Deep learning-based method for detecting inhalation of light-transmitting foreign matters in chest CT (Computed Tomography) image

Abstract

The invention relates to a method for detecting light-transmitting foreign matter inhalation in a chest CT image based on deep learning. The method comprises the following steps: executing an image preprocessing process; performing standardization processing on the input chest CT scanning image to obtain standardized image data; executing an automatic airway segmentation process; the probability that each voxel is an airway is obtained through output of the MedpSeg model; then carrying out thresholding, connected region extraction and structural integrity constraint on an output result of the MedpSeg model to obtain an airway tree structure; executing an airway tree structure correction process; judging fractured or isolated areas in the airway tree structure of the segmentation result; executing a foreign matter judgment process; the foreign matter existence probability is output through a ResNet-18-based specialized convolutional neural network; executing an output judgment process; and determining whether a foreign matter is detected in the current image. According to the invention, the missed diagnosis rate is reduced; the diagnosis workload of doctors is reduced, and the diagnosis efficiency is improved.

Description

Technical Field

[0001] This invention relates to the fields of structure, network information, and programming technology, and specifically to a method for detecting translucent foreign bodies in chest CT images based on deep learning. Background Technology

[0002] Computed tomography (CT) is an imaging technique widely used in the medical field, playing a vital role in clinical diagnosis and treatment. In the field of lung medical imaging, X-ray technology is frequently employed due to its high resolution and high contrast. These characteristics enable physicians to more accurately identify lung abnormalities, assess the extent of damage, and develop surgical plans.

[0003] Foreign body aspiration is a potentially life-threatening condition, primarily affecting children and the elderly. If left undiagnosed and untreated, it often leads to serious complications. Chest CT is the primary imaging modality for identifying airway foreign bodies, but detecting radiolucent foreign bodies remains challenging. Clinically, approximately 75% of airway foreign bodies in adult patients are radiolucent, making diagnosis difficult through traditional CT scans and often requiring manual interpretation.

[0004] In recent years, the rapid development of artificial intelligence (AI) technology has yielded significant breakthroughs in several important fields, particularly in its application in conjunction with CT technology. Among these, there has been considerable research on deep learning algorithms for CT images or chest X-ray images. Current methods primarily involve using CNN networks to build models for image classification or segmentation, such as ResNet, Inception, and UNet.

[0005] The shortcomings of existing technology are: 1. In clinical practice, up to 75% of airway foreign bodies in adult patients are radiopaque, making these cases difficult to detect with traditional CT scans, resulting in a high rate of missed diagnoses.

[0006] 2. Because the diagnosis and assessment of transmissible linear airway foreign bodies, which have a high rate of missed diagnosis in clinical practice, require manual interpretation of images for confirmation, this increases the workload of doctors and reduces diagnostic efficiency. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a deep learning-based method for detecting translucent foreign bodies in chest CT images, aiming to reduce the rate of missed diagnoses, decrease the workload of doctors, and improve diagnostic efficiency.

[0008] To solve the above problems, the technical solution provided by the present invention is as follows: A deep learning-based method for detecting translucent foreign body aspiration in chest CT images, comprising the following steps: S100. Execute the image preprocessing process; standardize the input chest CT scan image to obtain standardized image data; S200. Execute the automatic airway segmentation process; input the standardized image into the improved 3D U-Net architecture MedpSeg model, and obtain the probability of each voxel being an airway through the output of the MedpSeg model; then, the output of the MedpSeg model is thresholded, connected component extraction is performed, and structural integrity constraints are applied to obtain the airway tree structure. S300. Execute the airway tree structure correction process; determine the broken or isolated regions in the airway tree structure of the segmentation result in step S200, and then perform the following operations based on the determination result: If the determination result is that the broken or isolated region in the airway tree structure meets the criteria for the airway tree to be accepted, then the current region is retained. If the determination result is that the broken or isolated region in the airway tree structure does not meet the judgment criteria for the airway tree to be credible, then the current region is deleted. S400. Execute the foreign object detection process; input the multi-angle projection map generated on the airway tree surface into a specialized convolutional neural network based on ResNet-18, and then output the probability of the presence of a foreign object; S500. Execute the output determination process; when the output probability of the specialized convolutional neural network based on ResNet-18 is greater than or equal to the manually preset determination threshold, it is determined that there is a foreign object in the current image; when the output probability of the specialized convolutional neural network based on ResNet-18 is less than the determination threshold, it is determined that no foreign object is detected in the current image.

