Kidney cancer recurrence risk prediction method based on deep learning model

Abstract

The invention provides a kidney cancer recurrence risk prediction method based on a deep learning model, and relates to the technical field of deep learning, and the method comprises the steps: collecting an image data set for kidney cancer high recurrence risk prediction; carrying out registration on the collected multi-stage enhanced CT image; constructing and training a kidney tumor automatic detection and segmentation model; carrying out ROI positioning cutting and quality control; and constructing a deep learning model for renal cancer recurrence risk prediction based on the multi-modal convolutional neural network, and realizing renal cancer recurrence risk prediction through the constructed prediction network model. According to the method, the multi-phase enhanced CT image of the kidney cancer patient is analyzed through the deep learning model, the tumor postoperative recurrence risk is predicted, an objective basis is provided for a clinician to make an individualized follow-up visit scheme and an auxiliary treatment decision, and excessive treatment of a low-risk patient and insufficient treatment of a high-risk patient are avoided.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method for predicting the recurrence risk of kidney cancer based on a deep learning model. Background Art

[0002] In the field of renal cancer diagnosis and treatment, accurate assessment of the risk of postoperative recurrence has always been a major challenge facing the clinic. The current mainstream prognostic evaluation system mainly relies on postoperative pathological parameters, including traditional indicators such as TNM staging and WHO / ISUP grade. Although these methods provide a basic framework for clinical decision-making, they have obvious limitations: on the one hand, they simplify the complex biological behavior of tumors into limited classification criteria and cannot fully reflect the temporal and spatial heterogeneity of tumors; on the other hand, these assessments can only be obtained after surgery, making it difficult to provide a reference for the formulation of preoperative treatment plans. Although the molecular detection technology that has emerged in recent years can provide richer prognostic information, it is greatly limited in routine clinical applications due to the high cost of detection, complex procedures and significant batch differences.

[0003] Medical imaging, as an important noninvasive diagnostic tool, contains a wealth of underutilized prognostic information. In routine CT examinations, imaging findings such as tumor enhancement patterns, boundary characteristics, and internal heterogeneity reflect underlying biological properties. However, traditional manual image assessment relies primarily on the physician's subjective experience, with limited ability to identify subtle yet critical prognostic features. While existing radiomics methods attempt to extract these features through quantitative analysis, their reliance on manually designed features leads to a significant loss of underlying information, and the analysis process is cumbersome and time-consuming, limiting clinical practicality. The emergence of deep learning technology has opened up new possibilities for breakthroughs in this field. Such algorithms can automatically learn multi-level feature representations directly from raw image data, capturing imaging patterns closely related to prognosis without manual intervention. Of particular note, dynamic enhancement information provided by multi-phase contrast-enhanced CT examinations may contain richer prognostic value than single-phase imaging. By constructing end-to-end deep neural networks, it is possible to fully exploit the complex relationships between these spatiotemporal features, thereby establishing more accurate and robust recurrence risk prediction models. The advantage of this method lies not only in its improved predictive performance, but also in its clinical application value - it can non-invasively assess patient prognosis before surgery, provide an objective basis for the formulation of individualized treatment plans, the screening of postoperative adjuvant treatments, and the adjustment of follow-up intensity, ultimately achieving the goal of precision medicine. Summary of the Invention

[0004] In response to the defects in the existing technology, the purpose of the present invention is to provide a method for predicting the risk of renal cancer recurrence based on a deep learning model. By analyzing multi-phase enhanced CT images of renal cancer patients through a deep learning model, the risk of tumor recurrence after surgery is predicted, providing an objective basis for clinicians to formulate individualized follow-up plans and auxiliary treatment decisions, avoiding overtreatment of low-risk patients and undertreatment of high-risk patients.

[0005] In order to solve the above problems, the technical solution of the present invention is: A method for predicting the risk of renal cancer recurrence based on a deep learning model, comprising the following steps: Collect image datasets for high recurrence risk prediction of renal cancer; Register the collected multi-phase enhanced CT images; Build and train a renal tumor automatic detection and segmentation model; Perform ROI positioning, cropping and quality control; A deep learning model for predicting the risk of renal cancer recurrence is constructed based on a multimodal convolutional neural network, and the risk of renal cancer recurrence is predicted through the constructed prediction network model.

[0006] Preferably, the step of collecting an image data set for predicting a high recurrence risk of renal cancer specifically includes: collecting preoperative enhanced CT imaging data of patients who have undergone partial or radical nephrectomy, where the CT image data must include three types of enhanced CT images: arterial phase, venous phase, and plain scan phase; marking the recurrence risk based on the patient's postoperative follow-up information. Patients who have not experienced recurrence or metastasis during the follow-up period and whose follow-up time exceeds 5 years are defined as having a low recurrence risk. Patients who have experienced recurrence or metastasis during the follow-up period are defined as having a high recurrence risk of renal cancer.

