Target detection method based on region positioning

Abstract

The invention relates to a target detection method based on region positioning, and the method mainly comprises the steps: carrying out the chest region preprocessing of images in a CT image sequence, carrying out the pixel-level communication region recognition of each image, carrying out the label distribution and merging of the recognized communication regions, and carrying out the screening and similarity judgment of the communication regions, the thoracic cavity area of the CT image is determined, then a deep learning model and a database are established, and efficient and automatic CT image sequence target detection is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of medical image analysis, and in particular relates to a CT image analysis method based on deep learning, which is used to accurately locate and detect targets in key areas. Background Art

[0002] Chest computed tomography (CT) is an important tool for observing the structure of the chest cavity in clinical diagnosis. It can show the chest cavity structures such as the lungs, heart and large blood vessels in detail, and plays an important auxiliary role in assisting doctors to judge pneumonia, tumors and other chest diseases. With the rapid development of medical imaging technology, the speed and number of CT images have increased significantly, which not only provides rich information for clinical practice, but also brings challenges in data processing.

[0003] Traditional CT image analysis mainly relies on visual inspection by radiologists, which is not only time-consuming but also easily affected by subjective factors. Especially when faced with a large amount of image data, it may affect the efficiency and accuracy of doctors' diagnosis. In order to solve this problem, automated image processing technology, especially deep learning technology, has gradually been introduced into CT image analysis in order to improve processing speed and the accuracy of auxiliary diagnosis.

[0004] Deep learning, especially convolutional neural networks (CNNs), has shown excellent performance in image recognition and classification tasks. Applying deep learning technology to the analysis of CT images can effectively identify and locate key areas in the image, providing more accurate and objective diagnostic references. However, existing methods usually directly process the complete CT image and fail to make full use of the key areas in the image, which increases the complexity of model processing and limits its accuracy.

[0005] This paper proposes a target detection method based on regional positioning, which simplifies the complexity of deep learning processing objects by accurately positioning the key areas in the chest. By focusing on the key areas of the chest, this method helps to improve the learning efficiency of the model and the accuracy of auxiliary diagnosis, especially when processing large-scale CT image data, it can significantly improve the feasibility and effect of automated analysis. Summary of the invention

[0006] In view of the above-mentioned deficiencies of the prior art, the present invention proposes a target detection method based on regional positioning, comprising the following steps:

[0007] Step 1: Perform chest area preprocessing operations on the acquired CT images;

[0008] Step 2: Set the binarization threshold tht to binarize the CT image after dilation processing;

[0009] Step 3: Locate the chest area on the CT image, specifically:

[0010] Step 3.1: Identify connected regions at the pixel level on the CT image;

[0011] Step 3.2: Assign labels to the connected areas of the CT image;

[0012] Step 3.3: Merge all connected area labels in the binary list;

[0013] Step 3.4: Filter all connected areas in the merged binary list;

[0014] Step 3.5: Perform similarity judgment on the selected connected areas to obtain the chest area of ​​the CT image;

[0015] Step 4: Based on the extracted chest area of ​​the CT image, a neural network model is established to detect the CT image.

[0016] Furthermore, in step 3.1, the connected area identification is performed on the CT image at the pixel level, specifically: each pixel in the binarization processing result corresponding to the CT image is scanned, and the pixels in its eight neighborhoods are checked. If the pixel values ​​are the same, it is determined that these pixels belong to the connected area; if the pixel values ​​are different, it is determined that these pixels belong to the edge area; if the scanned pixel does not have a complete eight-neighborhood pixel, the pixel is discarded and not processed.

[0017] Furthermore, in step 3.2, labels are assigned to the connected areas of the CT image, specifically: a corresponding binarization list is set for each binarization processing result, and a unique connected area label is assigned to each identified connected area in each binarization processing result. Whenever a new connected area is found, its connected area label is added to the binarization list corresponding to the currently processed binarization processing result.

[0018] Furthermore, in the step 3.3, all connected area labels in the binary list are merged, specifically: first, when there is one or more pixel points with the same horizontal and vertical coordinate index in the connected areas corresponding to two connected area labels, the two connected areas are regarded as interconnected; then, when it is determined that the two connected areas are interconnected, the two areas will be merged into a new connected area, and the merging operation is to take the union of the two areas, that is, the new connected area includes all pixels in the original two areas; finally, the smaller label of the two connected areas before the merger is used as the label of the new merged area, and the larger label of the two connected areas before the merger is deleted.

