Lymphoma prognosis prediction method and device based on PET-CT marker and medium

Abstract

The invention relates to a lymphoma prognosis prediction method and device based on a PET-CT marker and a medium, and the method comprises the following steps: obtaining a PET / CT image of a patient, inputting the PET / CT image into a lymphoma prognosis prediction model based on image-clinical-molecular typing, and obtaining a prognosis risk category; the lymphoma prognosis prediction model based on image-clinical-molecular typing is constructed according to PET-CT markers, clinical indexes and genetic typing data, the PET-CT markers are obtained by calculating a segmentation result of a PET / CT image of a patient according to a bimodal lesion segmentation network, and prognosis significance verification is performed through a K-M survival curve; the bimodal lesion segmentation network comprises an encoder and a decoder, the encoder performs feature extraction through a plurality of convolutional layers, and a channel smoothing layer is introduced after the last convolutional layer to fuse the extracted features. Compared with the prior art, tumor information can be accurately captured from PET / CT images, and lymphoma prognosis prediction is carried out in combination with genomics and clinical medicine.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a lymphoma prognosis prediction method, device and medium based on PET-CT markers. Background Art

[0002] Diffuse large B-cell lymphoma (DLBCL) is the most common pathological subtype of non-Hodgkin's lymphoma (NHL). 25% to 40% of patients progress or relapse within one year after diagnosis. How to identify patients with poor prognosis early is a clinical problem that needs to be solved urgently. PET (positron emission tomography) / CT (computed tomography) integrates functional imaging and anatomical imaging. It can not only display anatomical information such as the morphology, size, and density of the lesion, but also provide functional information such as lesion metabolism and blood flow. It can effectively predict the poor prognosis of DLBCL patients and has important application value in the diagnosis and staging of benign and malignant tumor lesions, guiding treatment, and prognosis judgment.

[0003] With the continuous development of artificial intelligence, deep learning algorithms have made breakthrough progress in the field of medical image segmentation. The paper Automated Abdominal Segmentation of CT Scans for Body Composition Analysis Using Deep Learning (WESTON AD, KORFIATIS P, KLINE TL, et al., Radiology, 2019, 290 (3): 669-79) and the paper Fully-Automated Segmentation of Nasopharyngeal Carcinoma on Dual-Sequence MRI Using Convolutional Neural Networks (YE Y, CAI Z, HUANG B, et al., Front Oncol, 2020, 10: 166) have confirmed that the fully convolutional neural network (FCN) based on the U-Net architecture can achieve high-precision segmentation in CT and MRI images. However, its application in PET / CT images of DLBCL patients to automatically segment large-scale, diffuse lesions and extract corresponding indicators needs further verification. The Chinese invention patent with publication number CN112070809B proposes a precise diagnosis system for pancreatic cancer based on PET / CT dual-time imaging. It uses the significant differences in the uptake dynamics of inflammatory and malignant lesions to extract dual-time multimodal imaging features, maximum density projection features, retention coefficient features, and clinical features, ultimately improving the accuracy of clinical diagnosis and assisting doctors in quantitative analysis. However, due to the wide range of DLBCL lesions and their varying shapes and sizes, as well as the low resolution and high noise of PET images, how to conveniently and accurately extract relevant indicators from the PET / CT images of DLBCL patients to predict prognosis remains an extremely challenging task. Therefore, it is urgent to design a lymphoma prognosis prediction method that accurately captures tumor information from PET / CT images and incorporates it into clinical risk and prognosis assessments to more accurately distinguish high-risk patients who are ineffective for standard treatment. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a lymphoma prognosis prediction method, device and medium based on PET-CT markers, accurately capture tumor information from PET / CT images, and incorporate it into clinical risk and prognosis assessment to more accurately distinguish high-risk patients.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The present invention provides a method for predicting the prognosis of lymphoma based on PET-CT markers, comprising the following steps:

[0007] Obtain the patient's PET / CT images, input them into the lymphoma prognosis prediction model based on imaging-clinical-molecular classification, and obtain the prognostic risk category;

[0008] Among them, the lymphoma prognosis prediction model based on imaging-clinical-molecular classification is constructed according to PET-CT markers, clinical indicators and genotyping data. The PET-CT markers include multiple indicators, each of which is calculated based on the segmentation results of the patient's PET / CT images by the dual-modal lesion segmentation network, and the prognostic significance is verified by the KM survival curve; the dual-modal lesion segmentation network is built based on the nnU-net structure, including an encoder and a decoder connected by jump connections. The encoder gradually downsamples through multiple convolutional layers to extract features, and a channel smoothing layer is introduced after the last convolutional layer to fuse the extracted features.