[0007] Preferably, the step of registering the collected multi-phase enhanced CT images specifically includes: for the plain phase, arterial phase, and venous phase CT images of each patient collected clinically, the plain phase CT image and the venous phase CT image are used as floating images, and the arterial phase CT image is used as a reference image, and the registration and alignment of the multi-phase enhanced CT images are achieved by using affine transformation through the universal registration tool of 3D Slicer, and the enhanced CT images of the plain phase and venous phase are registered to the arterial phase CT image.

[0008] Preferably, the step of constructing and training a renal tumor automatic detection and segmentation model specifically includes: dividing a portion of the enhanced CT image data collected clinically as a training data set for the renal tumor detection model; for the arterial phase enhanced CT images in the training set, imaging experts perform pixel-level manual segmentation and annotation of the kidney and renal tumor areas, and the annotated areas include the kidney and tumor areas as the local training set of the segmentation model; deploying the nnU-Net network in a local GPU workstation or server; using the CT images in the local training set and the corresponding image annotations to train the nnU-Net network; and using the trained nnU-Net model as the renal tumor detection model.

[0009] Preferably, the steps of performing ROI positioning, cropping, and quality control specifically include: locating the axial slice containing the largest tumor area based on the renal mass detected according to the segmentation results of the renal tumor region, cropping and extracting an ROI with a physical scale of 14 cm × 14 cm based on the center of the tumor; in the case of multiple renal tumors, manually selecting the most invasive renal tumor for ROI cropping based on the pathology report and clinical information by the reviewer; and resampling the cropped ROI to a spatial resolution of 0.625 mm × 0.625 mm per pixel, ultimately obtaining an image block with a scale of 224 pixels × 224 pixels.

[0010] Preferably, the deep learning model for predicting the risk of renal cancer recurrence based on the multimodal convolutional neural network is constructed, and the step of predicting the risk of renal cancer recurrence is realized through the constructed prediction network model, specifically including: constructing a deep learning model for predicting the risk of renal cancer recurrence based on the multimodal convolutional neural network, and inputting the enhanced CT images of three phases into the prediction network model to obtain the probability value of the patient's renal cancer recurrence risk, thereby realizing accurate identification of renal cancer with high recurrence risk.

[0011] Preferably, the deep learning model includes three convolutional encoders, a feature fusion module based on deep feature vector stacking, and a classification prediction head based on an MLP structure, wherein the encoder networks for three different enhanced CT phases are independent of each other and all adopt a ResNet-18 network structure. For each CT axial plane ROI view of the arterial phase, venous phase, and plain scan phase, the input corresponding encoder will be mapped into a 512-dimensional feature vector respectively. The feature vectors of the three different phases are fed into the classification prediction head after feature stacking, and the output predicts the probability value of the current patient having a high risk of renal cancer recurrence and a low risk of recurrence, thereby realizing the prediction of the recurrence risk of renal cancer in a classified manner.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This method uses a deep learning model to analyze multi-phase contrast-enhanced CT images of renal cancer patients. It integrates the spatiotemporal characteristics of the arterial, venous, and plain scan phases through a three-dimensional convolutional neural network. Compared with traditional single-phase two-dimensional analysis methods, it can more comprehensively assess tumor biological characteristics and predict the risk of postoperative tumor recurrence. This provides an objective basis for clinicians to formulate individualized follow-up plans and auxiliary treatment decisions, thereby avoiding overtreatment of low-risk patients and undertreatment of high-risk patients. 2. This paper constructs a fully automatic renal cancer lesion segmentation model, which can accurately identify tumor areas without manual delineation, greatly improving analysis efficiency; 3. This paper adopts an improved deep residual network architecture and uses multi-scale feature fusion technology to simultaneously capture local details and global semantic features, which has stronger feature expression capabilities than traditional imaging omics methods; 4. The model was established based on large-sample, long-term follow-up data. The training cohort included more than 2,000 kidney cancer patients with complete prognostic follow-up information for more than 5 years, ensuring the model's predictive reliability. It covers a comprehensive range of kidney cancer pathological subtypes, including not only the common clear cell carcinoma, but also rare types such as papillary carcinoma, chromophobe cell carcinoma, and medullary carcinoma, as well as familial hereditary kidney cancer cases, making the model more widely applicable. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings: Figure 1 This is a flowchart of the method for predicting the risk of renal cancer recurrence based on a deep learning model of the present invention; Figure 2 This is a flow chart of the method for predicting the risk of renal cancer recurrence based on a deep learning model of the present invention; Figure 3 This is a schematic diagram of the network structure for renal tumor detection; Figure 4 Schematic diagram of the multi-view feature fusion convolutional neural network structure. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0015] Specifically, the present invention proposes a method for predicting the risk of renal cancer recurrence based on a deep learning model. Figure 1 and Figure 2 As shown, the method includes the following steps: S1: Collection of image datasets for high recurrence risk prediction of renal cancer; Specifically, preoperative enhanced CT imaging data from patients who underwent partial or radical nephrectomy were collected. CT image data included three types of enhanced CT images: arterial phase, venous phase, and plain phase. Recurrence risk was assessed based on the patient's postoperative follow-up information. Patients with no recurrence or metastasis during follow-up and a follow-up period of more than 5 years were defined as having a low recurrence risk. Patients with any recurrence or metastasis during follow-up were defined as having a high recurrence risk of kidney cancer.

