Auxiliary discrimination method for nasopharyngeal carcinoma radiation-induced brain injury image, device and medium

Abstract

The application discloses an auxiliary discrimination method for a nasopharyngeal carcinoma radiation-induced brain injury image, a device and a medium, relating to the technical field of medical information. The method comprises: preprocessing an MRI image, segmenting temporal lobe areas, using the MRI image temporal lobe areas and corresponding focus of infection labels to train a slice-level prediction model, and predicting the probability of a radiation-induced brain injury focus existing in a single temporal lobe image area; taking the probability of all slices in each subject containing the focus of infection as the input, wherein the probability is outputted by the slice-level prediction model, and taking whether the subject has radiation-induced brain injury as the output, training a patient-level prediction model; finally, using the trained slice-level prediction model and the patient-level prediction model to obtain the radiation-induced brain injury disease probability of a single nasopharyngeal carcinoma subject, and when a plurality of modes coexist, taking a plurality of modal prediction results to output a mean value as a final prediction result. The present application can determine, according to the input nasopharyngeal carcinoma brain MRI image, whether radiation-induced brain injury exist, so as to reduce missed diagnosis of nasopharyngeal carcinoma radiation-induced brain injury.

Description

Auxiliary identification method, equipment and medium for imaging of radiation-induced brain damage in nasopharyngeal carcinoma

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on November 7, 2023, with application number 202311470891.2 and invention name “Auxiliary identification method, equipment and medium for imaging of radiation-induced brain damage in nasopharyngeal carcinoma”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present application belongs to the field of medical information technology, and specifically relates to an auxiliary identification method, device and medium for imaging of radiation-induced brain damage in nasopharyngeal carcinoma. Background Art

[0003] Nasopharyngeal carcinoma (NPC) is an epithelial malignancy that originates in the nasopharyngeal mucosa, most commonly in the pharyngeal recesses. Due to the close proximity of the nasopharyngeal cavity to the skull base and the aggressive growth of NPC, even with intensity-modulated radiotherapy and volume-modulated arc radiation therapy, it is difficult to avoid the temporal lobe, particularly the bilateral anterior and inferior temporal lobes, which are prone to radiation-induced temporal lobe damage. Late-stage symptoms of radiation-induced temporal lobe damage are irreversible, manifesting as memory loss, cognitive impairment, and, in severe cases, temporal lobe necrosis. However, in clinical practice, MRIs often miss or delay the diagnosis of NPC radiation-induced brain damage due to physician attention, lesion size, and location, impacting patient prognosis and quality of life. In recent years, deep learning has shown promising prospects in medical image analysis, including disease diagnosis and lesion segmentation. Therefore, a deep learning-based diagnostic model for NPC radiation-induced brain damage can be developed to assist physicians in clinical decision-making. Summary of the Invention

[0004] The present application provides an auxiliary identification method, device and medium for nasopharyngeal carcinoma radiation brain damage images, which can determine whether there is radiation brain damage based on the input nasopharyngeal carcinoma brain MRI images, provide a reference for clinical doctors' diagnosis, and reduce the missed diagnosis of nasopharyngeal carcinoma radiation brain damage.

[0005] To achieve the above technical objectives, this application adopts the following technical solutions:

[0006] An auxiliary identification method for nasopharyngeal carcinoma radiation brain damage imaging, comprising:

[0007] Step 1: Collect brain MRI images of nasopharyngeal carcinoma patients and preprocess them;

[0008] Step 2, segmenting the temporal lobe region in the MRI image to obtain a temporal lobe MRI image;

[0009] Step 3: Using the temporal lobe MRI image plane as input and the expert-annotated lesion as output, the first deep learning network is trained to obtain a lesion recognition model for nasopharyngeal carcinoma radiation-induced brain damage.

[0010] Step 4: Input each layer of the temporal lobe MRI image into the nasopharyngeal carcinoma radiation-induced brain injury lesion recognition model, and output the probability of the presence of radiation-induced brain injury lesions in each layer.

[0011] Step 5: Using the probability of radiation-induced brain injury lesions at all levels of each patient's temporal lobe MRI images as input and the presence of radiation-induced brain necrosis as output, a second deep learning network is trained to obtain a diagnostic model for radiation-induced brain injury in nasopharyngeal carcinoma.

[0012] Step 6: Preprocess the newly acquired brain MRI images of NPC patients according to step 1, obtain segmented temporal lobe MRI images according to step 2, and use the radiation brain injury lesion recognition model according to step 4 to obtain the probability of the presence of NPC lesions in each temporal lobe layer. Then, input the radiation brain injury lesion probabilities of all temporal lobe layers into the NPC radiation brain injury diagnosis model to obtain the probability of radiation brain injury in each patient.