[0009] Furthermore, in the bimodal lesion segmentation network, each convolutional layer sequentially performs convolution operation, batch normalization operation and ReLU activation operation on the output of the previous convolutional layer.

[0010] Furthermore, in the bimodal lesion segmentation network, the channel smoothing layer performs convolution operation, batch normalization operation and ReLU activation operation on the output of the last convolution layer in sequence.

[0011] Furthermore, when training the bimodal lesion segmentation network, the network parameters are optimized using a gradient descent method with the goal of minimizing a loss function, and the loss function is constructed based on a cross entropy loss.

[0012] Furthermore, the PET-CT markers include maximum standard uptake value SUV max , maximum lesion tumor metabolic volume MAX_MTV, total metabolic tumor volume TMTV, total lesion metabolic volume TLG, maximum distance of homogenization SDmax patient and the maximum distance to the homogenized maximum lesion SDmax bulk The segmentation result of the dual-modality lesion segmentation network is all tumor lesion areas in the input PET / CT image. The formula for calculating the PET-CT marker based on the segmentation result is as follows:

[0013] Maximum Standard Uptake Value SUV max :

[0014] SUV max =max(V i ) / (DOS*Weight)

[0015] Among them, V i is the area of ​​the ith tumor lesion, max(V i ) is the maximum uptake activity, DOS is the dose of radiolabel injected into the patient, and Weight is the patient's weight;

[0016] Maximum lesion tumor metabolic volume MAX_MTV:

[0017] MAX_MTV=max(MTV)

[0018]

[0019] Among them, MTV is the metabolic volume of a single lesion, spacing is the pixel spacing, and N is the number of voxels in the ROI;

[0020] Total metabolic tumor volume TMTV:

[0021]

[0022] Among them, M is the number of lesions;

[0023] Total lesion metabolic volume TLG:

[0024] TLG=MTV×SUVmean

[0025] Among them, SUVmean is the mean SUV value within a single tumor lesion;

[0026] Maximum distance of uniformity SDmax patient :

[0027] SDmax patient =Standarized*D max

[0028]

[0029] D max =max(D1, D2, ..., D M*M-1 )

[0030]

[0031] Among them, Standarized is the standardized coefficient, D max is the maximum distance between lesions, D is the Euclidean distance between two lesions with three-dimensional coordinates (x1, y1, z1) and (x2, y2, z2), and Weight and Height are the weight and height of the patient, respectively;

[0032] The maximum distance of the homogenized maximum lesion SDmax bulk :

[0033] SDmax bulk =Standarized*D M-max

[0034]

[0035] Among them, D M-max is the maximum distance between the largest lesion and another lesion.

[0036] Furthermore, the specific process of verifying the prognostic significance of the PET-CT markers by using the KM survival curve is as follows:

[0037] Obtain lymphoma patient sample data and divide the data set through multiple K-fold cross-validation;

[0038] According to the division results, the average optimal cutoff value of each indicator in the PET-CT marker is obtained based on the Youden index, and each indicator is divided into a high group and a low group;

[0039] KM survival curves were drawn according to the grouped PET-CT markers, and the prognostic significance of each indicator was verified by logrank test.

[0040] Furthermore, clinical indicators, genotyping data and PET-CT markers are input into the Lasso-cox model to construct the lymphoma prognosis prediction model based on imaging-clinical-molecular typing. The lymphoma prognosis prediction model based on imaging-clinical-molecular typing obtains independent risk factors from the input through regression analysis, and scores the independent risk factors based on the z-score in the cox regression analysis, calculates the patient's prognostic score, and divides the prognostic risk category accordingly and outputs it.

