the formula: p is the dose distribution map predicted by the model, g is the actual dose distribution map, i and j are the coordinate indices of the pixels, w and h are the width and height of the image, respectively, and α and β are the weight coefficients of the MSE and MAE indices, respectively. In this embodiment, the values ​​are α = 0.25 and β = 0.75.

[0040] This embodiment uses a cervical cancer radiotherapy dataset collected by Sichuan Cancer Hospital.

[0041] This invention proposes a deep learning-based method for predicting brachytherapy dose in cervical cancer tumors, the flowchart of which is shown below. Figure 1 As shown, the method includes the following steps:

[0042] S1: Obtain the original DICOM data of cervical cancer radiotherapy from the hospital, preprocess the dataset, and then divide the preprocessed image dataset into a training set and a test set with a ratio of 7:3. Randomly distribute the data into the two datasets according to the ratio.

[0043] The preprocessing includes the following five steps:

[0044] S11: CT Image Acquisition and Preprocessing; The original CT image sequence is read from the DICOM data. Thresholding is used to correct the window width and window level of the CT image sequence. The window level is set to 50 and the window width is set to 400, so the truncation range is -150~250. Image data outside this range is set to -150 or 250. Then, the HU value is uniformly mapped to the [0,1] range to reduce the grayscale difference caused by different scanning devices. On this basis, the image is spatially cropped to 256×256 according to the location of the target area, so that the cropped area covers the complete target area and the necessary surrounding organs at risk. Finally, linear interpolation is used to resample the CT image and unify the voxel spacing to the preset equidistant resolution to ensure that all data are consistent in spatial structure.

[0045] S12: Target extraction; Parse the target contour from the file, convert the polygonal contour into a three-dimensional binary mask with the same resolution as the CT image, and perform spatial clipping and resampling operations in S11 on the generated target mask to make it completely aligned with the preprocessed CT image in terms of voxel spacing and spatial coordinates.

[0046] S13: Extraction of endangered organ regions; Parse the contours of each endangered organ from the file, convert the contour of each endangered organ into a three-dimensional binary mask consistent with the CT image, encode each endangered organ mask independently and merge the images to obtain a label image with multiple endangered organs, normalize the pixels of this label image to the [0,1] interval, and then perform the same spatial cropping and resampling operations as in S11 on the generated label image to make it completely aligned with the preprocessed CT image in terms of voxel spacing and spatial coordinates.

[0047] S14: Distance Image Generation; Based on the target area mask and the masks of each endangered organ obtained in S12 and S13, extract their voxel point sets in the coordinate system. Using the target area boundary voxels as the target points for distance calculation, calculate the minimum distance between any pixel point inside each endangered organ and the target point, thereby generating a distance image. The shortest distance from each endangered organ to the target area is calculated using Euclidean distance, and its expression in two dimensions is:

[0048]

[0049] Where (x1,y1) represents the two-dimensional coordinates of any pixel in the region of the organs at risk in the binary image, and (x2,y2) represents the two-dimensional coordinates of any pixel on the boundary of the target region in the binary image. The pixel values ​​of the generated distance image are normalized to the interval [0,1] to generate the distance image from the target region to each organ at risk.

[0050] S15: Extract dose images; extract the three-dimensional dose matrix data of the corresponding case from the DICOM file, and map the dose data to the voxel coordinate system consistent with the CT image according to its spatial location information to generate three-dimensional dose volume data corresponding to the anatomical structure. Then, normalize the dose image according to the maximum dose value to realize the construction of standardized dose images from the original dose data.

[0051] S2: Establish a cervical cancer dose prediction network model, characterized in that step S2 includes the following two sub-steps:

[0052] S21: Establish a dual encoder network; the dual encoder network consists of 4 dense convolutional coding layers, 4 hybrid downsampling layers, 4 guided attention modules, 1 feature fusion module, and 1 Transformer encoder module.

[0053] The structure of a dual encoder is as follows Figure 9 As shown, its workflow is as follows:

[0054] Two encoder branches are constructed, denoted as the first encoder branch and the second encoder branch. Each encoder branch contains four layers, where the first layer contains a dense convolutional coding block, and the second to fourth layers contain a downsampling block and a dense convolutional coding block. The output of each layer is represented by L. Let L be the value of n, where n takes the values ​​a and b, representing the first encoder branch and the second encoder branch respectively, and i takes the values ​​1, 2, 3, and 4, representing the outputs of the dense convolutional coding blocks from the first to the fourth layer respectively. The target region image, the organ at risk image, and the distance image, all with a size of 256×256 from the training set, are input from the first encoder branch, pass through the dense convolutional coding block of the first layer, and the output is L. A CT image with a size of 256×256 is input from the second encoder branch, passes through the dense convolutional coding block of the first layer, and outputs L. , will L and L The input is fed into the guided attention module to generate the guided attention graph G. 1 Then G 1 With L L Multiplying them separately yields L L Then L L These are used as inputs to the second-layer encoders of the two branches, respectively, and so on, to obtain L. L L L L L Each time an image passes through an encoder layer, its size is reduced by half compared to the previous layer. After three downsampling steps, the image size becomes 1 / 8 of the original image. L The input is fed into the feature fusion module to obtain output Z1, and so on, to obtain Z2, Z3, and Z4 respectively. Z4 is fed into the Transformer encoder module to extract global information. Then, this output is passed through a 3×3 convolution with a stride of 2, which reduces the feature map size to half of its original size, resulting in a feature map size of 16×16. This output is denoted as Z. out .

[0055] S22: Establish a decoder network, including 4 dual-channel collaborative attention modules, 4 progressive reconstruction modules, and 4 upsampling modules.

