CN-122025089-A - CT image-based lung cancer curative effect prediction system

CN122025089ACN 122025089 ACN122025089 ACN 122025089ACN-122025089-A

Abstract

The invention discloses a lung cancer curative effect prediction system based on CT images, which relates to the technical field of medical image analysis and comprises a data input and preprocessing module, an image quality grading and artifact identification module, an artifact correction module, an image histology feature extraction module and a lung cancer curative effect prediction module. The lung cancer curative effect prediction system based on CT images, provided by the invention, remarkably improves the accuracy and reliability of lung cancer curative effect prediction and solves the problem of insufficient multi-center data compatibility by integrating advanced data processing and deep learning technologies.

Inventors

SHI XIANGRONG
GE JIANJUAN
LU YI
HAO LI

Assignees

南通市肿瘤医院（南通市第五人民医院）

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (7)

1. The lung cancer curative effect prediction system based on CT image is characterized by comprising: the device comprises a data input and preprocessing module, an image quality grading and artifact identification module, an artifact correction module, an image group science feature extraction module and a lung cancer curative effect prediction module; The data input and preprocessing module is used for receiving CT image data and clinical auxiliary information of multiple centers and multiple devices, carrying out format standardization and automatic ROI extraction, and outputting influence data containing tumor areas; the image quality classification and artifact identification module is electrically connected with the data input and preprocessing module and is used for identifying image artifacts and quality grades based on the CNN model, screening qualified images and triggering supplementary scanning prompt for unqualified images; the artifact correction module is electrically connected with the image quality grading and artifact identification module and is used for correcting respiratory artifacts, metal artifacts and scanning parameter differences and outputting corrected images; The image group science feature extraction module is electrically connected with the artifact correction module and is used for extracting and outputting an image group science feature set through feature screening and consistency verification; The lung cancer curative effect prediction module is electrically connected with the image histology feature extraction module and is used for outputting an individualized curative effect prediction result through a prediction model according to the feature set and clinical auxiliary information.
2. The CT image-based lung cancer efficacy prediction system according to claim 1, wherein the data input and preprocessing module specifically comprises: The received multi-center and multi-equipment CT image data format is DICOM standard format, and the clinical auxiliary information comprises patient basic information, treatment scheme information, pathological diagnosis results and follow-up records; The format standardization processing comprises the steps of uniformly converting DICOM files output by different devices into NIfTI formats, and simultaneously cutting the pixel value range to-1000-400 HU; The automatic ROI extraction adopts a semantic segmentation model based on U-Net, takes a tumor region and lung parenchyma within a surrounding 5mm range as segmentation targets, and outputs three-dimensional image data containing a complete tumor region.
3. The CT image based lung cancer efficacy prediction system according to claim 2, wherein the data input and preprocessing module further comprises: performing post-processing optimization on the automatically extracted ROI region, removing a noise region through operation, and correcting the saw-tooth artifact of the segmentation edge by using a boundary smoothing algorithm; Carrying out structural processing on the clinical auxiliary information, extracting key fields of patient age, sex, tumor pathology type, TNM stage and therapeutic drug type, and storing the key fields as a structural data table after standardized field format; And taking the patient unique identification ID as an association key, carrying out association matching on the normalized CT image ROI data and the structured clinical auxiliary information, and generating an image and clinical integrated data set.
4. The CT image-based lung cancer efficacy prediction system according to claim 3, wherein the image quality classification and artifact identification module specifically comprises: the CNN model adopts a lightweight MobileNetV as a backbone network, and model parameters are reduced through depth separable convolution; The training data set is a multi-center labeling data set of a plurality of medical institutions, and comprises four common problem images of breathing artifact, metal artifact, motion artifact and low-dose noise, wherein each image is labeled with a quality grade and artifact region coordinates; the model training adopts a cross entropy loss function, and simultaneously optimizes a quality grading task and an artifact identification task.
5. The CT image-based lung cancer efficacy prediction system according to claim 4, wherein the image quality classification and artifact identification module further comprises: the quality grade judging standard is that 0 grade has no artifact, the texture of a tumor area is clear, the gray level is uniform, 1 grade artifact accounts for 10 percent, 2 grade artifact accounts for 10 percent to 30 percent, and 3 grade artifact accounts for 30 percent; the artifact positioning output is a two-dimensional/three-dimensional coordinate matrix, and the initial pixel point, the end pixel point and the influence range of the artifact region are marked; for the images judged to be 3 levels, the system automatically generates a supplementary scanning prompt containing the work order ID, the patient ID, the artifact type and the influence degree, and pushes the supplementary scanning prompt to a clinical image workstation through an HL7 interface.
6. The CT image based lung cancer therapy efficacy prediction system according to claim 5, wherein the artifact correction module corrects for respiratory/motion artifacts by: extracting contour key points of adjacent CT (computed tomography) layer tumor areas by a Canny edge detection algorithm, wherein each layer selects at least 10 characteristic key points; calculating the contour offset of the adjacent layers based on the coordinates of the key points, and fitting the motion track by using a polynomial fitting algorithm to obtain a dynamic displacement curve of the tumor in the scanning process; And (3) carrying out pixel realignment on the artifact region according to the displacement curve, matching similar gray blocks in 33 pixel neighborhood by using K nearest neighbor algorithm on the pixel points which are missing after realignment, and completing pixel repair by using a weighted interpolation method.
7. The CT image-based lung cancer efficacy prediction system according to claim 6, wherein the artifact correction module comprises: Identifying a metal region in the CT image by adopting an Otsu self-adaptive threshold segmentation algorithm, setting the gray threshold range to be 1000-4096 HU, and masking and shielding the metal region; the GAN model comprises a generator and a discriminator, wherein the generator adopts a U-Net structure, and takes normal lung tissue textures within a range of 5mm around metal as input; The arbiter adopts a convolutional neural network to distinguish the filling texture output by the generator from the real normal texture, and optimizes the performance of the generator through countermeasure training.