[0041] Furthermore, the PET-CT markers include maximum standard uptake value SUV max , maximum lesion tumor metabolic volume MAX_MTV, total metabolic tumor volume TMTV, total lesion metabolic volume TLG, maximum distance of homogenization SDmax patient and the maximum distance to the homogenized maximum lesion SDmax bulk ;

[0042] The clinical indicators include multiple clinical factors in the International Prognostic Index (IPI), specifically age>60, elevated breast dehydrogenase, decreased general physical status, clinical stage III-IV, and extranodal invasion sites>2;

[0043] The genotyping data include BN2-like, MCD-like, TP53 Mut, EZB-like-MYC+, EZB-like-MYC-, ST2-like, N1-like, and NOS typing;

[0044] The independent risk factors screened by regression analysis included increased total metabolic tumor volume TMTV, normalized maximum distance SDmax patient Elevated, elevated mammary dehydrogenase, EZB-like-MYC+, MCD-like and TP53 Mut Typing.

[0045] The present invention also provides an electronic device, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the above method when executing the program.

[0046] The present invention also provides a computer-readable storage medium on which a computer program is stored, and the program implements the above method when executed by a processor.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The present invention proposes a lymphoma prognosis prediction method based on PET-CT markers, which outputs a prognostic risk category according to the patient's PET / CT images through a lymphoma prognosis prediction model based on imaging-clinical-molecular typing; wherein the lymphoma prognosis prediction model based on imaging-clinical-molecular typing is constructed according to PET-CT markers, clinical indicators and genotyping data, and integrates medical image processing, genomics and clinical medicine, which can improve the comprehensiveness and accuracy of prognosis prediction; the PET-CT markers include multiple indicators, which are calculated based on the segmentation results of the patient's PET / CT images by the dual-modal lesion segmentation network. The accessibility of imaging indicators in actual clinical decision-making is improved. The dual-modal lesion segmentation network is built based on the nnU-net structure, including an encoder and a decoder connected by jump connections. The encoder gradually downsamples through multiple convolutional layers to extract features. A channel smoothing layer is introduced after the last convolutional layer to fuse the extracted features, which can fully exploit the long sequence scans of the input data and improve the accuracy of lesion segmentation. In addition, the prognostic significance of PET-CT markers is verified by the KM survival curve, which can identify the key factors affecting the patient's prognosis, which is of great significance for guiding clinical treatment and improving the patient's prognosis.

[0049] 2. The present invention divides the acquired lymphoma patient sample data into data sets through multiple K-fold cross-validation, obtains the average optimal cutoff value based on the division results, and verifies its prognostic significance. This can overcome the high variability of PET-CT markers to a certain extent and eliminate the problem of inaccurate and imprecise verification of prognostic significance after a single division of the data set in the past. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2 Schematic diagram of the structure of extracting PET-CT markers based on the dual-modality lesion segmentation network;

[0052] Figure 3 Schematic diagram to verify the prognostic significance of PET-CT markers;

[0053] Figure 4 This is a schematic diagram of the validation results of the prognostic significance of PET-CT markers.

[0054] Among them, (4a), (4b), (4c), (4d), (4e), and (4f) are the maximum standard uptake values ​​SUV max , maximum lesion tumor metabolic volume MAX_MTV, total metabolic tumor volume TMTV, total lesion metabolic volume TLG, maximum distance of homogenization SDmax patient and the maximum distance to the homogenized maximum lesion SDmax bulk The experimental results on the training and test sets, (4g), (4h), (4i), (4j), (4k), (4l) are the maximum standard uptake value SUV max , maximum lesion tumor metabolic volume MAX_MTV, total metabolic tumor volume TMTV, total lesion metabolic volume TLG, maximum distance of homogenization SDmax patient and the maximum distance to the homogenized maximum lesion SDmax bulk Experimental results on the validation set;

[0055] Figure 5 This is an experimental result diagram of the prognostic effect of the lymphoma prognosis prediction model based on imaging-clinical-molecular classification.

[0056] Among them, (5a) and (5b) are the experimental results of the lymphoma prognosis prediction model based on imaging-clinical-molecular classification and the international prognostic index IPI on the training and test sets, respectively; (5c) and (5d) are the experimental results of the lymphoma prognosis prediction model based on imaging-clinical-molecular classification and the international prognostic index IPI on the training and validation set 1, respectively; (5e) and (5f) are the experimental results of the lymphoma prognosis prediction model based on imaging-clinical-molecular classification and the international prognostic index IPI on the training and validation set 2, respectively. DETAILED DESCRIPTION

[0057] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0058] Example:

[0059] This embodiment provides a method for predicting lymphoma prognosis based on PET-CT markers, such as Figure 1 As shown, the following steps are included:

[0060] S1. Build and train a dual-modal lesion segmentation network.

