Search

CN-121812071-B - Nasopharyngeal carcinoma radiotherapy dosage prediction system and method of cascade network

CN121812071BCN 121812071 BCN121812071 BCN 121812071BCN-121812071-B

Abstract

The invention relates to the technical field of medical image processing, and discloses a nasopharyngeal carcinoma radiotherapy dosage prediction system and method of a cascade network, which solve the technical problems that in the prior art, the nasopharyngeal carcinoma radiotherapy dosage prediction lacks collaborative modeling of a global-local structure and self-adaptive cross-stage feature interaction; the feature activation distribution of the first stage is stabilized through example normalization, and is seamlessly integrated with the second stage through a convolution adaptation layer after being spliced with the original input, so that the problems of dimension mismatch and gradient saturation are solved, and flexible multi-stage feature multiplexing is supported.

Inventors

  • XIE SIQI
  • ZHAO WENJUAN
  • NIE SHENGDONG
  • LANG JINYI

Assignees

  • 四川省肿瘤医院

Dates

Publication Date
20260512
Application Date
20260310

Claims (9)

  1. 1. The nasopharyngeal carcinoma radiotherapy dosage prediction system of a cascade network is characterized by comprising a data reading module, a data preprocessing module, a three-dimensional U-Net network, an improved MedNeXt network, a residual fusion module and a loss function training module; The reading data module is used for reading the multi-mode medical image data of the nasopharyngeal carcinoma and outputting the multi-mode medical image data to the data preprocessing module; The data preprocessing module eliminates abnormal data from the multi-mode medical image data through spatial registration, uniformly resamples the multi-mode medical image data subjected to spatial registration to a three-dimensional matrix, constructs a channel for resampled data, normalizes resampled dose, combines the channel and the normalized data into three-dimensional data containing multi-feature channels, and outputs the three-dimensional data to a three-dimensional U-Net network and an improved MedNeXt network; the three-dimensional U-Net network performs downsampling processing on three-dimensional data of the multi-characteristic channels through a first encoder, performs upsampling processing through a first decoder, and outputs preliminary prediction to a residual fusion module through a first output layer; The improved MedNeXt network carries out dimension reduction processing on the splicing characteristics of the data preprocessing module and the three-dimensional U-Net network through an Adapter layer, carries out downsampling processing through a second encoder, carries out upsampling processing through a second decoder, and outputs fine prediction to a residual error fusion module through a second output layer and an output module; The residual fusion module obtains the radiotherapy dosage of the nasopharyngeal carcinoma according to the preliminary prediction and the fine prediction; the loss function training module trains the prediction of the radiotherapy dosage of the nasopharyngeal carcinoma according to the total loss function; the residual fusion module obtains the radiotherapy dosage of the nasopharyngeal carcinoma according to the preliminary prediction and the fine prediction, and the formula is as follows: ; Wherein, the In order to make a preliminary prediction, In order to make a fine prediction, A single-channel diagram output by a three-dimensional U-Net network, For the improved MedNeXt network output residual map, the tanh function activation constraint value is [ -1,1],0.3 is the scaling factor, For the cut-off operation Forced constraint to the [0,1] range, Is a radiotherapy dose for nasopharyngeal carcinoma.
  2. 2. The cascade network nasopharyngeal carcinoma radiotherapy dosage prediction system of claim 1, wherein said read data module comprises a CT image reading sub-module, a structural profile reading sub-module, a dosage distribution sub-module; The CT image reading sub-module is used for reading a DICOM file of each CT slice, extracting slice position labels from a DICOM file header, sequencing images according to slice positions, constructing a three-dimensional CT image, extracting geometric information and a reference coordinate system identifier of the CT image from the DICOM file header, and constructing a 4X 4 affine transformation matrix from a CT voxel grid to a patient coordinate system according to the geometric information; the structural contour reading submodule is used for reading RTSTRUCT files, extracting structural contour point sets and geometric reference information, and converting the structural contour point sets and the geometric reference information into a binary mask aligned with the CT image; the dose distribution submodule is used for reading a dose distribution file in a DICOM format, taking clinically approved three-dimensional actual dose distribution as a gold standard dose, extracting geometric information of the dose distribution from a DICOM file head, extracting a reference coordinate system identifier of the dose distribution file, and constructing a 4X 4 affine transformation matrix from a dose grid to a patient coordinate system according to the geometric information.
