A target detection method

Abstract

This invention relates to the field of target detection technology, and in particular to a target detection method based on human computed tomography (CT) images. The method includes: acquiring human medical digital imaging and communication image datasets of different objects; reconstructing three-dimensional CT images from sequence files in the medical digital imaging and communication datasets; generating two-dimensional digitally reconstructed radiographic images of the three-dimensional CT images in the coronal and horizontal planes respectively; labeling corresponding target regions on the two-dimensional digitally reconstructed radiographic images as training labels; constructing a YOLO target detection network based on the two-dimensional digitally reconstructed radiographic images; training the target detection network using the two-dimensional digitally reconstructed radiographic images and corresponding region labels to obtain a detection model; and using the obtained detection model to perform target detection on the three-dimensional CT images to obtain the localization information of the target regions. This invention improves the efficiency and accuracy of target detection in human computed tomography (CT) images.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a target detection method based on human computed tomography (CT) images. Background Technology

[0002] In traditional medicine, medical image data requires radiologists to expend considerable effort in screening and interpreting it. The repetitive and mechanical nature of this work can easily lead to fatigue and misdiagnosis. Object detection, however, can quickly locate objects by drawing appropriate bounding boxes around them. This allows for rapid localization of lesions, joints, and other areas of interest to doctors within large sequences of medical images, assisting in faster diagnosis and treatment. The development of deep learning has also significantly improved object detection in medical images. Training deep learning models requires large amounts of labeled datasets, necessitating substantial medical resources and financial investment. For 3D object detection, even more time and effort are required to accurately label training data, along with additional computing resources. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a target detection method that improves the efficiency and accuracy of target detection based on human computed tomography images.

[0004] The technical solution adopted by this invention to solve its technical problem is: a target detection method for locating regions of interest in human computed tomography images, comprising the following steps:

[0005] S1. Obtain human medical digital imaging and communication image datasets of different objects, and reconstruct three-dimensional CT images from the sequence files in the medical digital imaging and communication datasets.

[0006] S2. Generate two-dimensional digital reconstructed radiographic images in the coronal and horizontal directions of the three-dimensional CT images in step S1, and mark the corresponding target areas on the two-dimensional digital reconstructed radiographic images as training labels.

[0007] S3. Construct a YOLO target detection network based on the two-dimensional digitally reconstructed radiographic image from step S2. Train the target detection network using the two-dimensional digitally reconstructed radiographic image and corresponding region labels to obtain the detection model.

[0008] S4. Use the detection model obtained in step S3 to perform target detection on the three-dimensional CT image to obtain the location information of the target area;

[0009] The specific operation steps of step S2 are as follows:

[0010] First, a two-dimensional digital reconstructed radiographic image is generated in the horizontal direction of the three-dimensional CT image of the human body: the digital reconstructed radiographic image of the main body; the generated digital reconstructed radiographic image contains the main body region and interference regions, and the main body region of interest is manually labeled as a training label: the main body label; the labeling results of the previous step are mapped to the three-dimensional CT image, and the three-dimensional CT volume data of the main body of the human body is cropped out. A digital reconstructed radiographic image is then generated in the coronal direction of this volume data: the digital reconstructed radiographic image of the coronal plane, and the target region is labeled on this digital reconstructed radiographic image as a training label: the coronal plane label; then, the labeling results of the previous step are mapped to the three-dimensional CT volume data of the main body of the human body, and the three-dimensional CT volume data of the target region is cropped out. A digital reconstructed radiographic image is generated in the horizontal direction of this volume data: the digital reconstructed radiographic image of the horizontal plane, and the target region is labeled on this digital reconstructed radiographic image as a training label: the horizontal plane label;

[0011] The specific operation steps of step S3 are as follows:

[0012] A localization network for the human trunk was constructed using major digitally reconstructed radiographic images and major labels: main body localization, and the model was trained and saved; a localization network for the coronal plane was constructed using digitally reconstructed radiographic images and coronal plane labels: coronal plane target localization, and the model was trained and saved; a localization network for the horizontal plane was constructed using digitally reconstructed radiographic images and horizontal plane labels: horizontal plane target localization, and the model was trained and saved.