[0061] Based on nnU- ne t structure to build a dual-modal lesion segmentation network, such as Figure 2 As shown, it includes an encoder and a decoder. The encoder downsamples step by step through n convolutional layers. Symmetrically, the decoder upsamples step by step through n convolutional layers. The layers of the encoder and the decoder are connected by jump connections.

[0062] In the encoder, each convolutional layer performs convolution operation (Conv), batch normalization operation (BatchNorm) and ReLU activation operation on the output of the previous convolutional layer in sequence. For the i-th layer encoder, its output feature map can be expressed as:

[0063] E i =ReLU(BatchNorm(Conv(E i-1 ))), i = 0, 1, 2

[0064] In this embodiment, a channel smoothing layer is introduced after the last convolutional layer to fuse the extracted features to capture global information. Its output feature map can be expressed as:

[0065] M = ReLU(BatchNorm(Conv(E k )))

[0066] Here, k is the index of the last convolutional layer of the encoder.

[0067] The feature map of a certain layer of the encoder is connected to the feature map of the decoder layer at a symmetrical position through a jump connection, which can be expressed as:

[0068] U i =Concatenate(Conv(D i-1 , E i-1 ))

[0069] For the i-th layer decoder, its output feature map can be expressed as:

[0070] Di =ReLU(BatchNorm(Com([D i-1 , U i-1 ])))

[0071] The last layer of the decoder uses a convolutional layer to produce the final segmentation output, usually using a sigmoid or softmax activation function to generate pixel-level segmentation predictions. For a binary classification problem, the output layer can be expressed as:

[0072] O = Sigmoid(Conv(D n ))

[0073] When training the above-mentioned dual-modal lesion segmentation network, the goal is to minimize the loss function, and the gradient descent method is used to optimize the network parameters. The loss function L is constructed based on the cross entropy loss:

[0074]

[0075] Among them, N is the number of samples, C is the number of categories, yi,c is the actual label, is the predicted value of the model.

[0076] This step uses the channel smoothing layer method, which can fully mine the long sequence scans of the input data and ultimately achieve the effect of improving the classification accuracy. Subsequent experimental verification shows that the accuracy of this method is better than the mainstream algorithm on the AutoPET public dataset and is more suitable for lymphoma lesion segmentation.

[0077] S2. Input the patient's PET / CT image into the dual-modality lesion segmentation network and calculate the PET-CT markers based on the segmentation results.

[0078] The segmentation results of the dual-modality lesion segmentation network are all tumor lesion areas in the input PET / CT image. PET-CT markers include the maximum standard uptake value SUV max , maximum lesion tumor metabolic volume MAX_MTV, total metabolic tumor volume TMTV, total lesion metabolic volume TLG, maximum distance of homogenization SDmax patient and the maximum distance to the homogenized maximum lesion SDmax bulk The formula for calculating PET-CT markers based on the segmentation results is as follows:

[0079] (1) Maximum Standard Uptake Value (SUV) max

[0080] Maximum Standard Uptake Value SUV max It is a commonly used parameter in PET-CT to describe the highest uptake activity of tumors or tissues:

[0081] SUVmax =max(V i ) / (DOS*Weight)

[0082] Among them, V i is the area of ​​the ith tumor lesion, max(V i ) is the maximum uptake activity, DOS is the dose of radioactive marker injected into the patient, and Weight is the patient's weight.

[0083] (2) Maximum lesion tumor metabolic volume MAX_MTV

[0084] Metabolic Tumor Volume (MTV) is an indicator used to measure tumor volume in PET-CT:

[0085]

[0086] MAX_MTV=max(MTV)

[0087] Among them, MTV is the metabolic volume of a single lesion, spacing is the pixel spacing, and N is the number of voxels in the ROI. The above formula converts the volume of each voxel x, y, and z of the lesion identified by the segmentation model from pixel space to voxel physical space to obtain the tumor size.

[0088] (3) Total metabolic tumor volume TMTV

[0089]

[0090] Where M is the number of lesions.

