A model for semantic segmentation of pituitary tumors and prediction of intraoperative cerebrospinal fluid leakage

Abstract

The present invention provides a pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model, including a multi-task deep neural network, a clinical data feature screening module, and a multi-layer perceptron; the multi-task deep neural network takes the pre-processed MRI image sequence as input, uses a convolution block or a Transformer block to extract the multi-scale features of the pre-processed MRI image sequence, and fuses the multi-scale features. Based on the fused features, a pituitary tumor semantic segmentation head and a cerebrospinal fluid leakage classification head are constructed to obtain the semantic segmentation results of the pituitary tumor and the preliminary prediction results of cerebrospinal fluid leakage; the clinical data feature screening module is used to filter out the target clinical data features using the Pearson correlation coefficient and the distance correlation number; the multi-layer perceptron takes the target clinical data features and the preliminary prediction results of cerebrospinal fluid leakage as input, and outputs the final prediction results of cerebrospinal fluid leakage. The present invention can achieve effective semantic segmentation and improve the accuracy and sensitivity of cerebrospinal fluid leakage prediction based on MRI images.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a model for semantic segmentation of pituitary tumors and prediction of intraoperative cerebrospinal fluid leakage. Background Art

[0002] Pituitary tumor is a common neuroendocrine tumor. Endoscopic transnasal surgery has become one of the main means of treating pituitary tumor. Although endoscopic transnasal surgery has a higher tumor resection rate and better surgical field, it also brings a higher incidence of postoperative cerebrospinal fluid leakage. Intraoperative cerebrospinal fluid leakage is an important risk factor for postoperative cerebrospinal fluid leakage, which may cause serious complications such as meningitis, intracranial infection and cerebrospinal fluid hypotension syndrome, thus endangering the patient's life. In order to reduce the risk of postoperative cerebrospinal fluid leakage, a variety of skull base reconstruction techniques have been proposed by scholars at different stages, including the use of biological protein glue, autologous fat, fascia lata, in situ bone flap, dura mater suture, and vascularized nasal septum mucosal flap. Among them, vascularized nasal septum mucosal flap is the most widely used and has a significant effect, which mainly depends on whether cerebrospinal fluid leakage occurs during surgery. However, this method may lead to local nasal inflammation, olfactory disorders and increased risk of epistaxis after surgery, which seriously affects the patient's quality of life. Therefore, in order to reduce the unnecessary use of vascularized nasal septal mucosal flaps, clinical practice urgently needs a prediction model to determine whether cerebrospinal fluid leakage will occur during endoscopic transnasal pituitary tumor surgery.

[0003] The diagnosis of cerebrospinal fluid leakage mainly relies on endoscopic direct observation, clinical manifestations, laboratory tests and imaging examinations. Among them, endoscopic direct observation requires a high level of operating skills and experience. During endoscopic operation, limited visual field may be encountered, especially in narrow or tortuous anatomical areas, which may affect the accuracy of observation. The results of endoscopic direct observation may be affected by the subjective judgment of the operator, and different doctors may have different interpretations. Clinical manifestations can directly reflect the patient's condition and remind doctors of potential problems, but the patient's description is often highly subjective, and some symptoms (such as headache and fever) are non-specific and may be caused by a variety of reasons. In addition, mild cerebrospinal fluid leakage may not produce typical clinical symptoms and is easily overlooked. Although laboratory tests can provide objective evidence, their accuracy will be affected when the leakage is small or intermittent. Contamination problems during sample collection, such as mixing with nasal secretions, saliva or other body fluids, may also reduce the accuracy of the test. In imaging examinations, especially magnetic resonance imaging (MRI) images can intuitively show structural abnormalities and are currently a commonly used method.

[0004] However, the prediction of cerebrospinal fluid leakage based on MRI images still faces multiple challenges. The existing technology lacks effective semantic segmentation and cannot accurately define the boundaries and size of pituitary tumors. In addition, the accuracy and sensitivity of cerebrospinal fluid leakage prediction based on MRI images need to be improved. Summary of the invention

[0005] The purpose of the present invention is to provide a semantic segmentation model for pituitary tumors and a prediction model for intraoperative cerebrospinal fluid leakage, so as to achieve effective semantic segmentation and improve the accuracy and sensitivity of cerebrospinal fluid leakage prediction based on MRI images.

