An image target detection method based on multi-level judgment

Abstract

The present application relates to a kind of image target detection methods based on multistage judgment, mainly first to CT image sequence is preprocessed, extract thoracic cavity area and two cavity regions therein. Then, two-stage target existence possibility judgment is implemented: first-stage judgment is the basis of target screening, second-stage judgment is more refined analysis to the cavity region after screening by depth morphological processing and data enhancement technology, then, using the ResNet-50 model integrated by spatial transformation network according to the target existence possibility attribute of cavity area, the degree of local amplification in image is automatically adjusted, to optimize the accuracy of target detection.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, and in particular relates to a method for detecting targets in a computed tomography (CT) image sequence based on multi-level judgment. Background Art

[0002] Chest computed tomography (CT) is a core diagnostic tool in modern medicine, playing a crucial role in identifying and evaluating intrathoracic diseases such as lung lesions. Technological advances have not only increased CT scan speed but also generated vast amounts of image data, enabling early diagnosis and disease monitoring. However, as the volume of data increases, effectively processing and analyzing this data becomes a challenge.

[0003] Traditional CT image analysis methods rely primarily on radiologists to make diagnoses through visual inspection. This method is not only inefficient but also susceptible to the influence of personal experience and judgment. Furthermore, manual inspection is prone to fatigue when processing large amounts of data, which affects the accuracy of judgment. Therefore, researchers have begun to explore automated solutions to improve processing efficiency and diagnostic accuracy.

[0004] In recent years, deep learning techniques, particularly convolutional neural networks (CNNs), have shown great potential in medical image analysis. They can automatically identify complex patterns in images, providing faster and more accurate analysis results than traditional methods. However, most existing deep learning methods process the entire image without filtering, failing to fully utilize key image regions, such as specific target areas. This results in less refined analysis and sometimes introduces unnecessary background noise.

[0005] To address this issue, the present invention proposes a multi-level judgment-based image target detection method. This method not only considers the overall image but also pays special attention to key areas within the chest cavity (such as cavity areas A and B). Through detailed target presence probability assessment and the establishment of a deep learning model, the accuracy and efficiency of target detection are significantly improved. This method, by combining advanced image processing technology and deep learning, provides a more reliable reference aid for clinical practice and is particularly suitable for processing large-scale medical imaging data. Summary of the Invention

[0006] In view of the above shortcomings of the prior art, the present invention proposes an image target detection method based on multi-level judgment, which includes the following steps:

[0007] Step 1: Perform superpixel-based preprocessing on the CT image to obtain the chest region in the CT image, as well as two cavity regions in the chest region, cavity region A and cavity region B. Perform the first-level judgment to obtain the chest region key monitoring level attributes and the cavity region key monitoring level attributes;

[0008] Step 2: Perform a second-level judgment on the chest area in the CT image, including:

[0009] Step 2.1: Filter and sort the key monitoring level attributes of the cavity area in the thoracic region of the CT image according to their importance. Specifically, filter out abnormal images containing the key monitoring level attributes of the cavity area as high-low, medium-high, low-high, and low-medium. Then, sort the key monitoring level attributes of the cavity area from high to low according to their importance as high-high, high-medium, medium-medium, medium-low, and low-low.

[0010] Step 2.2: Set observation and control attributes for the two cavity areas in the thoracic region in the CT image. Specifically, compare the key monitoring level attributes of the two cavity areas, mark the cavity area with a higher importance ranking as the observation cavity area, and mark the cavity area with a lower importance ranking as the control cavity area. If the key monitoring level attributes of the two cavity areas are the same, then both cavity areas are marked as observation cavity areas.

[0011] Step 2.3: Perform a second-level judgment based on the marking conditions of cavity area A and cavity area B. Specifically, determine whether the marking conditions of cavity area A and cavity area B are the same. If they are not the same, execute step 2.3.1. If they are the same, directly perform a similarity judgment on the two cavity areas.

