Deep learning-based chest x-ray multi-lesion intelligent screening method and system

Abstract

The application discloses a chest X-ray multi-lesion intelligent screening method and system based on deep learning, and belongs to the technical field of medical image processing. The method comprises the following steps: performing contrast adaptive enhancement, lung field automatic segmentation and rib suppression standardized pretreatment on a chest frontal film; adopting a densely connected network with embedded channel and spatial double attention modules to extract multi-scale lesion features; simultaneously detecting seven common abnormalities and outputting probability scores and confidence levels through a multi-task learning architecture; generating a heat map by using a class activation map technology to realize lesion positioning visualization; and performing pre-screening and shunting according to the confidence level.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method and system for intelligent screening of multiple lesions on chest X-ray based on deep learning. Background Technology

[0002] Chest X-ray is one of the most widely used imaging techniques in clinical practice. Its advantages, such as ease of operation, low cost, and relatively controllable radiation dose, make it irreplaceable in the initial screening and follow-up examination of lung diseases. However, with the continuous increase in outpatient visits to primary healthcare institutions, radiologists often need to read hundreds or even thousands of chest X-rays daily. Meanwhile, there is a severe shortage of radiologists with extensive experience in interpreting X-rays in these institutions. This imbalance between supply and demand directly leads to high rates of missed diagnoses and misdiagnoses. Especially when dealing with lesions requiring meticulous differentiation, such as pulmonary nodules, early-stage pulmonary tuberculosis, and cardiomegaly, X-ray fatigue and lack of experience often become key factors affecting diagnostic quality.

[0003] To address the aforementioned issues, some existing research has attempted to introduce deep learning algorithms into the field of assisted diagnosis using chest X-ray images. For example, Chinese patent CN111062947A discloses a method and system for locating lesions on chest X-rays based on deep learning. This method acquires chest X-ray images, preprocesses them, and then inputs the images into a trained convolutional neural network to generate semantic segmentation images of the lesions. After morphological processing and binarization to filter false positives, the segmentation results are finally superimposed and rendered on the original image as a heatmap to achieve lesion localization. This approach shortens doctors' image review time to some extent, but it still has the following shortcomings: First, it uses a general convolutional neural network structure for feature extraction, lacking a specialized attention-focusing mechanism for the differences between lesion areas and normal tissue in chest X-ray images. This results in limited specificity and discriminative power of feature extraction, making it prone to missed detections of small lesions or lesions with insignificant density differences. Second, this approach is essentially a single semantic segmentation task framework, failing to achieve simultaneous classification and localization of multiple lesion types. In actual clinical scenarios, a single chest X-ray may simultaneously show pneumonia exudate, pleural effusion, and cardiac hypertrophy. The single-task model cannot meet the needs of parallel screening of multiple lesions due to various abnormal signs such as large lesions. Deploying an independent model for each type of lesion would lead to a significant decrease in inference efficiency and a significant increase in maintenance costs. Thirdly, the proposed scheme only performs routine format conversion and histogram equalization operations in the image preprocessing stage, without addressing rib occlusion suppression. The density projection formed by skeletal structures such as ribs and clavicles on chest X-rays can seriously interfere with lesion identification in the lung field. Related studies have shown that approximately 82% to 95% of missed lung nodules are directly related to skeletal projection occlusion. Therefore, effectively eliminating skeletal occlusion interference is a key prerequisite for improving the performance of chest X-ray-assisted diagnosis.

[0004] Furthermore, while existing technologies have made improvements from single dimensions such as multi-label classification or attention mechanisms, they typically focus only on improving classification performance while neglecting the elimination of skeletal occlusion interference in the preprocessing stage. Alternatively, although attention modules are introduced, they are not deeply coupled with a multi-task learning framework, lacking a collaborative optimization mechanism between different stages, making it difficult to effectively suppress false positive rates while maintaining high detection sensitivity. More importantly, most existing technologies treat classification and localization as two independent tasks, lacking a unified pre-screening triage mechanism to effectively integrate test results and transform them into clinically actionable grading recommendations, thus limiting their practical deployment value as a pre-screening triage tool in real-world clinical settings. Therefore, there is an urgent need for an intelligent screening method for multiple lesions on chest X-rays that can deeply integrate standardized preprocessing, attention-enhanced feature extraction, multi-label multi-task classification, and precise lesion localization. Summary of the Invention

[0005] To address the technical problems of existing chest X-ray-assisted diagnostic methods, such as insufficient feature extraction targeting, inability to simultaneously detect multiple lesion types, and lack of bone occlusion interference elimination, this invention provides a deep learning-based intelligent screening method and system for multiple lesions on chest X-rays.

[0006] The first aspect of this invention provides a deep learning-based intelligent screening method for multiple lesions on chest X-ray, comprising the following steps:

[0007] The input chest X-ray is subjected to standardized preprocessing, which includes adaptive contrast enhancement, automatic lung field segmentation, and rib suppression processing to obtain a standardized lung field image that removes bone occlusion interference.

[0008] Standardized lung field images are input into a feature extraction network to extract multi-scale lesion features. The feature extraction network uses a dense connection network as its backbone structure, and a dual attention module consisting of a channel attention module and a spatial attention module is embedded after each dense connection block to enable the feature extraction network to focus on the lesion area and suppress the response of normal tissue.

[0009] Multi-scale lesion features are input into a multi-label classification network to obtain multi-lesion classification results. The multi-label classification network adopts a multi-task learning architecture to simultaneously detect seven types of abnormalities, including pulmonary nodules, pneumonia, pulmonary tuberculosis, pneumothorax, pleural effusion, cardiac enlargement, and mediastinal widening. It also outputs a probability score and confidence level for each type of abnormality.

[0010] Based on multi-scale lesion features and multi-lesion classification results, heat maps corresponding to each lesion type are generated through class activation map technology. The heat maps are then overlaid on standardized lung field images, with suspicious areas highlighted in red, to obtain lesion localization visualization results.

[0011] Chest X-rays are pre-screened and triaged based on probability scores and confidence levels to generate structured screening reports.

[0012] Preferably, in the standardized preprocessing, the rib suppression processing uses the lung field mask obtained from the automatic lung field segmentation as a spatial constraint condition to guide the rib suppression network to perform skeletal component separation only within the lung field region, thereby avoiding misprocessing of the external lung field region.