[0009] Preferably, the image preprocessing procedure includes the following steps: Sa100. The voxel dimensions of the original chest CT volume data are resampled using a trilinear interpolation algorithm, expressed as follows: in: Used to characterize the resampled CT volume data; Used to characterize the CT volume data prior to resampling; Sa200. The numerical values ​​of the resampled CT volume data are intensity-trimmed and limited to the Hounsfield range of [–1000, +400], expressed as follows: in: The CT volume data is used to characterize the intensity-trimmed CT volume data; Sa300. The intensity-trimmed CT volume data is normalized and linearly mapped to the interval [-1, 1], expressed as follows: in: Used to characterize the normalized CT volume data; Sa400 adjusts the three-dimensional volume data to a fixed preset size by clipping or zero-filling the edges centered on the airway center point; Sa500. The CT volume data after step Sa400 is smoothed using a three-dimensional Gaussian filter, as expressed by the following formula: in: Used to characterize the CT volume data after smoothing; The standard deviation used to characterize the three-dimensional Gaussian filter; The input chest CT scan image is DICOM format medical image data, and the slice thickness is in the range of 0.5mm to 1.5mm.

[0010] Preferably, the preset size in step Sa300 is 128 × 128 × 128; the standard deviation of the three-dimensional Gaussian filter in step Sa500 is 1.

[0011] Preferably, the structure of the MedpSeg model specifically includes an encoder, an intermediate layer, a decoder, and a first output layer; The encoder comprises a 3D convolutional layer, a batch normalization layer, a linear rectified unit, and a max pooling layer. The 3D convolutional layer is used to extract local spatial features from chest CT volumetric images. The batch normalization layer is used to standardize the feature distribution, accelerate convergence, and prevent gradient vanishing. The linear rectified unit is used to enhance the model's expressive power. The max pooling layer is used to downsample in the spatial dimension, retaining key features and reducing computational cost. The intermediate layer contains two sets of residual blocks; the residual blocks maintain channel consistency by constructing short-path connections. The decoder includes a 3D deconvolution layer, a feature concatenation layer, the 3D convolution layer, a batch normalization layer, and a linear rectification unit. The 3D deconvolution layer is used to upsample low-resolution features to restore spatial resolution. The feature concatenation layer is used to concatenate the upsampled features with the features of the corresponding layer of the encoder to achieve information fusion. The 3D convolution layer, the batch normalization layer, and the linear rectification unit are used to further extract and optimize the fused features. The first output layer includes the 3D convolutional layer and the Sigmoid activation function; the 3D convolutional layer is used to map multi-channel features to a single channel; the Sigmoid activation function is used to map the output value to the [0,1] interval and finally generate an airway binary mask.

[0012] Preferably, the 3D convolutional layer of the encoder uses a 3×3×3 convolutional kernel; the number of output channels of the encoder is [32, 64, 128]; the number of output channels of the decoder is [64, 32, 16]; and the 3D convolutional layer of the output layer uses a 1×1×1 convolutional kernel.

[0013] Preferably, the thresholding decision step of the MedpSeg model is expressed by the following formula: in: Used to characterize the probability that each voxel in the network output is an airway; Used to characterize the thresholded binary mask result, where 1 indicates that the voxel belongs to the airway and 0 indicates that the voxel does not belong to the airway.

[0014] The connected region extraction step is used to remove isolated regions with a volume of less than 100 voxels; the structural integrity constraint is used to preserve the main airway framework connecting the tracheal initiation layer to the hilum of the lung.

[0015] Preferably, the criteria for determining the reliability of the airway tree in step S300 are as follows: the main airway segment must be continuous, and each sub-branch structure must have ≥ 50 voxels; if a broken area exists, the area with a distance <15 mm from the center of the main trunk and a diameter >2 mm is retained, and the area with a connection degree of less than 10% with the main airway segment is deleted.

