Chest radiograph image abnormal area detection method and program product

Abstract

The invention provides a chest radiograph abnormal region detection method and a program product, and the method comprises the steps: adding an FAHG module on the basis of YOLOv11, and carrying out the frequency domain enhancement and noise suppression of a feature map outputted to a detection head; in the FAHG module, Fourier transform is firstly carried out on the feature map; then low-pass, band-pass and high-pass filtering is executed respectively, and low-frequency, intermediate-frequency and high-frequency sub-band components are obtained through decomposition; enhancement processing is carried out on the three sub-band components, cross-channel information interaction and feature expression are enhanced for the low-frequency sub-band components and the intermediate-frequency sub-band components, and noise is suppressed for the high-frequency sub-band components; and after the enhanced low-frequency sub-band component, the enhanced intermediate-frequency sub-band component and the enhanced high-frequency sub-band component are respectively subjected to inverse Fourier transform, weighted fusion is carried out. Compared with other existing YOLO series models, the method for detecting the abnormal area of the chest radiograph image has high detection precision and generalization ability, the detection precision is remarkably improved under the condition that the calculated amount is not obviously increased, and good balance is achieved between efficiency and precision.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and computer-aided diagnosis technology, specifically relating to a method and program product for detecting abnormal areas in chest X-ray images. Background Technology

[0002] Chest X-ray is the most widely used clinical method for screening chest diseases. Currently, radiologists face a heavy workload in manually interpreting images, and are prone to missed or misdiagnosed cases due to fatigue and differences in experience. Computer vision-based automated detection technology can help alleviate this problem. However, the application of computer vision technology in chest X-ray lesion detection presents unique challenges: 1) Lesions vary greatly in size, ranging from tiny nodules to large-scale consolidations; 2) The contrast between lesions and normal tissue is often low, with blurred boundaries; 3) The image background is complex, severely interfered with by overlapping anatomical structures such as ribs and blood vessels; 4) Different types of lesions (such as nodules, exudates, and fibrosis) may appear simultaneously in the image and overlap with each other.

[0003] Current mainstream solutions are mostly based on convolutional neural networks (CNNs), especially the YOLO series of single-stage detectors. While these methods offer speed advantages, their feature extraction and fusion mechanisms are typically designed for general objects and fail to adequately adapt to the characteristics of chest X-ray images. Specifically, existing methods are largely limited to spatial domain operations, lacking the utilization of image frequency domain information to separate structure from noise; when fusing multi-scale features, the receptive field is fixed or singular, making it difficult to flexibly capture lesions with extremely large spans; when fusing features from different network levels, simple splicing or addition is often used, leading to the dilution of key features or the amplification of noise. Therefore, existing YOLO-based neural networks still exhibit insufficient accuracy and recall for detecting small-sized, low-contrast lesions in chest X-ray images. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and program for detecting abnormal regions in chest X-ray images. Based on improvements to YOLOv11, it enhances the detection accuracy of small-sized, low-contrast lesions in chest X-ray images.

[0005] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0006] A method for detecting abnormal regions in chest X-ray images, employing the following image detection model:

[0007] A new FAHG module has been added to YOLOv11 to enhance the frequency domain and suppress noise in the feature map output to the detection head;

[0008] The FAHG module includes the following processing:

[0009] S1, Perform a Fourier transform on the feature map;

[0010] S2, perform low-pass, band-pass, and high-pass filtering on the transformed features respectively, and decompose them into three sub-band components: low frequency, mid frequency, and high frequency.

[0011] S3 enhances the three sub-band components, specifically enhancing cross-channel information interaction and feature expression for the low-frequency and mid-frequency sub-band components, and suppressing noise for the high-frequency sub-band components.

[0012] S4 performs inverse Fourier transform on the enhanced low-frequency, mid-frequency, and high-frequency sub-band components respectively, and then performs weighted fusion.

[0013] Furthermore, in step S3, the low-frequency and mid-frequency sub-band components are enhanced by a channel aggregation module, wherein the processing procedure of the channel aggregation module is as follows:

[0014]

[0015] In the formula, and The input and output quantities of the respective modules. For weight parameters, This represents the GELU activation function. This represents a 1×1 convolution.

