Construction method, classification method and system of interpretable chest x-ray classification model

Abstract

The application discloses an interpretable X-ray film classification method and system based on image decoupling and double-flow evidence learning. The method comprises the following steps: acquiring chest X-ray film training images; decomposing the training images into low-frequency structural information flow and high-frequency texture information flow by using discrete wavelet transform; inputting the two deep learning models respectively to generate structural / texture positive evidence and negative evidence of each disease; constructing Beta distribution to obtain positive probability and confidence degree based on each information flow; inputting the spliced result into a classification multi-layer perception machine to generate a prediction label and a confidence degree; and jointly training the model by minimizing the combination of binary cross-entropy loss and variation loss. During classification, the trained model is used to perform the same processing on the image to be classified, and the final classification label and confidence degree of each disease are output. The application realizes internal interpretable chest X-ray film classification, can quantify cognitive uncertainty, avoids shortcut learning, and does not depend on expert knowledge.

Description

Technical Field

[0001] This invention relates to a method for constructing a classification model that can interpret chest X-rays, and also to a classification method for classifying chest X-rays using the classification model, belonging to the field of medical image processing technology. Background Technology

[0002] Chest X-rays (CXR) are one of the primary imaging methods for assessing lung diseases due to their low cost and widespread availability. However, the accurate interpretation and diagnosis of CXR images remains challenging for radiologists due to tissue overlap and the highly similar visual features between different diseases. CXR image classification technology aims to assist physicians in achieving faster, more objective, and accurate diagnoses by extracting key image features and constructing mapping models.

[0003] Benefiting from the release of large-scale public datasets, deep learning models with powerful representation learning capabilities have demonstrated performance comparable to or even surpassing that of human experts in tasks such as pneumonia and tuberculosis. Nevertheless, deep learning still faces significant interpretability challenges in clinical applications. First, models with massive parameters exhibit a typical "black box" structure, lacking intuitive understanding of their decision-making mechanisms and deviating from the fundamental principles of evidence-based medicine. Second, such models are prone to falling into the "shortcut learning" trap, tending to use non-pathological artifacts such as chest tubes and equipment markings instead of true pathological features for prediction, resulting in insufficient robustness. Finally, deep learning models often lack awareness of predictive uncertainty, easily exhibiting "overconfidence" when handling difficult samples, outputting erroneous conclusions with extremely high confidence. These shortcomings collectively limit the credibility and practical application of deep learning technology in medical scenarios.

[0004] Therefore, there is an urgent need for a chest X-ray classification method that does not rely on expert knowledge, is inherently interpretable, and can supervise the evidence collection process. Summary of the Invention

[0005] The primary technical problem to be solved by this invention is to provide a method for constructing a classification model that can interpret chest X-rays.

[0006] Another technical problem to be solved by the present invention is to provide a method for classifying interpretable chest X-rays.

[0007] Another technical problem to be solved by the present invention is to provide a classification system that can interpret chest X-rays.

[0008] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: According to a first aspect of the present invention, a method for constructing an interpretable chest X-ray classification model is provided, comprising the following steps: Acquire training images from multiple historical chest X-rays and their disease labels; The training image is decomposed into a low-frequency structural information stream representing the overall structure and a high-frequency texture information stream representing texture details using discrete wavelet transform. The low-frequency structural information stream is input into a first deep learning model to be trained to generate positive and negative structural evidence for each disease; the high-frequency texture information stream is input into a second deep learning model to be trained to generate positive and negative texture evidence for each disease. Based on the structural positive evidence and structural negative evidence, a corresponding first Beta distribution is constructed for each disease to obtain the first positive probability and first confidence level of the disease based on the structural information flow; based on the texture positive evidence and texture negative evidence, a corresponding second Beta distribution is constructed for each disease to obtain the second positive probability and second confidence level of the disease based on the texture information flow. The first positive probability, first confidence level, second positive probability, and second confidence level of each disease are concatenated into vectors and then input into a classification multilayer perceptron to generate the predicted classification label and predicted confidence level of the disease. By minimizing the total loss function, which is a weighted combination of binary cross-entropy loss and variational loss, the first deep learning model, the second deep learning model, and the classification multilayer perceptron are jointly trained to obtain a trained classification model.

[0009] Preferably, the total loss function The simplified form is: ; in, For binary cross-entropy loss, For variational loss, Here are the non-negative hyperparameters used to balance the two types of losses; The variational loss is derived by minimizing the negative value of the modified lower bound of evidence, which is obtained by replacing the posterior distribution with an uninformative prior, and is used to constrain the statistical reasonableness of the evidence generated by the model.