[0016] Preferably, the specialized convolutional neural network based on ResNet-18 in step S400 includes the following layers and parameters: Input layer: The input to the input layer consists of 12 multi-angle projection images of the airway tree, with a size of 224×224×3. 2D Convolutional Layer: The 2D convolutional layer uses a 7×7 convolutional kernel with a stride of 2, and the output feature map size is 112×112×64; First pooling layer: The first pooling layer uses 3×3 max pooling with a stride of 2, and the output feature map size is 56×56×64. Residual module layer: The residual module layer consists of 4 stages, each stage contains 2 residual blocks, and the number of output channels is 64, 128, 256 and 512 respectively, with corresponding feature map sizes of 56, 28, 14 and 7 respectively. Global average pooling layer: The output of the global average pooling layer is a 1×1×512 feature vector; Fully connected layer: The fully connected layer is used to linearly map the 1×1×512 feature vector output by the global average pooling layer, converting high-dimensional features into low-dimensional representations, thereby completing the mapping from features to the classification space; at the same time, a Dropout operation with a dropout rate of 0.5 is set in the fully connected layer to randomly drop some neuron connections to prevent overfitting and enhance the generalization ability of the model. Second output layer: The second output layer outputs the binary classification result through the Softmax function.

[0017] Preferably, the default value of the determination threshold in step S500 is 0.5, and it can be adaptively adjusted within the range of [0.3, 0.7] according to different datasets or task requirements.

[0018] Preferably, the detection method is trained using the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4 The detection method employs a Cosine Annealing learning rate scheduler and uses a Focal Loss function with parameters set to α=0.25 and γ=2.0.

[0019] Compared with the prior art, the present invention has the following advantages: 1. Because the 3D airway segmentation model of the present invention has high accuracy and stability and universality in different clinical scenarios, the rate of missed diagnosis is reduced.

[0020] 2. Because the deep learning model of this invention is significantly superior to the image reading of clinical radiology experts, it reduces the workload of doctors in diagnosis and improves diagnostic efficiency. Attached Figure Description

[0021] Figure 1 This is a schematic flowchart of a detection method according to a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the airway segmentation logic of the MedpSeg deep learning model training set data in a specific embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the workflow of image-based foreign body inhalation classification according to a specific embodiment of the present invention. Figure 4 This is a typical illustration comparing the accuracy of airway tree reconstruction using three methods: MedSeg, MedPSeg, and AG-UNet (connectivity sensing), as a specific embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0023] This invention application claims a method for detecting translucent foreign body aspiration in chest CT images based on deep learning, such as... Figure 1 As shown, it includes the following steps: S100. Execute the image preprocessing process; standardize the input chest CT scan image to obtain standardized image data.

[0024] S200. Execute the automatic airway segmentation process; input the standardized image into the improved 3D U-Net architecture MedpSeg model, and obtain the probability of each voxel being an airway through the output of the MedpSeg model; then, the output of the MedpSeg model is thresholded, connected component extraction is performed, and structural integrity constraints are applied to obtain the airway tree structure.

[0025] S300. Execute the airway tree structure correction process; determine the broken or isolated regions in the airway tree structure of the segmentation result in step S200, and then perform the following operations based on the determination result: If the determination result is that the broken or isolated region in the airway tree structure meets the criteria for the airway tree to be accepted, then the current region is retained.

[0026] If the determination result is that the broken or isolated region in the airway tree structure does not meet the criteria for the airway tree to be accepted, then the current region is deleted.

[0027] S400. Execute the foreign object detection process; input the multi-angle projection map generated on the surface of the airway tree into a specialized convolutional neural network based on ResNet-18, and then output the probability of the presence of a foreign object.

[0028] S500. Execute the output determination process; when the output probability of the specialized convolutional neural network based on ResNet-18 is greater than or equal to the manually preset determination threshold, it is determined that there is a foreign object in the current image; when the output probability of the specialized convolutional neural network based on ResNet-18 is less than the determination threshold, it is determined that no foreign object is detected in the current image.

[0029] It should be noted that the image preprocessing workflow includes the following steps: Sa100. The voxel dimensions of the original chest CT volume data are resampled using a trilinear interpolation algorithm, expressed as in equation (1): (1) in: Used to characterize CT volume data after resampling; Used to characterize CT volume data before resampling.

[0030] Sa200. The numerical values ​​of the resampled CT volume data are intensity-trimmed and the range is limited to the Hounsfield interval [–1000, +400], expressed by equation (2): (2) in: Used to characterize CT volume data after intensity clipping.

[0031] Sa300. The intensity-trimmed CT volume data is normalized and linearly mapped to the interval [-1, 1], expressed by equation (3): (3) in: Used to characterize normalized CT volume data.

[0032] Sa400 adjusts the three-dimensional volume data to a fixed preset size by cropping centered on the airway center point or using zero-fill edge.