[0056] The structure of the decoder network is as follows Figure 10 As shown, its workflow is as follows:

[0057] The encoder outputs a feature map Z of size 16×16. out Upsampling is performed using nonlinear interpolation, followed by a 1×1 convolution to adjust the number of channels to be the same as Z4, resulting in a 32×32 feature map U4. U4 and Z4 from step S21 are input into the dual-channel collaborative attention module to obtain a residual feature map A4. A4 and U4 are concatenated along the channel dimension. Subsequently, the concatenated feature map is passed sequentially through two identical convolutional units. Each unit contains a 3×3 convolutional layer, IN, and ReLU activation functions to adjust the number of channels. The convolutional output is then split into two paths: one path is input to the progressive reconstruction module, and the other path is used as input for the next upsampling. This process is repeated to obtain the output feature maps of the third and second layers. For the first layer, after processing by the two convolutional units, the output is no longer split into two paths but is directly input into the progressive reconstruction module of this layer. Finally, the reconstructed feature map of this layer is added to the reconstructed feature map of the second layer to obtain the final dose prediction map of the model.

[0058] S3: Use the training set to train the cervical cancer dose prediction network constructed in S2. According to the loss function, train the parameters of the cervical cancer dose prediction network model, obtain the loss value between the predicted dose and the clinical dose, use the Adam optimizer to update the model parameters, and finally generate the trained cervical cancer dose prediction network model.

[0059] S3 specifically includes the following two sub-steps:

[0060] S31: Set training parameters. The dose prediction model utilizes the PyTorch framework. Both training and testing are performed on an NVIDIA RTX 4090 with 24GB of RAM. The total number of training iterations is set to 800, and the batch size is set to 20 samples. The Adam optimizer is used to optimize network parameters, with the weight decay coefficient set to 3 × 10⁻⁶. -5 The learning rate employs a dynamic decay strategy, starting from an initial 3×10⁻⁶. -4 Gradually reduce to 3×10 -7 Furthermore, a deep supervision mechanism is introduced during the training process, including random rotation and flipping for data augmentation, thereby improving the stability of the training process.

[0061] S32: Train the cervical cancer dose prediction network, calculate the loss values ​​of the encoding and decoding modules using the loss function, and update the parameters in the network using the backpropagation method. After 800 iterations, the trained cervical cancer dose prediction network is obtained.

[0062] S4: Input the test set into the cervical cancer dose prediction network model trained in S3 to obtain the cervical cancer tumor dose prediction results. See [link to dose prediction results] for details. Figure 11 .

[0063] like Figure 11 As shown, this is a comparison chart of the cervical cancer tumor dose prediction label and the automatic dose prediction results of the model. It can be seen that the automatic dose prediction results obtained by the method described in this invention are basically the same as the label, achieving the effect of meeting practical application requirements.

[0064] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1.A deep learning-based dose prediction method for cervical cancer brachytherapy, characterized in that Comprise the following steps: S1: preprocessing of the data set, mainly including: CT image acquisition and preprocessing, target region extraction, organ at risk region extraction, organ at risk to target distance image generation, case dose image generation, dividing the preprocessed image data set into training set and test set, the proportion is set to 7:3, and randomly divided into two data sets according to the proportion; S2: establish a cervical cancer dose prediction network model, including a double encoder and a decoder network; Specifically, the following steps are included: S21: establish a double encoder network; The double encoder is composed of 4 dense convolutional encoding layers, 4 hybrid down-sampling layers, 4 guided attention modules, 1 feature fusion module and 1 transformer encoder module; S22: establish a decoder network; The decoder is composed of 4 double-channel collaborative attention modules, 4 progressive reconstruction modules and 4 up-sampling modules; S3: train the cervical cancer dose prediction network constructed in S2 using the training set, obtain the loss value of the predicted dose and the clinical dose according to the loss function, update the parameters of the model using the Adam optimizer, and finally generate a trained cervical cancer dose prediction network model; Specifically, the following steps are included: S31: set the training parameters. The total number of training iterations is set to 800, the sample size of each batch is set to 20, and a deep supervision mechanism is introduced during training, including random rotation, data augmentation by flipping; S32: train the cervical cancer dose prediction network, calculate the loss value of the encoding module and the decoding module using the loss function, and update the parameters in the network using the back propagation method, after 800 iterations, a trained cervical cancer dose prediction network is obtained; S4: input the new patient data in the test set into the dose prediction network model trained in S3 to obtain the dose prediction result of the cervical cancer tumor. 2.The deep learning-based dose prediction method for cervical cancer brachytherapy according to claim 1, wherein: After each dense convolutional encoder, the feature map size remains unchanged and the channel number increases by 64. After each hybrid down-sampling module, the feature map size becomes 1 / 2 of the original, and the channel number also increases by 64. After passing through the complete encoder, the feature map will become 1 / 16 of the original feature map size. 3.The deep learning-based dose prediction method for cervical cancer brachytherapy according to claim 1, wherein: The loss function of step S3 is a weighted loss function; Its expression is: ； In the formula: p is the dose distribution map predicted by the model, g is the actual dose distribution map, i and j are the pixel coordinate indices, w and h are the width and height of the image, respectively, and α and β are the weight coefficients of the MSE and MAE indices, respectively. The dose prediction model uses the PyTorch framework and employs the Adam optimizer to optimize the network parameters, with the weight decay coefficient set to 3×10. -5 The learning rate employs a dynamic decay strategy, starting from an initial 3×10⁻⁶. -4 Gradually reduce to 3×10 -7 .

Images