A method, system, device and storage medium for predicting the prognosis of cervical cancer

Abstract

The application discloses a cervical cancer prognosis prediction method, system, device and storage medium, which is applied to the technical field of medical image processing and prognosis prediction, and the method comprises the following steps: acquiring a CT image sequence of a patient, and pre-processing the CT image sequence; segmenting the pre-processed CT image by using a trained segmentation model to obtain a segmentation mask of a tumor region; extracting comprehensive features from the pre-processed CT image based on the segmentation mask, and acquiring clinical features of the patient; inputting the comprehensive features and the clinical features into a trained prognosis prediction model to output a prognosis prediction result of the patient; and the application can accurately predict the disease-free survival period of a patient receiving radical radiotherapy by combining multi-modal data, so that the risk patients can be identified in the early stage of treatment, and targeted intensive treatment or unnecessary excessive treatment can be avoided.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and prognosis prediction technology, and in particular to a method, system, device and storage medium for predicting the prognosis of cervical cancer. Background Technology

[0002] Cervical cancer is a malignant tumor, and the standard treatment for locally advanced cervical cancer (LACC) is concurrent chemoradiotherapy (CCRT). However, some patients still experience recurrence or metastasis after treatment. Currently, prognostic assessment of locally advanced cervical cancer mainly relies on the FIGO staging system. Clinical trials (such as the KEYNOTE-A18 and OUTBACK trials) typically use FIGO staging to screen high-risk patients for intensive treatment. However, FIGO staging is primarily based on anatomical extent and cannot accurately reflect the intrinsic biological heterogeneity of the tumor and differences in treatment response.

[0003] Existing research attempts to construct prognostic models by combining clinicopathological features, but the following problems still exist: traditional techniques rely too much on pre-treatment static images, ignoring the radiosensitivity information contained in the image changes during or after treatment; manually delineating regions of interest is inefficient and subject to subjective differences; and the model has limited generalization ability due to its singular feature extraction, making it difficult to apply stably in multi-center clinical settings.

[0004] To overcome these shortcomings, this application proposes a method, system, device, and storage medium for predicting the prognosis of cervical cancer. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, device and storage medium for predicting the prognosis of cervical cancer, which aims to solve the problems of low prediction accuracy and poor clinical applicability of existing cervical cancer prognosis assessment methods due to their reliance on static images, manual delineation and single features.

[0006] To solve the above-mentioned technical problems, one technical solution adopted in this application is: to provide a method for predicting the prognosis of cervical cancer, the steps of which include:

[0007] Acquire the patient's CT image sequence and preprocess the CT image sequence;

[0008] The pre-processed CT image is segmented using a trained segmentation model to obtain a segmentation mask for the tumor region;

[0009] Based on the segmentation mask, comprehensive features are extracted from the preprocessed CT image, and the patient's clinical features are obtained.

[0010] The comprehensive features and the clinical features are input into the trained prognostic prediction model, and the prognostic prediction results for the patient are output.

[0011] As a further improvement to this application, the step of segmenting the preprocessed CT image using a trained segmentation model to obtain a segmentation mask for the tumor region specifically includes:

[0012] A 3D U-Net network model is used as the segmentation model. The 3D U-Net network model is based on an encoder and decoder structure and includes skip connections.

[0013] The preprocessed CT image is input into the 3D U-Net network model. After encoding and decoding the input CT image, the segmentation mask is output.

[0014] As a further improvement to this application, the step of extracting comprehensive features from the preprocessed CT image based on the segmentation mask and obtaining the patient's clinical features specifically includes:

[0015] The integrated features include depth features and radiomics features;

[0016] The preprocessed CT images and segmentation masks are input into a medical image model pre-trained by self-supervised learning to extract high-dimensional depth features before, after, or during treatment; the high-dimensional depth features are then subjected to dimensionality reduction processing using principal component analysis to obtain depth features.

[0017] Within the region of interest defined by the segmentation mask, initial radiomics features are extracted; the initial radiomics features are then filtered to obtain radiomics features relevant to prognosis.

