A Prognostic Intelligent Analysis Method for Rectal Cancer Based on Two-Stage 3D Feature Extraction

Abstract

This invention discloses an intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction, comprising: acquiring MRI images of the target patient for prognosis; constructing a three-dimensional convolutional neural network based on a 3D-Unet architecture; training the three-dimensional convolutional neural network in stages to generate a rectal cancer tumor extraction model; using the preprocessed MRI 3D images as model input to segment the rectal cancer tumor images; and outputting 3D tumor images and risk scores in stages; combining the risk scores with the target patient's clinical data for prognostic assessment to obtain the target patient's survival rate. This invention can extract 3D tumor information from MRI images, further enhancing the accuracy of prognostic analysis and providing more precise and reliable prognostic results, thereby improving clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of image segmentation and processing technology, and more specifically, to an intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction. Background Technology

[0002] In the prognostic analysis of rectal cancer, magnetic resonance imaging (MRI) is an important assessment tool. Traditional image analysis methods typically rely on manually selected regions of interest (ROI) layers, which only utilize information from the largest cross-section of the tumor. However, this single-layer ROI analysis ignores the overall morphology and spatial distribution of the tumor, leading to biased prognostic assessment results. In recent years, deep learning technology has been increasingly applied to medical image analysis, improving the accuracy of analysis through automated feature extraction; however, most methods are still limited to single-layer ROI analysis.

[0003] Existing methods for prognostic analysis of rectal cancer MRI images mainly focus on processing single-layer regions of interest (ROIs), typically employing convolutional neural networks (CNNs) for feature extraction and analysis. While these methods can improve the efficiency of prognostic analysis to some extent, they have limitations in processing three-dimensional information. Typical implementations include processing single-layer images through two-dimensional convolution, but this method cannot fully utilize the three-dimensional information in MRI images, leading to insufficient understanding of the spatial structure and overall morphology of the tumor, thus affecting the accuracy and reliability of prognostic results. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes an intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction. The aim is to extract 3D tumor information from MRI images, thereby enabling more accurate prognostic analysis, providing more precise and reliable prognostic assessment results, and ultimately improving clinical decision-making.

[0005] This invention provides an intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction, comprising:

[0006] Step S102: Obtain MRI images of the target patient's prognosis and preprocess the MRI images;

[0007] Step S104: Construct a three-dimensional convolutional neural network based on the 3D-Unet architecture, perform staged training on the three-dimensional convolutional neural network, and generate a rectal cancer tumor extraction model.

[0008] Step S106: The preprocessed 3D MRI image is used as the model input, and the rectal cancer tumor extraction model is used to segment the rectal cancer tumor image to obtain the 3D tumor image and risk score.

[0009] Step S108: Combine the risk score with the clinical data of the target patient to conduct a prognostic assessment and obtain the survival rate of the target patient.

[0010] In this approach, MRI images of the target patient's prognosis are acquired, and these MRI images are preprocessed, specifically as follows:

[0011] Extract MRI images of the target patient's prognosis to obtain MRI grayscale images, and perform floating-point image conversion on the MRI grayscale images to obtain the converted MRI grayscale images;

[0012] The converted MRI grayscale image was corrected using N4 bias correction to obtain the corrected training MRI grayscale image;

[0013] Histogram matching is performed on the corrected training MRI grayscale image to obtain the preprocessed 3D MRI image.

[0014] In this scheme, a three-dimensional convolutional neural network based on the 3D-Unet architecture is constructed, and the three-dimensional convolutional neural network is trained in stages, specifically as follows:

[0015] A three-dimensional convolutional neural network is constructed based on the 3D-Unet architecture. The encoder part is trained in the first stage. In the encoder part, a three-dimensional convolutional kernel is used to progressively downsample the input three-dimensional image.

[0016] The encoder part is implemented by four downsampling operations, where each downsampling operation uses two convolutional modules, one sparse module and one max pooling module to extract feature information of the 3D image and avoid network overfitting.

[0017] The decoding part is implemented by four layers of upsampling operations. Each upsampling operation uses two convolution operations, a sparse module, and a deconvolution module to restore the original image size layer by layer. Finally, the segmentation output result is obtained through an activation function.