[0010] According to a second aspect of the present invention, a method for classifying interpretable chest X-rays is provided, which classifies chest X-rays to be classified using a classification model constructed by the above method, comprising the following steps: Obtain the original chest X-ray image to be classified; The original image is decomposed into a low-frequency structural information stream representing the overall structure and a high-frequency texture information stream representing texture details using discrete wavelet transform. The low-frequency structural information stream is input into the first deep learning model in the classification model to generate positive and negative structural evidence for each disease; the high-frequency texture information stream is input into the second deep learning model in the classification model to generate positive and negative texture evidence for each disease. Based on the structural positive evidence and structural negative evidence, a corresponding first Beta distribution is constructed for each disease to obtain the first positive probability and first confidence level of the disease based on the structural information flow; based on the texture positive evidence and texture negative evidence, a corresponding second Beta distribution is constructed for each disease to obtain the second positive probability and second confidence level of the disease based on the texture information flow. The first positive probability, first confidence level, second positive probability, and second confidence level of each disease are concatenated into vectors and then input into the classification multilayer perceptron in the classification model to generate the final classification label and final confidence level of the disease. Output interpretable classification results, which include: the final classification label for each disease and its corresponding final confidence score.

[0011] Preferably, the process of decomposition using discrete wavelet transform includes: The training image or the original image is decomposed into low-frequency components, horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components. The diagonal high-frequency components are filtered out to perform data denoising; The low-frequency component is used as the low-frequency structural information stream, and the horizontal high-frequency component and the vertical high-frequency component are used together as the high-frequency texture information stream.

[0012] Preferably, the first deep learning model and the second deep learning model are convolutional neural networks with the same or different structures, and the outputs of the first deep learning model and the second deep learning model are respectively connected to an evidence extraction multilayer perceptron with ELU as the activation function, for mapping the extracted features into scalar evidence greater than 1.

[0013] Preferably, the first Beta distribution is based on the structural positive evidence α for each disease. sk Negative evidence of structure β sk Parameterization, the first positive probability is y sk =α sk / (α sk +β sk The first confidence level is b. sk =(α sk +β sk -2) / (α sk +β sk ); The second Beta distribution is based on the positive evidence α of the texture of each disease. tk Negative evidence of texture β tk Parameterization, the second positive probability is y tk =α tk / (α tk +β tk The second confidence level is b. tk =(α tk +β tk -2) / (α tk +β tk ); Where, α sk β represents positive structural evidence for the k-th disease. sk α represents the structural negative evidence for the k-th disease. tk β represents positive evidence of texture for the k-th disease. tk The texture represents negative evidence for the k-th disease, k∈(1,K), where K is the total number of disease categories included in the training image or the original image; The first confidence level and the second confidence level are used to characterize the cognitive uncertainty of the first deep learning model and the second deep learning model, respectively. The lower the confidence level, the higher the cognitive uncertainty.

[0014] Preferably, the classification method further includes a consistency verification step, specifically including: Set a threshold, compare the first positive probability with the threshold to generate intermediate prediction labels for the structure flow, and compare the second positive probability with the threshold to generate intermediate prediction labels for the texture flow; Calculate the first Cohen's Kappa coefficient between the intermediate predicted label of the structure stream and the final classification label, and the second Cohen's Kappa coefficient between the intermediate predicted label of the texture stream and the final classification label; The first Cohen's Kappa coefficient is used to quantify the consistency between the evidence collected by the low-frequency structural information stream and the final classification result, while the second Cohen's Kappa coefficient is used to quantify the consistency between the evidence collected by the high-frequency texture information stream and the final classification result. The closer the coefficient value is to 1, the more reliable the evidence of the corresponding information stream.

[0015] According to a third aspect of the present invention, a classification system for interpreting chest X-rays is provided for performing the above-described classification method, comprising: The image decomposition module is used to acquire the original chest X-ray image and decompose it into a low-frequency structural information stream that represents the overall structure and a high-frequency texture information stream that represents the texture details using discrete wavelet transform. A dual-information-flow evidence extraction module is used to input the low-frequency structural information flow into a first deep learning model to generate positive structural evidence and negative structural evidence, and to input the high-frequency texture information flow into a second deep learning model to generate positive texture evidence and negative texture evidence; wherein, the output terminals of the first deep learning model and the second deep learning model are respectively connected to an evidence extraction multilayer perceptron with ELU as the activation function; The probability synthesis module is used to construct a first Beta distribution based on the structural positive evidence and structural negative evidence for each disease to obtain a first positive probability and a first confidence level for the disease; and to construct a second Beta distribution based on the texture positive evidence and texture negative evidence for each disease to obtain a second positive probability and a second confidence level for the disease. The classification module is used to concatenate the first positive probability, the first confidence level, the second positive probability, and the second confidence level into a classification multilayer perceptron, generate the final classification label and final confidence level for each disease, and output an interpretable classification result. The interpretable classification results include the final classification label for each disease and its corresponding final confidence level.

[0016] Among the more favorable locations are: The consistency verification module is used to set a threshold, compare the first positive probability with the threshold to generate a structural flow intermediate prediction label, compare the second positive probability with the threshold to generate a texture flow intermediate prediction label, and calculate the first Cohen's Kappa coefficient between the structural flow intermediate prediction label and the final classification label, and the second Cohen's Kappa coefficient between the texture flow intermediate prediction label and the final classification label. The first Cohen's Kappa coefficient is used to quantify the consistency between the evidence collected by the low-frequency structural information stream and the final classification result, while the second Cohen's Kappa coefficient is used to quantify the consistency between the evidence collected by the high-frequency texture information stream and the final classification result. The closer the coefficient value is to 1, the more reliable the evidence of the corresponding information stream.