[0016] Furthermore, in S3, the high-frequency subband components are enhanced by a squeeze-excitation module, wherein the processing procedure of the squeeze-excitation module is as follows:

[0017]

[0018] In the formula, and The input and output quantities of the respective modules. This represents the Sigmoid activation function. It is a fully connected layer. Represents the ReLU activation function. This indicates pooling processing.

[0019] Furthermore, S1 is a Fast Fourier Transform.

[0020] Furthermore, the image detection model is improved by the following steps for the original C3k2 module:

[0021] When the C3k2 module is in the False state, replace the Bottleneck module in the original C3k2 module with the MSPLC module;

[0022] When the C3k2 module is in the True state, the C3k module in the original C3k2 module will be replaced with the C3_MSPLC module;

[0023] In the MSPLC module: For the input quantity, it is first processed by batch normalization and two layers of convolution, and then processed in parallel by three dilated convolutions with different kernel sizes. The results are then spliced ​​together and fused by 1×1 convolution.

[0024] In the C3_MSPLC module: for the input quantity, it is first divided into two parallel paths. One path is processed by a 1×1 convolution and then by a series of multi-layer MSPLC modules; the other path is processed by a 1×1 convolution only; finally, the results of the two paths are concatenated and then fused by a 1×1 convolution.

[0025] Furthermore, the processing procedure of the MSPLC module is as follows:

[0026]

[0027]

[0028] In the formula, and These represent the module's input and output values, respectively. Indicates feature splicing, , , These represent dilated convolutions with kernel sizes of 3×3, 5×5, and 7×7, respectively. This represents a 5×5 convolution. This indicates batch normalization.

[0029] Furthermore, the dilation rate of the dilated convolutions is 3.

[0030] Furthermore, in the image detection model, the following feature fusion module is used to replace the feature splicing operation in the original neck network:

[0031] First, stitch the components together according to the channel dimension;

[0032] The concatenated feature map is then multiplied element-wise with the weight vector.

[0033] Furthermore, the weight vector is normalized using Softmax.

[0034] A computer program product includes a computer program that, when executed by a processor, implements the above-described method for detecting abnormal regions in chest X-ray images.

[0035] The beneficial effects of this invention are as follows:

[0036] (1) This invention provides a method and program product for detecting abnormal regions in chest X-ray images. The method uses an image detection model based on YOLOv11 to detect chest X-ray images. Compared with other existing YOLO models, it has strong detection accuracy and generalization ability, and significantly improves detection accuracy without significantly increasing the amount of computation, achieving a good balance between efficiency and accuracy. On the test set, it achieved a detection accuracy of 48.2%, a recall rate of 42.5%, a mAP@0.5 of 41.5%, and a mAP@0.5:0.95 of 22.6%, which comprehensively surpasses YOLOv5, YOLOv8, YOLOv10 and the baseline model YOLOv11n.

[0037] (2) Based on YOLOv11, this invention adds a frequency adaptive hybrid gate module (FAHG) to the neck network, which is used to enhance the frequency domain and suppress noise of the three scale feature maps output to the detection head respectively, so that the improved model has a certain noise resistance and can still effectively detect most lesions in low to medium noise environments.

[0038] (3) This invention replaces the C3k2 module in the original YOLOv11 with C3_MSPLC2 based on multi-scale parallel large convolutional blocks, thereby greatly expanding the receptive field of the model and enhancing its long-range modeling capabilities. It deploys three dilated convolutions of different sizes to capture contextual information at different scales. The small kernel focuses on local details, while the large kernel captures a wider range of pathological features.

[0039] (4) The present invention uses a feature fusion module to replace the simple feature splicing operation in the original YOLOv11 neck network and implements cross-level feature fusion. In this invention, the feature map after splicing is adjusted by adding a learnable weight vector to achieve adaptive calibration of features in different channels, strengthen the features related to the lesion detection task, and weaken irrelevant or interfering features. Attached Figure Description

[0040] Figure 1 This is a structural diagram of the improved neural network model for chest radiograph detection based on YOLOv11n in this invention;

[0041] Figure 2 This is a diagram of the original YOLOv11n neural network model structure.