[0018] As a further improvement to this application, the step of inputting the comprehensive features and the clinical features into the trained prognostic prediction model and outputting the patient's prognostic prediction result specifically includes:

[0019] The prognostic prediction model is a multilayer perceptron trained based on the DeepHit survival analysis framework.

[0020] The multilayer perceptron includes an input layer, a fusion layer, a hidden layer, and a prediction layer;

[0021] The deep features, radiomics features, and clinical features are input into the input layer and then concatenated through the fusion layer to obtain a fused feature vector.

[0022] The hidden layer maps the fused feature vector using a non-linear activation function; and learns the conditional probability distribution of patient survival time based on the DeepHit framework.

[0023] The prediction layer outputs the patient's individual risk score and the probability of disease-free survival at at least one preset time point in the future.

[0024] As a further improvement of this application, the CT image sequence includes: a first CT image sequence at a time point before treatment, and a second CT image sequence at a time point after treatment or during treatment.

[0025] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a cervical cancer prognosis prediction system, the system specifically including:

[0026] An image processing module is used to acquire the patient's CT image sequence and preprocess the CT image sequence;

[0027] The segmentation module is used to segment the preprocessed CT image using a trained segmentation model to obtain a segmentation mask for the tumor region.

[0028] The feature extraction module is used to extract comprehensive features from the preprocessed CT image based on the segmentation mask, and to obtain the patient's clinical features;

[0029] The prognosis prediction module is used to input the comprehensive features and the clinical features into the trained prognosis prediction model and output the prognosis prediction results for the patient.

[0030] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a computer device, the computer device including a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing a cervical cancer prognosis prediction method; the processor is used to execute the program instructions stored in the memory to implement a cervical cancer prognosis prediction.

[0031] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a computer-readable storage medium storing processor-executable program instructions, the program instructions being used to execute a cervical cancer prognosis prediction method.

[0032] This application provides a method, system, device, and storage medium for predicting the prognosis of cervical cancer, which has the following beneficial effects:

[0033] This application acquires and preprocesses time-series CT images of patients before and during treatment, and then uses a pre-trained and finely tuned 3D U-Net network to achieve fully automated segmentation of the clinical target area. Based on the depth features extracted from the segmentation mask, radiomics features, and acquired clinical features, a fusion feature vector is obtained by stitching them together. Prognostic prediction is completed through a multilayer perceptron model based on the DeepHit framework. This not only solves the problem of insufficient prediction accuracy when using only clinical features or pre-treatment radiomics models through a multimodal feature fusion strategy, but also introduces post-treatment CT images to capture changes in the tumor during radiotherapy, quantifying the biological heterogeneity and radiosensitivity of the tumor, overcoming the limitation of traditional static image models that cannot reflect the tumor treatment response. Furthermore, it realizes a fully automated process from CT image input to risk score output, eliminating the subjective error and time cost of manual target area delineation, and improving clinical usability. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a method for predicting the prognosis of cervical cancer according to Embodiment 1 of this application.

[0035] Figure 2 This is a flowchart of a cervical cancer prognosis prediction method according to Embodiment 1 of this application;

[0036] Figure 3 This is a schematic diagram of the predicted risk stratification results in Embodiment 1 of this application;

[0037] Figure 4 This is a schematic diagram of the results of different modal models in Embodiment 1 of this application;

[0038] Figure 5 This is a schematic diagram of the practice dependency prediction results in Embodiment 1 of this application;

[0039] Figure 6 This is a schematic diagram of the structure of a cervical cancer prognosis prediction system according to Embodiment 2 of this application;

[0040] Figure 7 This is a schematic diagram of the computer device structure according to Embodiment 3 of this application;

[0041] Figure 8 This is a schematic diagram of the storage medium structure of Embodiment 4 of this application. Detailed Implementation

[0042] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0043] The following analysis, based on relevant technologies, examines existing solutions.