[0033] Sa500. The CT volume data after step Sa400 is smoothed by a three-dimensional Gaussian filter, as expressed by equation (4): (4) in: Used to characterize CT volume data after smoothing; Used to characterize the standard deviation of a three-dimensional Gaussian filter.

[0034] The input chest CT scan images must be in DICOM format medical image data, with a slice thickness in the range of 0.5mm to 1.5mm and an image matrix size ≥ 512×512; the image type must be non-contrast CT, and multiple brands of scanners must be supported, including but not limited to Canon Aquilion, GE Revolution, Philips Brilliance, etc.

[0035] It should be further noted that the preset size in step Sa300 is 128 × 128 × 128. If the original image is larger than this size, it will be cropped with the center point of the airway as the center; otherwise, the edges will be zero-filled. The standard deviation of the three-dimensional Gaussian filter in step Sa500 is 1.

[0036] It should be noted that, as Figure 2As shown, the MedpSeg model's structural configuration specifically includes an encoder, intermediate layer, decoder, and first output layer.

[0037] The encoder consists of 3D convolutional layers, batch normalization layers, linear rectified units, and max pooling layers. The 3D convolutional layers are used to extract local spatial features from chest CT volumetric images. The batch normalization layers are used to standardize the feature distribution, accelerate convergence, and prevent gradient vanishing. The linear rectified units are used to enhance the expressive power of the model. The max pooling layers are used to downsample in the spatial dimension, retain the main features, and reduce computation.

[0038] The intermediate layer contains two sets of residual blocks; the residual blocks maintain channel consistency by constructing short-path connections.

[0039] The decoder includes a 3D deconvolution layer, a feature concatenation layer, a 3D convolution layer, a batch normalization layer, and a linear rectifier unit. The 3D deconvolution layer is used to upsample low-resolution features to restore spatial resolution. The feature concatenation layer is used to concatenate the upsampled features with the features of the corresponding layer of the encoder to achieve information fusion. The 3D convolution layer, batch normalization layer, and linear rectifier unit are used to further extract and optimize the fused features.

[0040] The first output layer contains a 3D convolutional layer and a Sigmoid activation function; the 3D convolutional layer is used to map multi-channel features to a single channel; the Sigmoid activation function is used to map the output value to the [0,1] interval and finally generate a binary mask for the airway.

[0041] In this specific embodiment, the encoder's 3D convolutional layer uses a 3×3×3 convolutional kernel; the encoder's output channels are [32, 64, 128]; the decoder's output channels are [64, 32, 16]; and the output layer's 3D convolutional layer uses a 1×1×1 convolutional kernel.

[0042] It should be further explained that the thresholding decision step of the MedpSeg model is expressed as equation (5): (5) in: Used to characterize the probability that each voxel in the network output is an airway; Used to characterize the thresholded binary mask result, where 1 indicates that the voxel belongs to the airway and 0 indicates that the voxel does not belong to the airway.

[0043] The connected region extraction step is used to remove isolated regions with a volume of less than 100 voxels; structural integrity constraints are used to preserve the main airway framework connecting the tracheal initiation layer to the hilum of the lung.

[0044] It should be noted that, as Figure 3As shown, the criteria for determining the reliability of the airway tree in step S300 are: the main airway segment must be continuous, and each sub-branch structure must be ≥ 50 voxels; if a broken area exists, the area with a distance <15 mm from the center of the main trunk and a diameter >2 mm is retained, and the area with a connection degree of less than 10% with the main airway segment is deleted.

[0045] It should be noted that the ResNet-18-based specialized convolutional neural network in step S400 includes the following layers and parameters: Input layer: The input to the input layer consists of 12 multi-angle projection images of the airway tree, with a size of 224×224×3.

[0046] 2D Convolutional Layer: The 2D convolutional layer uses a 7×7 convolutional kernel with a stride of 2, and the output feature map size is 112×112×64.

[0047] First pooling layer: The first pooling layer uses 3×3 max pooling with a stride of 2, and the output feature map size is 56×56×64.

[0048] Residual Module Layer: The residual module layer consists of 4 stages, each containing 2 residual blocks. The number of output channels is 64, 128, 256 and 512 respectively, and the corresponding feature map sizes are 56, 28, 14 and 7 respectively.

[0049] Global average pooling layer: The output of the global average pooling layer is a 1×1×512 feature vector.