[0012] Step 2.3.1: Perform deep morphological processing on the observation cavity area and the control cavity area to obtain multiple groups of morphological processing result images;

[0013] Step 2.3.2: Perform data enhancement on the morphological processing results of the observation cavity area in each set of morphological processing result images to obtain multiple sets of enhanced morphological processing result images of the observation cavity area;

[0014] Step 2.3.3: Filter out the enhanced observation cavity region morphological processing results with the highest similarity to the control cavity region morphological processing results in each group of morphological processing result images, and use them as the matching of the control cavity region morphological processing results;

[0015] Step 2.3.4: Perform a difference operation on the morphological processing result of the matched enhanced observation cavity area and the morphological processing result of the control cavity area to obtain a difference operation result image;

[0016] Step 2.3.5: Set a threshold value thh to perform binarization on the difference operation result image, calculate the ratio Rx of the number of pixels with a pixel value of 0 to the number of pixels with a pixel value of 1 in the binarization result image, set a threshold thx, analyze the chest area of ​​the CT image corresponding to the difference operation result image, and obtain the key monitoring level attribute of the cavity area in the CT image;

[0017] Step 3: Based on the key monitoring level attributes of CT images, the ResNet-50 model integrated with the spatial transformer network (STN) is used to perform object detection in CT images.

[0018] Furthermore, in step 1, the CT image is preprocessed based on superpixels, specifically: first, the CT image is normalized, the brightness and contrast are adjusted, and the image noise is reduced using noise reduction technology, and then the image is binarized using the Otsu method to help preliminarily identify the chest area; then, morphological operations are applied to the binarized image to further clarify the boundary of the cavity area, and two cavity areas in the chest area are obtained; then, the chest area is converted into a superpixel block, and the superpixel block comprehensively considers the proximity of color, texture and space to ensure that the consistency within each superpixel block is maximized, and the superpixel block is divided into background superpixel blocks and target superpixel blocks according to the pixel histogram within the superpixel block, and finally, the chest area thresholds th_SP1 and th_SP2 and the cavity area thresholds th_SP3 and th_SP4 are set, and the key monitoring level attributes of the chest area and the two cavity areas are judged based on the regional threshold to obtain the key monitoring level attributes of the chest area and the cavity area.

[0019] Furthermore, the chest area thresholds th_SP1 and th_SP2 and the cavity area thresholds th_SP3 and th_SP4 are finally set, and the key monitoring level attributes of the chest area and the two cavity areas are judged based on the regional thresholds to obtain the key monitoring level attributes of the chest area and the cavity area. Specifically, the ratio of the number of background superpixel blocks and target superpixel blocks in the chest area is first calculated, and then the ratio of the number of superpixel blocks in the chest area is compared with the chest area thresholds th_SP1 and th_SP2 to obtain the key monitoring level attributes of the chest area, which are high, medium, and low respectively. Then, the ratio of the number of background superpixel blocks and target superpixel blocks in the two cavity areas is calculated respectively, and then the ratio of the number of superpixel blocks in the cavity area is compared with the cavity area thresholds th_SP3 and th_SP4. Then, according to the key monitoring level attributes of the chest area to which the cavity area belongs, the key monitoring level attributes of the cavity area are fused, which are high-high, high-medium, high-low, medium-high, medium-medium, medium-low, low-high, low-medium, and low-low respectively.

[0020] Furthermore, in step 2.3.1, deep morphological processing is performed on the observation cavity area and the control cavity area, specifically: first, a circular structuring element with a radius of R pixels is used to continuously erode the image Rn times, and after each erosion, a circular structuring element with the same radius is immediately used to perform expansion processing. After each processing, a set of morphological processing result images including the morphological processing results of the observation cavity area and the morphological processing results of the control cavity area are obtained.