[0044] The existing technologies suffer from the following problems: First, traditional methods relying on anatomical staging such as FIGO fail to reflect the biological heterogeneity within tumors, leading to inaccurate prognostic assessments for patients at the same stage. Second, most imaging models only use static images before treatment, failing to utilize temporal imaging information during or after treatment to capture the dynamic treatment response of the tumor. Existing methods typically require physicians to manually delineate the tumor target area, a time-consuming and labor-intensive process that introduces subjective differences, hindering automated and standardized applications. Furthermore, most models rely on single-type features, failing to fully integrate deep features from images, traditional radiomics features, and clinical features, limiting the improvement of predictive performance. Due to limited training data and a lack of multi-center validation, existing models often have insufficient generalization ability under different acquisition devices or protocols, restricting their clinical applicability.

[0045] This application combines multimodal data, especially imaging changes and clinical characteristics before and after radiotherapy, to accurately predict the disease-free survival (DFS) of patients receiving radical radiotherapy, thereby identifying patients at high risk of relapse in the early stages of treatment so as to carry out targeted intensive treatment or avoid unnecessary overtreatment.

[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0047] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0048] Example 1

[0049] Please see Figure 1 This is a flowchart illustrating a cervical cancer prognosis prediction method according to Embodiment 1 of this application. It should be noted that if substantially the same result is obtained, the method of this invention is not necessarily identical. Figure 1 The illustrated process sequence is limited. For example... Figure 1 The steps of the cervical cancer prognosis prediction method include:

[0050] S100: Acquire the patient's CT image sequence and preprocess the CT image sequence.

[0051] Specifically, patients with pathologically confirmed locally advanced cervical cancer were selected as study subjects. Patients were required to meet the inclusion criterion of not having received any anti-tumor treatment other than concurrent chemoradiotherapy. Two sets of CT image sequences were acquired for each patient: the first set was the pre-treatment CT image sequence, specifically CT simulation localization images before radiotherapy, used to present the initial anatomical morphology and location information of the tumor; the second set was the post-treatment or mid-treatment CT image sequence, specifically follow-up CT images after the patient completed the entire course of radical radiotherapy, or when 50% of the total radiotherapy dose was completed, used to present the morphological changes of the tumor after radiation exposure. Both sets of CT image sequences covered the pelvic region, and the scanning mode used was the standard clinical spiral CT plain scan mode, ensuring that the images completely included the clinical target area consisting of the primary tumor, cervix, uterus, part of the vagina, and lymphatic drainage area.

[0052] The two sets of acquired CT image sequences were preprocessed to eliminate data heterogeneity caused by differences in different CT equipment, scanning parameters, and patient body shape. The specific steps are as follows:

[0053] A linear interpolation algorithm was used to uniformly resample the two sets of CT image sequences at voxel resolution, adjusting the voxel size of all CT images to be consistent, such as 1.5 × 1.5 × 3 mm³. This resampling parameter selection balances image spatial resolution and computational efficiency, preserving subtle anatomical information in the CTV region while avoiding the surge in computational complexity for subsequent feature extraction caused by high resolution. During resampling, the image coordinate system remained unchanged to ensure spatial consistency of corresponding anatomical locations in the pre-treatment and post-treatment / intermediate-treatment CT image sequences.

[0054] Based on the Henle units (HU) standardization of CT images, the HU values ​​of all pixels in the two sets of CT image sequences are first limited to the conventional range [-1024, 3071] for pelvic tissue analysis in medical imaging. Pixels exceeding this range are truncated. Subsequently, a linear normalization method is used to map the truncated HU values ​​to the numerical range [0, 1]. This operation makes the grayscale features of the two sets of time-series CT images comparable and avoids interference from grayscale value shifts caused by differences in equipment calibration that could affect subsequent feature extraction.

[0055] Based on prior knowledge of the human pelvic anatomy, the approximate coordinate range of the clinical target volume (CTV) in CT images is determined. Using this coordinate range as a benchmark, the smallest axial bounding box that can completely contain the CTV and the surrounding 5mm of normal tissue is selected, and the two sets of CT image sequences are simultaneously cropped. The cropping operation retains only the image region within the bounding box and removes redundant tissue regions in the image that are not related to the pelvic tumor. This reduces the computational load of the subsequent network model and reduces the interference of noise from irrelevant regions on tumor features.

[0056] This application differs from existing technologies that use only single-time-point CT images in acquiring CT image sequences, ensuring the effective extraction of dynamic changes in tumors before and after radiotherapy in subsequent steps.