[0013] Preferably, the outputs of the channel attention module and the spatial attention module in the dual attention module are cascaded and fused for feature recalibration. That is, the channel attention module first compresses and excites the global information of each channel to generate a channel weight vector, and then the spatial attention module further identifies the spatial location of the lesion on the feature map after channel recalibration to generate a spatial weight map.

[0014] Preferably, the multi-label classification network employs label smoothing regularization and mixed sample data augmentation strategies to alleviate the class imbalance problem, wherein the mixed sample data augmentation strategy performs linear interpolation on training sample pairs to generate virtual training samples.

[0015] A second aspect of this invention provides a deep learning-based intelligent screening system for multiple lesions on chest X-ray, comprising:

[0016] The standardized preprocessing module is configured to perform standardized preprocessing on the input chest X-ray to obtain standardized lung field images;

[0017] A dual-attention feature extraction module is configured to input standardized lung field images into a feature extraction network to extract multi-scale lesion features;

[0018] The multi-label classification module is configured to input multi-scale lesion features into a multi-label classification network to obtain multi-lesion classification results.

[0019] The lesion localization module is configured to generate a heat map based on multi-scale lesion features and multi-lesion classification results using class activation map technology, and then overlay it onto a standardized lung field image to obtain lesion localization visualization results.

[0020] The pre-screening and triage module is configured to pre-screen and triage chest X-rays based on probability scores and confidence levels and generate structured screening reports.

[0021] The beneficial effects of this invention are as follows: by cascading and fusing contrast adaptive enhancement, automatic lung field segmentation, and rib suppression processing, the interference of skeletal occlusion on lesion identification is effectively eliminated; by embedding channel and spatial dual attention modules in the densely connected network backbone structure, the focusing ability of the feature extraction network on lesion areas is significantly improved; the multi-task learning architecture enables simultaneous detection of seven common abnormalities, and the label smoothing and mixed sample enhancement strategies effectively alleviate the class imbalance problem; and the class activation map technology enables precise localization and visual annotation of lesion areas, assisting physicians in quickly locating suspicious areas. Validated on a multi-center dataset, this invention achieves a lung nodule detection sensitivity of 94%, a pneumonia identification specificity of 91%, and a single image reading time of less than 2 seconds, making it a valuable pre-screening and triage tool for radiology departments to effectively improve the image reading efficiency of primary healthcare institutions. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the intelligent screening method for multiple lesions on chest X-ray based on deep learning provided in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the architecture of the intelligent screening system for multiple lesions on chest X-ray based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 1 , Figure 1 This is a flowchart illustrating the intelligent screening method for multiple lesions on chest X-ray based on deep learning provided in this embodiment of the invention. In this embodiment, the method includes steps S1 to S5, which form a deeply coupled cascaded processing pipeline. The output of the previous step directly serves as the key input of the next step, and the feedback results of the subsequent steps can influence the parameter adjustment of the previous steps, thereby forming an end-to-end closed-loop optimization architecture.

[0026] Step S1: Perform standardized preprocessing on the input chest X-ray to obtain a standardized lung field image that removes bone occlusion interference.

[0027] In one embodiment of the present invention, the standardized preprocessing includes three sub-steps executed sequentially: contrast adaptive enhancement, automatic lung field segmentation, and rib suppression processing. There is a data dependency between the three sub-steps, that is, the output of the previous sub-step is used as the input for the subsequent sub-step.

[0028] Specifically, the contrast adaptive enhancement sub-step first performs contrast-limited adaptive histogram equalization on the input raw chest X-ray. In practice, the system first converts the input DICOM format medical image into a grayscale matrix representation and normalizes the pixel values ​​to the range of 0 to 1. Then, the grayscale image is divided into several non-overlapping local region blocks, preferably each region block is set to 8x8 pixels. Histogram equalization is performed independently within each region block, while an upper limit constraint is applied to the contrast enhancement amplitude to prevent excessive noise amplification. In one embodiment of the invention, the contrast limiting coefficient... The value is set to 2.5, and this value is chosen based on the following: when Below 2.0, the contrast enhancement of low-density lesions (such as ground-glass nodules) within the lung field is insufficient, making it difficult to effectively distinguish them from the surrounding normal lung tissue; when When the value exceeds 3.5, noise in the rib and mediastinal edge regions is significantly amplified, increasing interference in subsequent feature extraction stages. Preferably, The value ranges from 2.0 to 3.5. After adaptive contrast enhancement, the local contrast of the output image is uniformly improved, especially the tissue layers in the lung field edge region and the paramedian region are clearer, laying a good image quality foundation for subsequent lung field segmentation and lesion identification.

[0029] After completing the contrast adaptive enhancement, the automatic lung field segmentation sub-step is initiated. In one embodiment of this invention, the automatic lung field segmentation is implemented using a pre-trained U-Net semantic segmentation network. This U-Net network takes the enhanced grayscale image as input and outputs a binary lung field mask of the same size as the input, where a pixel value of 1 represents a lung field region and a pixel value of 0 represents a non-lung field region. In this embodiment, the encoder part of the U-Net network includes four downsampling stages, each consisting of two 3x3 convolutional layers and one 2x2 max-pooling layer, with 64, 128, 256, and 512 channels respectively. The decoder part symmetrically sets four upsampling stages and uses skip connections to concatenate and fuse the feature maps of corresponding encoder layers. This segmentation network is pre-trained on a dataset containing approximately 5000 chest anteroposterior images with labeled lung field boundaries, achieving a segmentation accuracy of over 0.97, measured by the Dice coefficient. Preferably, to further improve the boundary accuracy of the lung field mask, morphological closing operation processing is performed on the U-Net output, and a disk structure element with a radius of 5 pixels is used to fill the tiny voids that may exist at the edge of the lung field mask.

[0030] It is particularly noteworthy that the output of the automatic lung field segmentation in this invention is not only used in the subsequent rib suppression sub-step, but also participates as a spatial constraint in the generation of the lesion localization heatmap in step S4. This cross-step data reuse reflects the design concept of deep coupling between the steps in this invention—the lung field mask provides spatial guidance for the separation of skeletal components in the preprocessing stage and regional constraints for the suppression of false positives in the postprocessing stage, thereby establishing a bridge for information transmission between preprocessing and postprocessing.