[0013] In an exemplary embodiment, the pre-processing in step 1 includes image quality control and bias field correction.

[0014] In an exemplary embodiment, the first deep learning network adopts an improved EfficientNet network; specifically, an attention mechanism is added after the last convolutional layer of the original EfficientNet network, and the top fully connected layer is modified to four fully connected layers, with the number of neurons being 256, 128, 64 and 32 respectively, the activation function being ReLU, and the last being a sigmoid classification layer with one neuron; the first deep learning network is trained using a binary cross entropy loss function.

[0015] In an exemplary embodiment, L2 regularization and Dropout are added to each fully connected layer of the improved EfficientNet network.

[0016] In an exemplary embodiment, the second deep learning network uses a multilayer perceptron with 30, 16 and 8 neurons respectively, an activation function of ReLU, and a sigmoid classification layer with one neuron at the end; the second deep learning network is trained using a category cross entropy loss function.

[0017] In an exemplary embodiment, the brain MRI image includes three modality data: T1WI, T2WI, and CE-T1WI. Three lesion recognition models and three diagnosis models are obtained by training respectively using the three different modality data.

[0018] In an exemplary embodiment, the average value of the radiation-induced brain injury probabilities output by the diagnostic models corresponding to the three modalities is taken as the probability of the final diagnosis of radiation-induced brain injury.

[0019] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor implements any of the above-mentioned methods for auxiliary identification of nasopharyngeal carcinoma radiation-induced brain damage images.

[0020] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements any of the above-mentioned methods for assisting in distinguishing radiation-induced brain damage in nasopharyngeal carcinoma. Beneficial effects

[0021] This application utilizes artificial intelligence (AI) methods to extract deep convolutional features from temporal lobe MRI images, enabling the detection and diagnosis of radiation-induced brain damage in the temporal lobe of nasopharyngeal carcinoma. The incorporated attention mechanism improves lesion localization, forcing the model to focus on global lesion characteristics, further enhancing the diagnostic performance of radiation-induced brain damage in nasopharyngeal carcinoma. This deep learning model can be integrated into the daily workflow of radiologists. Research has shown that the use of a deep learning model has reduced the missed detection rate of radiation-induced brain damage in nasopharyngeal carcinoma to 9.8%, particularly improving the detection of small solid enhancing nodules ≤10 mm in diameter.

[0022] Figures in the specification

[0023] FIG1 is a technical roadmap of the method described in an embodiment of the present application.

[0024] Figure 2 is a schematic diagram of the network structure of EfficientNet-B4 in Figure 1.

[0025] FIG3 is a schematic diagram of the network structure of the ADL layer in FIG1 .

[0026] FIG4 is a schematic diagram of a multi-layer perceptron for the patient-level prediction model in FIG1 . DETAILED DESCRIPTION

[0027] The following is a detailed description of the embodiments of the present application. This embodiment is based on the technical solution of the present application, provides a detailed implementation method and a specific operation process, and further explains the technical solution of the present application.

[0028] Example 1

[0029] This embodiment provides an auxiliary identification method for radiation-induced brain damage in nasopharyngeal carcinoma images, as shown in FIG1 . The network structure of EfficientNet-B4 in FIG1 is shown in FIG2 , the network structure of the ADL layer in FIG1 is shown in FIG3 , and the multi-layer perceptron of the patient-level prediction model in FIG1 is shown in FIG4 , including the following steps:

[0030] Step 1: Collect brain MRI images of nasopharyngeal carcinoma patients and preprocess them.

[0031] The MRI images of nasopharyngeal carcinoma patients collected in this example include three modal data: CE-T1WI, T1WI, and T2WI. Data were collected from Xiangya Hospital of Central South University, Chenzhou First Hospital, the First Affiliated Hospital of Nanhua University, the Second Affiliated Hospital of Nanhua University, and other institutions, totaling 3,842 patients with nasopharyngeal carcinoma radiation-induced brain injury. The diagnostic criteria for nasopharyngeal carcinoma radiation-induced brain injury are as follows: (1) white matter lesions (finger-like lesions with increased T2WI signal intensity); (2) enhancing lesions (small nodules or large necrotic foci on CE-T1WI); (3) cysts (round or oval lesions on T2WI with signal intensity similar to cerebrospinal fluid); (4) hemorrhage with high signal on T1WI and hemosiderin deposition with low signal on T2WI. When any of these manifestations appear, nasopharyngeal carcinoma radiation-induced brain injury can be diagnosed.