[0057] S200: The pre-processed CT image is segmented using a trained segmentation model to obtain a segmentation mask for the tumor region.

[0058] Specifically, a 3D U-Net network model, pre-trained on a public medical imaging dataset and then fine-tuned using cervical cancer CT images, is used as the segmentation model. The structure and training process of this model include:

[0059] The 3D U-Net network model consists of three parts: an encoder module, a decoder module, and skip connections. The encoder module contains four downsampling units, each consisting of two 3×3×3 3D convolutional layers, one batch normalization layer, and one ReLU activation function layer. Each downsampling unit ends with a 3D max-pooling layer with a stride of 2, achieving four downsampling operations on the input CT image to progressively extract multi-scale spatial features from local texture to global anatomical structures. The decoder module contains four upsampling units, each using a 3D transposed convolutional layer with a stride of 2 to upsample the feature map, restoring it to the same resolution as the corresponding encoder layer. Skip connections concatenate the high-resolution local features output from each encoder downsampling unit with the low-resolution global features output from the corresponding decoder upsampling unit, channel-wise, compensating for feature loss during upsampling and ensuring the segmentation model's ability to recognize subtle structures at the target region edges.

[0060] Furthermore, the 3D U-Net network model was pre-trained using a public dataset for medical image segmentation. A weighted sum of the Dice and cross-entropy loss functions was used as the loss function, with AdamW as the optimizer, an initial learning rate of 1e-4, and a batch size of 2. Iterative training was conducted until the loss function converged, resulting in a pre-trained model capable of medical image segmentation. CT images of pathologically diagnosed locally advanced cervical cancer patients and corresponding clinical target area annotations were obtained to fine-tune the pre-trained model. During fine-tuning, the clinical target areas (including the primary tumor, cervix, uterus, part of the vagina, and lymphatic drainage area) in the annotation data were used as supervision information. The learning rate was adjusted to 1e-5, and iterative training continued until the Dice similarity coefficient on the validation set stabilized above 0.85, thus obtaining a network model for cervical cancer clinical target area segmentation.

[0061] The preprocessed CT images from step S100 are input into the trained 3D U-Net segmentation model: the preprocessed CT images are downsampled sequentially by the encoder module to generate four feature maps of different resolutions, completing the extraction of multi-scale spatial features; the feature maps of each resolution are upsampled sequentially by the decoder module, and the high-resolution local features output by the encoder are fused through skip connections to gradually restore the feature maps to the same spatial resolution as the input CT images; the final feature maps output by the decoder module are compressed by a 1×1×1 3D convolutional layer to map the feature maps into dual-channel outputs, corresponding to the foreground and background of the target area respectively; the dual-channel outputs are processed using the Sigmoid activation function to map the output values ​​to the [0,1] interval, and the pixel value threshold is set to 0.5. Areas with pixel values ​​greater than or equal to 0.5 are identified as clinical target areas and assigned a value of 1; areas with pixel values ​​less than 0.5 are identified as non-target background areas and assigned a value of 0, generating a binary segmentation mask with the same resolution as the input CT image.

[0062] This application proposes a pre-trained and fine-tuned 3D U-Net network model that significantly improves segmentation accuracy and generalization ability in small sample scenarios, eliminates subjective differences between observers, and provides standardized and repeatable regions of interest for subsequent multimodal feature extraction.

[0063] S300: Based on the segmentation mask, extract comprehensive features from the preprocessed CT image and obtain the patient's clinical features.

[0064] Specifically, the integrated features include deep features and radiomics features. A basic medical imaging model pre-trained using self-supervised learning on a public medical imaging dataset is selected as the backbone network for feature extraction. This model possesses the ability to represent features of general medical images and does not require de novo training for small-sample cervical cancer data.

[0065] First, the pre-treatment CT images and the corresponding segmentation mask are concatenated along the channel dimension and used as input to the backbone network. The semantic features of the images are extracted layer by layer through the network's convolutional and pooling layers, outputting an 8192-dimensional high-dimensional depth feature vector. The same method is used to process the post-treatment / intermediate-term CT images and the segmentation mask to obtain another set of 8192-dimensional high-dimensional depth feature vectors.