[0019] Furthermore, in the step 3.4, all connected areas in the merged binary list are screened, specifically: first, the pixel values ​​of the pixels in the connected areas are judged, and the connected areas with pixel values ​​of 0 are extracted, and then the number of pixels contained in each extracted connected area is judged, and the connected areas with the number of pixels greater than the lower chest threshold thdn and less than the upper chest threshold thup are screened.

[0020] Furthermore, in the step 3.5, similarity judgment is performed on the screened connected areas to obtain the chest area of ​​the CT image, specifically: the screened connected areas are mapped to the CT image, and histograms are established for the corresponding connected areas on the CT image, and the KL distance between the two histograms is calculated, and the two connected areas on the CT image corresponding to the two histograms with the shortest KL distance are selected, recorded as chest area A and chest area B, and the two are collectively used as the chest area of ​​the CT image.

[0021] Furthermore, in step 4, based on the extracted chest area of ​​the CT image, a neural network model is established to detect the CT image, specifically: first, a CT image database is established; second, a neural network model is established; third, based on the CT image database, the neural network model is trained and evaluated; finally, the trained neural network model is applied to actual CT image data for target detection.

[0022] Furthermore, in step 1, the collected CT images are subjected to chest region preprocessing operations, specifically: first, a circular structure element of size dx is selected to perform corrosion processing on the image; then a circular structure element of size dy is selected to perform expansion processing on the corrosion processing result.

[0023] Furthermore, the establishing of the CT image database specifically includes: extracting chest regions from the labeled CT image data set, and marking these regions into two categories: target presence and target absence.

[0024] Furthermore, 10. The establishment of the neural network model is specifically as follows: first, a pre-trained ResNet-50 model is used as the basic framework for feature extraction, and then the fully connected layer of the original model is removed from the tail of the model, and a fully connected layer is added, containing 256 neurons, using the ReLU activation function, and then a Dropout layer is added with a dropout rate of 0.5, and finally an output layer is added, which is a fully connected layer containing 2 neurons, and a softmax activation function is used for binary classification.

[0025] Compared with the traditional detection method, the present invention has the following advantages, thereby solving the corresponding technical problems:

[0026] 1. By locating the chest area for each image in the CT image sequence, the complexity of the data that the deep learning model needs to process is significantly reduced, allowing the model to focus more on potential target areas and enhancing the accuracy of the model in practical applications.

[0027] 2. The accuracy of target detection is greatly improved by performing precise pixel-level connected region identification and label assignment for each image in the CT image sequence. This method first carefully analyzes the spatial relationship between each pixel and its surroundings, thereby effectively distinguishing connected regions and edges, ensuring that only pixels that truly belong to the same region are classified together. Subsequently, a unique label is assigned to each identified connected region, and through an intelligent merging strategy, the contacting or close regions are further integrated to form a larger continuous region. This method not only improves the accuracy of regional positioning, but also optimizes the quality of input data for subsequent deep learning model processing, reduces the model's sensitivity to background noise and non-target structures, and thus improves the diagnostic efficiency and accuracy of the entire system. Through this precise image preprocessing, the model can focus more on the key chest area, effectively improving the reliability and speed of diagnosis. For the cloud particle images of varying sizes in CPI images, a segmentation method based on connected region labeling is proposed to provide support for subsequent classification and recognition.

[0028] 3. The present invention further enhances the accuracy and effectiveness of target detection by finely screening all connected areas in the merged binary list. By setting a specific pixel number threshold, the connected areas that meet the predetermined size criteria are effectively distinguished and screened out. This strategy ensures that only chest areas with potential clinical significance are retained for further analysis. This screening mechanism not only reduces the computational resources wasted in processing unnecessary or too small areas, but also improves the efficiency and accuracy of the entire system when processing large data sets.

[0029] 4. The present invention significantly enhances the recognition accuracy of the target area by performing advanced similarity judgment on the connected areas. Specifically, by establishing a histogram and calculating the KL distance, the present method can accurately compare and identify the chest area in the image, and effectively select the most representative and similar chest area from multiple connected areas. This step not only optimizes the selection process of the chest area, ensures the consistency of the selected area in visual and statistical characteristics, but also improves the relevance and efficiency of the subsequent deep learning model processing.