[0042] Figure 3 This is a diagram of the network structure of the Multi-Scale Parallel Large Convolutional Block (MSPLC) in this invention.

[0043] Figure 4 This is a network structure diagram of the Frequency Adaptive Hybrid Gate Module (FAHG) in this invention;

[0044] Figure 5This is a network structure diagram of the feature fusion module (FF) in this invention;

[0045] Figure 6 The curves showing the changes in the loss function and validation set evaluation metrics during the training process of the VinDr-CXR dataset are shown below.

[0046] Figure 7 This is an example of a detection comparison between the present invention and the original YOLOv11n on a test set;

[0047] Figure 8 This is an example of a detection comparison between the present invention and several existing neural network models;

[0048] Figure 9 This is a comparative example of detection under different intensities of Gaussian noise interference according to the present invention. Detailed Implementation

[0049] Embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein similar or identical reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0050] I. Technical Solution

[0051] This invention uses a neural network based on YOLOv11 as an image detection model for detecting abnormal areas in chest X-ray images, thereby assisting doctors in efficiently identifying the lesion sites in patients.

[0052] Figure 1 The diagram shows the improved network structure of this invention. Figure 2 The diagram shows the original YOLOv11n network structure; combined with... Figure 1 and Figure 2 The image detection model of this invention is an improvement on YOLOv11 with the following modifications:

[0053] Improvement 1: A new Frequency Adaptive Hybrid Gate (FAHG) module is added to perform frequency domain enhancement and noise suppression on the three scale feature maps output to the head in the neck network.

[0054] Improvement 2: Design C3_MSPLC2 based on multi-scale parallel large convolutional blocks (MSPLC) to replace the C3k2 module in the original YOLOv11, so as to significantly expand the receptive field of the model and enhance its long-range modeling capabilities.

[0055] Improvement 3: Design a feature fusion module (FF_Concat) to replace the simple feature concatenation operation (Concat) in the original YOLOv11 neck network, in order to implement cross-level feature fusion.

[0056] 1. Frequency Adaptive Hybrid Gate Module (FAHG)

[0057] like Figure 4 The FAHG module shown performs the following operations on the feature map input to this module:

[0058] S1, first use Fast Fourier Transform (FFT) to transform the input feature map from the spatial domain to the frequency domain;

[0059] S2 uses low-pass (LPF), band-pass (BPF), and high-pass (HPF) filters respectively to decompose the transformed frequency domain signal into three sub-band components: low frequency (LF), mid frequency (MF), and high frequency (HF). The low frequency component mainly contains global structural information such as organ contours, the mid frequency component contains detailed information such as tissue texture, and the high frequency component contains edges and noise.

[0060] S3.1, for the low-frequency subband component (LF) and the mid-frequency subband component (MF), the cross-channel information interaction and feature representation are enhanced through the channel aggregation module (CABlock), respectively, to obtain the enhanced low-frequency subband component (LF). * ) and intermediate frequency subband components (MF) * );

[0061] The processing procedure of the channel aggregation module (CA Block) is as follows:

[0062] For the input (i.e., LF or MF), it is first passed through a 1×1 convolution (with 1 channel) and then activated by GELU. The difference is then calculated with the original input. After that, the difference is multiplied by the learnable weight parameters and added to the original input to obtain the output.

[0063] The corresponding function expression is:

[0064]

[0065] In the formula, and These represent the input and output values ​​of the channel aggregation module, respectively. This is a learnable weight parameter used to adjust the weights of the feature map difference portion. This represents the GELU activation function. This represents a 1×1 convolution.

[0066] S3.2, for the high-frequency subband component (HF), the noise-dominant channel is adaptively suppressed and the effective edge feature channel is enhanced through the squeeze-excitation module (SE Block), thus obtaining the enhanced high-frequency subband component (HF). * );

[0067] The processing procedure of the squeeze-excitation module (SE Block) is as follows:

[0068] For the input (i.e., HF), it is first processed by pooling to a scale of 1×1, and then sequentially processed by fully connected (Linear), ReLU activation, fully connected, and Sigmoid activation. Finally, it is multiplied by the original input to obtain the output result.