[0013] The specific operation steps of step S4 are as follows:

[0014] The sequence files of the medical digital imaging and communication dataset are reconstructed into 3D CT volume data. A 2D digitally reconstructed radiographic image of the main body is generated from the CT volume data in the horizontal plane. The main body localization model is used to detect the main body trunk region on the main body digitally reconstructed radiographic image. Based on the main body localization results, 3D volume data containing only the main body trunk region is cropped from the original CT image: main data. Then, a digitally reconstructed radiographic image is generated from the main body data in the coronal plane. A coronal plane target localization model is used to detect the starting position and length and width information of the region of interest (ROI) on the x-axis and z-axis in the 3D coordinate system on the coronal plane digitally reconstructed radiographic image. Based on the results of the previous step, volume data containing the target region and other regions is cropped from the main body data: coronal volume data. A digitally reconstructed radiographic image is generated from the coronal plane volume data in the horizontal plane. A horizontal plane target localization model is used to detect the starting position and height information of the ROI on the y-axis on the horizontal plane digitally reconstructed radiographic image. This information is mapped onto the original 3D CT image of the human body, and the bounding box of the target region is drawn, completing the target detection.

[0015] Furthermore, in step S1, the sequence files in the medical digital imaging and communication image dataset are reconstructed to generate three-dimensional CT images based on the image orientation, image position, pixel spacing, and slice thickness information included in the header file.

[0016] The beneficial effects of this invention are:

[0017] The algorithm for generating two-dimensional digital reconstructed radiographic images from three-dimensional CT volume data in the method provided by this invention ensures the speed and effectiveness of the algorithm.

[0018] The YOLO algorithm in the method provided by this invention ensures the accuracy of the algorithm;

[0019] The method provided by this invention utilizes two-dimensional digitally reconstructed radiographic images from different directions to detect and locate three-dimensional targets, thereby improving the efficiency of three-dimensional target detection and saving computational costs. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Figure 1 This is a flowchart of a target detection method according to the present invention.

[0022] Figure 2 This is a schematic diagram of a target detection method according to the present invention. Detailed Implementation

[0023] The present invention will now be further described in conjunction with the accompanying drawings and preferred embodiments. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention in a schematic manner, and therefore only show the components relevant to the invention.

[0024] like Figure 1 As shown, a target detection method for locating regions of interest in human computed tomography (CT) images includes the following steps:

[0025] This study acquires human medical digital imaging and communication (DICOM) image datasets of different subjects and reconstructs 3D CT images from sequence files in the DICOM dataset. Specifically, the 3D CT images are reconstructed from sequence files in the DICOM dataset based on information such as image orientation, image position, pixel spacing, and slice thickness in the header file.

[0026] Two-dimensional digital reconstructed radiographs (DRRs) are generated from 3D CT images in both the coronal and horizontal planes. Target regions are then labeled on these DRRs as training tags. The specific steps are as follows:

[0027] First, a two-dimensional digital reconstructed radiograph (DRR) is generated in the horizontal direction of the 3D CT image of the human body: the main body's digital reconstructed radiograph (drr_main). The generated main body DRR (drr_main) includes the main body region of the human body and other interfering regions such as the CT table. The main body region of interest is manually labeled as a training label: main body label (label_main). The labeling results of the previous step are mapped to the 3D CT image, and the 3D CT volume data of the main body of the human body is cropped. A digital reconstructed radiograph is then generated in the coronal direction of this volume data: the coronal digital reconstructed radiograph (drr_coronal). The target region is labeled on this DRR as a training label: coronal label (label_coronal). Then, the labeling results of the previous step are mapped to the 3D CT volume data of the main body of the human body, and the 3D CT volume data of the target region is cropped. A digital reconstructed radiograph is generated in the horizontal direction of this volume data: the horizontal digital reconstructed radiograph (drr_axial). The target region is labeled on this DRR as a training label: horizontal label (label_axial).