[0021] Furthermore, in step 2.3.2, data enhancement operation is performed on the morphological processing results of the observation cavity area in each group of morphological processing result images, specifically: first, the morphological processing result image of the observation cavity area is rotated 180 degrees to match the orientation of the control cavity area; then, a slight angle adjustment is made to the rotated result, with the angle ranging from -5 degrees to +5 degrees, changing by 1 degree each time; and each adjusted result is scaled in series, with the scaling ratio ranging from 0.95 to 1.05, changing by 0.01 each time. Finally, the morphological processing results of the observation cavity area in each group of morphological processing result images are obtained to obtain a set of enhanced enhanced morphological processing result images of the observation cavity area.

[0022] Furthermore, in the step 2.3.3, the enhanced observation cavity area morphological processing results with the highest similarity to the control cavity area morphological processing results in each group of morphological processing result images are screened out. Specifically, the structural similarity index evaluation method is used to measure the similarity between the enhanced morphological processing results of the observation cavity area and the morphological processing results of the control cavity area, and the enhanced observation cavity area morphological processing result with the highest score is selected as the match of the control cavity area morphological processing result.

[0023] Furthermore, in step 2.3.5, a threshold value thh is set to perform binarization processing on the difference operation result image, and a ratio Rx of the number of pixel points with a pixel value of 0 to the number of pixel points with a pixel value of 1 in the binarization result image is calculated. A threshold value thx is set, and the chest area of ​​the CT image corresponding to the difference operation result image is analyzed to obtain the key monitoring level of the cavity area in each CT image in the CT image sequence. Specifically, when the ratio Rx is greater than thx, the key monitoring level of the cavity area corresponding to the observed cavity area is judged to be high, and a high evaluation is added to its existing key monitoring level attribute. Then, the key monitoring level of the cavity area corresponding to the control cavity area is judged to be medium, and a medium evaluation is added to its existing key monitoring level attribute. When the ratio Rx is less than thx, the key monitoring level of the cavity area corresponding to the observed cavity area is judged to be medium, and a medium evaluation is added to its existing key monitoring level attribute. Then, the key monitoring level of the cavity area corresponding to the control cavity area is judged to be low, and a low evaluation is added to its existing key monitoring level attribute.

[0024] Furthermore, in step 3.3, if they are the same, the two cavity areas are directly judged for similarity, specifically: if the two cavity areas are judged to have a high degree of similarity, an evaluation that is the same as the second evaluation in their key monitoring level attributes is added on the basis of their existing key monitoring level attributes.

[0025] Furthermore, in step 3, based on the key monitoring level attributes of the CT image, the ResNet-50 model of the integrated spatial transformation network is used to perform target detection on the CT image. Specifically, the preprocessed CT image is first input into the ResNet-50 model of the integrated spatial transformation network. Secondly, according to the key monitoring level attributes of the cavity area, the integrated spatial transformation network module automatically adjusts the local magnification degree of the identified cavity areas A and B in the image through local affine transformation. The image adjusted by the integrated spatial transformation network is again transmitted to the subsequent layer of ResNet-50 for feature extraction. Finally, target detection is performed through the binary classification output layer of the model to determine whether the target exists.

[0026] Furthermore, according to the key monitoring level attributes of the cavity area, the integrated spatial transformation network module automatically adjusts the local magnification degree of the identified cavity areas A and B in the image through local affine transformation. Specifically, when the key monitoring levels of the cavity areas A and B include 3 highs, a local magnification of 1.5 times is performed; when it includes 2 highs and 1 medium, a local magnification of 1.4 times is performed; when it includes 1 high and 2 mediums, a local magnification of 1.3 times is performed; when it includes 3 mediums, a local magnification of 1.2 times is performed; when it includes 2 mediums and 1 low, a local magnification of 1.15 times is performed; when it includes 1 medium and 2 lows, a local magnification of 1.1 times is performed; when it includes 3 lows, the original size is maintained without magnification.

[0027] Compared with the traditional method, the present invention has the following advantages, thereby solving the corresponding technical problems:

[0028] 1. By implementing a two-level target presence probability assessment, the accuracy and reliability of target region identification in CT image sequences are significantly improved. First, a preliminary target presence probability assessment quickly identifies areas of focus. Subsequently, a more refined second-level assessment allows for in-depth analysis and verification of these initially identified areas. This multi-level assessment strategy not only enhances detection accuracy but also optimizes the processing flow, making the method more efficient and accurate when processing large-scale data.