[0066] Principal component analysis (PCA) was performed on both sets of high-dimensional deep feature vectors for dimensionality reduction. A 95% threshold for principal component contribution rate was set, retaining low-dimensional deep feature vectors that could encompass core feature information while removing redundant information. The dimensionality difference between the corresponding low-dimensional deep feature vectors before and after treatment was calculated to construct a subset of radiotherapy response deep features. This subset quantifies the characteristic changes of tumors under radiation, providing dynamic biological evidence for subsequent prognostic prediction. The dimensionality-reduced pre-treatment deep features, post-treatment / intermediate-treatment deep feature vectors, and radiotherapy response deep feature subset were concatenated to form a complete deep feature set.

[0067] Furthermore, it also includes image omics feature extraction and screening based on regions of interest segmented masks.

[0068] Using the segmentation mask output in step S200 as the boundary, the region of interest (ROI) in the preprocessed CT image is determined. Within this region, radiomics features are extracted and screened: using medical radiomics analysis tools, three categories of initial radiomics features, totaling 214, are extracted in parallel within the ROI. These initial radiomics features are then screened using variance selection, removing low-discrimination features with variance less than 0.05 and retaining candidate features with variance greater than 0.05. This results in 147 radiomics features significantly correlated with prognosis. The screened radiomics features are then z-score standardized to eliminate dimensional differences between features, ensuring that feature values ​​are on the same order of magnitude.

[0069] Furthermore, obtain the patient's clinicopathological data from step S100, and extract clinical features closely related to the prognosis of cervical cancer, specifically including:

[0070] The extracted clinical features include patient age, FIGO stage, maximum tumor diameter, hemoglobin level, squamous cell carcinoma antigen (SCC) level, and lymph node metastasis. The clinical features were categorized: for continuous features (such as age, maximum tumor diameter, hemoglobin level, and SCC level), min-max normalization was used to map the values ​​to the [0,1] interval; for discrete features (such as FIGO stage and lymph node metastasis), one-hot encoding was used to convert them into numerical vectors.

[0071] Compared to traditional approaches that extract image features at a single time point, this application can capture the dynamic changes in tumor radiotherapy sensitivity. By using a screening strategy that combines variance selection with correlation testing in survival analysis, the correlation between radiomics features and prognosis is improved, laying a data foundation for subsequent multimodal feature fusion and outcome prediction.

[0072] S400: Input the comprehensive features and the clinical features into the trained prognostic prediction model and output the prognostic prediction results for the patient.

[0073] In this embodiment, the prognostic prediction model is a multilayer perceptron trained based on the DeepHit survival analysis framework. The multilayer perceptron includes an input layer, a fusion layer, a hidden layer, and a prediction layer. Deep features, radiomics features, and clinical features are input into the input layer and concatenated by the fusion layer to obtain a fused feature vector. The hidden layer maps the fused feature vector using a non-linear activation function. Based on the DeepHit framework, the model learns the conditional probability distribution of patient survival time. Finally, the prediction layer outputs the patient's individual risk score and the probability of disease-free survival at at least one preset time point in the future.

[0074] Specifically, a weighted fusion strategy is used to integrate multidimensional features, and the DeepHit survival analysis framework is used to accurately predict the disease-free survival of patients. The prognostic prediction model is a multilayer perceptron trained based on the DeepHit survival analysis framework, and the model structure includes an input layer, a fusion layer, a hidden layer, and a prediction layer. The input layer dimension matches the feature vector dimension, which is composed of deep features, radiomics features, and clinical features. The fusion layer calculates the Pearson correlation coefficient between each feature subset and the patient's disease-free survival based on a weighted fusion operation of feature correlation. Different weights are assigned according to the absolute value of the correlation coefficient, with the weight coefficient of the deep feature subset of radiotherapy response being higher than that of deep features, radiomics features, and clinical features at a single time point. The hidden layer uses a two-layer fully connected structure, with 256 neurons in the first layer and 128 neurons in the second layer. The activation function is ReLU. A dropout layer with a dropout probability of 0.2 is added after each hidden layer to avoid overfitting. The prediction layer contains two types of output neurons, which are used to output the patient's individual risk score and the disease-free survival probability at 12 months, 36 months, and 60 months, respectively.