  3. 3. The cascade network nasopharyngeal carcinoma radiotherapy dose prediction system of claim 2, wherein the set of structural contour points comprises a target region contour, an organ-at-risk contour, and a planned organ-at-risk volume; the target contour comprises a planned target 1, a planned target 2, a nasopharyngeal target and a cervical lymph node target; the organ-at-risk profile includes brain stem, left eyeball, right eyeball, left lens, right lens, left optic nerve, right optic nerve, optic crossover, left parotid gland, right parotid gland, spinal cord, left temporal lobe, right temporal lobe, larynx-trachea, thyroid; the planned organ-at-risk volume includes a brainstem extension safety margin and a spinal cord extension safety margin.
  4. 4. The cascade network nasopharyngeal carcinoma radiotherapy dosage prediction system of claim 1, wherein the data preprocessing module comprises a spatial registration verification sub-module, a resampling sub-module, a feature construction sub-module, and a dosage normalization sub-module; The space registration verification submodule reads 4 multiplied by 4 affine transformation matrixes corresponding to the CT image and the dose distribution respectively, marks whether the CT image, the structural outline and the dose distribution are in the same patient coordinate system or not through the reference coordinate system of the CT image, the structural outline and the dose distribution, if the CT image, the structural outline and the dose distribution are in the same patient coordinate system, the difference between the 4 multiplied by 4 affine transformation matrixes corresponding to the CT image and the dose distribution respectively is detected to exceed a preset threshold, if the difference exceeds the preset threshold, the registration abnormality is judged, and all data of the patient are removed; The resampling sub-module resamples the grid size and the voxel value of the CT image and the dose distribution by adopting linear interpolation, resamples the grid size and the voxel value of the binary mask by adopting nearest neighbor interpolation, uniformly resamples the three-dimensional matrix to 128 multiplied by 128 voxels, and outputs the three-dimensional matrix to the feature construction sub-module and the dose normalization sub-module; the feature construction submodule takes a binary mask and an expanded binary mask as target area channels, takes a organs at risk outline and a planned organs at risk volume as organ structure channels, takes CT image gray values as intensity channels after cutting and linear normalization, takes the normalized distance of the target area outline as a geometric distance channel, takes a high gradient area as a dangerous area mask channel, and outputs three-dimensional tensors integrated by all channels; The dose normalization module performs normalization processing on voxel values of the resampled dose distribution, so that all voxel value ranges are unified in a [0,1] interval, the voxel values are used as one-dimensional tensors and three-dimensional tensors of 29 channels to be combined to form four-dimensional tensors of 29 channels, and the four-dimensional tensors of 29 channels are output to a three-dimensional U-Net network.
  5. 5. The cascade network nasopharyngeal carcinoma radiotherapy dose prediction system of claim 1, wherein said three-dimensional U-Net network is constructed based on a first encoder, a first decoder, a first output layer; The first encoder, the first decoder and the first output layer are sequentially connected, each layer of encoding module gradually increases the characteristic channel layer by layer to be [16,32,64,128,256] through downsampling, the receptive field is increased layer by layer through convolution operation, instance normalization and activation function, the first decoder performs upsampling through three-line interpolation, the characteristic channel is gradually reduced to be [128,64,32,16] after upsampling, each layer of decoding module performs channel splicing on the upsampled characteristic graph and the characteristic graph transferred by the corresponding layer of the encoding module through jump connection, information loss in upsampling is compensated, spatial resolution is recovered layer by layer through convolution operation and activation function, the first output layer changes the number of the characteristic channel to be 1 through convolution operation, instance normalization and activation function, and preliminary prediction is output to the residual fusion module.