[0016] S2: registering the collected multi-phase enhanced CT images; Specifically, for the three enhanced CT images (plain phase, arterial phase, and venous phase CT images) of each patient collected clinically, the plain phase CT image and the venous phase CT image are used as floating images, and the arterial phase CT image is used as the reference image. The general registration tool (General Registration) of 3D Slicer uses affine transformation to achieve registration and alignment of multi-phase enhanced CT images, and the enhanced CT images of the plain phase and venous phase are registered to the arterial phase CT image for the framing process of the region of interest (ROI).

[0017] S3: Build and train a renal tumor automatic detection and segmentation model; Specifically, a portion of the enhanced CT image data collected clinically is divided as a training data set for the renal tumor detection model. For the arterial phase enhanced CT images in the training set, imaging experts manually segment and annotate the kidney and renal tumor areas at the pixel level. The annotated areas include the kidney and tumor areas, which serve as the local training set for the segmentation model. In one embodiment, a local GPU workstation or server is deployed as follows: Figure 3 The nnU-Net network shown in the figure inherits the weights of the pre-trained model in the KiTS dataset. The nnU-Net network is fine-tuned using the CT images and corresponding image annotations in the local training set. The trained nnU-Net model is then deployed on a GPU server or workstation to perform direct inference on the remaining arterial phase CT images, outputting automatic delineation results for the kidney and renal tumor regions corresponding to the CT images.

[0018] S4: Perform ROI positioning, cropping and quality control; Specifically, based on the renal mass detected by segmentation of the renal tumor region, the axial slice containing the largest tumor area is located. In one embodiment, a senior imaging expert is assigned to review all slices and the corresponding segmentation result images to ensure that the center of the renal tumor is accurately located using the segmentation mask. In the event of segmentation errors, the imaging expert manually corrects the errors and manually marks the tumor center. Subsequently, a ROI with a physical size of 14 cm × 14 cm is cropped from the tumor center. In cases involving multiple renal tumors, the reviewer manually selects the most invasive renal tumor for ROI cropping based on the pathology report and clinical information. In all cases, the cropped ROI needs to be spatially resampled to a spatial resolution of 0.625 mm × 0.625 mm per pixel, resulting in an image block size of 224 pixels × 224 pixels.

[0019] S5: A deep learning model for predicting the risk of renal cell carcinoma recurrence was constructed based on a multimodal convolutional neural network, and the risk of renal cell carcinoma recurrence was predicted through the constructed prediction network model.

[0020] Specifically, a deep learning model for predicting the risk of renal cancer recurrence using a multimodal convolutional neural network was constructed, such as Figure 4 As shown, the deep learning model includes three convolutional encoders, a feature fusion module based on deep feature vector stacking, and a classification prediction head based on an MLP structure, which classifies and distinguishes high and low recurrence risks of renal cancer: the encoder networks for the three different enhanced CT phases are independent of each other, all using the ResNet-18 network structure and pre-trained on ImageNet. For each CT axial plane ROI view of the arterial phase, venous phase, and plain scan phase, the input to the corresponding encoder will be mapped into a 512-dimensional feature vector. The feature vectors of the three different phases are stacked and sent to the classification prediction head, which outputs the probability value of predicting the current patient's high and low recurrence risks of renal cancer, thereby predicting the recurrence risk of renal cancer in a classified manner.

[0021] The trained prediction network model described above is deployed in the local GPU server. By inputting the enhanced CT images of the three phases into the prediction network model, the probability value of the patient's renal cancer recurrence risk can be directly obtained, thereby achieving accurate identification of renal cancer with a high recurrence risk.