[0075] During the predictive model training phase, multimodal feature data and corresponding follow-up survival data of patients with locally advanced cervical cancer from multiple centers were first acquired. The survival data needed to include the patients' disease-free survival time and truncation status identifiers. The dataset was divided into training, validation, and test sets in a 7:1:2 ratio. The loss function adopted was a weighted sum of log-likelihood loss and risk stratification loss from the DeepHit framework, with the weight ratio of the two types of losses set to 1:1. This was to simultaneously optimize the model's learning accuracy of the conditional probability distribution of survival time and its ability to distinguish between high- and low-risk groups. The optimizer was Adam, with an initial learning rate of 1e-4 and a batch size of 16. The learning rate was decayed to 0.1 after every 10 iterations. The C-index value of the validation set was used as the model performance evaluation index. Training was stopped when the C-index of the validation set showed no improvement for three consecutive iterations.

[0076] Please see Figure 2This is a flowchart of a cervical cancer prognosis prediction method according to Embodiment 1 of this application. First, paired CT image data of the patient before and after treatment (or during the treatment phase) are acquired. A deep neural network is used to achieve fully automated segmentation of the lesion area, eliminating subjective errors from manual delineation. Then, by constructing a multimodal feature fusion architecture, deep features, radiomics features, and clinical features reflecting dynamic changes in the tumor are extracted and integrated. Under the nonlinear mapping of a multilayer perceptron, the above-mentioned multidimensional heterogeneous data is transformed into an accurate prediction of the patient's disease-free survival (DFS). This process is fully automated and objective, capable of quantifying the tumor's radiosensitivity, thereby non-invasively identifying high-risk individuals for recurrence in the early stages of treatment and guiding the development of individualized treatment plans.

[0077] Please see Figure 3 This is a schematic diagram of the predicted risk stratification results in Embodiment 1 of this application; Figure 4 This is a schematic diagram of the results of different modal models in Embodiment 1 of this application; Figure 5 This is a schematic diagram of the practice dependency prediction results of Embodiment 1 of this application.

[0078] The method proposed in this application has been validated by clinical patient data from Peking Union Medical College Hospital and other institutions, and the results demonstrate its high feasibility and superiority. Figure 3 Experiments show that the proposed multimodal fusion model consistently outperforms the single-modal model, fully demonstrating the effectiveness of the multidimensional feature fusion strategy. Figure 4 In terms of time-dependent prediction ( Figure 5 The model's AUC values ​​at 12, 36, and 60 months on the training set were 0.96, 0.92, and 0.87, respectively, and it also maintained a high level of predictive accuracy of 0.79, 0.80, and 0.80 on the internal validation set. Furthermore, Kaplan-Meier survival curve analysis showed that the model could successfully divide patients into high-risk and low-risk groups, and the survival difference between the two groups was statistically significant (P<0.001), demonstrating clear clinical stratification application value.

[0079] In summary, Embodiment 1 of this application acquires and preprocesses temporal CT images of patients before and after / intermediate treatment. A pre-trained and finely tuned 3D U-Net network is used to achieve fully automated segmentation of the clinical target area. Depth features, radiomics features, and clinical features are extracted based on the segmentation mask. Then, a multilayer perceptron model based on the DeepHit framework is used to complete weighted fusion and prognostic prediction, ultimately outputting the patient's individual risk score and disease-free survival probability at multiple time points. This method overcomes the technical shortcomings of existing technologies, such as reliance on anatomical staging leading to high heterogeneity in prognostic assessment, neglect of dynamic response characteristics of tumor radiotherapy, strong subjectivity in manual target area delineation, and weak generalization ability of single-modal models. It achieves fully automated and standardized operation from image input to prognostic stratification.

[0080] Example 2

[0081] Please see Figure 6 This is a schematic diagram of the structure of a cervical cancer prognosis prediction system according to Embodiment 2 of this application. Figure 6 As shown, the cervical cancer prognosis prediction system includes an image processing module 100, a segmentation module 200, a feature extraction module 300, and a prognosis prediction module 400.