[0028] A YOLO target detection network is constructed based on two-dimensional digitally reconstructed radiographic images (DRR). The target detection network is trained using two-dimensional DRR images and corresponding region labels to obtain the detection model. The specific operation steps are as follows:

[0029] Before model training, the dataset needs to be preprocessed and augmented. Data preprocessing mainly involves pseudo-color processing of the generated digital reconstructed radiographs (DRRs), converting the two-dimensional DRRs containing only grayscale information into pseudo-color images with color information, suitable for YOLO model training and inference. Then, data augmentation is performed on the pseudo-color processed DRRs and their corresponding labels using methods such as rotation, flipping, cropping, and scaling. A human trunk localization network, "Focus_Main," is constructed using the main body's DRR (drr_main) and main body label (label_main), and the model is trained and saved. A coronal localization network, "Focus_Coronal," is constructed using the coronal DRR (drr_coronal) and coronal label (label_coronal), and the model is trained and saved. A horizontal localization network, "Focus_Axial," is constructed using the horizontal DRR (drr_axial) and horizontal label (label_axial), and the model is trained and saved.

[0030] The obtained detection model is used to perform target detection on 3D CT images to obtain the localization information of the target area. The specific operation steps are as follows:

[0031] The sequence files of the Medical Digital Imaging and Communications (DICOM) dataset were reconstructed into 3D CT volume data. A 2D digitally reconstructed radiographic image (drr_main) of the main body was generated from the CT volume data in the horizontal plane. The main body localization (focus_main) model was used to detect the main body trunk region on the drr_main digitally reconstructed radiographic image. Based on the focus_main results, 3D volume data containing only the main body trunk region was cropped from the original CT images: the main data (main_Volume). Then, a digitally reconstructed radiographic image (drr_main) was generated from the main data (main_Volume) in the coronal plane. The focus_coronal model was used to detect the target of interest on the coronal digitally reconstructed radiographic image (drr_coronal). The region's starting position and length and width information (x-position, z-position, width, height) along the x and z axes in a 3D coordinate system are determined. Based on the results of the previous step, volume data containing the target region and other regions is cropped from the main data (main_Volume): coronal volume data (coronal_Volume). A digital reconstructed radiographic image (drr_axial) is generated from the coronal volume data (coronal_Volume) in the horizontal direction. The horizontal target localization (focus_axial) model is used to detect the region of interest's starting position and height information (y-position, length) along the y-axis on the horizontal digital reconstructed radiographic image (drr_axial). This information (x-position, y-position, z-position, width, height, length) is mapped onto the original human 3D CT image to draw the bounding box (bbox) of the target region, thus completing the target detection.

[0032] YOLO is a convolutional neural network (CNN) used for real-time object detection. The algorithm applies a single neural network to the entire image, then divides the image into multiple regions and predicts bounding boxes and probabilities for each region, weighted by the predicted probabilities. YOLO is a regression-based algorithm; instead of selecting interesting parts of the image, it predicts classes and bounding boxes across the entire image. Instead of searching the input image for regions of interest that might contain objects, it divides the image into cells, each responsible for predicting N bounding boxes. The class with the highest probability is selected and assigned to that specific cell. After predicting class probabilities, the next step is non-maximum suppression, which helps the algorithm eliminate unnecessary anchors. Once this is complete, the algorithm finds the bounding box with the next highest class probability and repeats the same process until all distinct bounding boxes remain. The algorithm finally outputs the desired vector, showing the details of the bounding boxes for each class.

[0033] DICOM (Digital Imaging and Communications in Medicine) is an international standard for medical images and related information, widely used in radiology, cardiovascular imaging, and diagnostic radiology equipment (X-ray, CT, MRI, ultrasound, etc.). All patient medical images are stored in DICOM file format, which contains patient information and other image-related information. Diagnostic radiology equipment performs cross-sectional scans around a part of the body, resulting in multi-layered images. These layers are then stacked along the z-axis to form a three-dimensional image. Each layer is stored in a DICOM file. A DICOM file typically consists of a DICOM header and a DICOM dataset. The DICOM header contains information identifying the dataset, and each DICOM file includes a header.

[0034] DRR (Digitally Reconstructed Radiograph) is a technique that uses simulated light to pass through the human body to perform a three-dimensional scan and reconstruct grayscale images that can represent the various internal organs of the human body.

[0035] like Figure 2As shown, the principle of the target detection method is as follows: A two-dimensional DRR image of the main body is generated in the horizontal plane from the reconstructed 3D human CT image. This image is used for main body localization, specifically for locating the human trunk. Based on the localization result, the main body data containing the human trunk is cropped from the original volume data. A two-dimensional DRR image of the coronal plane is generated from the main body data in the coronal plane. This image is used for coronal plane target localization, confirming the target's position (x-position, z-position, width, height) in the 3D coordinate system. Based on the localization result, coronal volume data containing the target area and other areas is cropped from the main body data. A two-dimensional DRR image of the coronal volume data is generated in the horizontal plane. This image is used for horizontal plane target localization, confirming the target's position (y-position, length) in the 3D coordinate system. The bounding box (bbox) of the target area is drawn in the original CT image using (x-position, y-position, z-position, width, length, height), thus completing the target detection.