[0006] A model for semantic segmentation of pituitary tumors and prediction of intraoperative cerebrospinal fluid leakage, including a multi-task deep neural network, a clinical data feature screening module, and a multi-layer perceptron;

[0007] The multi-task deep neural network is a multi-scale neural network. The multi-task deep neural network takes the preprocessed MRI image sequence as input, uses a convolution block or a Transformer block to extract the multi-scale features of the preprocessed MRI image sequence, and fuses the multi-scale features to obtain fused features. Based on the fused features, a pituitary tumor semantic segmentation head and a cerebrospinal fluid leakage classification head are constructed to obtain a semantic segmentation result of the pituitary tumor and a preliminary prediction result of the cerebrospinal fluid leakage.

[0008] The clinical data feature screening module is used to screen out target clinical data features using Pearson correlation coefficient and distance correlation number;

[0009] The multilayer perceptron takes the target clinical data features and the preliminary prediction results of cerebrospinal fluid leakage as input, and outputs the final prediction results of cerebrospinal fluid leakage.

[0010] According to the pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model provided by the present invention, a multi-task deep neural network is designed, and the MRI image sequence is used to simultaneously realize the semantic segmentation of the pituitary tumor and the preliminary identification of cerebrospinal fluid leakage. Through the semantic segmentation of the pituitary tumor, the boundary and size of the pituitary tumor can be accurately defined, which can help surgeons to formulate a more accurate surgical plan and reduce surgical risks. The semantic segmentation task can also be used to improve the performance of the cerebrospinal fluid leakage classification task; in addition, in order to improve the accuracy of cerebrospinal fluid leakage identification, the present application integrates the patient's target clinical data features on the basis of the preliminary prediction results of cerebrospinal fluid leakage output by the multi-task deep neural network, and constructs a multi-layer perceptron based on the fusion of imaging genomics and clinical data, which can learn deeper features, help improve the accuracy and sensitivity of cerebrospinal fluid leakage prediction, and reduce misdiagnosis and missed diagnosis of cerebrospinal fluid leakage. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1It is a structural schematic diagram of the pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model of the present invention. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0013] See also Figure 1 The pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model provided by the embodiment of the present invention includes a multi-task deep neural network, a clinical data feature screening module, and a multi-layer perceptron.

[0014] The input of the multi-task deep neural network is the patient's MRI image sequence, and the output is the results of two subtasks, namely the semantic segmentation results of pituitary tumors and the preliminary prediction results of cerebrospinal fluid leakage. The multi-task learning designed in this application allows the network to share underlying features, which can capture the intrinsic connection between different tasks and make the two tasks promote each other, so as to simultaneously improve the performance of the two subtasks of pituitary tumor segmentation and cerebrospinal fluid leakage classification.

[0015] Specifically, the multi-task deep neural network is a multi-scale neural network. The multi-task deep neural network takes the preprocessed MRI image sequence as input, uses a convolution block or a Transformer block to extract the multi-scale features of the preprocessed MRI image sequence, and fuses the multi-scale features to obtain fused features. Based on the fused features, a pituitary tumor semantic segmentation head and a cerebrospinal fluid leakage classification head are constructed to obtain the semantic segmentation results of the pituitary tumor and the preliminary prediction results of the cerebrospinal fluid leakage.

[0016] In practical applications, imaging technicians and surgeons often set the number of layers of MRI image sequences based on specific clinical needs and research purposes. It is difficult for deep neural networks to learn consistent feature representations from data with diverse numbers of layers. This may lead to poor generalization of neural networks, especially when testing new data. Therefore, in order to improve the generalization of neural networks, it is necessary to standardize the training data and test data. Specifically, in the multi-task deep neural network, the format of MRI image sequences with different numbers of layers is standardized by trilinear interpolation, and then a scaling operation is performed to further improve the generalization of cerebrospinal fluid leakage classification, thereby obtaining a preprocessed MRI image sequence.