[0031] The rib suppression sub-step is the most innovative step in the standardized preprocessing. In one embodiment of this invention, the rib suppression process is implemented using a skeletal component separation algorithm based on a conditional generative adversarial network (GAN). Specifically, this GAN uses the stitched result of the enhanced grayscale image and the lung field mask as input conditions, and outputs a soft tissue image with skeletal component suppression through a generator network. In this embodiment, the generator network adopts a ResNet architecture, containing 9 residual blocks, with 2 input channels (grayscale image channel and lung field mask channel) and 1 output channel (soft tissue image channel). The discriminator network adopts a PatchGAN architecture, judging the authenticity of the input image within a 70x70 pixel local receptive field.

[0032] In one embodiment of the present invention, the training data for the rib suppression network comes from a set of paired images obtained by dual-energy subtraction angiography. Specifically, the same subject is exposed using a dual-energy X-ray device under both high-energy (e.g., 120kVp) and low-energy (e.g., 60kVp) tube voltage conditions. A weighted logarithmic subtraction algorithm is then used to obtain images of the skeletal and soft tissue components. Using the original chest X-ray as input and the soft tissue image as the supervision label, a conditional generative adversarial network is trained to learn the mapping relationship from the original chest X-ray to the soft tissue image. In this embodiment, the training dataset contains approximately 2000 pairs of paired dual-energy images. During training, the following method is used: The combination of reconstruction loss and adversarial loss is used as the objective function, where Loss weighting coefficient Set to 100, the weighting coefficient against loss. The weighting is set to 1, and this weighting ratio is based on the following: The loss function ensures pixel-level consistency between the generated image and the real soft tissue image, while the adversarial loss function ensures realism in texture detail. Weights can prevent the generator from producing too many artifacts.

[0033] After rib suppression processing, the density projection of skeletal structures in the output standardized lung field image is effectively eliminated, and the soft tissue structures (including vascular texture, bronchial course, and various lesion signs) within the lung field region are presented more clearly. In one embodiment of the present invention, after the standardized preprocessing is completed and before being input into the feature extraction network, online data augmentation operations are performed on the standardized lung field image to improve the model's generalization ability and robustness. The online data augmentation operations include random horizontal flipping (probability of 0.5), random rotation (rotation angle range of ±10 degrees), random brightness adjustment (adjustment coefficient range of 0.9 to 1.1), and random contrast adjustment (adjustment coefficient range of 0.9 to 1.1). The above augmentation operations are only performed during the training phase, and no data augmentation is performed during the inference phase. Compared with the existing preprocessing scheme that only uses histogram equalization, the present invention effectively reduces false positive detection results caused by skeletal occlusion by introducing rib suppression processing, especially in the lung apex region and the region above the diaphragm, where the confusion between rib projection and lung nodules is significantly improved.

[0034] Step S2: Input the standardized lung field image into the feature extraction network to extract multi-scale lesion features.

[0035] In one embodiment of this invention, the feature extraction network uses DenseNet-121 as its backbone structure and is innovatively modified for screening multiple lesions on chest X-rays. The core design concept of DenseNet-121 lies in its dense connection mechanism, where the input of any layer within each dense connection block contains the concatenated output feature maps of all preceding layers. This dense connection method effectively promotes feature reuse and alleviates the gradient vanishing problem. In this embodiment, the DenseNet-121 backbone structure contains four dense connection blocks, each containing 6, 12, 24, and 16 dense connection layers, respectively, with growth rates of [missing information]. The value is set to 32, meaning that each densely connected layer outputs a feature map with 32 channels.

[0036] The innovative modification of this invention lies in embedding a dual attention module, consisting of a channel attention module and a spatial attention module, at the output end of each densely connected block, forming a cascaded processing unit of dense connection-channel attention-spatial attention. The technical motivation for this design is that in the multi-channel feature map output by the densely connected block, different channels correspond to different levels of semantic information. Some channels may primarily respond to normal lung texture or bone residue information, while others primarily respond to abnormal signs in the lesion area. By adaptively weighting the importance of each channel through the channel attention module, redundant channel responses unrelated to the lesion can be suppressed. Furthermore, the spatial attention module further identifies the location information of the lesion in the spatial dimension, focusing the model's perception capabilities on the spatial region where the lesion is located.

[0037] Specifically, the processing flow of the channel attention module is as follows: First, the input feature map is processed... (in For the number of channels, and Global average pooling and global max pooling operations are performed along the spatial dimension (the height and width of the feature map, respectively) to obtain the channel description vectors. and In one embodiment of the present invention, channel attention weights The calculation formula is:

[0038] ,

[0039] in: The sigmoid activation function, with an output range of (0,1), is used to normalize the attention weights to between 0 and 1; the MLP is a multilayer perceptron with shared parameters, consisting of two fully connected layers. The first fully connected layer will... Dimensional input reduction to dimension( In this embodiment, the compression ratio is... The value is set to 16, and this value is chosen based on... A good balance was achieved between the number of model parameters and the ability to model attention. When the number of parameters increases but the performance improvement is limited, (When the value is too large, too much interaction information is lost between channels), and after processing by the ReLU activation function, the second fully connected layer will... Restored to dimension; This is the global average pooling vector, which encodes the average activation intensity information of each channel, in units of dimensionless normalized values; This is the global max-pooling vector, encoding the most salient activation response information for each channel. Channel attention weights. The input feature map is obtained by applying element-wise multiplication. ,in This indicates a channel-by-channel broadcast multiplication operation.

[0040] Feature map after channel recalibration Based on this, the spatial attention module performs the following processing: First, it performs average pooling and max pooling operations along the channel dimension to obtain two single-channel spatial description maps. and Then, the two are concatenated along the channel dimension, and a spatial attention weight map is generated using a 7x7 convolutional layer and a sigmoid activation function. In one embodiment of the present invention, the spatial attention weight... The calculation formula is:

[0041] ,

[0042] in: For a convolution operation with a kernel size of 7x7, the number of input channels is 2 (the concatenation of the average pooling result and the max pooling result), and the number of output channels is 1. The reason for choosing a kernel size of 7x7 instead of 3x3 is that a larger receptive field can capture a wider range of spatial context information, which helps the spatial attention module to fully cover larger lesions (such as pneumonia exudation areas or pleural effusion areas). This indicates a splicing operation along the channel dimension; The activation function is Sigmoid. Spatial attention weights. The feature map after channel recalibration is applied through element-wise multiplication. The feature map after dual attention enhancement is obtained. .