[0091] (4) Total lesion metabolic volume TLG

[0092] Total lesion glycolysis (TLG) is a metric that takes into account both the volume and metabolic activity of lesions identified on PET scans and can provide a more comprehensive assessment than single metrics such as maximum standardized uptake value (SUVmax) or metabolic tumor volume (MTV):

[0093] TLG=MTV×SUVmean

[0094] Among them, SUVmean is the mean SUV value within a single tumor lesion.

[0095] (5) Maximum distance of normalization SDmax patient

[0096] The Euclidean distance D between two lesions with three-dimensional coordinates (x1, y1, z1) and (x2, y2, z2) is:

[0097]

[0098] Calculate the distances between all possible lesions, the maximum distance D max for:

[0099] D max =max(D1, D2, ..., D M*M-1 )

[0100] Perform body surface homogenization to obtain the distance index after homogenization:

[0101]

[0102] SDmax patient =Standarized*D max

[0103] Among them, Standarized is the standardized coefficient (i.e. BSA), Weight and Height are the patient's weight and height, respectively.

[0104] (6) Maximum distance of the normalized maximum lesion SDmax bulk

[0105] SDmax bulk =Standarized*D M-max

[0106]

[0107] Among them, i is the largest index in the tumor lesion.

[0108] This step realizes the automated extraction of PET-CT markers related to lymphoma prognosis that have appeared in previous literature, improving the accessibility of imaging indicators in actual clinical decision-making.

[0109] S3. Verify the prognostic significance of PET-CT markers by KM survival curve, such as Figure 3 As shown, the specific process is as follows:

[0110] S301. In this embodiment, 580 lymphoma patient sample data from a hospital are obtained, and the data set is divided through 100 10-fold cross validations.

[0111] S302. The time-dependent receiver operating characteristic curve (ROC) is used to calculate the average optimal cutoff value (Mean cutoff) of each indicator in the PET-CT marker based on the Youden index to divide each indicator into a high group and a low group.

[0112] The average optimal cutoff value is the best node for the continuous variable to appear when the ROC curve is used to convert the continuous variable into a categorical variable and reach the specified outcome variable. The value greater than this value is the high group, otherwise it is the low group.

[0113] S303. Draw a KM survival curve based on the grouped PET-CT markers, and verify the prognostic significance of each indicator through the logrank test.

[0114] KM survival curve was used to verify the prognostic significance of each marker. Logrank p<0.05 indicated that the indicator was an effective indicator for distinguishing progression-free survival (PFS), and the mean area under the ROC curve (Mean AUC) ≥0.6 indicated that the indicator performed better.

[0115] This step can overcome the high variability of PET-CT markers to a certain extent and eliminate the problem of inaccurate and imprecise prognostic significance verification after a single division of the data set.

[0116] S4. Construct a lymphoma prognosis prediction model based on imaging-clinical-molecular typing according to PET-CT markers, clinical indicators and genotyping data, and output the prognostic risk category.

[0117] The International Prognostic Index (IPI) is a recognized prognostic scoring system for DLBCL patients. The clinical indicators in this example are derived from the International Prognostic Index (IPI), including age>60, elevated breast dehydrogenase, decreased general physical condition, clinical stage III-IV, and extranodal invasion sites>2. LymphPlex molecular typing based on gene sequencing also plays an important role in the diagnosis and treatment of DLBCL. The genotyping data in this example include BN2-like, MCD-like, TP53 Mut , EZB-like-MYC+, EZB-like-MYC-, ST2-like, N1-like and NOS. Clinical indicators, genotyping data and PET-CT markers (including SUV max , MAX_MTV, TMTV, TLG, SDmax patient and SDmax bulk ) was input into the Lasso-cox model to construct a lymphoma prognosis prediction model based on imaging-clinical-molecular classification.

[0118] When the minimum contraction (λ) was reached, the optimal value of the model was obtained. Finally, 11 variables were screened, including 2 PET-CT markers (increased TMTV, SDmax patientThe 11 variables screened were included in the multivariate Cox regression analysis, and the results showed that TMTV increase, breast dehydrogenase increase, EZB-like-MYC+, MCD-like, TP53Mut, and NOS typing were independent risk factors for recurrence in DLBCL patients within 2 years.