[0069] The corresponding function expression is:

[0070]

[0071] In the formula, and These represent the input and output quantities of the extrusion-excitation module, respectively. This represents the Sigmoid activation function. It is a fully connected layer. Represents the ReLU activation function. This indicates pooling processing.

[0072] S4, for the enhanced low frequency (LF) * ), intermediate frequency (MF) * ), high frequency (HF) * The three sub-band components are subjected to inverse Fourier transforms to reconstruct the spatial domain features of each component. Finally, the feature maps of the three components are weighted and fused to output the enhanced feature map.

[0073] 2. Multi-scale parallel large convolutional blocks (MSPLC)

[0074] 2.1, combined with Figure 1 and Figure 2 As shown, in the image detection model of this invention, C3_MSPLC2 has two states: False and True. These states correspond one-to-one with the False and True states of the C3k2 module in the original YOLOv11n. That is, when the C3k2 module in the original YOLOv11n is False, C3_MSPLC2 in the improved image detection model of this invention is also False.

[0075] Among them, combined Figure 2 and Figure 3 As shown, the C3_MSPLC2 module is:

[0076] 1) In the False state, replace the Bottleneck module in the original C3k2 module with the MSPLC module newly designed in this invention;

[0077] 2) In the True state, replace the C3k module in the original C3k2 module with the newly designed C3_MSPLC module of this invention.

[0078] 2.2, such as Figure 3 The MSPLC module shown in Figure (a):

[0079] For the input, preliminary feature extraction is first performed through batch normalization and two convolutional layers (in this embodiment, the convolutional kernels are 1×1 and 5×5 respectively). Then, three dilated convolutions with different kernel sizes capture contextual information at different scales (in this embodiment, the convolutional kernel sizes are 3×3, 5×5, and 7×7, with an expansion rate of 3 for each). The smaller kernel focuses on local details, while the larger kernel captures a wider range of pathological patterns. Finally, the results from the three branches are concatenated along the channel dimension and then fused through a 1×1 convolution to make the number of channels consistent with the input, forming a multi-scale fused feature (i.e., the output result).

[0080] The corresponding function expression is:

[0081]

[0082]

[0083] In the formula, and These represent the input and output quantities of the MSPLC module, respectively. Indicates feature splicing, , , These represent dilated convolutions with kernel sizes of 3×3, 5×5, and 7×7, respectively. This represents a 5×5 convolution. This indicates batch normalization.

[0084] 2.3, such as Figure 3 The C3_MSPLC module shown in Figure (b):

[0085] For the input, it is first divided into two parallel paths. One path is first processed by a 1×1 convolution and then processed by a series of multiple (n-layer) MSPLC modules; the other path is processed only by a 1×1 convolution. Finally, the results of the two paths are concatenated and then fused by a 1×1 convolution to obtain the output result.

[0086] The corresponding function expression is:

[0087]

[0088]

[0089]

[0090] In the formula, and These represent the input and output quantities of the C3_MSPLC module, respectively. This indicates the MSPLC module. This indicates processing via n nested MSPLC modules.

[0091] 3. Feature Fusion Module (FF_Concat)

[0092] like Figure 5 The feature fusion module shown is for input scales H×W×C. l and H×W×C k Two feature maps:

[0093] S1, first splice along the channel dimension to obtain a scale of H×W×(C l +C k The splicing feature map;

[0094] S2 introduces a learnable channel weight vector [w1, w2, ..., w N The dimension of the vector is the same as the number of channels in the concatenated feature map (i.e., N = C). l +C k );

[0095] S3, after the channel weight vector is normalized by Softmax, it is multiplied element-wise with the spliced ​​feature map according to the channel dimension, thereby realizing adaptive recalibration of different channel features, strengthening features related to the lesion detection task, and weakening other irrelevant or interfering features.