[0017] The multi-task deep neural network takes the preprocessed MRI image sequence as input, and uses a convolutional block (such as a 3D U-Net network, a SegResNet network, or a 3D UX-Net network) or a Transformer block (such as a Swin UNETR network) to extract multi-scale features of the preprocessed MRI image sequence. In this embodiment, the multi-task deep neural network takes the preprocessed MRI image sequence as input, and uses the Swin UNETR network to construct four feature extraction modules of different scales, which are respectively used to extract features ,feature ,feature ,feature ,feature The dimension is ,feature The dimension is ,feature The dimension is ,feature The dimension is In addition, convolution operation is used to extract features from the preprocessed MRI image sequence ,feature The dimension is , where H and W represent the height and width of the image in the preprocessed MRI image sequence, respectively, and C represents the number of layers of the preprocessed MRI image sequence;

[0018] The multi-task deep neural network uses sampling and addition operations to fuse features ,feature ,feature ,feature ,feature , get the features ,feature ,feature ,feature ,feature , the expression is as follows:

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] Among them, US and DS represent 2x upsampling operation and 2x downsampling operation respectively;

[0025] Next, the features ,feature ,feature ,feature ,feature Convolution operations are used to obtain features ,feature ,feature ,feature ,feature , then, for the feature ,feature ,feature ,feature ,feature Use convolution operation to get features ,feature ,feature ,feature ,feature , and then use sampling and addition operations to fuse features ,feature ,feature ,feature ,feature , get the features ,feature ,feature ,feature ,feature Finally, for the feature ,feature ,feature ,feature ,feature Use convolution operation to get the fused features , the fused features , the fused features , the fused features , the fused features .

[0026] In this embodiment, the total loss L of the multi-task deep neural network satisfies the following conditional formula:

[0027] ;

[0028] Among them, L 0 , L 1 , L 2 , L 3 and L 4 Represent the fused features respectively , the fused features , the fused features , the fused features And the fused features The training loss on the corresponding branch, , , , and Respectively represent L 0 , L 1 , L 2 , L 3 and L 4 The corresponding weight.

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] in, Represents the fused features The semantic segmentation loss of the branch, Represents the fused features The cerebrospinal fluid leakage classification loss of the branch, represents the BCE loss, represents the Dice loss, represents the weighted cross entropy loss, represents the focal loss, Indicates i The predicted labels for CSF leakage classification of the preprocessed MRI image sequences, N represents the total number of preprocessed MRI image sequences, Indicates i The loss weight of the category to which the sample belongs, Indicates i The true labels of cerebrospinal fluid leakage classification for the preprocessed MRI image sequences;

[0034] When i When the sample belongs to the positive class, The value of ; When the i When the sample belongs to the negative class, The value of ,and =1, represents the loss weight corresponding to the positive sample, Represents the loss weight corresponding to the negative class samples;

[0035] in, Satisfy the following formula:

[0036] ;

[0037] in, represents the number of samples corresponding to patients with cerebrospinal fluid leakage, represents the number of samples corresponding to patients without cerebrospinal fluid leakage, and , represents a hyperparameter.

[0038] Similarly, we can calculate L 1 , L 2 , L 3 and L 4 .

[0039] In the actual research and application of intraoperative cerebrospinal fluid leakage prediction, high-quality annotation needs to be completed manually by experienced physicians or experts, which is a time-consuming and costly process. Therefore, there are usually fewer MRI image sequences with manually annotated information. In addition, due to the low incidence of cerebrospinal fluid leakage, the proportion of positive examples (cerebrospinal fluid leakage) in the training data of MRI image sequence samples is very low, that is, the class is unbalanced. Class imbalance may cause the machine learning model to tend to predict the majority class (ie, non-cerebrospinal fluid leakage), thereby reducing the recognition ability of the minority class (ie, cerebrospinal fluid leakage). The cerebrospinal fluid leakage classification loss designed in this application By weighted cross entropy loss And focal loss The weighted cross entropy loss and focal loss are mainly used to solve the problem of class imbalance, that is, to prevent the prediction result of cerebrospinal fluid leakage from being too biased towards the negative class (no cerebrospinal fluid leakage). The core idea of ​​focal loss is to reduce the weight of easy-to-classify samples and increase the weight of difficult-to-classify samples, so as to improve the model's recognition ability for minority classes. The weighted cross entropy loss gives different weights to samples of different categories. The samples of the minority class (corresponding to the positive class in this application) have a greater weight than the samples of the majority class (corresponding to the negative class in this application), so that the multi-task network pays more attention to the positive class in the classification of cerebrospinal fluid leakage.