[0022] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A method for predicting the recurrence risk of renal cancer based on a deep learning model, characterized in that: The method comprises the following steps: Collect image datasets for high recurrence risk prediction of renal cancer; Register the collected multi-phase enhanced CT images; Build and train a renal tumor automatic detection and segmentation model; Perform ROI positioning, cropping and quality control; A deep learning model for predicting the risk of renal cancer recurrence is constructed based on a multimodal convolutional neural network, and the risk of renal cancer recurrence is predicted through the constructed prediction network model.

2. The method for predicting renal cancer recurrence risk based on a deep learning model according to claim 1, characterized in that: The steps of collecting an image dataset for predicting a high recurrence risk of renal cancer specifically include: collecting preoperative enhanced CT imaging data of patients who have undergone partial or radical nephrectomy, where the CT image data must include three types of enhanced CT images: arterial phase, venous phase, and plain scan phase; and marking the recurrence risk based on the patient's postoperative follow-up information. Patients who have not experienced recurrence or metastasis during the follow-up period and whose follow-up time exceeds 5 years are defined as having a low recurrence risk. Patients who experience recurrence or metastasis during the follow-up period are defined as having a high recurrence risk of renal cancer.

3. The method for predicting the recurrence risk of renal cancer based on a deep learning model according to claim 1, characterized in that: The step of registering the collected multi-phase enhanced CT images specifically includes: for the plain phase, arterial phase, and venous phase CT images of each patient collected clinically, the plain phase CT image and the venous phase CT image are used as floating images, and the arterial phase CT image is used as a reference image. The universal registration tool of 3D Slicer is used to use affine transformation to achieve registration and alignment of the multi-phase enhanced CT images, and the plain phase and venous phase enhanced CT images are registered to the arterial phase CT image.

4. The method for predicting renal cancer recurrence risk based on a deep learning model according to claim 1, wherein: The steps of constructing and training the automatic detection and segmentation model for renal tumors specifically include: allocating a portion of the enhanced CT image data collected clinically as a training data set for the renal tumor detection model; for the arterial phase enhanced CT images in the training set, having imaging experts manually segment and annotate the kidney and renal tumor areas at the pixel level, where the annotated areas include the kidney and tumor areas as the local training set for the segmentation model; deploying the nnU-Net network in a local GPU workstation or server; training the nnU-Net network using the CT images in the local training set and the corresponding image annotations; and using the trained nnU-Net model as the renal tumor detection model.

5. The method for predicting renal cancer recurrence risk based on a deep learning model according to claim 1, wherein: The steps of performing ROI positioning, cropping, and quality control specifically include: locating the axial slice containing the largest tumor area based on the renal mass detected according to the segmentation results of the renal tumor region, cropping and extracting a ROI with a physical scale of 14 cm × 14 cm based on the center of the tumor; in the case of multiple renal tumors, the reviewer manually selects the most invasive renal tumor for ROI cropping based on the pathology report and clinical information; and resampling the cropped ROI to a spatial resolution of 0.625 mm × 0.625 mm per pixel, ultimately obtaining an image block with a scale of 224 pixels × 224 pixels.

6. The method for predicting renal cancer recurrence risk based on a deep learning model according to claim 1, characterized in that: The deep learning model for predicting the risk of renal cancer recurrence based on a multimodal convolutional neural network is constructed, and the steps of predicting the risk of renal cancer recurrence are realized through the constructed prediction network model. Specifically, the deep learning model for predicting the risk of renal cancer recurrence based on a multimodal convolutional neural network is constructed, and the prediction network model has been trained in a local GPU server. By inputting enhanced CT images of three phases into the prediction network model, the probability value of the patient's renal cancer recurrence risk is obtained, thereby realizing accurate identification of renal cancer with a high recurrence risk.

7. The method for predicting renal cancer recurrence risk based on a deep learning model according to claim 6, characterized in that: The deep learning model includes three convolutional encoders, a feature fusion module based on deep feature vector stacking, and a classification prediction head based on an MLP structure. The encoder networks for three different enhanced CT phases are independent of each other and all adopt the ResNet-18 network structure. For each CT axial plane ROI view of the arterial phase, venous phase, and plain scan phase, the input of the corresponding encoder will be mapped into a 512-dimensional feature vector respectively. The feature vectors of the three different phases are fed into the classification prediction head after feature stacking, and the output predicts the probability value of the current patient having a high risk of renal cancer recurrence or a low risk of recurrence, thereby predicting the risk of renal cancer recurrence in a classified manner.

Images