[0032] During MR scanning, the bias field causes nonuniform magnetic field strength, causing MR intensity values ​​to vary between images acquired from the same scanner, the same patient, and even the same tissue. In this example, the N4BiasFieldCorrection function in the Python package SimpleITK was used to address this issue, and ANTS was used to align the multimodal imaging data.

[0033] Step 2: Segment all temporal lobe regions in the brain MRI image. The segmentation method is as follows: import the collected image data into ITKSNAP, outline the temporal lobe boundary, and export the original image data image and the segmented image in ".nii" format.

[0034] The original image was multiplied with the corresponding manually segmented temporal lobe mask to obtain a region containing only the temporal lobe, which was then cropped with the minimum horizontal bounding rectangle to contain the minimum horizontal rectangle of the region and resized to a temporal lobe MRI image of 380 × 380 square pixels.

[0035] Since this embodiment collects thick-layer images and the scan range is inconsistent (some are head and neck, some are only head), manual segmentation is adopted. If the collected thin-layer images are also available, automatic segmentation can be achieved through registration (registering to a template, and then using the template to segment the area).

[0036] In step 3, the temporal lobe MRI image layer is used as input and the expert annotation of whether the lesion is contained is used as output to train the first deep learning network and obtain a lesion recognition model for nasopharyngeal carcinoma radiation-induced brain damage.

[0037] Among them, the method used by the film reading experts to mark whether the temporal lobe MRI image layer contains lesions is: the experts browse the original images of each modality and mark the radiation brain injury lesions at the level of the transverse scan, with the skull base as the starting level.

[0038] In step 3, the data set is divided into a training set, a validation set, and a test set based on patients. The training set and validation set data come from Xiangya Hospital of Central South University, and the data from other hospitals constitute the test set.

[0039] In step 4, each layer of the temporal lobe MRI image is input into the nasopharyngeal carcinoma radiation brain injury lesion recognition model, and the output is the probability of the presence of radiation brain injury lesions in each layer.

[0040] In step 5, the second deep learning network is trained using the probability of radiation-induced brain injury lesions at all levels of each patient's temporal lobe MRI images as input and the presence of radiation-induced brain necrosis as output to obtain a diagnostic model for radiation-induced brain injury in nasopharyngeal carcinoma.

[0041] In steps 3-5 above, first train the lesion recognition model using all training data in step 3. Then, in step 4, use the trained lesion recognition model to obtain the probability of lesions in each layer. Finally, in step 5, use the training data to train the nasopharyngeal carcinoma radiation-induced brain injury diagnosis model. The training of the lesion recognition and diagnosis models for nasopharyngeal carcinoma radiation-induced brain injury is done in Python using the Keras framework. The training graphics card is a Quadro GV100 with 16GB of memory. During training, image augmentation is performed on the training samples, such as random axis flipping, translation, and random cropping, to enrich the training sample set.

[0042] The nasopharyngeal carcinoma radiation-induced brain injury lesion detection model uses a modified EfficientNet-B4 network. The network input is a preprocessed temporal lobe MRI image slice, and the output is whether the slice contains a radiation-induced brain injury lesion. This network achieves a balance across all dimensions of network width, depth, and resolution, resulting in improved performance. Furthermore, to force the network to more comprehensively learn the characteristics of nasopharyngeal carcinoma radiation-induced brain injury lesions, an Attention-based Dropout Layer (ADL) is added after the last convolutional layer of the base EfficientNet-B4 network (without the last fully connected layer). This attention mechanism is applied to each feature map, guiding the model to learn the complete lesion region without increasing network parameters. Finally, four fully connected layers and a Sigmoid layer are added after the ADL layer for classification. The four fully connected layers have 256, 128, 64, and 32 neurons, respectively. L2 regularization and a Dropout layer are added to each fully connected layer to reduce the risk of overfitting. The network uses a binary cross-entropy loss function.

[0043] Among them, N is the number of slices, y is the gold standard corresponding to the slice (y changes with the change of i), that is, whether the level contains nasopharyngeal carcinoma radiation brain injury lesions, The probability of including lesions output by the detection model ( changes with i).

[0044] The nasopharyngeal carcinoma radiation-induced brain injury diagnosis model uses a multi-layer perceptron network. The input is the probability output of the nasopharyngeal carcinoma radiation-induced brain injury detection model for each patient's temporal lobe slice, and the output is whether the individual patient has been diagnosed with nasopharyngeal carcinoma radiation-induced brain injury. The number of neurons in each layer is 32, 16, 8, and 1, respectively. The activation function for the last layer is Sigmoid, and the activation functions for the remaining layers are Reluctant Units (ReLUs). The loss function used for network training is the binary cross-entropy loss function.