[0043] It is important to emphasize that the cascaded application order of channel attention and spatial attention in this invention is not arbitrary, but based on the following technical considerations: the channel attention module first semantically filters out channels related to the lesion, filtering out channels encoding normal tissue or noise information. Only then can the spatial attention module more accurately locate the spatial position of the lesion in the purified feature space. This cascaded strategy of first filtering channels and then determining location has higher positioning accuracy than parallel fusion strategies or reverse cascaded strategies. Experimental verification shows that this cascaded order improves the positioning accuracy of lung nodule detection by approximately 3.2 percentage points compared to the reverse cascaded strategy.

[0044] In one embodiment of the present invention, before inputting the standardized lung field image into the DenseNet-121 backbone structure, the single-channel grayscale image is first copied as a 3-channel input to adapt to the DenseNet-121 input interface, and the image size is adjusted to 512x512 pixels. Preferably, the DenseNet-121 backbone structure is initialized using weight parameters pre-trained on the ImageNet large-scale natural image dataset. This transfer learning strategy allows the feature extraction network to inherit the low-level visual features (such as edges, textures, and shapes) learned by the pre-trained model. Based on this, it can quickly adapt to specific feature patterns in the field of medical imaging through fine-tuning training on a chest X-ray dataset. In this embodiment, the fine-tuning training process is divided into two stages: the first stage freezes the parameters of the first two densely connected blocks of DenseNet-121, and only trains the parameters of the last two densely connected blocks and all dual attention modules. The training is repeated for 5 epochs to bring the parameters of the newly introduced attention modules to a reasonable range; the second stage unfreezes all parameters and performs end-to-end joint training, continuing training until the early stopping condition is met. This phased training strategy effectively prevents randomly initialized attention module parameters from causing destructive interference to pre-trained features in the early stages of training. The output of the feature extraction network is the feature map of the fourth densely connected block after processing by the dual attention module, with a spatial size of 16x16 pixels and 1024 channels. This multi-scale lesion feature map serves as a shared input for both the classification network in step S3 and the localization module in step S4. This feature sharing mechanism is one of the core designs of the multi-task learning architecture of this invention, enabling the classification and localization tasks to be jointly optimized in the same feature space.

[0045] Step S3: Input the multi-scale lesion features into the multi-label classification network to obtain the multi-lesion classification results.

[0046] In one embodiment of the present invention, the multi-label classification network adopts a multi-task learning architecture to process the multi-scale lesion feature maps output in step S2 in parallel to simultaneously detect seven common chest abnormalities. Unlike the single semantic segmentation task architecture in the prior art, the multi-label classification network of the present invention can simultaneously output seven independent binary classification results in one forward inference, each result corresponding to the probability of the existence of a type of chest abnormality.

[0047] Specifically, the multi-label classification network first processes the feature map output in step S2. Performing global average pooling compresses the spatial dimension to 1x1, resulting in a 1024-dimensional global feature vector. Subsequently, the global feature vector is reduced to 512 dimensions through a shared fully connected layer, and then input into seven independent classification heads. Each classification head consists of a fully connected layer and a sigmoid activation function, outputting a probability score between 0 and 1, representing the likelihood of the corresponding lesion type. In this embodiment, the seven classification heads correspond to pulmonary nodules, pneumonia, pulmonary tuberculosis, pneumothorax, pleural effusion, cardiomegaly, and mediastinal widening, respectively.

[0048] In one embodiment of the present invention, in order to address the severe class imbalance problem among different lesion types in the training dataset (for example, in a typical multicenter chest X-ray dataset, the positive sample ratio of pneumonia may be as high as 15%, while the positive sample ratio of pneumothorax may be only about 2%), a combined scheme of label smoothing regularization and mixed sample data augmentation strategy is adopted.

[0049] The core idea of ​​label smoothing regularization is to replace hard labels (i.e., 0 and 1) in the training labels with soft labels to prevent the model from becoming overconfident in the training samples. In one embodiment of this invention, the label smoothing coefficient... Setting it to 0.1 means the original label value 1 is replaced with... The original label value 0 was replaced with ,in This represents the total number of lesion types. The value range is usually from 0.05 to 0.2. The smoothing effect is not significant when the size is too small, but when the size is too small, the smoothing effect is not significant. An excessively large value will reduce the discriminative power of the model output. In this embodiment, 0.1 is selected as a compromise value.

[0050] Mixup data augmentation strategy: Randomly select 2 training samples in each training batch. and Virtual training samples are generated through linear interpolation. In one embodiment of the present invention, the formula for generating virtual samples is:

[0051] ,

[0052] ,

[0053] in: The generated virtual input image has the same size as the original input image, 512x512 pixels; The generated virtual label vector has a dimension of 7, and each element takes a continuous value between 0 and 1. and These are two original input images randomly sampled from the training batch; and These are the corresponding multi-label vectors; The mixing coefficients follow a Beta distribution. ,in The shape parameter of the Beta distribution is given in this embodiment. The value is set to 0.4, and this value is chosen based on the fact that... Approaching 0 Tends to 0 or 1 (i.e., hardly mixes), when When equal to 1 It follows a uniform distribution (excessive mixing may compromise the semantic integrity of medical images). Make Most of the time the values ​​are concentrated between 0.6 and 1.0, introducing moderate interpolation perturbations while maintaining the semantic coherence of the training samples.

[0054] In one embodiment of the present invention, the training loss function of the multi-label classification network adopts the weighted binary cross-entropy loss, and its calculation formula is as follows: ,in: This represents the multi-label classification loss value, expressed as a dimensionless numerical value. The smaller the value, the higher the consistency between the model's prediction and the label. In this embodiment, the number of samples in the training batch is... Set to 32; This represents the total number of lesion types. For the first The first sample The soft label value of the lesion after label smoothing; For the first The first sample The predicted probability value of the lesion type is output by the Sigmoid activation function, and the value range is (0,1). For the first The weighting coefficient for positive samples with similar lesions is calculated as follows: ,in The total number of samples in the training set. For the first Number of positive samples; For the first The weighting coefficient for negative samples of lesion-like lesions is calculated as follows: ,in For the first Number of negative samples. This is determined by introducing class-related weighting coefficients. and This amplifies the loss contribution of rare classes during training, thereby effectively alleviating the problem of model bias towards the majority class caused by class imbalance.