  6. 6. The cascade network nasopharyngeal carcinoma radiotherapy dosage prediction system of claim 1, wherein said modified MedNeXt network comprises an Adapter layer, a second encoder, a second decoder, a bottleneck layer, a second output layer, an output module; The Adapter layer reduces the splicing characteristic to 24 channels and outputs the channels to a second encoder, four layers of codes of the second encoder are used for downsampling, medNeXt modules are sequentially stacked on each layer, a residual error connection and a GRN mechanism are matched, the downsampling convolution is carried out by matching with MedNeXt downsampling modules and convolution kernels, the number of channels is increased to [48,96,192,384] step by step, the space size is halved step by step and is output to a bottleneck layer, the bottleneck layer optimizes multi-scale representation and outputs the multi-scale representation to a second decoder by sequentially stacking MedNeXt modules, the residual error connection and the GRN mechanism are matched, four layers of codes of the second decoder are used for upsampling and upsampling, each layer sequentially stacks MedNeXt modules, the residual error connection and the GRN mechanism are matched, and the upsampling convolution is carried out by matching with MedNeXt upsampling modules and convolution kernels, so that the number of channels is reduced to [192,96,48,24] step by step, the second output layer carries out convolution operation, instance normalization and activation functions to change the number of characteristic channels to 1 after deconvolution, and outputs fine prediction to the residual error fusion module.
  7. 7. The cascade network nasopharyngeal carcinoma radiotherapy dose prediction system of claim 6, wherein the MedNeXt module comprises a deep convolution layer, a normalization layer, an extended convolution layer, a global response normalization layer, a compressed convolution layer; MedNeXt the downsampling module comprises a convolution downsampling layer, a normalization layer, an expansion convolution layer, a global response normalization layer and a compression convolution layer; the MedNeXt up-sampling module comprises a convolution up-sampling layer, a normalization layer, an extended convolution layer, a global response normalization layer and a compressed convolution layer.
  8. 8. The cascade network nasopharyngeal carcinoma radiotherapy dosage prediction system of claim 1, wherein the total loss function module constructs a total loss function from a base reconstruction loss function and a DVH constraint loss function, and the formula is: ; Wherein, the As a function of the total loss, 、 As the weight of the stage is to be determined, 、 As a weight of the task(s), The loss function is constrained for DVH at the three-dimensional U-Net network stage, The loss function is constrained for DVH at the improved MedNeXt network stage.
  9. 9. A method for predicting the dose of radiation therapy for nasopharyngeal carcinoma in a cascade network, for implementing the system for predicting the dose of radiation therapy for nasopharyngeal carcinoma in a cascade network according to any one of claims 1-8, comprising the steps of: reading multi-mode medical image data of nasopharyngeal carcinoma; removing abnormal data from the multi-modal medical image data through spatial registration, uniformly resampling the multi-modal medical image data after spatial registration to a three-dimensional matrix, constructing a channel for the resampled data, normalizing the resampled dose, and combining the channel and the normalized data to form three-dimensional data containing multi-feature channels; The three-dimensional U-Net network carries out downsampling processing on three-dimensional data of the multi-characteristic channels through a first encoder, carries out upsampling processing through a first decoder, and outputs preliminary prediction through a first output layer; The improved MedNeXt network carries out dimension reduction processing on the splicing characteristics of the data preprocessing module and the three-dimensional U-Net network through an Adapter layer, downsampling processing is carried out through a second encoder, upsampling processing is carried out through a second decoder, and fine prediction is output through a second output layer and an output module; Obtaining the radiotherapy dosage of the nasopharyngeal carcinoma according to the preliminary prediction and the fine prediction; and training the prediction of the radiotherapy dosage of the nasopharyngeal carcinoma according to the total loss function.