[0045] Where M is the number of patients, x is the gold standard corresponding to the patient (x changes with the change of j), that is, whether the patient is diagnosed with nasopharyngeal carcinoma radiation brain injury, is the probability of illness output by the diagnosis model ( changes with the change of j).

[0046] Step 6: Preprocess the newly acquired brain MRI images of NPC patients according to step 1, obtain segmented temporal lobe MRI images according to step 2, and use the radiation brain injury lesion recognition model according to step 4 to obtain the probability of the presence of NPC lesions in each temporal lobe layer. Then, input the radiation brain injury lesion probabilities of all temporal lobe layers into the NPC radiation brain injury diagnosis model to obtain the probability of radiation brain injury in each patient.

[0047] Since the MRI images in this embodiment include three modal data types: CE-T1WI, T1WI, and T2WI, in order to improve diagnostic accuracy, three lesion recognition models and three diagnostic models are obtained by training using three different modal data types according to steps 3 to 5. Since the lesion recognition model in this embodiment uses a three-channel input structure, when the model uses only one modal data type, the three channels of the lesion recognition model are input using a single slice and three-channel stacking method. Furthermore, when judging the newly acquired brain MRI images of nasopharyngeal carcinoma patients, the average value of the probability of radiation-induced brain injury output by the diagnostic models corresponding to the three modalities is taken as the probability of the final diagnosis of radiation-induced brain injury.

[0048] Through the deep learning-based nasopharyngeal carcinoma radiation brain injury detection and diagnosis method of this application, the following can be obtained from brain MRI images: (1) the probability and corresponding location of a single temporal lobe layer containing a radiation brain injury lesion; and (2) the probability of whether a single patient has radiation brain injury.

[0049] Example 2

[0050] This embodiment provides an electronic device, including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor implements the method described in Embodiment 1.

[0051] Example 3

[0052] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method described in Embodiment 1 is implemented.

[0053] The above embodiments are preferred embodiments of the present application. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present application, these changes or improvements should fall within the scope of protection required by the present application.

Claims

1. A method for assisting in distinguishing radiation-induced brain damage in nasopharyngeal carcinoma, characterized in that: include: Step 1, collecting brain MRI images of NPC patients and preprocessing them; Step 2, segmenting the temporal lobe region in the MRI image to obtain a temporal lobe MRI image; Step 3, using the temporal lobe MRI image level as input and the expert's annotation of whether the lesion is included as output, the first deep learning network is trained to obtain a lesion recognition model for nasopharyngeal carcinoma radiation brain damage; Step 4, input each layer of the temporal lobe MRI image into the nasopharyngeal carcinoma radiation brain injury lesion recognition model, and output the probability of the presence of radiation brain injury lesions in each layer; Step 5, using the probability of radiation brain injury lesions at all levels of each patient's temporal lobe MRI images as input and whether the patient has radiation brain injury necrosis as output, train the second deep learning network to obtain a diagnostic model for nasopharyngeal carcinoma radiation brain injury; Step 6, for the newly acquired brain MRI image of the NPC patient, preprocess it according to step 1, obtain the segmented temporal lobe MRI image according to step 2, and use the radiation brain injury lesion recognition model to obtain the probability of the presence of NPC lesions in each temporal lobe layer according to step 4, and then input the radiation brain injury lesion probabilities of all temporal lobe layers into the NPC radiation brain injury diagnosis model to obtain the probability of radiation brain injury in each patient.

2. The method according to claim 1, characterized in that The preprocessing described in step 1 includes image quality control and bias field correction.

3. The method according to claim 1, characterized in that The first deep learning network adopts an improved EfficientNet network; specifically, an attention mechanism is added after the last convolutional layer of the original EfficientNet network, and the top fully connected layer is modified to four fully connected layers, with the number of neurons being 256, 128, 64 and 32 respectively, the activation function being ReLU, and finally a sigmoid classification layer with one neuron; the first deep learning network is trained using a binary cross entropy loss function.

4. The method according to claim 3, characterized in that L2 regularization and Dropout are added to each fully connected layer of the improved EfficientNet network.

5. The method according to claim 1, characterized in that The second deep learning network uses a multilayer perceptron with 30, 16 and 8 neurons respectively, an activation function of ReLU, and finally a sigmoid classification layer of one neuron; the second deep learning network is trained using a category cross entropy loss function.

6. The method according to claim 1, characterized in that The brain MRI images include three modality data: T1WI, T2WI and CE-T1WI. Three lesion recognition models and three diagnosis models are obtained by training respectively using the three different modality data.

7. The method according to claim 6, characterized in that The average value of the radiation brain injury probabilities output by the diagnostic models corresponding to the three modalities was taken as the probability of the final diagnosis of radiation brain injury.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the computer program is executed by the processor, the processor is caused to implement the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images