[0119] Based on the z-score in the cox regression analysis, the risk factors were scored. The lowest z-score item was used as the reference (in this model, breast dehydrogenase elevation). The z-score of each other risk factor was compared with the reference, and the obtained comparison value was rounded to obtain the score of each risk factor (TMTV elevation -2 points, breast dehydrogenase elevation -1 point, EZB-like-MYC+ -1 point, MCD-like -1 point, TP53Mut typing -2 points). The prognostic score outputted based on this ranges from 0 to 5. According to the KM survival curve, if there is no statistical difference between the survival curves of different scores, they are merged into one group, otherwise they are grouped separately. The lymphoma prognosis prediction model based on imaging-clinical-molecular typing finally outputs 4 groups: 0: low risk, 1-2: medium-low risk, 3: medium-high risk, 4-5: high risk.

[0120] In order to verify the performance of the dual-modal lesion segmentation network constructed in step S1, this embodiment conducted experiments on the AutoPET public data set, and analyzed and compared the convolutional neural network method using a channel smoothing layer and the convolutional neural network method without a channel smoothing layer. The experimental results are shown in Table 1. The experiments were trained and tested in accordance with the experimental regulations of the corresponding data set. It can be seen from Table 1 that the convolutional neural network with a channel smoothing layer has a higher segmentation accuracy than the traditional convolutional neural network without a channel smoothing layer.

[0121] Table 1 Experimental data of AutoPET dataset

[0122]

[0123] To verify the prognostic significance of the PET-CT markers extracted in step S2, this embodiment conducted experiments on data from 580 patients in a hospital and verified it on data from 50 patients in another hospital. The results are as follows: Figure 4 As shown, it can be seen that the area under the curve of each indicator reached more than 0.6, indicating that the prognostic significance is better.

[0124] To verify the effect of the lymphoma prognosis prediction model based on imaging-clinical-molecular classification constructed in step S4, this experiment was conducted on data of 580 patients from a hospital, and the prognostic significance of the model was further verified based on data of 102 and 40 patients from different clinical studies (validation set 1 and validation set 2, respectively). The experimental results are as follows: Figure 5 and as shown in Table 2. The model predicted the progression-free survival time with an area under the ROC curve of 0.718, 0.648, and 0.826.

[0125] Table 2 Comparison of the lymphoma prognostic prediction model based on imaging-clinical-molecular classification and the International Prognostic Index (IPI)

[0126]

[0127] In Table 2, the area under the receiver operating characteristic curve A measure of the predictive power of the model, defined as the probability that the predicted survival is consistent with the actual survival. Values ​​of 0.5, 0.7, and 1.0 indicate that the model's ability to distinguish short-term and long-term survival times is completely random, acceptable, or perfect, respectively.

[0128] The Akaike Information Criterion* provides a relative measure of model fit; smaller values ​​indicate better model fit. A difference between models <2 indicates no significant improvement in fit, a difference >2 but <10 indicates some improvement in fit, and a difference ≥10 indicates a significant improvement in model fit.

[0129] If the above method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0130] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for predicting the prognosis of lymphoma based on PET-CT markers, characterized in that: The following steps are involved: Obtain the patient's PET / CT images, input them into the lymphoma prognosis prediction model based on imaging-clinical-molecular classification, and obtain the prognostic risk category; Among them, the lymphoma prognosis prediction model based on imaging-clinical-molecular classification is constructed according to PET-CT markers, clinical indicators and genotyping data. The PET-CT markers include multiple indicators, each of which is calculated based on the segmentation results of the patient's PET / CT images by the dual-modal lesion segmentation network, and the prognostic significance is verified by the KM survival curve; the dual-modal lesion segmentation network is built based on the nnU-net structure, including an encoder and a decoder connected by jump connections. The encoder gradually downsamples through multiple convolutional layers to extract features, and a channel smoothing layer is introduced after the last convolutional layer to fuse the extracted features.

2. The method for predicting lymphoma prognosis based on PET-CT markers according to claim 1, characterized in that: In the dual-modal lesion segmentation network, each convolutional layer sequentially performs convolution operation, batch normalization operation and ReLU activation operation on the output of the previous convolutional layer.

3. The method for predicting lymphoma prognosis based on PET-CT markers according to claim 1, characterized in that: In the dual-modal lesion segmentation network, the channel smoothing layer sequentially performs convolution operation, batch normalization operation and ReLU activation operation on the output of the last convolution layer.