Description

CT image-based lung cancer curative effect prediction system Technical Field The invention relates to the technical field of medical image analysis, in particular to a lung cancer curative effect prediction system based on CT images. Background In clinical diagnosis and treatment of lung cancer, CT images have become a core imaging means for tumor screening, focus positioning and treatment effect evaluation by virtue of the advantages of high resolution, noninvasive property and the like. At present, the fusion of a medical image analysis technology and an artificial intelligence algorithm provides possibility for mining tumor pathological features and physiological state information hidden in CT images, and the early judgment of lung cancer treatment response is hopefully realized by extracting image histology features and constructing a prediction model, so that data support is provided for clinical decisions. However, the prior related art still faces a plurality of problems to be optimized in practical application, namely that multi-center data are sourced from different devices, the problems of non-uniform formats, large scanning parameter difference and the like exist, data compatibility is insufficient, CT images are easily influenced by factors such as respiration, motion and metal implants to generate artifacts, the existing artifact processing means are not strong in pertinence, the feature extraction accuracy is influenced, meanwhile, the stability and consistency of image histology features lack of system verification, part of features have large variation in multi-center data sets, and the precise requirements of individual prediction are difficult to meet. In addition, the integration of clinical auxiliary information and image data is not efficient enough, and the synergistic effect of the two is not fully exerted. These problems limit the reliability and generalization ability of efficacy predictions to some extent. Disclosure of Invention In order to solve the technical problems, the lung cancer curative effect prediction system based on CT images is provided. In order to achieve the above purpose, the invention adopts the following technical scheme: A lung cancer efficacy prediction system based on CT images, comprising: the device comprises a data input and preprocessing module, an image quality grading and artifact identification module, an artifact correction module, an image group science feature extraction module and a lung cancer curative effect prediction module; The data input and preprocessing module is used for receiving CT image data and clinical auxiliary information of multiple centers and multiple devices, carrying out format standardization and automatic ROI extraction, and outputting influence data containing tumor areas; the image quality classification and artifact identification module is electrically connected with the data input and preprocessing module and is used for identifying image artifacts and quality grades based on the CNN model, screening qualified images and triggering supplementary scanning prompt for unqualified images; the artifact correction module is electrically connected with the image quality grading and artifact identification module and is used for correcting respiratory artifacts, metal artifacts and scanning parameter differences and outputting corrected images; The image group science feature extraction module is electrically connected with the artifact correction module and is used for extracting and outputting an image group science feature set through feature screening and consistency verification; The lung cancer curative effect prediction module is electrically connected with the image histology feature extraction module and is used for outputting an individualized curative effect prediction result through a prediction model according to the feature set and clinical auxiliary information. Preferably, the data input and preprocessing module specifically includes: The received multi-center and multi-equipment CT image data format is DICOM standard format, and the clinical auxiliary information comprises patient basic information, treatment scheme information, pathological diagnosis results and follow-up records; The format standardization processing comprises the steps of uniformly converting DICOM files output by different devices into NIfTI formats, and simultaneously cutting the pixel value range to-1000-400 HU; The automatic ROI extraction adopts a semantic segmentation model based on U-Net, takes a tumor region and lung parenchyma within a surrounding 5mm range as segmentation targets, and outputs three-dimensional image data containing a complete tumor region. Preferably, the data input and preprocessing module further comprises: performing post-processing optimization on the automatically extracted ROI region, removing a noise region through operation, and correcting the saw-tooth artifact of the segmentation edge by using a boundary s