[0040] The clinical data feature screening module is used to screen out target clinical data features using the Pearson correlation coefficient and the distance correlation number.

[0041] Since there is less training data for cerebrospinal fluid leakage, a high feature dimension can easily lead to model overfitting. Therefore, in order to improve the generalization ability of the model, this application adopts a sampling feature selection algorithm with features that have low correlation with cerebrospinal fluid leakage. The Pearson correlation coefficient is a classic method to measure the relationship between features and target variables. In some cases, even if the Pearson correlation coefficient is 0, there may be a complex nonlinear relationship between the two variables, and the distance correlation coefficient can capture this relationship.

[0042] Specifically, the clinical data feature screening module calculates the Pearson correlation coefficient between each clinical data feature and the true label of cerebrospinal fluid leakage , and the distance correlation coefficient between each clinical data feature and the true label of cerebrospinal fluid leakage ;

[0043] Will ,and The clinical data features of the target are eliminated to screen out the target clinical data features. Represents the coefficient threshold.

[0044] In this embodiment, the feature type and coding form of the screened target clinical data features are:

[0045] Gender: binary coding, male is 0, female is 1;

[0046] Age: normalized to a continuous value between 0 and 1;

[0047] Blood pressure: normalized to a continuous value between 0 and 1;

[0048] Visual acuity: visual acuity values ​​above 0.8 are marked as 0, visual acuity values ​​between 0.4 and 0.8 are marked as 0.5, and visual acuity values ​​less than 0.4 are marked as 1;

[0049] Headache: Use binary coding;

[0050] Whether acromegaly: using binary coding;

[0051] Whether the thyroid function is normal: using binary coding;

[0052] Whether the eyeball is moving; using binary coding;

[0053] Whether the tumor resection rate is greater than 95%: binary coding is used;

[0054] Knosp grading: Knosp grading of 0-2 is marked as 0, and Knosp grading of 3-4 is marked as 1;

[0055] Hardy classification: According to the invasion range and cytological characteristics of the tumor, invasive pituitary tumors are divided into four levels, namely subdural, intradural, skull base invasion and extraperiosteal invasion. These four levels are encoded in the form of one-hot coding;

[0056] The maximum diameter of the tumor in each direction in three dimensions: each diameter value is normalized to a continuous value between 0 and 1;

[0057] Whether the tumor invades the cavernous sinus: binary coding is used;

[0058] Whether the tumor invades the dorsum sella: binary coding is used;

[0059] Whether the tumor invades the sphenoid sinus: binary coding is used;

[0060] Whether the tumor invades the clivus: binary coding is used;

[0061] Whether the tumor is lobed: using binary coding.

[0062] The multilayer perceptron takes the target clinical data features and the preliminary prediction results of cerebrospinal fluid leakage as input, and outputs the final prediction results of cerebrospinal fluid leakage.

[0063] The hidden layer of the multilayer perceptron uses the ReLU function as the activation function, and the output layer of the multilayer perceptron uses the Sigmoid function as the activation function;

[0064] In the multilayer perceptron, the hidden layers are L Layer, for the connection from the input layer to the first hidden layer, it is expressed as follows:

[0065] ;

[0066] in, The dimension is The result of the forward propagation of the hidden layer nodes is represents the number of neurons in the first hidden layer; The dimension is The input vector contains c Target clinical data characteristics and 1 preliminary prediction result of cerebrospinal fluid leakage; The dimension is The weight matrix of The dimension is The bias vector of

[0067] For Hidden layers to The connection of the hidden layer is expressed as follows:

[0068] ;

[0069] in, The dimension is The result of the forward propagation of the hidden layer nodes is Indicates The number of neurons in the hidden layer, It is The output vector of the hidden layer; The dimension is The weight matrix of Indicates The number of neurons in the hidden layer; The dimension is The bias vector of

[0070] For the connection from the hidden layer to the output layer, it is expressed as follows:

[0071] ;

[0072] in, The dimension is The result of the forward propagation of the output layer nodes is It is L The output vector of the hidden layer; The dimension is The weight matrix of Indicates L The number of neurons in the hidden layer; The dimension is The bias vector of .