[0017] Compared with the prior art, the present invention has the following technical effects: (1) Internal interpretability This invention decouples chest X-ray images into low-frequency structural information streams and high-frequency texture information streams, and independently constructs positive / negative evidence and confidence levels for each information stream. The model can transparently display its decision-making basis—which structural features support the disease, which texture features oppose the disease, and the confidence level of each information stream, making the diagnostic process traceable and understandable.

[0018] (2) Statistically supervised evidence This invention introduces variational loss to impose statistical rationality constraints on the positive / negative evidence generated by the model, enabling the model to learn how to collect evidence reasonably, rather than just learning classification boundaries, effectively avoiding shortcut learning and artifact interference.

[0019] (3) Quantification of cognitive uncertainty By constructing a Beta distribution, this invention can quantify the model's cognitive uncertainty (i.e., confidence level) for each prediction. When evidence is insufficient, the confidence level naturally decreases, avoiding the model's "overconfidence" in outputting erroneous conclusions, which is more in line with clinical decision-making needs.

[0020] (4) It does not rely on expert knowledge and has good scalability and stability. Unlike concept bottleneck models that require expert-defined concepts, this invention learns entirely from data, eliminating the need for manual annotation of intermediate concepts and exhibiting excellent scalability and versatility. Furthermore, experiments have demonstrated that the dual-stream evidence framework provided by this invention helps the model converge faster and significantly improves training stability. Attached Figure Description

[0021] Figure 1 The overall flowchart of a method for constructing an interpretable chest X-ray classification model provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the decomposition result of a CXR image in the TBX11K dataset in the first embodiment of the present invention; Figure 3 This is a schematic diagram of the probabilistic graphical model in the first embodiment of the present invention; Figure 4 This is a schematic diagram of a dual-information-flow evidence modeling framework based on discrete wavelet transform in the first embodiment of the present invention. Figure 5 A flowchart illustrating an interpretable chest X-ray classification method provided in the second embodiment of the present invention; Figure 6 A schematic diagram of a classification system for interpretable chest X-rays provided in the third embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device provided in the fourth embodiment of the present invention. Detailed Implementation

[0022] The technical content of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0023] The core idea of ​​this invention is to mimic the diagnostic logic of radiologists by using Discrete Wavelet Transform (DWT) to decouple chest X-ray images into low-frequency structural information and high-frequency texture information, while filtering out diagonal high-frequency components that represent noise. Then, independent deep learning evidence-gathering networks are established for the structural and texture streams, respectively. Each network outputs positive evidence (strength of evidence supporting the disease) and negative evidence (strength of evidence opposing the disease) for each disease. Each evidence stream is parameterized as a Beta distribution to obtain the positive probability and confidence level based on that information stream. Finally, the probabilities and confidence levels obtained from both streams are concatenated and fed into a classification multilayer perceptron to obtain the final classification label and final confidence level for each disease.

[0024] In the training process of the classification model, in addition to the conventional binary cross-entropy loss, a variational loss derived from variational Bayes is introduced. This loss constrains the statistical rationality of the evidence generated by the model by minimizing the negative value of the modified lower bound of evidence, ensuring that the evidence collection process itself is interpretable and conforms to the theory of probabilistic graphical models.

[0025] The present invention will be described in detail below, from the specific construction method of the classification model, the classification method based on the classification model, and the classification system.

[0026] First Embodiment like Figure 1 As shown, the first embodiment of the present invention provides a method for constructing an interpretable chest X-ray classification model, which specifically includes the following steps: S1: Obtain training images.

[0027] First, training images and their disease labels from multiple historical chest X-ray films are obtained from a historical database for model training.

[0028] In one embodiment of the present invention, the TBX11K dataset is used, which contains 11,200 CXR images and is divided into three categories: healthy, non-tuberculous, and tuberculous.

[0029] S2: Image decomposition based on discrete wavelet transform.

[0030] In this embodiment, the two-dimensional discrete wavelet transform (DWT) is used to transform each training image. Decomposed into four components: low-frequency information Horizontal high-frequency information Vertical high-frequency information and diagonal high-frequency information The specific form is as follows:

[0031] Among them, training images yes The matrix, c is the training image The dimension of the matrix. and It is an orthogonal wavelet low-pass filter and high-pass filter The resulting circular matrix has a size of [missing information]. The specific form is as follows:

[0032] This represents the filter factor of the low-pass filter. This represents the filter factor of the high-pass filter, and m represents the matrix element index. For a set of integers, ={-1,0,1}.

[0033] A training video After DWT decomposition, low-frequency information It reflects the overall structure of the image and high-frequency information. and It reflects the texture details and high-frequency information of the image. This represents noise in the image. This embodiment uses DWT as the image decomposition module and selects Haar as the wavelet function family for the DWT module, with its low-pass filter being... The high-pass filter is When CXR images After the technical framework is input, it is first decomposed by DWT into , , and , It reflects the overall structural information of the image (such as organ outlines, heart shadow size, etc.). and The components reflect the texture details of the image (such as lung texture, nodule edges, etc.). The component primarily represents high-frequency noise. Therefore, this embodiment filters out... Components are used to achieve data denoising, and As a low-frequency structural information flow, As a high-frequency texture information stream.