[0050] Fully connected layer: The fully connected layer is used to linearly map the 1×1×512 feature vector output by the global average pooling layer, converting high-dimensional features into low-dimensional representations, thereby completing the mapping from features to the classification space; at the same time, a dropout operation with a dropout rate of 0.5 is set in the fully connected layer to randomly drop some neuron connections to prevent overfitting and enhance the generalization ability of the model.

[0051] Second output layer: The second output layer outputs the binary classification result through the Softmax function.

[0052] It should be further explained that the criteria for terminating training are as follows: 1. Maximum number of iterations: 100 epochs 2. Conditions for early termination: The F1 score on the validation set does not improve for ≥ 10 epochs.

[0053] The verification accuracy is ≥ 0.90 and the verification loss decreases and stagnates for 10 consecutive times.

[0054] 3. The final model selection is based on equation (6): (6).

[0055] In this specific embodiment, the output probability range of the model is [0,1]; the default value of the judgment threshold in step S500 is 0.5, and it can be adaptively adjusted within the range of [0.3,0.7] according to different datasets or task requirements.

[0056] In this specific embodiment, the model training platform is: Ubuntu 20.04, Python 3.9, PyTorch 2.0, and the GPU is an RTX A5000 (24GB VRAM); the detection method is trained using the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4 The Cosine Annealing learning rate scheduler was used; the loss function of the detection method was the Focal Loss function, with parameters set to α=0.25 and γ=2.0.

[0057] It should be noted that the following is the specific implementation process for the sample patients: Patient basic information: Male, 68 years old, with a long-term cough. In March 2022, a CT scan showed a suspected foreign body, which was later confirmed by bronchoscopy to be a chicken bone.

[0058] Process Reproduction: Input CT: Original DICOM image → Resample to 1.0mm 3 → Crop to 128 3 .

[0059] MedpSeg segmentation: Output probability map → 0.5 threshold binarization → Preserve main airway structure.

[0060] Generate projection images: airway surface reconstruction, generating 12 snapshot images.

[0061] ResNet-18 classification: The output is as shown in equation (7): (7).

[0062] The result was determined to be: there was a foreign object.

[0063] Subsequent procedures: Bronchoscopy confirmed that a chicken bone was lodged in the right middle lobe bronchus.

[0064] Patient cases were collected from three independent hospitals and were divided into an internal modeling dataset, an external validation dataset, and an independent evaluation dataset. The internal modeling dataset was used to build the model, the external validation dataset was used to test the generalizability of the model, and the independent evaluation dataset was used to evaluate the diagnostic performance of the model in different clinical settings in the real world.

[0065] Radiology experts outside the study manually reviewed the CT scans of the cases and compared them with the diagnostic results of the model.

[0066] It should be noted that the 3D airway segmentation model is accurate and has stability and universality in different clinical scenarios.

[0067] The MedpSeg model demonstrates outstanding performance in airway segmentation, consistently outperforming MedSeg and AG UNet on key evaluation metrics (e.g., ...). Figure 4 As shown in the illustration, a typical comparison of the accuracy of airway trees reconstructed using three methods—MedSeg, MedPSeg, and AG-UNet (connectivity-aware)—is presented. In the visualization, red represents model predictions, blue represents the overlap between model predictions and the gold standard, and green represents the gold standard. In the internal modeling cohort, MedpSeg exhibited the highest DSC (Dice similarity coefficient) at 87.48% and the lowest ASSD (mean symmetry surface distance) at 0.71 mm, indicating high segmentation and boundary accuracy. In the external validation cohort, the model maintained robust performance with a DSC of 86.58% and an ASSD of 0.75 mm, further confirming its universality. Compared to MedSeg and AG-UNet, MedpSeg demonstrated higher accuracy in capturing complex airway structures, especially in cases involving partial obstruction. MedpSeg's segmentation results were closest to the gold standard (clinical fact), demonstrating minimal false negatives and false positives. These findings highlight the potential of MedpSeg to improve airway segmentation accuracy, providing a reliable basis for subsequent foreign body detection and localization. The model’s consistency across internal and external datasets highlights its robustness and applicability in different clinical settings.

[0068] It should be noted that deep learning models significantly outperform clinical radiology experts in interpreting images.