[0029] 2. The second-level target presence probability determination method includes detailed observation and comparison attribute analysis of the cavity area and deep morphological processing, which provides a highly refined analysis method. By performing detailed comparisons of the observation and comparison cavity areas, and through deep morphological processing and data enhancement operations, the present invention can accurately distinguish subtle differences, thereby providing more precise positioning and assessment of the presence of targets. This detailed regional analysis is often overlooked in traditional methods, and the present method therefore has a significant advantage in recognition accuracy.

[0030] 3. By integrating multi-level judgment results into the ResNet-50 model of the spatial transformer network, the present invention can automatically adjust the magnification of local areas based on the different levels of target presence probability. This adaptive local magnification strategy ensures that the model focuses more closely on areas that may contain critical lesions, thereby improving detection accuracy while also optimizing the model's learning and operational efficiency. This intelligent image preprocessing greatly enhances the detection capabilities of subsequent deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of an image target detection method based on multi-level judgment; 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 superpixel-based preprocessing on the CT image to obtain the chest region in the CT image, as well as two cavity regions in the chest region, cavity region A and cavity region B. Perform the first-level judgment to obtain the chest region key monitoring level attributes and the cavity region key monitoring level attributes;

[0034] Step 2: Perform a second-level judgment on the chest area in the CT image, including:

[0035] Step 2.1: Filter and sort the key monitoring level attributes of the cavity area in the thoracic region of the CT image according to their importance. Specifically, filter out abnormal images containing the key monitoring level attributes of the cavity area as high-low, medium-high, low-high, and low-medium. Then, sort the key monitoring level attributes of the cavity area from high to low according to their importance as high-high, high-medium, medium-medium, medium-low, and low-low.

[0036] Step 2.2: Set observation and control attributes for the two cavity areas in the thoracic region in the CT image. Specifically, compare the key monitoring level attributes of the two cavity areas, mark the cavity area with a higher importance ranking as the observation cavity area, and mark the cavity area with a lower importance ranking as the control cavity area. If the key monitoring level attributes of the two cavity areas are the same, then both cavity areas are marked as observation cavity areas.

[0037] Step 2.3: Perform a second-level judgment based on the marking conditions of cavity area A and cavity area B. Specifically, determine whether the marking conditions of cavity area A and cavity area B are the same. If they are not the same, execute step 2.3.1. If they are the same, directly perform a similarity judgment on the two cavity areas.

[0038] Step 2.3.1: Perform deep morphological processing on the observation cavity area and the control cavity area to obtain multiple groups of morphological processing result images;

[0039] Step 2.3.2: Perform data enhancement on the morphological processing results of the observation cavity area in each set of morphological processing result images to obtain multiple sets of enhanced morphological processing result images of the observation cavity area;

[0040] Step 2.3.3: Filter out the enhanced observation cavity region morphological processing results with the highest similarity to the control cavity region morphological processing results in each group of morphological processing result images, and use them as the matching of the control cavity region morphological processing results;

[0041] Step 2.3.4: Perform a difference operation on the morphological processing result of the matched enhanced observation cavity area and the morphological processing result of the control cavity area to obtain a difference operation result image;

[0042] Step 2.3.5: Set a threshold value thh to perform binarization on the difference operation result image, calculate the ratio Rx of the number of pixels with a pixel value of 0 to the number of pixels with a pixel value of 1 in the binarization result image, set a threshold thx, analyze the chest area of ​​the CT image corresponding to the difference operation result image, and obtain the key monitoring level attribute of the cavity area in the CT image;

[0043] Step 3: Based on the key monitoring level attributes of CT images, the ResNet-50 model integrated with the spatial transformer network (STN) is used to perform object detection in CT images.