[0034] Figure 2The image shown is the decomposition result of a CXR image from the TBX11K dataset in this embodiment. In CXR images, the diagonal high-frequency components mainly originate from electronic noise, scattered radiation, and quantization noise during film digitization in the imaging system. These noise components lack any structured information helpful for disease diagnosis. This embodiment utilizes the mathematical properties of DWT to concentrate and filter out multidirectional noise in the diagonal high-frequency sub-band. This removes noise interference while fully preserving the low-frequency sub-band carrying anatomical structural information and the horizontal and vertical high-frequency sub-bands carrying texture details. Compared to traditional denoising methods such as Gaussian filtering, this DWT-based denoising strategy does not cause image edge blurring and can more accurately protect the subtle texture features crucial for tasks such as lung cancer detection.

[0035] Furthermore, because the Discrete Wavelet Transform (DWT) performs interval sampling on the input image during decomposition, the spatial resolution of its output low-frequency components and each high-frequency component is reduced to half that of the original image. This means that in the technical framework provided by this invention, the actual input feature map size processed by the subsequent first and second deep learning models is only one-quarter of the original input image (height and width are halved). Compared to directly extracting features from the original high-resolution image, this built-in downsampling mechanism significantly reduces the computational load and memory usage of the convolutional neural network. Thanks to this design, under the same hardware conditions, enabling the DWT module can double the training batch size (e.g., from 32 to 64), thereby accelerating model training while maintaining or even improving the stability of the training process. Therefore, image decomposition based on Discrete Wavelet Transform not only provides structural / texture decoupling that conforms to diagnostic logic for dual-information-flow evidence modeling, but also significantly reduces the computational complexity of the entire technical framework through spatial downsampling, which is beneficial for deployment and application in clinical environments with limited computing resources.

[0036] S3: Dual-information-flow evidence modeling.

[0037] like Figure 3 As shown, the low-frequency structural information stream S from step S2 is input to the first deep learning model (denoted as Nets), and the high-frequency texture information stream T is input to the second deep learning model (denoted as Nett). Nets and Nett can be convolutional neural networks (CNNs) with the same or different structures, such as DenseNet121, ResNet50, or MobileNetV2. Furthermore, the output of each model is connected to an evidence extraction multilayer perceptron (MLP) with ELU (Exponential Linear Unit) as the activation function. This MLP maps the features extracted by the CNN into scalar evidence greater than 1.

[0038] Specifically, for the k-th disease (k = 1, ..., K, where K is the total number of disease categories), the above low-frequency structural information is... and high-frequency texture information The features are fed into two separate deep learning models (e.g., CNN) for feature extraction, resulting in feature maps containing all disease evidence features with dimensions of [missing value]. d's evidence vector Subsequently, the above evidence vector is input into a multilayer perceptron with an Exponential Linear Unit (ELU) as the activation function, which outputs scalar evidence α greater than 1. sk β sk α tk and β tk The specific form of the ELU activation function used is as follows: .

[0039] Where n is the input of the ELU activation function, which can be any real number. In this embodiment, n is specifically the output of the evidence extraction MLP.

[0040] Ultimately, positive evidence α for the Nets output structure sk Negative evidence of structure β sk Nett output texture positive evidence α tk Negative evidence of texture β tk These pieces of evidence are all positive real numbers greater than 1, satisfying the definition of the subsequent Beta distribution.

[0041] It is worth noting that the technical framework provided by this invention has a non-invasive characteristic: the first and second deep learning models can be any existing convolutional neural network (such as DenseNet121, ResNet50, or MobileNetV2). Only the final classification layer needs to be replaced with an evidence vector extraction layer to complete the upgrade from black-box classification to interpretable evidence modeling, without any modification to the internal structure of the backbone network. This characteristic enables this invention to endow existing medical imaging diagnostic systems with inherent interpretability at extremely low modification costs.

[0042] S4: Construct a Beta distribution and calculate the positive probability and confidence level.

[0043] In this embodiment, for each disease k, based on structural evidence (α) sk , β sk Construct the first Beta distribution to obtain the first positive probability y. sk And the first confidence level b sk Furthermore, based on texture evidence (α)tk , β tk Construct a second Beta distribution to obtain the second positive probability y. tk Second confidence level b tk .

[0044] Specifically, let For the true labels of K diseases, for the label of the kth disease , Assume it follows a binomial distribution, i.e. ,in, It represents the probability of the k-th disease. To further model... Due to the uncertainty, this embodiment assumes a probability parameter Let be a continuous random variable in the range [0,1], and satisfy the following conditions: Among them, parameters and Considered as structural information Based on the positive and negative evidence for the k-th disease, therefore, It can be used to measure based on structural information The first positive probability of diagnosing the k-th disease. It can be used to measure the first confidence level of a diagnosis.

[0045] Similarly, parameters and Considered as Based on the positive and negative evidence for the k-th disease, similarly... and Able to measure based on texture information The second positive probability and second confidence level for diagnosing the kth disease.