[0082] The image processing module 100 is used to acquire the patient's CT image sequence and preprocess the CT image sequence;

[0083] The segmentation module 200 is used to segment the preprocessed CT image using a trained segmentation model to obtain a segmentation mask for the tumor region.

[0084] The feature extraction module 300 is used to extract comprehensive features from the preprocessed CT image based on the segmentation mask and obtain the patient's clinical features;

[0085] The prognosis prediction module 400 is used to input the comprehensive features and the clinical features into the trained prognosis prediction model and output the prognosis prediction results of the patient.

[0086] In this embodiment, the image processing module 100 acquires and preprocesses the patient's CT image sequence. Specifically, it selects patients with locally advanced cervical cancer diagnosed by pathology and receiving only concurrent chemoradiotherapy. The module acquires CT simulation localization image sequences before radiotherapy and CT image sequences after or during radiotherapy. Both sets of images cover the pelvic cavity to fully encompass the clinical target area. Simultaneously, the acquired CT image sequences undergo resampling, grayscale normalization, and lesion region cropping operations to provide high-quality input data for subsequent modules.

[0087] The segmentation module 200 is used to accurately segment preprocessed CT images using a trained segmentation model to obtain a segmentation mask for the tumor region. The segmentation model is a 3D U-Net network model pre-trained on a public medical imaging dataset and then fine-tuned using cervical cancer CT images. It includes an encoder, a decoder, and skip connections. The encoder extracts multi-scale spatial features through downsampling, the decoder restores resolution through upsampling, and the skip connections fuse high-resolution local features with low-resolution global features. The preprocessed CT image is input into the 3D U-Net model, and after encoding and decoding, a dual-channel feature map is output. Then, binarization is performed using a pixel value threshold of 0.5 to generate a clinical target region segmentation mask with the same resolution as the input image, replacing manual delineation to eliminate inter-observer differences.

[0088] The feature extraction module 300 is used to extract comprehensive features based on the segmentation mask and obtain the patient's clinical characteristics. Specifically, the preprocessed CT image and the corresponding segmentation mask are input into the self-supervised learning pre-trained medical imaging basic model to extract 8192-dimensional high-dimensional depth features. After dimensionality reduction by principal component analysis, the feature difference before and after treatment is calculated to construct a subset of radiotherapy response depth features. At the same time, initial radiomics features are extracted within the region of interest defined by the segmentation mask, and prognostic radiomics features are obtained through a variance selection method.

[0089] The prognostic prediction module 400 is used to input comprehensive features and clinical features into the trained prognostic prediction model to output the patient's prognostic prediction results. The module includes a multilayer perceptron model trained based on the DeepHit survival analysis framework, which includes an input layer, a fusion layer, a hidden layer, and a prediction layer. The module ultimately outputs the patient's individual risk score and the disease-free survival probability at 12 months, 36 months, and 60 months.

[0090] For further details regarding the implementation techniques of each module in the cervical cancer prognosis prediction system described in the above embodiments, please refer to the description in the cervical cancer prognosis prediction method in the above embodiments, which will not be repeated here.

[0091] Example 3

[0092] Please see Figure 7 This is a schematic diagram of the computer device structure according to Embodiment 3 of this application. The computer device 50 includes a processor 51 and a memory 52 coupled to the processor 51.

[0093] The memory 52 stores program instructions for implementing the above-described method for predicting the prognosis of cervical cancer.

[0094] The processor 51 is used to execute program instructions stored in the memory 52 to implement a cervical cancer prognosis prediction.

[0095] The processor 51 can also be referred to as a CPU (Central Processing Unit).

[0096] Processor 51 may be an integrated circuit chip with signal processing capabilities. Processor 51 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0097] Example 4

[0098] Please see Figure 8 This is a schematic diagram of the storage medium in Embodiment 4 of this application. The storage medium in this embodiment stores a program file 61 capable of implementing all the above methods. This program file 61 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or devices such as computers, servers, mobile phones, and tablets.

[0099] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0100] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0101] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

[0102] Of course, the present invention may have many other embodiments. Based on this embodiment, other embodiments obtained by those skilled in the art without any creative effort are all within the scope of protection of the present invention.