[0018] Rectal cancer diagnosis and tumor region division results are read from historical rectal cancer medical records as label information. Based on the data label information and rectal cancer images, a first-stage data sample is constructed and divided into training data, validation data and test data according to a preset ratio.

[0019] The encoder is pre-trained using the Dice loss function to obtain the encoder weights.

[0020] In this solution, the Dice loss function is used to calculate the loss. The Dice loss function is expressed as:

[0021]

[0022] Where X represents the predicted 3D tumor image, and Y represents the label information.

[0023] In this solution, the second-stage network construction and training of the three-dimensional convolutional neural network specifically includes:

[0024] Obtain the encoder structure and weights of the three-dimensional convolutional neural network after the first stage of training, and fill the decoder structure with four fully connected layers to replace the decoder and skip connections trained in the first stage.

[0025] Data label information is generated by reading whether the patient’s recurrence event occurred and the follow-up time from historical rectal cancer medical records. Based on the data label information and historical risk scores, a second-stage data sample is constructed and divided into training data, validation data and test data according to a preset ratio.

[0026] The weights of the first-stage encoder are used as the pre-training weights of the second-stage encoder. The model is trained by calculating the loss using the Cox partial likelihood loss function based on the training data. During the training process, the hyperparameters are adjusted and the selected models are tested using validation data. The selected models are then tested using test data.

[0027] In this scheme, the loss is calculated using the Cox partial likelihood loss function, which is expressed as:

[0028]

[0029] Where n represents the size of the 3D convolutional neural network batch, a batch contains n training data samples, each sample contains a risk score s, a follow-up time t, and whether a relapse event e occurred, k represents the number of relapse events in a batch sample, and i and j represent the number of items in a batch sample.

[0030] In this protocol, risk scores are combined with clinical data of the target patients for prognostic assessment to obtain the survival rate of the target patients, specifically as follows:

[0031] Obtain the risk score set corresponding to the second-stage test data of the three-dimensional convolutional neural network, perform ROC curve analysis on the risk score set, obtain the optimal cutoff point, and classify patients into high-risk and low-risk groups based on the optimal cutoff point.

[0032] Obtain the risk score of the target patient, and generate risk score features based on the judgment result of the risk score and the risk score threshold corresponding to the optimal cutoff point;

[0033] The clinical characteristics and risk score characteristics of the target patients were obtained, and KM survival analysis was used to obtain low-risk and high-risk score curves, and the predicted survival value of the target patients was output.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] This invention proposes a prognostic assessment model based on 3D-UNet. Through two-stage training, this invention achieves the ability to automatically extract ROI images, eliminating the need for manual intervention and solving the problem of relying on manually cropped ROI images in traditional tumor prognostic models.

[0036] Furthermore, this invention can extract 3D tumor information from MRI images, further enhancing the accuracy of prognostic analysis. The purpose of this invention is to provide more accurate and reliable prognostic results through the comprehensive utilization of three-dimensional image information, thereby improving clinical decision-making. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0038] Figure 1 A flowchart of a two-stage 3D feature extraction-based intelligent prognostic analysis method for rectal cancer is shown.

[0039] Figure 2 A structural diagram of the three-dimensional convolutional neural network in the embodiment is shown;

[0040] Figure 3 A flowchart of the sparse module in the embodiment is shown;

[0041] Figure 4 A structural diagram of the rectal cancer tumor extraction model in the embodiment is shown;

[0042] Figure 5 A schematic diagram of the KM survival analysis in the example is shown;

[0043] Figure 6 The flowchart illustrating the feature acquisition and KM survival analysis in this embodiment is shown.

[0044] Figure 7 A block diagram of a rectal cancer prognostic intelligent analysis system based on two-stage 3D feature extraction is shown. Detailed Implementation

[0045] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0047] Figure 1 A flowchart of a two-stage 3D feature extraction-based intelligent prognostic analysis method for rectal cancer is shown.

[0048] like Figure 1 As shown, this embodiment provides an intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction, including:

[0049] S102, acquire MRI images of the target patient's prognosis, and preprocess the MRI images;

[0050] S104, Construct a three-dimensional convolutional neural network based on the 3D-Unet architecture, perform staged training on the three-dimensional convolutional neural network, and generate a rectal cancer tumor extraction model;

[0051] S106, The preprocessed 3D MRI image is used as the model input, and the rectal cancer tumor extraction model is used to segment the rectal cancer tumor image to obtain the 3D tumor image and risk score.