[0046] In this embodiment, the values ​​of the first confidence level and the second confidence level are both in the range [0,1). or When the first or second confidence level is 0, it indicates complete uncertainty; when →∞ or As the confidence level approaches infinity, either the first or second confidence level approaches 1, indicating a high degree of certainty. This embodiment uses a subtraction of 2 to ensure that... or A confidence level of 0 is more intuitive.

[0047] S5: Evidence fusion and model prediction.

[0048] like Figure 4 As shown, based on the calculation results of step S4 above, the four vectors (y) for each disease are... sk , b sk , ytk ,b tk The vectors are concatenated to obtain a fused vector of dimension 4K. This fused vector is then input into a classification multilayer perceptron (called a fusion MLP), which outputs the predicted classification label for the k-th disease. and prediction confidence level b k .

[0049] Among them, predicted labels A value of 0 or 1 indicates that the model determines whether the patient has the disease; the prediction confidence level b k A real number between [0,1] represents the overall confidence level of the model in the prediction.

[0050] S6: Joint training and model parameter tuning.

[0051] In this embodiment, a weighted combination of binary cross-entropy loss and variational loss is used as the total loss function to optimize the first deep learning network Net. s The second deep learning network, Net t And by integrating MLP for joint training, an interpretable classification model can be obtained.

[0052] Specifically, in order to estimate the probability of disease Furthermore, to supervise the construction of the evidence vector, this embodiment employs the Variational Bayesian (VB) method, aiming to utilize tractable variational distributions. To approximate the true posterior distribution .

[0053] In this embodiment, following the principle of variational inference, the marginal log-likelihood of the observation label is used. Its lower bound can be determined by the Evidence Lower Bound (ELBO). Therefore, by replacing the tractable posterior distribution with an uninformative prior... Thus, the revised ELBO was derived:

[0054] in, The inequality is the balance coefficient. By minimizing the negative value of the corrected ELBO, this embodiment derives the variational loss of the provided technical framework.

[0055] in, It is a double gamma function. , Representation function In this embodiment, the input is... Specifically, α sk β sk α tk and β tk Any one of them.

[0056] Therefore, the total loss function Set as a weighted combination of binary cross-entropy (BCE) loss and variational loss:

[0057] Among them, hyperparameters It is a non-negative real scalar used to balance the BCE loss and variational loss. This total loss function... The formula can be simplified as follows: ;in, For binary cross-entropy loss, This is the variational loss. The first term here ensures the accuracy of the diagnosis; while the second term plays a supervisory role in evidence construction, constraining the positive and negative evidence generated by the model to meet the statistical consistency required by variational inference, thereby supervising the rationality of the evidence construction process.

[0058] The underlying mechanism of introducing variational loss here lies in the fact that traditional binary cross-entropy loss only focuses on the accuracy of the classification boundary. The model may minimize empirical risk by capturing spurious correlations in the training data (such as non-pathological artifacts like chest tubes and equipment labels), resulting in the so-called shortcut learning phenomenon. Variational loss, however, imposes explicit distributional constraints on the evidence components, forcing the model to collect statistically consistent positive and negative evidence for each prediction. When the model attempts to classify based on artifact features, the distribution of its generated evidence will deviate from the theoretical form established by variational inference, thus being penalized by the variational loss term. This mechanism fundamentally inhibits shortcut learning behavior, guiding the model to focus on true pathological features.

[0059] In a preferred embodiment of the present invention, the Adam optimizer is used during training, and the weight decay is set to 10. -4 The gradient clipping maximum norm is 1.0, and the learning rate is adjusted using a cosine annealing strategy. The model is trained for 200 epochs on the TBX11K dataset, resulting in an interpretable classification model.

[0060] Second Embodiment like Figure 5 As shown, based on the first embodiment described above, the second embodiment of the present invention further provides a method for classifying interpretable chest X-rays, used to classify a new chest X-ray using the trained classification model described above, specifically including the following steps: S10: Obtain the image to be classified.

[0061] Obtain the raw chest X-ray image to be classified. test For example, X-ray images in DICOM format obtained from a hospital's PACS system.

[0062] S20: Model output.

[0063] Specifically, the image to be classified in step S10 is input into the classification model trained in the first embodiment, and the classification model is used to sequentially perform image decomposition based on discrete wavelet transform → dual information flow evidence extraction → constructing a Beta distribution and calculating positive probability and confidence level → evidence fusion and outputting the final result.

[0064] In this embodiment, the classification model classifies each disease by (y sk , b sk , y tk , b tk After being concatenated, the data is input into a trained classification multilayer perceptron (fusion MLP), which outputs the final classification label for each disease. (0 or 1) and final confidence level b k (Real numbers between 0 and 1).

[0065] Therefore, the interpretable classification results finally output by the classification model include: the final classification label for each disease and its corresponding final confidence level.

[0066] In the above embodiments, preferably, S30 is further included: consistency verification.

[0067] In the above embodiments, preferably, in order to verify the reliability of the collected evidence, the following consistency verification steps can be performed (this step is usually used in the model validation stage or model interpretation stage to evaluate the reliability of the evidence, and is not a necessary step for regular reasoning): S301: Set a threshold (e.g., 0.5) and set the first positive probability y sk Compare with the threshold: If y sk If the probability is ≥ 0.5, the intermediate predicted label of the generated structure flow is 1; otherwise, it is 0. Similarly, the second positive probability y is... tk The intermediate predicted labels of the texture stream are generated by comparing them with a threshold.