[0096] II. Testing and Verification

[0097] 1. Data preparation and partitioning

[0098] The open-source dataset VinDr-CXR was used as the primary training and evaluation data. This dataset contains 15,000 chest X-ray images (DICOM format), annotated by professional radiologists with 14 categories of chest abnormalities, totaling 22,719 bounding boxes. The images were randomly partitioned in an 8:1:1 ratio into a training set (12,000 images), a validation set (2,250 images), and a test set (750 images). To alleviate class imbalance issues in the dataset (e.g., sparse samples for categories like "atelectasis" and "consolidation"), a weighted upsampling strategy was implemented.

[0099] 1) Count the number of occurrences of each anomaly category in the training set.

[0100] 2) According to the formula Calculate the weight for each category. ,in Let T be the number of occurrences of the i-th class, and T be the total number of classes; the fewer classes that appear, the higher their weight.

[0101] 3) For an image containing multiple bounding boxes, the sampling weight is the sum of the weights of the categories to which all bounding boxes in the image belong, and then normalized.

[0102] 4) When loading training data, sampling with replacement is performed based on the normalized weights of the images, which indirectly increases the frequency of rare class samples.

[0103] 2. Model Construction and Initialization

[0104] Based on the three improvements mentioned above, an image detection model is built using YOLOv11. All model weights are randomly initialized using the Kaiming normal distribution method, and the bias is initialized to 0. This embodiment does not use pre-trained weights from datasets such as ImageNet.

[0105] 3. Model Training

[0106] Training was conducted on a server equipped with an NVIDIA RTX 4080 GPU, using PyTorch 2.0.1 as the software environment. Key training hyperparameter settings are as follows:

[0107] 1) Optimizer: Stochastic gradient descent, initial learning rate = 0.01, momentum = 0.937, weight decay = 0.0005.

[0108] 2) Learning rate scheduling: A cosine annealing strategy with warm-up is adopted. The learning rate is warmed up in the first 3 epochs, increasing linearly from 0.0 to 0.01.

[0109] 3) Batch size: 16.

[0110] 4) Training cycle: 300.

[0111] 5) Loss function: YOLOv11n composite loss is adopted, including bounding box regression loss (CIoU Loss, weight 7.5), classification loss (BCE Loss, weight 0.5) and distribution focus loss (DFL Loss, weight 1.5).

[0112] 6) Data augmentation: Mosaic and MixUp augmentations, as well as random rotation, scaling, and color jitter, are used to improve the model's generalization ability.

[0113] 7) Monitor the training process, including training loss and validation set mAP@0.5. For example... Figure 6 As shown, the model converges well after about 250 epochs.

[0114] 4. Model Testing and Performance Analysis

[0115] The trained final model was evaluated on an independent test set. Key evaluation metrics included precision (P), recall (R), F1 score, mean average precision (mAP@0.5 and map@0.5:0.95), number of parameters, and floating-point operations per second (FLOPs). The improved image detection model of this invention was compared with current mainstream object detection models, and the results are shown in Table 1; the first four rows represent existing YOLO-based neural network models, rows 5-10 represent models based on YOLOv11n but with only some improvements from this invention, and row 11 represents models with all three improvements.

[0116] Table 1: Performance Comparison of Models on the VinDr-CXR Dataset

[0117]

[0118] Experimental data show that, compared with the baseline model YOLOv11n, the improved model of this invention improves accuracy (P) by 14.2%, recall (R) by 9.8%, mAP@0.5 by 7.0%, and mAP@0.5:0.95 by 5.6%, which strongly verifies the performance breakthrough brought about by the module innovation.

[0119] A selection of test examples from the above test data are provided for visual demonstration, including:

[0120] Figure 7 The image detection model of this invention is shown as a comparison example with the original benchmark model YOLOv11n. Column (a) shows three original images with labels, and the lesion sites are marked with pre-made boxes of different colors. Column (b) shows the corresponding detection results of YOLOv11n. Column (c) shows three sets of corresponding detection results of this invention. In the first row of images, both this invention and YOLOv11n successfully detected all three lesion areas. In the second row of images, YOLOv11n had 4 missed detections, while this invention had 2 missed detections. In the third row of images, YOLOv11n had more missed detections and misidentifications compared to this invention.