[0044] Furthermore, in step 1, the CT image is preprocessed based on superpixels, specifically: first, the CT image is normalized, the brightness and contrast are adjusted, and the image noise is reduced using noise reduction technology, and then the image is binarized using the Otsu method to help preliminarily identify the chest area; then, morphological operations are applied to the binarized image to further clarify the boundary of the cavity area, and two cavity areas in the chest area are obtained; then, the chest area is converted into a superpixel block, and the superpixel block comprehensively considers the proximity of color, texture and space to ensure that the consistency within each superpixel block is maximized, and the superpixel block is divided into background superpixel blocks and target superpixel blocks according to the pixel histogram within the superpixel block, and finally, the chest area thresholds th_SP1 and th_SP2 and the cavity area thresholds th_SP3 and th_SP4 are set, and the key monitoring level attributes of the chest area and the two cavity areas are judged based on the regional threshold to obtain the key monitoring level attributes of the chest area and the cavity area.

[0045] Furthermore, the chest area thresholds th_SP1 and th_SP2 and the cavity area thresholds th_SP3 and th_SP4 are finally set, and the key monitoring level attributes of the chest area and the two cavity areas are judged based on the regional thresholds to obtain the key monitoring level attributes of the chest area and the cavity area. Specifically, the ratio of the number of background superpixel blocks and target superpixel blocks in the chest area is first calculated, and then the ratio of the number of superpixel blocks in the chest area is compared with the chest area thresholds th_SP1 and th_SP2 to obtain the key monitoring level attributes of the chest area, which are high, medium, and low respectively. Then, the ratio of the number of background superpixel blocks and target superpixel blocks in the two cavity areas is calculated respectively, and then the ratio of the number of superpixel blocks in the cavity area is compared with the cavity area thresholds th_SP3 and th_SP4. Then, according to the key monitoring level attributes of the chest area to which the cavity area belongs, the key monitoring level attributes of the cavity area are fused, which are high-high, high-medium, high-low, medium-high, medium-medium, medium-low, low-high, low-medium, and low-low respectively.

[0046] Furthermore, in step 2.3.1, deep morphological processing is performed on the observation cavity area and the control cavity area, specifically: first, a circular structuring element with a radius of R pixels is used to continuously erode the image Rn times, and after each erosion, a circular structuring element with the same radius is immediately used to perform expansion processing. After each processing, a set of morphological processing result images including the morphological processing results of the observation cavity area and the morphological processing results of the control cavity area are obtained.

[0047] Furthermore, in step 2.3.2, data enhancement operation is performed on the morphological processing results of the observation cavity area in each group of morphological processing result images, specifically: first, the morphological processing result image of the observation cavity area is rotated 180 degrees to match the orientation of the control cavity area; then, a slight angle adjustment is made to the rotated result, with the angle ranging from -5 degrees to +5 degrees, changing by 1 degree each time; and each adjusted result is scaled in series, with the scaling ratio ranging from 0.95 to 1.05, changing by 0.01 each time. Finally, the morphological processing results of the observation cavity area in each group of morphological processing result images are obtained to obtain a set of enhanced enhanced morphological processing result images of the observation cavity area.

[0048] Furthermore, in the step 2.3.3, the enhanced observation cavity area morphological processing results with the highest similarity to the control cavity area morphological processing results in each group of morphological processing result images are screened out. Specifically, the structural similarity index evaluation method is used to measure the similarity between the enhanced morphological processing results of the observation cavity area and the morphological processing results of the control cavity area, and the enhanced observation cavity area morphological processing result with the highest score is selected as the match of the control cavity area morphological processing result.