4. The method for predicting lymphoma prognosis based on PET-CT markers according to claim 1, characterized in that: When training the bimodal lesion segmentation network, the network parameters are optimized by using a gradient descent method with the goal of minimizing a loss function, and the loss function is constructed based on a cross entropy loss.

5. The method for predicting lymphoma prognosis based on PET-CT markers according to claim 1, characterized in that: The PET-CT markers include maximum standardized uptake value SUV max , maximum lesion tumor metabolic volume MAX_MTV, total metabolic tumor volume TMTV, total lesion metabolic volume TLG, maximum distance of homogenization SDmax patient and the maximum distance to the homogenized maximum lesion SDmax bulk The segmentation result of the dual-modality lesion segmentation network is all tumor lesion areas in the input PET / CT image. The formula for calculating the PET-CT marker based on the segmentation result is as follows: Maximum Standard Uptake Value SUV max : SUV max =max(V i ) / (DOS*Weight) Among them, V i is the area of ​​the ith tumor lesion, max(V i ) is the maximum uptake activity, DOS is the dose of radiolabel injected into the patient, and Weight is the patient's weight; Maximum lesion tumor metabolic volume MAX_MTV: MAX_MTV=max(MTV) Among them, MTV is the metabolic volume of a single lesion, spacing is the pixel spacing, and N is the number of voxels in the ROI; Total metabolic tumor volume TMTV: Among them, M is the number of lesions; Total lesion metabolic volume TLG: TLG=MTV×SUVmean Among them, SUVmean is the mean SUV value within a single tumor lesion; Normalized maximum distance SDmax patient : SDmax patient =Standarized*D max D max =max(D1,D2,…,D M*M-1 ) Among them, Standarized is the standardized coefficient, D max is the maximum distance between lesions, D is the Euclidean distance between two lesions with three-dimensional coordinates (x1, y1, z1) and (x2, y2, z2), and Weight and Height are the weight and height of the patient, respectively; The maximum distance of the homogenized maximum lesion SDmax bulk : SDmax bulk =Standarized*D M-max Among them, D M-max is the maximum distance between the largest lesion and another lesion.

6. The method for predicting lymphoma prognosis based on PET-CT markers according to claim 1, characterized in that: The specific process of verifying the prognostic significance of the PET-CT markers by using the KM survival curve is as follows: Obtain lymphoma patient sample data and divide the data set through multiple K-fold cross-validation; According to the division results, the average optimal cutoff value of each indicator in the PET-CT marker is obtained based on the Youden index, and each indicator is divided into a high group and a low group; KM survival curves were drawn according to the grouped PET-CT markers, and the prognostic significance of each indicator was verified by logrank test.

7. The method for predicting lymphoma prognosis based on PET-CT markers according to claim 1, characterized in that: Clinical indicators, genotyping data and PET-CT markers are input into the Lasso-cox model to construct the lymphoma prognosis prediction model based on imaging-clinical-molecular typing. The lymphoma prognosis prediction model based on imaging-clinical-molecular typing obtains independent risk factors from the input through regression analysis, and scores the independent risk factors based on the z-score in the cox regression analysis, calculates the patient's prognostic score, and divides the prognostic risk category accordingly and outputs it.

8. The method for predicting lymphoma prognosis based on PET-CT markers according to claim 7, characterized in that: The PET-CT markers include maximum standardized uptake value SUV max , maximum lesion tumor metabolic volume MAX_MTV, total metabolic tumor volume TMTV, total lesion metabolic volume TLG, maximum distance of homogenization SDmax patient and the maximum distance to the homogenized maximum lesion SDmax bulk ; The clinical indicators include multiple clinical factors in the International Prognostic Index (IPI), specifically age>60, elevated breast dehydrogenase, decreased general physical status, clinical stage III-IV, and extranodal invasion sites>2; The genotyping data include BN2-like, MCD-like, TP53 Mut , EZB-like-MYC+, EZB-like-MYC-, ST2-like, N1-like, and NOS typing; The independent risk factors screened by regression analysis included increased total metabolic tumor volume TMTV, normalized maximum distance SDmax patient Elevated, elevated mammary dehydrogenase, EZB-like-MYC+, MCD-like and TP53 Mut Typing.

9. An electronic device comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images