[0030] 5. By using the pre-trained ResNet-50 model structure, targeted optimization and customization, the ability to accurately extract key features from CT images has been significantly improved. The original fully connected layer is removed and new customized layers are introduced, such as the fully connected layer and Dropout layer with ReLU activation function, which enhances the adaptability and anti-overfitting ability of the model. In addition, by adding a binary classification output layer for specific tasks, the application effect and accuracy of the model on actual medical data are directly improved, which greatly improves the automation level and diagnostic efficiency of CT image target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of a target detection method based on regional positioning; DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solution in the embodiment of the present invention in conjunction with the accompanying drawings in the embodiment of the present invention. The method includes the following steps:

[0033] Step 1: Perform chest area preprocessing operations on the acquired CT images;

[0034] Step 2: Set the binarization threshold tht to binarize the CT image after dilation processing;

[0035] Step 3: Locate the chest area on the CT image, specifically:

[0036] Step 3.1: Identify connected regions at the pixel level on the CT image;

[0037] Step 3.2: Assign labels to the connected areas of the CT image;

[0038] Step 3.3: Merge all connected area labels in the binary list;

[0039] Step 3.4: Filter all connected areas in the merged binary list;

[0040] Step 3.5: Perform similarity judgment on the selected connected areas to obtain the chest area of ​​the CT image;

[0041] Step 4: Based on the extracted chest area of ​​the CT image, a neural network model is established to detect the CT image.

[0042] Furthermore, in step 3.1, the connected area identification is performed on the CT image at the pixel level, specifically: each pixel in the binarization processing result corresponding to the CT image is scanned, and the pixels in its eight neighborhoods are checked. If the pixel values ​​are the same, it is determined that these pixels belong to the connected area; if the pixel values ​​are different, it is determined that these pixels belong to the edge area; if the scanned pixel does not have a complete eight-neighborhood pixel, the pixel is discarded and not processed.

[0043] Furthermore, in step 3.2, labels are assigned to the connected areas of the CT image, specifically: a corresponding binarization list is set for each binarization processing result, and a unique connected area label is assigned to each identified connected area in each binarization processing result. Whenever a new connected area is found, its connected area label is added to the binarization list corresponding to the currently processed binarization processing result.

[0044] Furthermore, in the step 3.3, all connected area labels in the binary list are merged, specifically: first, when there is one or more pixel points with the same horizontal and vertical coordinate index in the connected areas corresponding to two connected area labels, the two connected areas are regarded as interconnected; then, when it is determined that the two connected areas are interconnected, the two areas will be merged into a new connected area, and the merging operation is to take the union of the two areas, that is, the new connected area includes all pixels in the original two areas; finally, the smaller label of the two connected areas before the merger is used as the label of the new merged area, and the larger label of the two connected areas before the merger is deleted.

[0045] Furthermore, in the step 3.4, all connected areas in the merged binary list are screened, specifically: first, the pixel values ​​of the pixels in the connected areas are judged, and the connected areas with pixel values ​​of 0 are extracted, and then the number of pixels contained in each extracted connected area is judged, and the connected areas with the number of pixels greater than the lower chest threshold thdn and less than the upper chest threshold thup are screened.

[0046] Furthermore, in the step 3.5, similarity judgment is performed on the screened connected areas to obtain the chest area of ​​the CT image, specifically: the screened connected areas are mapped to the CT image, and histograms are established for the corresponding connected areas on the CT image, and the KL distance between the two histograms is calculated, and the two connected areas on the CT image corresponding to the two histograms with the shortest KL distance are selected, recorded as chest area A and chest area B, and the two are collectively used as the chest area of ​​the CT image.

[0047] Furthermore, in step 4, based on the extracted chest area of ​​the CT image, a neural network model is established to detect the CT image, specifically: first, a CT image database is established; second, a neural network model is established; third, based on the CT image database, the neural network model is trained and evaluated; finally, the trained neural network model is applied to actual CT image data for target detection.

[0048] Furthermore, in step 1, the collected CT images are subjected to chest region preprocessing operations, specifically: first, a circular structure element of size dx is selected to perform corrosion processing on the image; then a circular structure element of size dy is selected to perform expansion processing on the corrosion processing result.

[0049] Furthermore, the establishing of the CT image database specifically includes: extracting chest regions from the labeled CT image data set, and marking these regions into two categories: target presence and target absence.

[0050] Furthermore, 10. The establishment of the neural network model is specifically as follows: first, a pre-trained ResNet-50 model is used as the basic framework for feature extraction, and then the fully connected layer of the original model is removed from the tail of the model, and a fully connected layer is added, containing 256 neurons, using the ReLU activation function, and then a Dropout layer is added with a dropout rate of 0.5, and finally an output layer is added, which is a fully connected layer containing 2 neurons, and a softmax activation function is used for binary classification.