[0052] S108 combines risk scores with clinical data of target patients to conduct prognostic assessments and obtain the survival rate of target patients.

[0053] It should be noted that the process involves extracting MRI grayscale images from the target patient's prognosis, converting these grayscale images to floating-point numbers, and then applying N4 bias correction to the converted grayscale images to obtain corrected training MRI grayscale images. Histogram matching is then performed on these corrected training MRI grayscale images to obtain preprocessed 3D MRI images. Applying N4 bias correction and histogram matching to the MRI images during image preprocessing helps avoid noise interference and model design redundancy, thus improving the efficiency of subsequent feature extraction.

[0054] According to an embodiment of the present invention, a three-dimensional convolutional neural network is constructed based on the 3D-Unet architecture, such as... Figure 2As shown, the encoder part undergoes a first-stage training process. In this encoder part, a 3D convolutional kernel is used to progressively downsample the input 3D image to extract multi-scale features and deeply mine image information, while avoiding network overfitting. The decoding part is implemented by four layers of upsampling operations. Each upsampling operation uses two convolutional operations, a sparse module, and a deconvolutional module to progressively restore the image size to the original size. Finally, the segmentation output is obtained through an activation function. By fusing the shallow features from the encoding stage with a skip connection mechanism, not only are rich spatial details preserved, but the semantic expressive ability of the model is also significantly enhanced, ensuring the consistency between the output and input image sizes and promoting accurate segmentation.

[0055] Sparse modules effectively enhance the generalization ability of networks by utilizing multi-branch feature extraction, feature fusion, and batch normalization through multi-scale convolutional operations. For example... Figure 3 As shown, the input to the sparse module is a tensor C×D×H×W, which undergoes feature processing via two parallel paths. In the first path, the input first passes through a 2×2×2 3D max pooling layer to downsample the spatial dimension and capture salient region features. Subsequently, a 1×1×1 3D convolutional layer compresses the channel dimension, reducing computational complexity while preserving important feature information. To recover the spatial dimension reduced by pooling, a 3×3×3 3D deconvolution (ConvTranspose3D) layer is used to reconstruct spatial information; this operation restores the spatial structure and captures fine-grained features. A batch normalization (BN) layer is added at the end of this path to ensure training stability and prevent overfitting.

[0056] In another path, the input tensor first passes through a 1×1×1 3D convolutional layer to compress the number of channels and extract global features. This path then splits into three branches. In the first branch, a 3×3×3 3D convolutional layer is applied to capture local spatial features. In the second branch, a 1×1×1 3D convolutional layer is applied to further compress the number of channels and extract features from different dimensions. The third branch directly passes through a batch normalization (BN) layer, skipping the convolution operation, to normalize the features and ensure stable feature fusion across different scales. The features output from these three branches are concatenated along the channel dimension. Through this concatenation, the network can fuse multi-scale information from different branches, enhancing its ability to represent complex input data. On the concatenated feature map, a 3×3×3 3D convolutional layer is applied to further integrate and reorganize the previously extracted features, and batch normalization is applied again to ensure training stability.

[0057] Finally, the output feature maps from the two main branches are concatenated along the channel dimension and then further integrated with multi-scale and multi-dimensional information through a 3×3×3 3D convolutional layer, ensuring effective fusion of features extracted from different branches. Subsequently, the concatenated feature map is normalized through a batch normalization (BN) layer to further improve training stability and prevent gradient vanishing or exploding. This continuous processing of convolution and normalization not only strengthens the expressive power of the fused features but also effectively suppresses the model's overfitting tendency, ensuring precise feature extraction and the network's generalization ability.

[0058] Rectal cancer diagnosis and tumor region division results are read from historical rectal cancer medical records as label information. Based on the data label information and rectal cancer images, a first-stage data sample is constructed and divided into training data, validation data and test data according to a preset ratio. The encoder is pre-trained using the Dice loss function to obtain encoder weights.