[0068] S302: Calculate the first Cohen's Kappa coefficient κ between the intermediate predicted labels and the final classification labels of the structured flow. s And the second Cohen's Kappa coefficient κ between the intermediate predicted labels of the texture flow and the final classification labels. t The Kappa coefficient ranges from -1 to 1, with values ​​closer to 1 indicating stronger consistency.

[0069] Understandably, κs quantifies the consistency between the evidence collected by the structural information flow and the final classification result, while κt quantifies the consistency between the evidence collected by the texture information flow and the final classification result. The closer the coefficient value is to 1, the more reliable the evidence of the corresponding information flow. The reliability of the collected evidence is quantified by the first and second Cohen's Kappa coefficients. When the coefficient value is lower than a preset reliability threshold, an alarm is triggered or the final classification result is rejected to avoid making diagnostic decisions based on unreliable evidence.

[0070] To verify the effectiveness of the embodiments of the present invention, the inventors conducted the following experimental verifications on the TBX11K public dataset. Specifically, these include: (1) TBX11K dataset The tuberculosis screening dataset contains 11,200 CXR images (5,000 healthy, 5,000 non-tuberculous, and 1,200 tuberculous). Healthy and non-tuberculous cases are merged into a single "non-tuberculosis" category, using 6,600 images in the training set and 1,800 images in the validation set.

[0071] (2) Experimental setup The backbone network uses DenseNet121, ResNet50, and MobileNetV2, with the final classification layer modified to a 128-dimensional evidence feature extractor. All models are trained from scratch (without pre-training). The balance coefficient λ in the loss function is 0.01. The Adam optimizer (weight decay of 1e-4), gradient clipping with a maximum norm of 1.0, and cosine annealing are used for the learning rate. Training is performed for 200 epochs, with the learning rate decaying from 1e-3 to 1e-6. The batch size is 64 when the DWT module is enabled, and 32 otherwise. Implemented in PyTorch 2.8, it runs on an NVIDIA RTX 4090 GPU.

[0072] Evaluation metrics included AUC, ACC, F1 score (95% confidence interval for 1000 Bootstrap estimates), area under the learning curve (ALC), and total variation (TV). Consistency validation used Cohen's Kappa coefficient with a threshold of 0.5.

[0073] (3) Ablation test setup To comprehensively evaluate the contributions of the two core components of the provided technical framework—wavelet-based image decomposition (WID) and evidence-based deep learning (EDL)—this experiment compared four different architectural configurations: ①Baseline: A single-stream CNN trained using standard BCE.

[0074] ②WID-only: Introduce DWT decomposition into two streams, using only BCE for supervision.

[0075] ③EDL-only: Single-stream evidence learning, no DWT, using variational loss.

[0076] ④WID+EDL (this invention): Complete framework (DWT+Dual-stream EDL+Combination loss).

[0077] (4) Experimental results In all three scenarios, the present invention outperformed the baseline in AUC, ACC, and F1 by 2 / 3, 2 / 3, and 3 / 3 times respectively, outperforming other variants and demonstrating the most consistent superiority (see Table 1 for details).

[0078] Table 1 Summary of the two-category indicators for ablation experiments

[0079] (5) Verification of interpretability: ① Consistency analysis: Taking DenseNet121 as an example, the Kappa coefficients of the structured flow and texture flow in the TBX11K task are 0.9558 and 0.9853, respectively, and the ACC are 98.83% and 99.28%, respectively, which are close to the final accuracy of 99.44%, indicating that the two-stream evidence is highly reliable.

[0080] ②Visualization: The technical framework provided in this embodiment can collect sufficient negative evidence for disease-free images, collect positive evidence for specific diseases, and maintain transparency of the evidence path when multiple abnormalities coexist.

[0081] The above experiments fully demonstrate the significant advantages of the embodiments of the present invention in terms of classification performance, convergence efficiency, training stability, and inherent interpretability.

[0082] Third Embodiment like Figure 6 As shown, based on the second embodiment described above, the third embodiment of the present invention further provides a classification system for interpretable chest X-ray images, used to perform the above classification method. This classification system includes: an image decomposition module 1, a dual-information-flow evidence extraction module 2, a probability synthesis module 3, a classification module 4, and a consistency verification module 5.

[0083] The working process of each module unit will be explained in detail below: Image decomposition module 1 is used to acquire the original chest X-ray image and decompose it into a low-frequency structural information stream and a high-frequency texture information stream using discrete wavelet transform. Specifically, this module executes the image decomposition process described in the first and second embodiments above, including decomposing it into four components: ll, lh, hl, and hh. hh is filtered out, ll is used as the structural stream, and lh and hl are used as the texture streams.

[0084] Dual-information-stream evidence extraction module 2 is connected to image decomposition module 1 and is used to input low-frequency structural information stream into the first deep learning model (Net). s This is used to generate positive and negative structural evidence, and to input high-frequency texture information streams into a second deep learning model (Net). t This is used to generate positive and negative texture evidence. Among them, Net... s and Net t The output terminals are connected to evidence extraction multilayer perceptrons with ELU as the activation function.