[0055] Preferably, the present invention also employs a cosine annealing learning rate scheduling strategy during the training process. The initial learning rate is set to... The minimum learning rate is The training period was set to 10 epochs. The optimizer used was AdamW, and the weight decay coefficient was set to... The total number of training rounds is 100, and an early stopping strategy is adopted. Training is terminated when the average AUC on the validation set fails to improve for 15 consecutive rounds.

[0056] Furthermore, during the inference phase of the multi-label classification network, the probability score for each type of lesion needs to be converted into a binary classification result through a corresponding decision threshold. In one embodiment of the present invention, the decision threshold for each type of lesion is not a uniform 0.5, but rather an optimal threshold is determined separately on the validation set based on maximizing the Youden exponent. Specifically, receiver operating characteristic (ROC) curves are plotted for each type of lesion, and the Youden exponent corresponding to each possible threshold is calculated. Select to make The maximized threshold is used as the optimal decision threshold for this type of lesion. In this embodiment, after threshold optimization on the validation set, the optimal decision thresholds are 0.42 for pulmonary nodules, 0.38 for pneumonia, 0.45 for pulmonary tuberculosis, 0.52 for pneumothorax, 0.35 for pleural effusion, 0.40 for cardiac enlargement, and 0.48 for mediastinal widening.

[0057] Step S4: Based on the multi-scale lesion features and multi-lesion classification results, generate heat maps corresponding to each lesion type using class activation map technology, and overlay the heat maps onto the standardized lung field image to highlight suspicious areas in red, thereby obtaining lesion localization visualization results.

[0058] In one embodiment of the present invention, the lesion localization module employs gradient-weighted activation map technology to achieve precise localization and visual annotation of the lesion region. Unlike existing technologies that directly generate segmentation maps through semantic segmentation, the present invention uses gradient information from a classification network to infer the spatial region that contributes most to the classification decision. This gradient-based localization method does not require additional pixel-level annotation data and can achieve weakly supervised localization solely based on image-level classification labels.

[0059] Specifically, for the first The generation process of the gradient-weighted class activation map for lesion-like lesions is as follows: First, calculate the... Predictive probability of lesions The feature map output in step S2 The Middle gradient of each channel Then, global average pooling is performed on the gradient along the spatial dimension to obtain the first... The first channel is for the first Importance weight of lesion-like features In one embodiment of the present invention, channel importance weights The calculation formula is: ,in: For the first The first channel is for the first The importance weight of the lesion classification result is expressed as a dimensionless value. The larger the value, the greater the contribution of the channel to the discrimination of that type of lesion. and These are the height and width of the feature map, respectively (in pixels). For the first The predicted probability of the lesion type is obtained from the Sigmoid output of the corresponding classification head in step S3; Feature maps after dual attention enhancement The Middle The spatial location of each channel The activation value at that location.

[0060] After obtaining the importance weights of all channels, the first... Activation map of lesion-like structures The following is obtained by weighted summation of the feature maps of all channels followed by ReLU activation: ,in: For the first The activation map of the lesion-like structure is 16x16 pixels in size; is the total number of channels in the feature map; the ReLU activation function is used to filter out negative responses (negative values ​​indicate an inhibitory effect rather than an activating effect on this type of lesion), retaining only the positive response regions.

[0061] In one embodiment of the present invention, due to the class activation graph The spatial size of the activation map (16x16 pixels) is much smaller than that of the input image (512x512 pixels), so bilinear interpolation is needed to upsample the activation map to the same size as the input image. After the upsampled activation map is normalized to the range of 0 to 1 using a minimum-maximum normalization process, it is mapped to a heatmap using the JET color scheme. Specifically, areas with normalization values ​​close to 0 are mapped to blue (indicating that the area is normal tissue), and areas with normalization values ​​close to 1 are mapped to red (indicating that the area is a high-confidence suspicious lesion area). Finally, the heatmap is rendered on top of the standardized lung field image in a semi-transparent overlay, with an overlay weight set to 0.4 (i.e., the heatmap accounts for 40% and the original image accounts for 60%), so that physicians can still see the anatomical details of the original image while observing the lesion localization markings.

[0062] It is worth noting that this invention introduces a lung field mask constraint mechanism during the heatmap generation process. Specifically, the lung field mask obtained in step S1 is multiplied element-wise with the upsampled activation map, forcibly setting the activation values ​​outside the lung field region to zero. This design aims to prevent the activation map from generating false highlighting in areas outside the lung field (such as the mediastinum and subcutaneous soft tissue regions), thereby effectively reducing the false positive rate of lesion localization. This is another manifestation of the lung field mask being reused across steps in step S1, further highlighting the deep coupling and synergy between the steps of this invention.

[0063] Step S5: Pre-screen and triage chest X-rays based on probability scores and confidence levels to generate structured screening reports.

[0064] In one embodiment of the present invention, the pre-screening and triage module, based on the probability scores of the seven types of lesions output in step S3, categorizes the screening results of each chest X-ray into different processing priority queues according to a preset confidence level classification rule. Preferably, the confidence level is divided into three levels: when the probability score of a certain type of lesion is higher than the high confidence threshold corresponding to that category... When the probability score is within the low confidence threshold, it is judged as a high-confidence positive and included in the emergency review queue; when the probability score is within the low-confidence threshold... With high confidence threshold When the score is between these thresholds, it is judged as medium confidence questionable and included in the regular review queue; when the probability score is below the low confidence threshold... If the result is negative with low confidence, it is classified into the queue that does not require retesting.

[0065] In one embodiment of the present invention, a pulmonary nodule is used as an example. Set to 0.75. The value is set to 0.25. In multi-label scenarios, if any type of lesion in a chest X-ray is determined to be positive with high confidence, the entire chest X-ray is included in the emergency review queue; if there is no positive lesion with high confidence but there is a susceptibility with moderate confidence, it is included in the regular review queue.