[0059] The Dice loss function calculates the overlap between the predicted and actual results and optimizes the model by minimizing the difference between them. It is suitable for class imbalance scenarios and can reduce the negative impact of an imbalance between foreground and background ratios in the samples. The Dice loss function is expressed as follows:

[0060]

[0061] Where X represents the predicted 3D tumor image, and Y represents the label information.

[0062] Obtain the encoder structure and weights of the 3D convolutional neural network after the first stage of training, such as... Figure 4 As shown, the decoder structure is filled with four fully connected layers of different sizes, replacing the decoder and skip connections in the first-stage training. Data labels are generated by reading patient recurrence events and follow-up times from historical rectal cancer case data. Based on these labels and historical risk scores, a second-stage data sample is constructed and divided into training, validation, and test data according to a preset ratio. The weights of the first-stage encoder are used as pre-training weights for the second-stage encoder. The model is trained using the Cox partial likelihood loss function based on the training data. The loss calculation includes all features within a batch (risk score, whether a patient recurrence event occurred, and follow-up time). During training, validation data is used to adjust hyperparameters and test selected models. The selected models are then tested using test data.

[0063] The loss is calculated using the Cox partial likelihood loss function, which is expressed as:

[0064]

[0065] Where n represents the size of the 3D convolutional neural network batch, a batch contains n training data samples, each sample contains a risk score s, a follow-up time t, and whether a relapse event e occurred, k represents the number of relapse events in a batch sample, and i and j represent the number of items in a batch sample.

[0066] The rectal cancer tumor extraction model is configured based on the encoder weights and prognostic score weights after training. The preprocessed 3D MRI images are used as input to the model. The rectal cancer tumor extraction model is used to segment the rectal cancer tumor images to obtain 3D tumor images and risk scores.

[0067] According to an embodiment of the present invention, a risk score set corresponding to the second-stage test data of a three-dimensional convolutional neural network is obtained. ROC curve analysis is performed on the risk score set to obtain the optimal cutoff point. Based on the optimal cutoff point, patients are divided into high-risk and low-risk groups. The risk score of the target patient is obtained, and risk score features are generated based on the judgment result of the risk score and the risk score threshold corresponding to the optimal cutoff point. KM survival analysis is performed on the clinical characteristics and risk score features of the target patient to obtain low-risk and high-risk score curves. The predicted survival rate of the target patient is output. The low-risk score curve (pred0) and the high-risk score curve (pred1) are shown below. Figure 5 As shown.

[0068] It should be noted that the risk score of the target patient is obtained, and a risk score feature is generated based on the judgment result of the risk score and the risk score threshold corresponding to the optimal cutoff point. This risk score feature represents either a high-risk or low-risk population. Figure 6 As shown, features such as recurrence events and follow-up time in clinical information are extracted as clinical information features.

[0069] In another preferred embodiment of the present invention, a prognostic database for target patients is constructed. This database records the treatment response and survival rate of target patients. A prognostic profile of the target patients is constructed based on the treatment response and survival rate. Similarity is searched in historical rectal cancer medical record data based on the prognostic profile. Historical rectal cancer medical record data meeting a preset similarity threshold is selected. Historical rectal cancer medical record data with a survival rate greater than a preset survival rate threshold is marked. The recovery status of patients in the marked data is analyzed. Rectal cancer recovery indicators are retrieved using big data methods. The recovery status of patients is evaluated. Marked data whose recovery evaluation results meet health standards are selected. Corresponding prognostic methods are extracted and implemented for the target patients. The recovery and health results are visualized and provided to the target users to improve the treatment confidence of the target patients.

[0070] Figure 7 A block diagram of a rectal cancer prognostic intelligent analysis system based on two-stage 3D feature extraction is shown.

[0071] A second aspect of the present invention also provides a rectal cancer prognostic intelligent analysis system 7 based on two-stage 3D feature extraction. The system includes a memory 71 and a processor 72, wherein the memory and the processor store and execute a program for a rectal cancer prognostic intelligent analysis method based on two-stage 3D feature extraction.

[0072] A third aspect of the present invention provides a computer-readable storage medium comprising a program for an intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction. When the program for the intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction is executed by a processor, it implements the steps of the intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction.