[0069] The deep learning model achieved a 90.0% accuracy, 71.4% precision, 76.9% sensitivity, and an F1 score of 74.0% for patients with radiolucent airway foreign bodies on CT scans. In contrast, radiologists demonstrated 100% accuracy, but a lower accuracy of 35.7% and an F1 score of 52.6%. The deep learning model outperformed experienced radiologists in both precision (71.4% vs. 35.7%) and F1 score (74.0% vs. 52.6%). The F1 score, as the harmonic mean of precision and recall, provides a balanced metric, which is particularly useful in imbalanced datasets. These clinical performance evaluation results highlight the ability of deep learning models to significantly reduce false positives across various clinical scenarios.

[0070] In the detailed description above, various features are combined together in a single embodiment to simplify this disclosure. This approach to disclosure should not be construed as reflecting an intention that embodiments of the claimed subject matter require more features than are explicitly stated in each claim. Rather, as reflected in the appended claims, the invention is presented with fewer features than all of the features in a single disclosed embodiment. Therefore, the appended claims are hereby explicitly incorporated into the detailed description, with each claim representing a separate preferred embodiment of the invention.

[0071] The disclosed embodiments have been described above to enable any person skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments without departing from the spirit and scope of this disclosure. Therefore, this disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in this application.

[0072] The foregoing description includes examples of one or more embodiments. It is certainly impossible to describe all possible combinations of components or methods in order to describe the above embodiments, but those skilled in the art will recognize that further combinations and arrangements of the various embodiments are possible. Therefore, the embodiments described herein are intended to cover all such changes, modifications, and variations that fall within the scope of the appended claims. Furthermore, the term "comprising" as used in the specification or claims is interpreted in a manner similar to the term "including," as interpreted when used as a conjunction in the claims. Additionally, the use of any term "or" in the specification of the claims is intended to mean "non-exclusive or."

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting translucent foreign body aspiration in chest CT images based on deep learning, characterized in that: Includes the following steps: S100. Execute the image preprocessing procedure; standardize the input chest CT scan image to obtain standardized image data; S200. Execute the automatic airway segmentation process; input the standardized image into the improved 3D U-Net architecture MedpSeg model, and obtain the probability of each voxel being an airway through the output of the MedpSeg model; then, the output of the MedpSeg model is thresholded, connected component extraction is performed, and structural integrity constraints are applied to obtain the airway tree structure. S300. Execute the airway tree structure correction process; The broken or isolated regions in the airway tree structure of the segmentation result in step S200 are determined, and then the following operations are performed based on the determination result: If the determination result is that the broken or isolated region in the airway tree structure meets the criteria for the airway tree to be accepted, then the current region is retained. If the determination result is that the broken or isolated region in the airway tree structure does not meet the judgment criteria for the airway tree to be credible, then the current region is deleted. S400. Execute the foreign object detection process; input the multi-angle projection map generated on the airway tree surface into a specialized convolutional neural network based on ResNet-18, and then output the probability of the presence of a foreign object; S500. Execute the output determination process; When the output probability of the specialized convolutional neural network based on ResNet-18 is greater than or equal to a manually preset judgment threshold, it is determined that there is a foreign object in the current image; when the output probability of the specialized convolutional neural network based on ResNet-18 is less than the judgment threshold, it is determined that no foreign object is detected in the current image.

2. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 1, characterized in that: The image preprocessing procedure includes the following steps: Sa100. The voxel dimensions of the original chest CT volume data are resampled using a trilinear interpolation algorithm, expressed as follows: in: Used to characterize the resampled CT volume data; Used to characterize the CT volume data prior to resampling; Sa200. The numerical values ​​of the resampled CT volume data are intensity-trimmed and limited to the Hounsfield range of [–1000, +400], expressed as follows: in: The CT volume data is used to characterize the intensity-trimmed CT volume data; Sa300. The intensity-trimmed CT volume data is normalized and linearly mapped to the interval [-1, 1], expressed as follows: in: Used to characterize the normalized CT volume data; Sa400 adjusts the three-dimensional volume data to a fixed preset size by clipping or zero-filling the edges centered on the airway center point; Sa500. The CT volume data after step Sa400 is smoothed using a three-dimensional Gaussian filter, as expressed by the following formula: in: Used to characterize the CT volume data after smoothing; The standard deviation used to characterize the three-dimensional Gaussian filter; The input chest CT scan image is DICOM format medical image data, and the slice thickness is in the range of 0.5mm to 1.5mm.

3. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 2, characterized in that: The preset size in step Sa300 is 128 × 128 × 128; the standard deviation of the three-dimensional Gaussian filter in step Sa500 is 1.

4. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 1, characterized in that: The specific structural configuration of the MedpSeg model includes an encoder, an intermediate layer, a decoder, and a first output layer; The encoder comprises a 3D convolutional layer, a batch normalization layer, a linear rectified unit, and a max pooling layer. The 3D convolutional layer is used to extract local spatial features from chest CT volumetric images. The batch normalization layer is used to standardize the feature distribution, accelerate convergence, and prevent gradient vanishing. The linear rectified unit is used to enhance the expressive power of the model. The max pooling layer is used to downsample in the spatial dimension, retaining key features and reducing computational cost. The intermediate layer contains two sets of residual blocks; the residual blocks maintain channel consistency by constructing short-path connections. The decoder includes a 3D deconvolution layer, a feature concatenation layer, the 3D convolution layer, the batch normalization layer, and the linear rectifier unit; the 3D deconvolution layer is used to upsample low-resolution features to restore spatial resolution; The feature concatenation layer is used to concatenate the upsampled features with the features of the corresponding layer of the encoder to achieve information fusion; the 3D convolutional layer, the batch normalization layer, and the linear rectifier unit are used to further extract and optimize the fused features; The first output layer includes the 3D convolutional layer and the Sigmoid activation function; the 3D convolutional layer is used to map multi-channel features to a single channel; the Sigmoid activation function is used to map the output value to the [0,1] interval and finally generate an airway binary mask.

5. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 4, characterized in that: The encoder's 3D convolutional layer uses a 3×3×3 convolutional kernel; the encoder's output channel count is [32, 64, 128]; the decoder's output channel count is [64, 32, 16]; and the output layer's 3D convolutional layer uses a 1×1×1 convolutional kernel.

6. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 5, characterized in that: The thresholding decision step of the MedpSeg model is expressed by the following formula: in: Used to characterize the probability that each voxel in the network output is an airway; Used to characterize the thresholded binary mask result, where 1 indicates that the voxel belongs to the airway and 0 indicates that the voxel does not belong to the airway; The connected region extraction step is used to remove isolated regions with a volume of less than 100 voxels; the structural integrity constraint is used to preserve the main airway framework connecting the tracheal initiation layer to the hilum of the lung.

7. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 1, characterized in that: The criteria for determining the reliability of the airway tree in step S300 are as follows: the main airway segment must be continuous, and each sub-branch structure must be ≥ 50 voxels; if a broken area exists, the area with a distance < 15 mm from the center of the main trunk and a diameter > 2 mm is retained, and the area with a connection degree of less than 10% with the main airway segment is deleted.

8. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 1, characterized in that: The ResNet-18-based specialized convolutional neural network in step S400 includes the following layers and parameters: Input layer: The input to the input layer consists of 12 multi-angle projection images of the airway tree, with a size of 224×224×3. 2D Convolutional Layer: The 2D convolutional layer uses a 7×7 convolutional kernel with a stride of 2, and the output feature map size is 112×112×64; First pooling layer: The first pooling layer uses 3×3 max pooling with a stride of 2, and the output feature map size is 56×56×64. Residual module layer: The residual module layer consists of 4 stages, each stage contains 2 residual blocks, and the number of output channels is 64, 128, 256 and 512 respectively, with corresponding feature map sizes of 56, 28, 14 and 7 respectively. Global average pooling layer: The output of the global average pooling layer is a 1×1×512 feature vector; Fully connected layer: The fully connected layer is used to linearly map the 1×1×512 feature vector output by the global average pooling layer, converting high-dimensional features into low-dimensional representations, thereby completing the mapping from features to the classification space; at the same time, a Dropout operation with a dropout rate of 0.5 is set in the fully connected layer to randomly drop some neuron connections to prevent overfitting and enhance the generalization ability of the model. Second output layer: The second output layer outputs the binary classification result through the Softmax function.

9. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 1, characterized in that: The default value of the determination threshold in step S500 is 0.5, and it can be adaptively adjusted within the range of [0.3, 0.7] according to different datasets or task requirements.

10. The method for detecting translucent foreign bodies in chest CT images based on deep learning according to claim 1, characterized in that: The detection method is trained using the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4 The CosineAnnealing learning rate scheduler is used; the loss function of the detection method is the Focal Loss function, where the parameters are set to α=0.25 and γ=2.0.

Images