[0066] The structured screening report includes the following information: basic information of the examinee, probability scores and confidence levels of various lesions, lesion localization results marked with heat maps, and triage recommendations. In addition, the report includes a brief textual description of each type of detected lesion, such as a high-density, roundish shadow with a diameter of approximately 1.2 cm visible in the left upper lung field, with a lung nodule probability score of 0.87, suggesting further confirmation based on clinical information, so that radiologists can quickly understand the key information of the system's detection results. The report is output in a standardized JSON format, facilitating integration with hospital information systems and image archiving and communication systems. Preferably, the system also records the processing time for each screening. In this embodiment, the end-to-end processing time from input to generation of a complete screening report for a single chest X-ray is less than 2 seconds (measured under an NVIDIA Tesla V100 GPU environment), meeting the real-time requirements of high-throughput pre-screening and triage in the daily work of the radiology department.

[0067] A key technical feature of this invention lies in the closed-loop structure of bidirectional information transmission between each step. Specifically, the lung field mask output in step S1 not only constrains the spatial range of rib suppression during the preprocessing stage but also constrains the spatial range of the heatmap in step S4 to suppress false positives; the multi-scale lesion feature map output in step S2 is shared by steps S3 and S4, achieving joint optimization of classification and localization; the pre-screening and triage results in step S5 can be fed back for hard sample mining during the training process—marking medium-confidence suspicious samples as hard samples and increasing their sampling weight in the next round of training, thereby continuously improving the model's discriminative ability on boundary samples. This multi-step deep coupling and closed-loop collaborative mechanism results in a synergistic effect of 1+1>2 between each step, and the overall system performance is significantly better than the linear superposition of simple serial steps.

[0068] Please see Figure 2 , Figure 2 This is a schematic diagram of the architecture of a deep learning-based intelligent screening system for multiple lesions on chest X-ray provided in an embodiment of the present invention. In this embodiment, the system includes a standardized preprocessing module 1, a dual attention feature extraction module 2, a multi-label classification module 3, a lesion localization module 4, and a pre-screening and triage module 5. The above five modules correspond one-to-one with steps S1 to S5 in the method embodiment. The modules achieve high-speed data transmission and collaborative work through a data bus and shared storage space.

[0069] The standardization preprocessing module 1 is configured to perform standardization preprocessing on the input chest X-ray to obtain a standardized lung field image. In one embodiment of the present invention, the standardization preprocessing module 1 further includes a contrast enhancement submodule, a lung field segmentation submodule, and a rib suppression submodule. The contrast enhancement submodule is configured to perform contrast-limited adaptive histogram equalization processing on the input image, as described in step S1 of the method embodiment, where the contrast limiting coefficient is... The value is set to 2.5. The lung field segmentation submodule is configured to generate a binary lung field mask using a pre-trained U-Net semantic segmentation network. This lung field mask is stored in a shared storage space for subsequent use by the lesion localization module 4. The rib suppression submodule is configured to use the stitching result of the enhanced grayscale image and the lung field mask as input conditions, and output a soft tissue image with bone component suppression as a standardized lung field image through a conditional generative adversarial network. Preferably, the execution order of the three submodules in the standardization preprocessing module 1 is contrast enhancement, lung field segmentation, and rib suppression, and the output result of lung field segmentation is also used as the spatial constraint input for rib suppression.

[0070] The dual-attention feature extraction module 2 is configured to input the standardized lung field image output by the standardized preprocessing module 1 into the feature extraction network to extract multi-scale lesion features. As described in step S2 of the method embodiment, the feature extraction network uses DenseNet-121 as its backbone structure, and embeds a dual-attention module consisting of a channel attention module and a spatial attention module at the output end of each densely connected block. The output of the dual-attention feature extraction module 2 is a multi-scale lesion feature map with 1024 channels and a size of 16x16 pixels. This feature map is provided to both the multi-label classification module 3 and the lesion localization module 4 through shared storage space, thereby achieving deep coupling between the classification task and the localization task at the feature level.

[0071] The multi-label classification module 3 is configured to input the multi-scale lesion features output by the dual-attention feature extraction module 2 into the multi-label classification network to obtain multi-lesion classification results. As described in step S3 of the method embodiment, the multi-label classification network includes a global average pooling layer, a shared fully connected layer, and seven independent classification heads, which output probability scores and confidence levels for seven types of chest abnormalities, respectively. During the training phase, the multi-label classification module 3 employs label smoothing regularization and mixed sample data augmentation strategies to alleviate class imbalance. During the inference phase, it converts the probability scores into binary classification results through independently optimized decision thresholds for each class. Preferably, the multi-label classification module 3 also integrates a threshold adaptive calibration submodule. This submodule dynamically fine-tunes the decision thresholds for each class based on the statistical distribution information of recently processed samples to adapt to the changes in prior probability caused by differences in the disease spectrum distribution of patient groups between different medical institutions, ensuring that the system can maintain stable screening performance in different deployment environments.

[0072] The lesion localization module 4 is configured to generate heatmaps corresponding to each lesion type based on the multi-scale lesion features output by the dual attention feature extraction module 2 and the multi-lesion classification results output by the multi-label classification module 3, using gradient-weighted class activation map technology. As described in step S4 of the method embodiment, the lesion localization module 4 reads the lung field mask stored in the standardized preprocessing module 1 from the shared storage space and applies it as a spatial constraint to the class activation map to suppress false positive annotations in the outer regions of the lung fields. Preferably, the heatmap output by the lesion localization module 4 is visualized and encoded using the JET color scheme and rendered on the original image in a semi-transparent overlay manner.

[0073] The pre-screening and triage module 5 is configured to pre-screen and triage chest X-rays based on the probability scores and confidence levels output by the multi-label classification module 3 and generate a structured screening report. As described in step S5 of the method embodiment, the pre-screening and triage module 5 categorizes the screening results into three priority queues: emergency review, routine review, and no review required, and outputs a structured screening report in JSON format for integration with the hospital information system. Preferably, the pre-screening and triage module 5 is also configured with a difficult sample feedback interface, which transmits the identification information of medium-confidence suspicious samples to the training management subsystem. This allows for increased sampling weights for these difficult samples in subsequent model iterations, thereby forming a closed-loop continuous improvement mechanism of inference-feedback-optimization.