[0036] The above embodiments are only for illustrating the technical concept and features of the present invention. Their purpose is to enable those skilled in the art to understand the content of the present invention and implement it. They should not be used to limit the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A target detection method for locating regions of interest in human computed tomography (CT) images, characterized in that, Includes the following steps: S1. Obtain human medical digital imaging and communication image datasets of different objects, and reconstruct three-dimensional CT images from the sequence files in the medical digital imaging and communication datasets. S2. Generate two-dimensional digital reconstructed radiographic images in the coronal and horizontal directions of the three-dimensional CT images in step S1, and mark the corresponding target areas on the two-dimensional digital reconstructed radiographic images as training labels. S3. Construct a YOLO target detection network based on the two-dimensional digitally reconstructed radiographic image from step S2. Train the target detection network using the two-dimensional digitally reconstructed radiographic image and corresponding region labels to obtain the detection model. S4. Use the detection model obtained in step S3 to perform target detection on the three-dimensional CT image to obtain the location information of the target area; The specific steps of step S2 are as follows: First, a two-dimensional digital reconstructed radiographic image is generated in the horizontal direction of the 3D CT image of the human body: the digital reconstructed radiographic image of the main body; the generated digital reconstructed radiographic image includes the main body region and interference regions, and the main body region of interest is manually labeled as a training label: the main body label; the labeling results of the previous step are mapped to the 3D CT image, and the 3D CT volume data of the main body of the human body is cropped. A digital reconstructed radiographic image is then generated in the coronal direction of this volume data: the digital reconstructed radiographic image of the coronal plane, and the target region is labeled on this digital reconstructed radiographic image as a training label: the coronal plane label; then, the labeling results of the previous step are mapped to the 3D CT volume data of the main body of the human body, and the 3D CT volume data of the target region is cropped. A digital reconstructed radiographic image is generated in the coronal direction of this volume data: the digital reconstructed radiographic image of the horizontal plane, and the target region is labeled on this digital reconstructed radiographic image as a training label: the horizontal plane label; The specific operation steps of step S3 are as follows: A localization network for the human trunk is constructed using digitally reconstructed radiographic images of the main body and main body labels: main body localization, model training and saving; a localization network for the coronal plane is constructed using digitally reconstructed radiographic images of the coronal plane and coronal plane labels: coronal plane target localization, model training and saving; a localization network for the horizontal plane is constructed using digitally reconstructed radiographic images of the horizontal plane and horizontal plane labels: horizontal plane target localization, model training and saving. The specific operation steps of step S4 are as follows: The sequence files of the medical digital imaging and communication dataset are reconstructed into 3D CT volume data. A 2D digitally reconstructed radiographic image of the main body is generated from the CT volume data in the horizontal plane. The main body localization model is used to detect the main body trunk region on the main body digitally reconstructed radiographic image. Based on the main body localization results, 3D volume data containing only the main body trunk region is cropped from the original CT image: main body data. Then, the main body data is scanned in the coronal plane to generate a digitally reconstructed radiographic image. The coronal plane target localization model is used to detect the starting position and length and width information of the region of interest (ROI) on the x and z axes in the 3D coordinate system on the coronal plane digitally reconstructed radiographic image. Based on the results of the previous step, volume data containing the target region and other regions is cropped from the main body data: coronal plane volume data. A digitally reconstructed radiographic image is generated from the coronal plane volume data in the horizontal plane. The horizontal plane target localization model is used to detect the starting position and height information of the ROI on the y axis on the horizontal plane digitally reconstructed radiographic image. This information is mapped onto the original 3D CT image of the human body, and the bounding box of the target region is drawn, completing the target detection.

2. The target detection method according to claim 1, characterized in that: In step S1, the sequence files in the medical digital imaging and communication image dataset are used to reconstruct three-dimensional CT images based on the image orientation, image position, pixel spacing, and slice thickness information included in the header file.

Images