[0121] Figure 8The following are comparative examples of the present invention with several other existing YOLO-based models: (a) the original image with labels, (b) the detection results of YOLOv5nu, (c) the detection results of YOLOv8n, (d) the detection results of YOLOv10n, (e) the detection results of YOLOv11n, and (f) the detection results of the present invention. The above comparison clearly shows that the detection results of the improved model of the present invention are superior to all other existing YOLO-based models.

[0122] 5. Noise robustness test

[0123] To comprehensively evaluate the performance of the image detection model of this invention on high-noise chest X-ray images, randomly selected... Figure 9 The chest X-ray shown in (a) was used for testing. The image contains four nodular masses, one aortic enlargement, and one pulmonary fibrosis lesion. The test results were compared in three scenarios by adding low-intensity and medium-intensity Gaussian noise to simulate interference: no additional noise, low-intensity noise, and high-intensity noise.

[0124] like Figure 9 As shown in (b), the present invention correctly detected three targets (including a small target) when no additional noise was added. Figure 9 (c) After adding low-intensity Gaussian noise, the model can still detect and correctly identify four targets, but the confidence of most detection results decreases. Figure 9 The nodular masses that were missed in both (b) and 9(c) were located in the upper right region of the chest radiograph, and their indistinct intensity boundaries made them difficult to identify. Figure 9 (d) When stronger Gaussian noise is introduced, the detection capability of the present invention is further reduced.

[0125] Experimental results show that the improved model of this invention has a certain noise resistance capability and can still effectively detect most lesions in low to medium noise environments, which confirms the effectiveness of the FAHG module in noise suppression.

[0126] 6. Ablation experiment

[0127] To quantitatively evaluate the individual contributions of the three core modules—FAHG, MSPLC, and FF—and their impact on model inference efficiency, ablation experiments were designed and executed. Model variants containing only a single module, pairwise combinations, and all modules were trained under identical training settings. Performance comparisons are shown in Table 1, and the comparison of inference efficiency versus resource consumption is shown in Table 2.

[0128] Table 2: Comparison of inference efficiency and resource consumption after the introduction of different modules

[0129]

[0130] As shown in Table 1: ① The introduction of any module can bring performance gains; ② The FF module contributes a significant improvement in accuracy at the cost of a very small number of parameters; ③ When all three modules exist simultaneously, the model reaches peak performance, verifying the synergy and completeness of the module design in the technical solution of this invention.

[0131] As shown in Table 2, the baseline model YOLOv11n achieves an inference time of 7.0 milliseconds (equivalent to 143 FPS) per image on current hardware. The three newly introduced modules collectively introduce an additional 4.8 milliseconds of overhead: the MSPLC module, due to its parallel large-kernel dilated convolution structure, is the most computationally intensive component, adding 2.9 milliseconds of latency; the FAHG module, although involving frequency domain operations, only adds 0.9 milliseconds of inference time due to the highly optimized FFT implementation on modern GPUs; and the FF module is the most lightweight, adding only 0.8 milliseconds.

[0132] In terms of memory consumption, the fully improved model requires 950.8 MB of GPU memory during inference, which can be easily handled by most modern GPUs. Compared to the baseline model YOLOv11n, this model achieves a performance breakthrough with a 14.2% improvement in accuracy, a 9.8% improvement in recall, a 7.0% improvement in mAP@0.5, a 5.6% improvement in mAP@0.5:0.95, and an 11.8% improvement in F1 score—a trade-off considered acceptable in the field of medical image diagnosis where accuracy is paramount. Furthermore, the overall inference speed of 12.8 milliseconds (78 FPS) still far exceeds the real-time requirements of clinical deployment.

[0133] 7. External dataset validation

[0134] To further evaluate the model's generalization ability, external validation was performed using the independently sourced SZ-HospCXR dataset. This dataset differs from VinDr-CXR in terms of patient population, imaging equipment, and annotation specifications. First, its annotations were mapped to the 10 common anomaly categories supported by this model, and then testing was conducted. The results are shown in Table 3.

[0135] Table 3: Performance comparison of common anomalies detection on internal and external validation datasets.