[0074] In one embodiment of the present invention, the data transfer process between the above five modules is as follows: the standardized preprocessing module 1 receives the original DICOM format chest X-ray, and outputs a standardized lung field image after contrast enhancement, lung field segmentation and rib suppression processing, while storing the lung field mask in a shared buffer; the dual attention feature extraction module 2 obtains the standardized lung field image from the standardized preprocessing module 1, and outputs a multi-scale lesion feature map in the shared buffer after extraction by the DenseNet-121 backbone structure and the dual attention module; the multi-label classification module 3 and the lesion localization module 4 simultaneously read the multi-scale lesion feature map from the shared buffer, and perform classification inference and heatmap generation respectively; the lesion localization module 4 also reads the lung field mask from the shared buffer to perform spatial constraint operations; the pre-screening and triage module 5 integrates the output results of the multi-label classification module 3 and the lesion localization module 4 to generate the final structured screening report.

[0075] In one embodiment of the present invention, the system is deployed on a server platform equipped with an NVIDIA Tesla V100 GPU, 32GB of video memory, an Intel Xeon Gold 6248 processor, and 128GB of RAM. The system receives DICOM format chest X-rays from a medical image archiving and communication system, processes them serially through various modules, and then sends a structured screening report back to the hospital information system. Preferably, the system supports batch processing mode, where a single GPU can simultaneously process parallel inference requests for four chest X-rays. Under this configuration, the system's average throughput is approximately 1800 images per hour. To meet the deployment needs of different medical institutions, the system also supports a lightweight deployment scheme based on the NVIDIA Jetson Xavier NX edge computing platform. Through model quantization and knowledge distillation techniques, the model parameter size is compressed from approximately 8.5M to approximately 2.1M, and the inference time per image at the edge is approximately 4.5s, still meeting the application scenarios of non-real-time pre-screening in primary healthcare institutions.

[0076] To verify the effectiveness of the method of the present invention, a systematic comparative experiment was conducted on a multi-center labeled dataset.

[0077] In one embodiment of this invention, the training data came from three tertiary-level hospitals and two primary healthcare institutions, totaling approximately 150,000 chest X-rays. Of these, approximately 85,000 were labeled as containing at least one type of abnormality, and approximately 65,000 were labeled as normal. The labeling was independently completed by at least two attending radiologists with more than five years of experience. In cases where the labeling results of two radiologists differed, an associate chief physician arbitrated the dispute. The dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio, with the division performed on a patient-by-patient basis (i.e., all chest X-rays taken by the same patient at different time points appeared in only the same subset), to strictly avoid data leakage caused by images of the same patient being scattered across different subsets.

[0078] Quantitative evaluation results on the test set show that the detection performance of the method of the present invention in seven types of chest abnormalities is as follows: lung nodule detection sensitivity is 94.2%, specificity is 89.7%, and AUC is 0.961; pneumonia identification sensitivity is 92.8%, specificity is 91.3%, and AUC is 0.954; pulmonary tuberculosis detection sensitivity is 91.5%, specificity is 93.6%, and AUC is 0.948; and pneumothorax detection sensitivity is 96.1% and specificity is 95.8%. The mean AUC for the detection of pleural effusion was 93.4%, specificity was 92.1%, and AUC was 0.957; the mean AUC for cardiomegaly was 90.6%, specificity was 94.2%, and AUC was 0.945; and the mean AUC for mediastinal widening was 88.3%, specificity was 95.7%, and AUC was 0.936. The mean AUC for the seven abnormalities was 0.955, the mean sensitivity was 92.4%, and the mean specificity was 93.2%.

[0079] To verify the contributions of each technical module of this invention, the following ablation experiments were designed. Experiment 1: With rib suppression removed, classification was performed using only images with contrast enhancement and lung field segmentation. The results showed that the average AUC for the seven abnormalities decreased from 0.955 to 0.932, especially the sensitivity for lung nodules, which decreased from 94.2% to 87.6%, demonstrating that rib suppression has a significant effect on reducing bone occlusion interference. Experiment 2: With the dual attention module removed, feature extraction was performed using the original DenseNet-121 (without attention enhancement). The results showed that the average AUC decreased from 0.955 to 0.938, with a particularly significant decrease in detection performance for pneumonia and pleural effusion, demonstrating that the dual attention module can effectively guide the feature extraction network to focus on the lesion area. Experiment 3: The multi-label, multi-task architecture was replaced with seven independent binary classification models. The results showed that the average AUC decreased from 0.955 to 0.941, and the inference time per image increased from 1.8s to 6.3s, demonstrating that the multi-task shared architecture not only improved classification performance but also significantly reduced inference time. Experiment 4: Removing the label smoothing and mixed sample enhancement strategy resulted in a decrease in detection sensitivity of rare categories such as pneumothorax and mediastinal widening by approximately 5.3 and 4.8 percentage points, respectively, demonstrating the important role of this combined strategy in alleviating class imbalance. Experiment 5: Removing the spatial constraint of the lung field mask from the heatmap generation process increased the false positive rate for lesion localization from 8.6% to 14.2%, especially with a large number of false highlighting marks appearing in the paramedian and subdiaphragmatic regions, proving that the cross-step reuse mechanism of the lung field mask plays a crucial role in suppressing false positives in localization. The above ablation experimental results fully demonstrate that the overall performance gain generated by the deep coupling and synergistic cooperation among the various technical modules in this invention is far greater than the simple sum of the individual contributions of each module.

[0080] Compared with the scheme disclosed in CN111062947A, the method of this invention has significant advantages in the following aspects: In terms of the number of lesion types detected, this invention can simultaneously detect 7 common chest abnormalities, while CN111062947A only supports the localization of a single type of lesion; In terms of eliminating skeletal occlusion interference, this invention introduces rib suppression processing based on conditional generative adversarial networks, while CN111062947A does not involve skeletal occlusion suppression; In terms of feature extraction capability, this invention enhances the discriminativeness and localization accuracy of features through a dual attention module of channels and space, while CN111062947A uses a general convolutional neural network structure; In terms of inference efficiency, this invention achieves simultaneous output of 7 types of detection results in a single inference through a multi-task sharing architecture, with a single image processing time of less than 2 seconds, possessing the capability for real-time clinical screening.

[0081] Furthermore, to verify the effectiveness of the method of this invention in actual clinical applications, a prospective application evaluation was conducted for three months in the radiology department of a primary healthcare institution. During the evaluation period, the department processed approximately 12,000 chest X-rays, which were pre-screened and triaged by the system of this invention before being finally diagnosed by two radiologists. The results showed that the missed diagnosis rate for images pre-screened by the system and not requiring review was only 0.8% upon physician confirmation, and the positive concordance rate for images pre-screened for urgent review and confirmed by physicians reached 87.5%. Compared with before the introduction of the system, the average time for reviewing each chest X-ray in the department was reduced from approximately 45 seconds to approximately 18 seconds (approximately 2 seconds for system processing and approximately 16 seconds for physician review), improving review efficiency by approximately 60%, and reducing missed diagnoses due to fatigue by approximately 72%. The above clinical application results fully demonstrate the practical value of the method of this invention as a pre-screening and triage tool in radiology departments, especially suitable for primary healthcare institutions with limited radiologist resources.