[0049] Furthermore, in step 2.3.5, a threshold value thh is set to perform binarization processing on the difference operation result image, and a ratio Rx of the number of pixel points with a pixel value of 0 to the number of pixel points with a pixel value of 1 in the binarization result image is calculated. A threshold value thx is set, and the chest area of ​​the CT image corresponding to the difference operation result image is analyzed to obtain the key monitoring level of the cavity area in each CT image in the CT image sequence. Specifically, when the ratio Rx is greater than thx, the key monitoring level of the cavity area corresponding to the observed cavity area is judged to be high, and a high evaluation is added to its existing key monitoring level attribute. Then, the key monitoring level of the cavity area corresponding to the control cavity area is judged to be medium, and a medium evaluation is added to its existing key monitoring level attribute. When the ratio Rx is less than thx, the key monitoring level of the cavity area corresponding to the observed cavity area is judged to be medium, and a medium evaluation is added to its existing key monitoring level attribute. Then, the key monitoring level of the cavity area corresponding to the control cavity area is judged to be low, and a low evaluation is added to its existing key monitoring level attribute.

[0050] Furthermore, in step 3.3, if they are the same, the two cavity areas are directly judged for similarity, specifically: if the two cavity areas are judged to have a high degree of similarity, an evaluation that is the same as the second evaluation in their key monitoring level attributes is added on the basis of their existing key monitoring level attributes.

[0051] Furthermore, in step 3, based on the key monitoring level attributes of the CT image, the ResNet-50 model of the integrated spatial transformation network is used to perform target detection on the CT image. Specifically, the preprocessed CT image is first input into the ResNet-50 model of the integrated spatial transformation network. Secondly, according to the key monitoring level attributes of the cavity area, the integrated spatial transformation network module automatically adjusts the local magnification degree of the identified cavity areas A and B in the image through local affine transformation. The image adjusted by the integrated spatial transformation network is again transmitted to the subsequent layer of ResNet-50 for feature extraction. Finally, target detection is performed through the binary classification output layer of the model to determine whether the target exists.

[0052] Furthermore, according to the key monitoring level attributes of the cavity area, the integrated spatial transformation network module automatically adjusts the local magnification degree of the identified cavity areas A and B in the image through local affine transformation. Specifically, when the key monitoring levels of the cavity areas A and B include 3 highs, a local magnification of 1.5 times is performed; when it includes 2 highs and 1 medium, a local magnification of 1.4 times is performed; when it includes 1 high and 2 mediums, a local magnification of 1.3 times is performed; when it includes 3 mediums, a local magnification of 1.2 times is performed; when it includes 2 mediums and 1 low, a local magnification of 1.15 times is performed; when it includes 1 medium and 2 lows, a local magnification of 1.1 times is performed; when it includes 3 lows, the original size is maintained without magnification.

[0053] Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without making creative work are all within the scope of protection of the present invention.

Claims

1. A method for detecting an image target based on multi-level judgment, comprising the following steps: Step 1: Perform superpixel-based preprocessing on the CT image to obtain the chest region in the CT image, as well as two cavity regions within the chest region, cavity region A and cavity region B. Perform a first-level judgment to obtain the chest region key monitoring level attributes and the cavity region key monitoring level attributes. Step 2: Perform a second-level judgment on the chest area in the CT image, including: Step 2.1: Filter and sort the key monitoring level attributes of the cavity area in the thoracic region of the CT image according to their importance. Specifically, filter out abnormal images containing the key monitoring level attributes of the cavity area as high-low, medium-high, low-high, and low-medium. Then, sort the key monitoring level attributes of the cavity area from high to low according to their importance as high-high, high-medium, medium-medium, medium-low, and low-low. Step 2.2: Set observation and control attributes for the two cavity areas in the thoracic region in the CT image. Specifically, compare the key monitoring level attributes of the two cavity areas, mark the cavity area with a higher importance ranking as the observation cavity area, and mark the cavity area with a lower importance ranking as the control cavity area. If the key monitoring level attributes of the two cavity areas are the same, then both cavity areas are marked as observation cavity areas. Step 2.3: Perform a second-level judgment based on the markings of cavity area A and cavity area B. Specifically, determine whether the markings of cavity area A and cavity area B are the same. If they are not the same, execute step 2.3.