[0073] In summary, according to the pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model provided by the present invention, a multi-task deep neural network is designed, and the MRI image sequence is used to simultaneously realize the semantic segmentation of pituitary tumors and the preliminary identification of cerebrospinal fluid leakage. Through the semantic segmentation of pituitary tumors, the boundaries and sizes of pituitary tumors can be accurately defined, which can help surgeons formulate more accurate surgical plans and reduce surgical risks. The semantic segmentation task can also be used to improve the performance of the cerebrospinal fluid leakage classification task; in addition, in order to improve the accuracy of cerebrospinal fluid leakage identification, this application integrates the patient's target clinical data features on the basis of the preliminary prediction results of cerebrospinal fluid leakage output by the multi-task deep neural network, and constructs a multi-layer perceptron based on the fusion of imaging genomics and clinical data, which can learn deeper features, help improve the accuracy and sensitivity of cerebrospinal fluid leakage prediction, and reduce misdiagnosis and missed diagnosis of cerebrospinal fluid leakage.

[0074] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0075] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A model for semantic segmentation of pituitary tumors and prediction of intraoperative cerebrospinal fluid leakage, characterized in that: Including multi-task deep neural network, clinical data feature screening module, multi-layer perceptron; The multi-task deep neural network is a multi-scale neural network. The multi-task deep neural network takes the preprocessed MRI image sequence as input, uses a convolution block or a Transformer block to extract the multi-scale features of the preprocessed MRI image sequence, and fuses the multi-scale features to obtain fused features. Based on the fused features, a pituitary tumor semantic segmentation head and a cerebrospinal fluid leakage classification head are constructed to obtain a semantic segmentation result of the pituitary tumor and a preliminary prediction result of the cerebrospinal fluid leakage. The clinical data feature screening module is used to screen out target clinical data features using Pearson correlation coefficient and distance correlation number; The multilayer perceptron takes the target clinical data features and the preliminary prediction results of cerebrospinal fluid leakage as input, and outputs the final prediction results of cerebrospinal fluid leakage.

2. The pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model according to claim 1, characterized in that: In the multi-task deep neural network, the formats of MRI image sequences of different layers are standardized by trilinear interpolation, and then a scaling operation is performed to obtain a preprocessed MRI image sequence; The multi-task deep neural network takes the preprocessed MRI image sequence as input and uses the Swin UNETR network to construct four feature extraction modules of different scales, which are used to extract features. ,feature ,feature ,feature ,feature The dimension is ,feature The dimension is ,feature The dimension is ,feature The dimension is In addition, convolution operation is used to extract features from the preprocessed MRI image sequence ,feature The dimension is , where H and W represent the height and width of the image in the preprocessed MRI image sequence, respectively, and C represents the number of layers of the preprocessed MRI image sequence; The multi-task deep neural network uses sampling and addition operations to fuse features ,feature ,feature ,feature ,feature , get the features ,feature ,feature ,feature ,feature , the expression is as follows: ； ； ； ； ； Among them, US and DS represent 2x upsampling operation and 2x downsampling operation respectively; Next, the features ,feature ,feature ,feature ,feature Convolution operations are used to obtain features ,feature ,feature ,feature ,feature , then, for the feature ,feature ,feature ,feature ,feature Use convolution operation to get features ,feature ,feature ,feature ,feature , and then use sampling and addition operations to fuse features ,feature ,feature ,feature ,feature , get the features ,feature ,feature ,feature ,feature Finally, for the feature ,feature ,feature ,feature ,feature Use convolution operation to get the fused features , the fused features , the fused features , the fused features , the fused features .