[0082] In summary, this invention constructs a complete intelligent screening solution for multiple lesions on chest X-rays by deeply coupling and collaborating five core steps: standardized preprocessing, dual-attention feature extraction, multi-label multi-task classification, lesion localization visualization, and pre-screening triage. This solution significantly outperforms existing technologies in terms of detection performance, inference efficiency, and clinical applicability, effectively alleviating the practical problems of high reading pressure and high missed diagnosis rates in radiology departments of primary healthcare institutions.

[0083] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent screening method for multiple lesions on chest X-ray, characterized in that, Includes the following steps: The input chest X-ray is subjected to standardized preprocessing, which includes contrast adaptive enhancement, automatic lung field segmentation and rib suppression processing to obtain a standardized lung field image that removes bone occlusion interference. The rib suppression processing uses the lung field mask obtained by the automatic lung field segmentation as a spatial constraint to guide the rib suppression network to perform bone component separation only within the lung field region. The standardized lung field image is input into a feature extraction network to extract multi-scale lesion features. The feature extraction network uses a dense connection network as the backbone structure and embeds a dual attention module consisting of a channel attention module and a spatial attention module after each dense connection block. The channel attention module compresses and excites the global information of each channel to generate a channel weight vector, and the spatial attention module identifies the spatial location of the lesion on the feature map after channel recalibration to generate a spatial weight map. The multi-scale lesion features are input into a multi-label classification network to obtain multi-lesion classification results. The multi-label classification network adopts a multi-task learning architecture to simultaneously detect seven types of abnormalities, including pulmonary nodules, pneumonia, pulmonary tuberculosis, pneumothorax, pleural effusion, cardiac enlargement, and mediastinal widening, and outputs a probability score and confidence level for each type of abnormality. Based on the multi-scale lesion features and the multi-lesion classification results, a heat map corresponding to each lesion type is generated using class activation map technology. The heat map is then superimposed on the standardized lung field image, and suspicious areas are highlighted in red to obtain the lesion localization visualization result. In this process, the lung field mask and the heat map are multiplied element-wise to force the activation values ​​outside the lung field region to be set to zero. The chest X-ray is pre-screened and triaged based on the probability score and the confidence level, and a structured screening report is generated.

2. The method according to claim 1, characterized in that, The contrast adaptive enhancement employs contrast-limited adaptive histogram equalization, where the region block size is 8x8 pixels and the contrast limiting coefficient ranges from 2.0 to 3.

5.

3. The method according to claim 1, characterized in that, The automatic lung field segmentation uses a pre-trained U-Net semantic segmentation network. The encoder of the U-Net semantic segmentation network contains four downsampling stages with 64, 128, 256 and 512 channels, respectively, and the segmentation accuracy, measured by the Dice coefficient, reaches over 0.

97.

4. The method according to claim 1, characterized in that, The rib suppression network employs a conditional generative adversarial network, using the stitching result of the enhanced grayscale image and the lung field mask as input conditions. The generator uses a ResNet architecture containing 9 residual blocks, and the discriminator uses a PatchGAN architecture with a receptive field of 70x70 pixels.

5. The method according to claim 1, characterized in that, The channel attention module obtains channel description vectors by performing global average pooling and global max pooling on the input feature map along the spatial dimension. These vectors are then processed by a multilayer perceptron with shared parameters and summed. Finally, a channel weight vector is generated by the sigmoid activation function. The compression ratio of the multilayer perceptron is 16.

6. The method according to claim 1, characterized in that, The spatial attention module performs average pooling and max pooling along the channel dimension on the feature map after channel recalibration to obtain two single-channel spatial description maps, which are then concatenated along the channel dimension and generated into a spatial weight map through a 7x7 convolutional layer and a Sigmoid activation function.

7. The method according to claim 1, characterized in that, The multi-label classification network employs label smoothing regularization and mixed sample data augmentation strategies to alleviate class imbalance. The label smoothing coefficient is 0.1, and the mixing coefficient in the mixed sample data augmentation strategy follows a Beta distribution with a shape parameter of 0.

4.

8. The method according to claim 1, characterized in that, The training loss function of the multi-label classification network adopts weighted binary cross-entropy loss, in which the weight coefficients of positive samples and negative samples of each type of lesion are calculated based on the number of samples of the corresponding category in the training set to compensate for class imbalance.

9. The method according to claim 1, characterized in that, The pre-screening triage divides the confidence level into three levels: high confidence positive, medium confidence doubtful, and low confidence negative, and assigns the chest X-ray to the emergency review queue, the routine review queue, and the no-review queue, respectively.

10. A deep learning-based intelligent screening system for multiple lesions on chest X-ray, used to implement the method described in any one of claims 1-9, characterized in that, include: The standardized preprocessing module is configured to perform standardized preprocessing on the input chest anteroposterior radiograph to obtain a standardized lung field image. The standardized preprocessing includes contrast adaptive enhancement, automatic lung field segmentation, and rib suppression processing, wherein the rib suppression processing uses the lung field mask as a spatial constraint. The dual attention feature extraction module is configured to input the standardized lung field image into the feature extraction network to extract multi-scale lesion features. The feature extraction network adopts a dense connection network as the backbone structure and embeds a dual attention module consisting of a channel attention module and a spatial attention module after each dense connection block. The multi-label classification module is configured to input the multi-scale lesion features into the multi-label classification network to obtain the multi-lesion classification result. The multi-label classification network adopts a multi-task learning architecture to simultaneously detect 7 types of chest abnormalities and output probability scores and confidence levels. The lesion localization module is configured to generate a heat map based on the multi-scale lesion features and the multi-lesion classification results using class activation graph technology, and then overlay it onto the standardized lung field image to obtain a lesion localization visualization result. The pre-screening and triage module is configured to perform pre-screening and triage based on the probability score and the confidence level, and generate a structured screening report.

Images