1. If they are the same, directly perform a similarity judgment on the two cavity areas. Step 2.3.1: Perform deep morphological processing on the observation cavity area and the control cavity area to obtain multiple groups of morphological processing result images; Step 2.3.2: Perform data enhancement on the morphological processing results of the observation cavity area in each set of morphological processing result images to obtain multiple sets of enhanced morphological processing result images of the observation cavity area; Step 2.3.3: Filter out the enhanced observation cavity region morphological processing results with the highest similarity to the control cavity region morphological processing results in each group of morphological processing result images, and use them as the matching of the control cavity region morphological processing results; Step 2.3.4: Perform a difference operation on the morphological processing result of the matched enhanced observation cavity area and the morphological processing result of the control cavity area to obtain a difference operation result image; Step 2.3.5: Set a threshold value thh to perform binarization on the difference operation result image, calculate the ratio Rx of the number of pixels with a pixel value of 0 to the number of pixels with a pixel value of 1 in the binarization result image, set a threshold thx, analyze the chest area of ​​the CT image corresponding to the difference operation result image, and obtain the key monitoring level attribute of the cavity area in the CT image; Step 3: Based on the key monitoring level attributes of CT images, the ResNet-50 model integrated with the spatial transformer network (STN) is used to perform object detection in CT images.

2. The image target detection method based on multi-level judgment according to claim 1, characterized in that: In step 1, the CT image is preprocessed based on superpixels, specifically: first, the CT image is normalized, the brightness and contrast are adjusted, and the image noise is reduced using noise reduction technology, and then the image is binarized using the Otsu method to help preliminarily identify the chest area; Then, morphological operations are applied on the binary image to further clarify the boundary of the cavity area, and two cavity areas in the chest area are obtained; The chest area is then converted into superpixel blocks. The superpixel blocks comprehensively consider the proximity of color, texture and space to ensure that the consistency within each superpixel block is maximized, and the superpixel blocks are divided into background superpixel blocks and target superpixel blocks according to the pixel histogram within the superpixel blocks. Finally, the chest area thresholds th_SP1 and th_SP2 and the cavity area thresholds th_SP3 and th_SP4 are set, and the key monitoring level attributes of the chest area and the two cavity areas are judged based on the regional thresholds to obtain the key monitoring level attributes of the chest area and the cavity area.

3. The image target detection method based on multi-level judgment according to claim 2, characterized in that: Finally, the chest area thresholds th_SP1 and th_SP2 and the cavity area thresholds th_SP3 and th_SP4 are set, and the key monitoring level attributes of the chest area and the two cavity areas are judged based on the regional thresholds to obtain the key monitoring level attributes of the chest area and the cavity area. Specifically, the ratio of the number of background superpixel blocks and target superpixel blocks in the chest area is first calculated, and then the chest area superpixel fast number ratio is compared with the chest area thresholds th_SP1 and th_SP2 to obtain the chest area key monitoring level attributes as high, medium, and low respectively. Then, the ratio of the number of background superpixel blocks and target superpixel blocks in the two cavity areas is calculated respectively, and then the cavity area superpixel block number ratio is compared with the cavity area thresholds th_SP3 and th_SP4. Then, according to the key monitoring level attributes of the chest area to which the cavity area belongs, the key monitoring level attributes of the cavity area are fused and obtained as high-high, high-medium, high-low, medium-high, medium-medium, medium-low, low-high, low-medium, and low-low respectively.

4. The image target detection method based on multi-level judgment according to claim 1, characterized in that: In step 2.3.1, deep morphological processing is performed on the observation cavity area and the control cavity area. Specifically, the image is first eroded Rn times using a circular structuring element with a radius of R pixels. After each erosion, a circular structuring element with the same radius is immediately used to perform dilation processing. After each processing, a set of morphological processing result images including the morphological processing results of the observation cavity area and the morphological processing results of the control cavity area are obtained.