[0085] The probability synthesis module 3 is connected to the dual-information-flow evidence extraction module 2. Based on subjective logic theory, it constructs a first Beta distribution according to the structural positive and structural negative evidence for each disease, obtaining the first positive probability and first confidence level for that disease. Furthermore, it constructs a second Beta distribution according to the texture positive and texture negative evidence for each disease, obtaining the second positive probability and second confidence level for that disease. The specific calculation formulas are described in the first and second embodiments above and will not be repeated here.

[0086] The classification module 4 is connected to the probability synthesis module and is pre-set with a classification model trained by the method of the first embodiment described above. This model is used to input the first positive probability, first confidence level, second positive probability and second confidence level of each disease into the classification multilayer perceptron after vector concatenation, generate the final classification label and final confidence level of each disease, and output an interpretable classification result. The interpretable classification result includes the final classification label of each disease and its corresponding final confidence level.

[0087] Furthermore, the consistency verification module 5 is used to set thresholds, compare the first positive probability with the threshold to generate intermediate prediction labels for the structure flow, and compare the second positive probability with the threshold to generate intermediate prediction labels for the texture flow. Then, the first Cohen's Kappa coefficient between the intermediate prediction labels for the structure flow and the final classification label, and the second Cohen's Kappa coefficient between the intermediate prediction labels for the texture flow and the final classification label are calculated for model validation and reliability assessment.

[0088] It is understood that the connection and synergistic relationship of the above-mentioned module units in this embodiment is only one specific implementation method for the second embodiment described above. In other embodiments, the connection and synergistic relationship of each module unit can also be adaptively adjusted as needed, and no specific limitation is made here.

[0089] Fourth embodiment like Figure 7As shown, based on the second embodiment described above, the fourth embodiment of the present invention further provides an electronic device. This electronic device includes one or more processors 100 and a memory 200. The memory 200 is coupled to the processor 100 and is used to store one or more programs. When the programs are executed by the processor 100, the one or more processors 100 implement the interpretable chest X-ray classification method described in the above embodiments.

[0090] The processor 100 controls the overall operation of the electronic device to complete all or part of the steps of the aforementioned method for classifying interpretable chest X-ray images. The processor 100 may be a central processing unit (CPU), graphics processing unit (GPU), field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), digital signal processing (DSP), etc. The memory 200 stores various types of data to support the operation of the electronic device. This data may include, for example, instructions for any application or method operating on the electronic device, as well as application-related data. The memory 200 may be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, etc.

[0091] In one exemplary embodiment, the electronic device may be implemented by a computer chip or physical entity, or by a product with a certain function, for performing the above-described method for classifying interpretable chest X-rays and achieving the same technical effect as the method described above. A typical embodiment is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, an in-vehicle human-machine interface device, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0092] In another exemplary embodiment, the present invention also provides a computer-readable storage medium including program instructions that, when executed by a processor 100, implement the steps of the method for classifying interpretable chest X-ray images in any of the above embodiments. For example, the computer-readable storage medium may be the memory 200 including the program instructions, which may be executed by the processor 100 of an electronic device to complete the method for classifying interpretable chest X-ray images and achieve the same technical effects as the method described above.

[0093] It should be noted that the above embodiments are merely illustrative examples. The technical solutions of each embodiment can be combined, and all are within the protection scope of this invention.

[0094] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0095] The foregoing has provided a detailed description of the construction method, classification method, and system for the interpretable chest X-ray classification model provided by this invention. Any obvious modifications made by those skilled in the art without departing from the essence of this invention will constitute an infringement of the patent rights of this invention and will incur corresponding legal liability.

Claims

1. A method for constructing an interpretable chest X-ray classification model, characterized in that... Includes the following steps: Acquire training images from multiple historical chest X-rays and their disease labels; The training image is decomposed into a low-frequency structural information stream representing the overall structure and a high-frequency texture information stream representing texture details using discrete wavelet transform. The low-frequency structural information stream is input into a first deep learning model to be trained to generate positive and negative structural evidence for each disease; the high-frequency texture information stream is input into a second deep learning model to be trained to generate positive and negative texture evidence for each disease; wherein the first and second deep learning models are convolutional neural networks with the same or different structures, and the outputs of the first and second deep learning models are respectively connected to an evidence extraction multilayer perceptron with ELU as the activation function, for mapping the extracted features into scalar evidence greater than 1; Based on the structural positive evidence and structural negative evidence, a corresponding first Beta distribution is constructed for each disease to obtain the first positive probability and first confidence level of the disease based on the structural information flow; based on the texture positive evidence and texture negative evidence, a corresponding second Beta distribution is constructed for each disease to obtain the second positive probability and second confidence level of the disease based on the texture information flow; wherein, the first Beta distribution is formed by the structural positive evidence α for each disease. sk Negative evidence of structure β sk Parameterization, the first positive probability is y sk =α sk / (α sk +β sk The first confidence level is b. sk =(α sk +β sk -2) / (α sk +β sk The second Beta distribution is composed of positive evidence α for the texture of each disease. tk Negative evidence of texture β tk Parameterization, the second positive probability is y tk =α tk / (α tk +β tk The second confidence level is b. tk =(α tk +β tk -2) / (α tk +β tk );α sk β represents positive structural evidence for the k-th disease. sk α represents the structural negative evidence for the k-th disease. tk β represents positive evidence of texture for the k-th disease. tk The texture represents negative evidence for the k-th disease, k∈(1,K), where K is the total number of disease categories included in the training or original image; The first positive probability, first confidence level, second positive probability, and second confidence level of each disease are concatenated into vectors and then input into a classification multilayer perceptron to generate the predicted classification label and predicted confidence level of the disease. By minimizing the total loss function, which is a weighted combination of binary cross-entropy loss and variational loss, the first deep learning model, the second deep learning model, and the classification multilayer perceptron are jointly trained to obtain a trained classification model.