[0051] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

Claims

1. A target detection method based on regional positioning, comprising the following steps: Step 1: Perform chest area preprocessing operations on the acquired CT images; Step 2: Set the binarization threshold tht to binarize the CT image after dilation processing; Step 3: Locate the chest area on the CT image, specifically: Step 3.1: Identify connected regions at the pixel level on the CT image; Step 3.2: Assign labels to the connected areas of the CT image; Step 3.3: Merge all connected area labels in the binary list; Step 3.4: Filter all connected areas in the merged binary list; Step 3.5: Perform similarity judgment on the selected connected areas to obtain the chest area of ​​the CT image; Step 4: Based on the extracted chest area of ​​the CT image, a neural network model is established to detect the CT image.

2. The target detection method based on regional positioning according to claim 1, characterized in that: In the step 3.1, the connected area of ​​the CT image is identified at the pixel level, specifically: each pixel in the binarization processing result corresponding to the CT image is scanned, and the pixels in its eight neighborhoods are checked. If the pixel values ​​are the same, it is determined that these pixels belong to the connected area; if the pixel values ​​are different, it is determined that these pixels belong to the edge area; if the scanned pixel does not have a complete eight-neighborhood pixel, the pixel is discarded and not processed.

3. The target detection method based on regional positioning according to claim 1, characterized in that: In the step 3.2, labels are assigned to the connected areas of the CT image, specifically: a corresponding binarization list is set for each binarization processing result, and a unique connected area label is assigned to each identified connected area in each binarization processing result. Whenever a new connected area is found, its connected area label is added to the binarization list corresponding to the currently processed binarization processing result.

4. The target detection method based on regional positioning according to claim 1, characterized in that: In the step 3.3, all the connected area labels in the binary list are merged, specifically: first, when there is one or more pixel points with the same horizontal and vertical coordinate index in the connected areas corresponding to two connected area labels, the two connected areas are regarded as interconnected; then, when it is determined that the two connected areas are interconnected, the two areas will be merged into a new connected area, and the merging operation is to take the union of the two areas, that is, the new connected area includes all the pixels in the original two areas; finally, the smaller label of the two connected areas before the merger is used as the label of the new merged area, and the larger label of the two connected areas before the merger is deleted.

5. The target detection method based on regional positioning according to claim 1, characterized in that: In the step 3.4, all the connected areas in the merged binary list are screened, specifically: first, the pixel values ​​of the pixels in the connected areas are judged, and the connected areas with pixel values ​​of 0 are extracted, and then the number of pixels contained in each extracted connected area is judged, and the connected areas with the number of pixels greater than the lower chest threshold thdn and less than the upper chest threshold thup are screened.

6. The target detection method based on regional positioning according to claim 1, characterized in that: In the step 3.5, similarity judgment is performed on the screened connected areas to obtain the chest area of ​​the CT image, specifically: the screened connected areas are mapped to the CT image, and histograms are established for the corresponding connected areas on the CT image, and the KL distance between the two histograms is calculated, and the two connected areas on the CT image corresponding to the two histograms with the shortest KL distance are selected, recorded as chest area A and chest area B, and the two are collectively used as the chest area of ​​the CT image.

7. The target detection method based on regional positioning according to claim 1, characterized in that: In step 4, based on the extracted chest area of ​​the CT image, a neural network model is established to detect the CT image, specifically: first, a CT image database is established; second, a neural network model is established; third, based on the CT image database, the neural network model is trained and evaluated; finally, the trained neural network model is applied to the actual CT image data for target detection.

8. The target detection method based on regional positioning according to claim 1, characterized in that: In step 1, the collected CT images are preprocessed in the chest area, specifically: first, a circular structure element with a size of dx is selected to corrode the image; then a circular structure element with a size of dy is selected to dilate the corrosion result.

9. The target detection method based on regional positioning according to claim 7, characterized in that: The establishing of the CT image database specifically includes: extracting chest areas from the labeled CT image data set, and marking these areas into two categories: target presence and target absence.

10. The target detection method based on regional positioning according to claim 7, characterized in that: The neural network model is established as follows: first, a pre-trained ResNet-50 model is used as the basic framework for feature extraction, then the fully connected layer of the original model is removed from the tail of the model, a fully connected layer is added, including 256 neurons, a ReLU activation function is used, and then a Dropout layer is added with a dropout rate of 0.5, and finally an output layer is added, which is a fully connected layer including 2 neurons, and a softmax activation function is used for binary classification.

Images