3. The pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model according to claim 2, characterized in that: The total loss L of the multi-task deep neural network satisfies the following conditional formula: ； Among them, L0, L1, L2, L3 and L4 represent the fused features respectively. , the fused features , the fused features , the fused features And the fused features The training loss on the corresponding branch, , , , and Represent the weights corresponding to L0, L1, L2, L3 and L4 respectively; ； ； ； ； in, Represents the fused features The semantic segmentation loss of the branch, Represents the fused features The cerebrospinal fluid leakage classification loss of the branch, represents the BCE loss, represents the Dice loss, represents the weighted cross entropy loss, represents the focal loss, Indicates i The predicted labels for CSF leakage classification of the preprocessed MRI image sequences, N represents the total number of preprocessed MRI image sequences, Indicates i The loss weight of the category to which the sample belongs, Indicates i The true labels of cerebrospinal fluid leakage classification for the preprocessed MRI image sequences; When i When the sample belongs to the positive class, The value of ; When the i When the sample belongs to the negative class, The value of ,and =1, represents the loss weight corresponding to the positive sample, Represents the loss weight corresponding to the negative class samples; in, Satisfy the following formula: ； in, represents the number of samples corresponding to patients with cerebrospinal fluid leakage, represents the number of samples corresponding to patients without cerebrospinal fluid leakage, and , represents a hyperparameter.

4. The pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model according to claim 3, characterized in that: The clinical data feature screening module calculates the Pearson correlation coefficient between each clinical data feature and the true label of cerebrospinal fluid leakage , and the distance correlation coefficient between each clinical data feature and the true label of cerebrospinal fluid leakage ; Will ,and The clinical data features of the target are eliminated to screen out the target clinical data features. Represents the coefficient threshold.

5. The pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model according to claim 4, characterized in that: The characteristic type and coding form of the target clinical data characteristics are: Gender: binary coding, male is 0, female is 1; Age: normalized to a continuous value between 0 and 1; Blood pressure: normalized to a continuous value between 0 and 1; Visual acuity: visual acuity values ​​above 0.8 are marked as 0, visual acuity values ​​between 0.4 and 0.8 are marked as 0.5, and visual acuity values ​​less than 0.4 are marked as 1; Headache: Use binary coding; Whether acromegaly: using binary coding; Whether the thyroid function is normal: using binary coding; Whether the eyeball is moving; using binary coding; Whether the tumor resection rate is greater than 95%: binary coding is used; Knosp grading: Knosp grading of 0-2 is marked as 0, and Knosp grading of 3-4 is marked as 1; Hardy classification: According to the invasion range and cytological characteristics of the tumor, invasive pituitary tumors are divided into four levels, namely subdural, intradural, skull base invasion, and extraperiosteal invasion. These four levels are encoded in the form of one-hot coding; The maximum diameter of the tumor in each direction in three dimensions: each diameter value is normalized to a continuous value between 0 and 1; Whether the tumor invades the cavernous sinus: binary coding is used; Whether the tumor invades the dorsum sella: binary coding is used; Whether the tumor invades the sphenoid sinus: binary coding is used; Whether the tumor invades the clivus: binary coding is used; Whether the tumor is lobed: using binary coding.

6. The pituitary tumor semantic segmentation and intraoperative cerebrospinal fluid leakage prediction model according to claim 5, characterized in that: The hidden layer of the multilayer perceptron uses the ReLU function as an activation function, and the output layer of the multilayer perceptron uses the Sigmoid function as an activation function; In the multilayer perceptron, the hidden layers are L Layer, for the connection from the input layer to the first hidden layer, it is expressed as follows: ； in, The dimension is The result of the forward propagation of the hidden layer nodes is represents the number of neurons in the first hidden layer; The dimension is The input vector contains c Target clinical data characteristics and 1 preliminary prediction result of cerebrospinal fluid leakage; The dimension is The weight matrix of The dimension is The bias vector of For Hidden layers to The connection of the hidden layer is expressed as follows: ； in, The dimension is The result of the forward propagation of the hidden layer nodes is Indicates The number of neurons in the hidden layer, It is The output vector of the hidden layer; The dimension is The weight matrix of Indicates The number of neurons in the hidden layer; The dimension is The bias vector of For the connection from the hidden layer to the output layer, it is expressed as follows: ； in, The dimension is The result of the forward propagation of the output layer nodes is It is L The output vector of the hidden layer; The dimension is The weight matrix of Indicates L The number of neurons in the hidden layer; The dimension is The bias vector of .

Images