Claims

1. A method for predicting the prognosis of cervical cancer, characterized in that, include: Acquire the patient's CT image sequence and preprocess the CT image sequence; The pre-processed CT image is segmented using a trained segmentation model to obtain a segmentation mask for the tumor region; Based on the segmentation mask, comprehensive features are extracted from the preprocessed CT image, and the patient's clinical features are obtained. The comprehensive features and the clinical features are input into the trained prognostic prediction model, and the prognostic prediction results for the patient are output. The step of extracting comprehensive features from the preprocessed CT image based on the segmentation mask and obtaining the patient's clinical features specifically includes: The integrated features include depth features and radiomics features; The preprocessed CT images and segmentation masks are input into a medical image model pre-trained by self-supervised learning to extract high-dimensional depth features before, after, or during treatment; the high-dimensional depth features are then subjected to dimensionality reduction processing using principal component analysis to obtain depth features. Within the region of interest defined by the segmentation mask, initial radiomics features are extracted; the initial radiomics features are then filtered to obtain radiomics features related to prognosis. The step of inputting the comprehensive features and the clinical features into the trained prognostic prediction model and outputting the patient's prognostic prediction result specifically includes: The prognostic prediction model is a multilayer perceptron trained based on the DeepHit survival analysis framework. The multilayer perceptron includes an input layer, a fusion layer, a hidden layer, and a prediction layer; The deep features, radiomics features, and clinical features are input into the input layer and then concatenated through the fusion layer to obtain a fused feature vector. The hidden layer maps the fused feature vector using a non-linear activation function; and learns the conditional probability distribution of patient survival time based on the DeepHit framework. The prediction layer outputs the patient's individual risk score and the probability of disease-free survival at at least one preset time point in the future; The CT image sequence includes: a first CT image sequence at the time point before treatment, and a second CT image sequence at the time point after treatment or during treatment; In this process, the pre-treatment CT images and the corresponding segmentation mask are concatenated along the channel dimension. The semantic features of the images are extracted layer by layer through the convolutional and pooling layers of the network to output a high-dimensional depth feature vector. The same method is used to process the post-treatment or mid-term CT images and the segmentation mask to obtain another set of high-dimensional depth feature vectors. Principal component analysis was performed on the two sets of high-dimensional depth feature vectors to reduce their dimensionality. The difference in the corresponding dimensions of the low-dimensional depth feature vectors before and after treatment was calculated. A subset of deep features of radiotherapy response was constructed. The depth features before treatment, the depth feature vectors after treatment or in the middle stage, and the subset of deep features of radiotherapy response after the dimensionality reduction were concatenated to form a complete set of deep features.

2. The cervical cancer prognosis prediction method according to claim 1, characterized in that, The step of segmenting the preprocessed CT image using a trained segmentation model to obtain a segmentation mask for the tumor region specifically includes: A 3D U-Net network model is used as the segmentation model. The 3D U-Net network model is based on an encoder and decoder structure and includes skip connections. The preprocessed CT image is input into the 3D U-Net network model. After encoding and decoding the input CT image, the segmentation mask is output.

3. A cervical cancer prognosis prediction system, characterized in that, For performing the cervical cancer prognostic prediction method according to any one of claims 1 to 2, the cervical cancer prognostic prediction system comprises: An image processing module is used to acquire the patient's CT image sequence and preprocess the CT image sequence; The segmentation module is used to segment the preprocessed CT image using a trained segmentation model to obtain a segmentation mask for the tumor region. The feature extraction module is used to extract comprehensive features from the preprocessed CT image based on the segmentation mask, and to obtain the patient's clinical features; The prognosis prediction module is used to input the comprehensive features and the clinical features into the trained prognosis prediction model and output the prognosis prediction results for the patient.

4. A computer device, characterized in that, The computer device includes a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing the cervical cancer prognosis prediction method according to any one of claims 1-2; the processor is used to execute the program instructions stored in the memory to implement cervical cancer prognosis prediction.

5. A computer-readable storage medium, characterized in that, The device stores processor-executable program instructions for performing the cervical cancer prognosis prediction method according to any one of claims 1-2.

Images