[0073] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0074] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0075] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0076] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0078] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent prognostic analysis of rectal cancer based on two-stage 3D feature extraction, characterized in that, Includes the following steps: Step S102: Obtain MRI images of the target patient's prognosis and preprocess the MRI images; Step S104: Construct a three-dimensional convolutional neural network based on the 3D-Unet architecture, perform phased training on the three-dimensional convolutional neural network, and generate a rectal cancer tumor extraction model. Step S106: The preprocessed 3D MRI image is used as the model input, and the rectal cancer tumor extraction model is used to segment the rectal cancer tumor image to obtain the 3D tumor image and risk score. Step S108: Combine the risk score with the clinical data of the target patient to conduct a prognostic assessment and obtain the survival rate of the target patient; A three-dimensional convolutional neural network based on the 3D-Unet architecture is constructed, and the three-dimensional convolutional neural network is trained in stages, specifically as follows: A three-dimensional convolutional neural network is constructed based on the 3D-Unet architecture. The encoder part is trained in the first stage. In the encoder part, a three-dimensional convolutional kernel is used to progressively downsample the input three-dimensional image. The encoder part is implemented by four downsampling operations, where each downsampling operation uses two convolutional modules, one sparse module and one max pooling module to extract feature information of the 3D image and avoid network overfitting. The decoder part is implemented by four layers of upsampling operations. Each upsampling operation uses two convolution operations, a sparse module, and a deconvolution module to restore the original image size layer by layer. Finally, the segmentation output result is obtained through an activation function. Rectal cancer diagnosis and tumor region division results are read from historical rectal cancer medical records as label information. Based on the label information and rectal cancer images, a first-stage data sample is constructed and divided into training data, validation data and test data according to a preset ratio. The encoder is pre-trained using the Dice loss function to obtain the encoder weights; The second stage of network construction and training of the three-dimensional convolutional neural network is as follows: Obtain the encoder structure and weights of the three-dimensional convolutional neural network after the first stage of training, and fill the decoder part with four fully connected layers to replace the decoder and skip connections trained in the first stage. Data label information is generated by reading whether the patient’s recurrence event occurred and the follow-up time from historical rectal cancer medical records. Based on the data label information and historical risk scores, a second-stage data sample is constructed and divided into training data, validation data and test data according to a preset ratio. The weights of the first-stage encoder are used as the pre-training weights of the second-stage encoder. The model is trained by calculating the loss using the Cox partial likelihood loss function based on the training data. During the training process, the hyperparameters are adjusted and the selected models are tested using validation data. The selected models are then tested using test data.

2. The intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction according to claim 1, characterized in that, Obtain MRI images of the target patient's prognosis, and preprocess the MRI images, specifically as follows: Extract MRI images of the target patient's prognosis to obtain MRI grayscale images, and perform floating-point image conversion on the MRI grayscale images to obtain the converted MRI grayscale images; The converted MRI grayscale image was corrected using N4 bias correction to obtain the corrected training MRI grayscale image; Histogram matching is performed on the corrected training MRI grayscale image to obtain the preprocessed 3D MRI image.

3. The intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction according to claim 1, characterized in that, The loss is calculated using the Dice loss function, which is expressed as follows: Where X represents the predicted 3D tumor image, and Y represents the label information.

4. The intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction according to claim 1, characterized in that, The loss is calculated using the Cox partial likelihood loss function, which is expressed as: Where n represents the size of the 3D convolutional neural network batch, a batch contains n training data samples, each sample contains a risk score s, a follow-up time t, and whether a relapse event e occurred, k represents the number of relapse events in a batch sample, and i and j represent the number of items in a batch sample.

5. The intelligent prognostic analysis method for rectal cancer based on two-stage 3D feature extraction according to claim 1, characterized in that, The risk score is combined with the clinical data of the target patients for prognostic assessment to obtain the survival rate of the target patients, specifically: Obtain the risk score set corresponding to the second-stage test data of the three-dimensional convolutional neural network, perform ROC curve analysis on the risk score set, obtain the optimal cutoff point, and classify patients into high-risk and low-risk groups based on the optimal cutoff point. Obtain the risk score of the target patient, and generate risk score features based on the judgment result of the risk score and the risk score threshold corresponding to the optimal cutoff point; The clinical characteristics and risk score characteristics of the target patients were obtained, and KM survival analysis was used to obtain low-risk and high-risk score curves, and the predicted survival value of the target patients was output.

Images