5. The image target detection method based on multi-level judgment according to claim 1, characterized in that: In the step 2.3.2, a data enhancement operation is performed on the morphological processing results of the observation cavity area in each group of morphological processing result images, specifically: first, the morphological processing result images of the observation cavity area are rotated 180 degrees to match the orientation of the control cavity area; then, a slight angle adjustment is performed on the rotated results, with the angle ranging from -5 degrees to +5 degrees, changing by 1 degree each time; and each adjusted result is scaled in series, with the scaling ratio ranging from 0.95 to 1.05, changing by 0.01 each time. Finally, the morphological processing results of the observation cavity area in each group of morphological processing result images are obtained to obtain a set of enhanced enhanced morphological processing result images of the observation cavity area.

6. The image target detection method based on multi-level judgment according to claim 1, characterized in that: In the step 2.3.3, the enhanced observation cavity area morphological processing results with the highest similarity to the control cavity area morphological processing results in each group of morphological processing result images are screened out. Specifically, the structural similarity index evaluation method is used to measure the similarity between the enhanced morphological processing results of the observation cavity area and the morphological processing results of the control cavity area, and the enhanced observation cavity area morphological processing result with the highest score is selected as the match of the control cavity area morphological processing result.

7. The image target detection method based on multi-level judgment according to claim 1, characterized in that: In the step 2.3.5, a threshold value thh is set to perform binarization processing on the difference operation result image, and a ratio Rx of the number of pixel points with a pixel value of 0 to the number of pixel points with a pixel value of 1 in the binarization result image is calculated. A threshold value thx is set, and the chest area of ​​the CT image corresponding to the difference operation result image is analyzed to obtain the key monitoring level of the cavity area in each CT image in the CT image sequence. Specifically, when the ratio Rx is greater than thx, the key monitoring level of the cavity area corresponding to the observed cavity area is judged to be high, and a high evaluation is added to its existing key monitoring level attribute. Then, the key monitoring level of the cavity area corresponding to the control cavity area is judged to be medium, and a medium evaluation is added to its existing key monitoring level attribute. When the ratio Rx is less than thx, the key monitoring level of the cavity area corresponding to the observed cavity area is judged to be medium, and a medium evaluation is added to its existing key monitoring level attribute. Then, the key monitoring level of the cavity area corresponding to the control cavity area is judged to be low, and a low evaluation is added to its existing key monitoring level attribute.

8. The image target detection method based on multi-level judgment according to claim 1, characterized in that: In step 2.3, if they are the same, the two cavity areas are directly judged for similarity, specifically: if the two cavity areas are judged to have a high degree of similarity, an evaluation that is the same as the second evaluation in their key monitoring level attributes is added to their existing key monitoring level attributes.

9. The image target detection method based on multi-level judgment according to claim 1, characterized in that: In step 3, based on the key monitoring level attributes of the CT image, the ResNet-50 model of the integrated spatial transformation network is used to perform target detection on the CT image. Specifically, the preprocessed CT image is first input into the ResNet-50 model of the integrated spatial transformation network. Secondly, according to the key monitoring level attributes of the cavity area, the integrated spatial transformation network module automatically adjusts the local magnification degree of the identified cavity areas A and B in the image through local affine transformation. The image adjusted by the integrated spatial transformation network is again transmitted to the subsequent layer of ResNet-50 for feature extraction. Finally, target detection is performed through the binary classification output layer of the model to determine whether a target exists.

10. The image target detection method based on multi-level judgment according to claim 9, characterized in that: According to the key monitoring level attributes of the cavity area, the integrated spatial transformation network module automatically adjusts the local magnification degree of the identified cavity areas A and B in the image through local affine transformation. Specifically, when the key monitoring levels of the cavity areas A and B include 3 highs, a local magnification of 1.5 times is performed; when it includes 2 highs and 1 medium, a local magnification of 1.4 times is performed; when it includes 1 high and 2 mediums, a local magnification of 1.3 times is performed; when it includes 3 mediums, a local magnification of 1.2 times is performed; when it includes 2 mediums and 1 low, a local magnification of 1.15 times is performed; when it includes 1 medium and 2 lows, a local magnification of 1.1 times is performed; when it includes 3 lows, the original size is maintained without magnification.

Images