[0136]

[0137] As shown in Table 3, the overall performance of this invention on this external dataset (mAP@0.5: 28.9%) is still significantly better than the baseline model YOLOv11n (26.2%) and other comparative models, which fully demonstrates that the method proposed in this invention has strong versatility and clinical application potential.

[0138] III. Program Products

[0139] This application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the chest radiograph abnormality region detection method of the present invention.

[0140] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0141] This invention is not limited to the embodiments described above. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of this invention are within the scope of protection of this invention.

Claims

1. A method for detecting abnormal regions in chest X-ray images, characterized in that: The following image detection model is used: A new FAHG module has been added to YOLOv11 to enhance the frequency domain and suppress noise in the feature map output to the detection head; The FAHG module includes the following processing: S1, Perform a Fourier transform on the feature map; S2, perform low-pass, band-pass, and high-pass filtering on the transformed features respectively, and decompose them into three sub-band components: low frequency, mid frequency, and high frequency. S3 enhances the three sub-band components, specifically enhancing cross-channel information interaction and feature expression for the low-frequency and mid-frequency sub-band components, and suppressing noise for the high-frequency sub-band components. S4 performs inverse Fourier transform on the enhanced low-frequency, mid-frequency, and high-frequency sub-band components respectively, and then performs weighted fusion.

2. The method for detecting abnormal regions in chest X-ray images according to claim 1, characterized in that: In step S3, the low-frequency and mid-frequency sub-band components are enhanced by a channel aggregation module, wherein the processing procedure of the channel aggregation module is as follows: In the formula, and The input and output quantities of the respective modules. For weight parameters, This represents the GELU activation function. This represents a 1×1 convolution.

3. The method for detecting abnormal regions in chest X-ray images according to claim 1, characterized in that: In step S3, the high-frequency subband components are enhanced by a squeezing-excitation module, wherein the processing procedure of the squeezing-excitation module is as follows: In the formula, and The input and output quantities of the respective modules. This represents the Sigmoid activation function. It is a fully connected layer. Represents the ReLU activation function. This indicates pooling processing.

4. The method for detecting abnormal regions in chest X-ray images according to claim 1, characterized in that: S1 is the Fast Fourier Transform.

5. The method for detecting abnormal regions in chest X-ray images according to claim 1, characterized in that: In the image detection model, the original C3k2 module is improved as follows: When the C3k2 module is in the False state, replace the Bottleneck module in the original C3k2 module with the MSPLC module; When the C3k2 module is in the True state, the C3k module in the original C3k2 module will be replaced with the C3_MSPLC module; In the MSPLC module: For the input quantity, it is first processed by batch normalization and two layers of convolution, and then processed in parallel by three dilated convolutions with different kernel sizes. The results are then spliced ​​together and fused by 1×1 convolution. In the C3_MSPLC module: for the input quantity, it is first divided into two parallel paths. One path is processed by a 1×1 convolution and then by a series of multi-layer MSPLC modules; the other path is processed by a 1×1 convolution only; finally, the results of the two paths are concatenated and then fused by a 1×1 convolution.

6. The method for detecting abnormal regions in chest X-ray images according to claim 5, characterized in that: The processing procedure of the MSPLC module is as follows: In the formula, and These represent the module's input and output values, respectively. Indicates feature splicing, , , These represent dilated convolutions with kernel sizes of 3×3, 5×5, and 7×7, respectively. This represents a 5×5 convolution. This indicates batch normalization.

7. The method for detecting abnormal regions in chest X-ray images according to claim 6, characterized in that: The dilation rate of the dilated convolutions is 3.

8. The method for detecting abnormal regions in chest X-ray images according to claim 1, characterized in that: In the image detection model, the following feature fusion module is used to replace the feature concatenation operation in the original neck network: First, stitch the components together according to the channel dimension; The concatenated feature map is then multiplied element-wise with the weight vector.

9. The method for detecting abnormal regions in chest X-ray images according to claim 8, characterized in that: The weight vector is normalized using Softmax.

10. A computer program product, characterized in that: The method includes a computer program that, when executed by a processor, implements the chest radiograph abnormality region detection method as described in any one of claims 1 to 9.

Images