Description

Nasopharyngeal carcinoma radiotherapy dosage prediction system and method of cascade network Technical Field The invention relates to the technical field of medical image processing, in particular to a nasopharyngeal carcinoma radiotherapy dosage prediction system and method of a cascade network. Background Nasopharyngeal carcinoma is a common malignancy of the head and neck, and radiotherapy is the main treatment means. Intensity Modulated Radiation Therapy (IMRT) and volume rotational intensity modulated radiation therapy (VMAT) techniques can maximally protect surrounding normal tissue while irradiating a tumor target volume with high doses. However, conventional radiotherapy planning is done by a medical physicist through repeated trial and error optimization, is highly dependent on personal experience, usually takes hours to days, and has significant individual differences in plan quality. The rise of deep learning techniques provides a new solution to dose prediction. The early method mainly adopts a three-dimensional U-Net and other standard convolution networks to directly predict the dose distribution from CT images and contour masks, but faces a plurality of technical problems when processing a plurality of prescription doses (generally 54-70Gy spans a plurality of PTVs) and complex organ-endangering constraints of nasopharyngeal carcinoma, and the specific technical problems are that (1) a single network is difficult to process global anatomy structure and local dose gradient details simultaneously, so that the predicted dose is not maintained enough in steep gradient at the boundary of a target area, and (2) a multi-stage feature fusion strategy lacks self-adaptability, and the problem of feature dimension mismatch among stages is easy to cause unstable training and gradient saturation. Disclosure of Invention The invention aims to solve the technical problems that in the prior art, the nasopharyngeal carcinoma radiotherapy dosage prediction lacks collaborative modeling of a global-local structure and self-adaptive cross-stage feature interaction, and provides a nasopharyngeal carcinoma radiotherapy dosage prediction system and method of a cascade network. In order to achieve the above object, the embodiment of the present invention provides the following technical solutions: A nasopharyngeal carcinoma radiotherapy dosage prediction system of cascade network comprises a data reading module, a data preprocessing module, a three-dimensional U-Net network, an improved MedNeXt network, a residual fusion module and a loss function training module; The reading data module is used for reading the multi-mode medical image data of the nasopharyngeal carcinoma and outputting the multi-mode medical image data to the data preprocessing module; The data preprocessing module eliminates abnormal data from the multi-mode medical image data through spatial registration, uniformly resamples the multi-mode medical image data subjected to spatial registration to a three-dimensional matrix, constructs a channel for resampled data, normalizes resampled dose, combines the channel and the normalized data into three-dimensional data containing multi-feature channels, and outputs the three-dimensional data to a three-dimensional U-Net network and an improved MedNeXt network; the three-dimensional U-Net network performs downsampling processing on three-dimensional data of the multi-characteristic channels through a first encoder, performs upsampling processing through a first decoder, and outputs preliminary prediction to a residual fusion module through a first output layer; The improved MedNeXt network carries out dimension reduction processing on the splicing characteristics of the data preprocessing module and the three-dimensional U-Net network through an Adapter layer, carries out downsampling processing through a second encoder, carries out upsampling processing through a second decoder, and outputs fine prediction to a residual error fusion module through a second output layer and an output module; The residual fusion module obtains the radiotherapy dosage of the nasopharyngeal carcinoma according to the preliminary prediction and the fine prediction; The loss function training module trains the prediction of the nasopharyngeal carcinoma radiotherapy dosage according to the total loss function. Aiming at the technical problem that the steep gradient of the predicted dose at the boundary of the target region is insufficient due to the fact that a single network is difficult to process the details of the global anatomical structure and the local dose gradient simultaneously, the cognitive division is realized by adopting a two-stage cascade architecture through functional decoupling. The first stage U-Net focuses on the extraction of anatomical global features to establish the integral mapping of organ space relation and dose distribution, the second stage MedNeXt focuses on the optimization of local dose details in