2. The construction method as described in claim 1, characterized in that: The total loss function The simplified form is: ; in, For binary cross-entropy loss, For variational loss, Here are the non-negative hyperparameters used to balance the two types of losses; The variational loss is derived by minimizing the negative value of the modified lower bound of evidence, which is obtained by replacing the posterior distribution with an uninformative prior, and is used to constrain the statistical reasonableness of the evidence generated by the model.

3. A classification method for interpreting chest X-rays, characterized in that... The classification model constructed using the method described in claim 1 is used to classify chest X-ray images to be classified, including the following steps: Obtain the original chest X-ray image to be classified; The original image is decomposed into a low-frequency structural information stream representing the overall structure and a high-frequency texture information stream representing texture details using discrete wavelet transform. The low-frequency structural information stream is input into the first deep learning model in the classification model to generate positive and negative structural evidence for each disease; the high-frequency texture information stream is input into the second deep learning model in the classification model to generate positive and negative texture evidence for each disease; wherein the first and second deep learning models are convolutional neural networks with the same or different structures, and the outputs of the first and second deep learning models are respectively connected to an evidence extraction multilayer perceptron with ELU as the activation function, for mapping the extracted features into scalar evidence greater than 1; Based on the structural positive evidence and structural negative evidence, a corresponding first Beta distribution is constructed for each disease to obtain the first positive probability and first confidence level of the disease based on the structural information flow; based on the texture positive evidence and texture negative evidence, a corresponding second Beta distribution is constructed for each disease to obtain the second positive probability and second confidence level of the disease based on the texture information flow; wherein, the first Beta distribution is formed by the structural positive evidence α for each disease. sk Negative evidence of structure β sk Parameterization, the first positive probability is y sk =α sk / (α sk +β sk The first confidence level is b. sk =(α sk +β sk -2) / (α sk +β sk The second Beta distribution is composed of positive evidence α for the texture of each disease. tk Negative evidence of texture β tk Parameterization, the second positive probability is y tk =α tk / (α tk +β tk The second confidence level is b. tk =(α tk +β tk -2) / (α tk +β tk );α sk β represents positive structural evidence for the k-th disease. sk α represents the structural negative evidence for the k-th disease. tk β represents positive evidence of texture for the k-th disease. tk The texture represents negative evidence for the k-th disease, k∈(1,K), where K is the total number of disease categories included in the training or original image; The first positive probability, first confidence level, second positive probability, and second confidence level of each disease are concatenated into vectors and then input into the classification multilayer perceptron in the classification model to generate the final classification label and final confidence level of the disease. Output interpretable classification results, which include: the final classification label for each disease and its corresponding final confidence score.

4. The classification method as described in claim 3, characterized in that... The process of decomposition using discrete wavelet transform specifically includes: The training image or the original image is decomposed into low-frequency components, horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components. The diagonal high-frequency components are filtered out to perform data denoising; The low-frequency component is used as the low-frequency structural information stream, and the horizontal high-frequency component and the vertical high-frequency component are used together as the high-frequency texture information stream.

5. The classification method as described in claim 3, characterized in that: The first confidence level and the second confidence level are used to characterize the cognitive uncertainty of the first deep learning model and the second deep learning model, respectively. The lower the confidence level, the higher the cognitive uncertainty.

6. The classification method as described in claim 3, characterized in that... It also includes a consistency verification step, specifically including: Set a threshold, compare the first positive probability with the threshold to generate intermediate prediction labels for the structure flow, and compare the second positive probability with the threshold to generate intermediate prediction labels for the texture flow; Calculate the first Cohen's Kappa coefficient between the intermediate predicted label of the structure stream and the final classification label, and the second Cohen's Kappa coefficient between the intermediate predicted label of the texture stream and the final classification label; The first Cohen's Kappa coefficient is used to quantify the consistency between the evidence collected by the low-frequency structural information stream and the final classification result, while the second Cohen's Kappa coefficient is used to quantify the consistency between the evidence collected by the high-frequency texture information stream and the final classification result. The closer the coefficient value is to 1, the more reliable the evidence of the corresponding information stream.

7. A classification system for interpreting chest X-rays, used to perform the classification method of claim 3, characterized in that... include: The image decomposition module is used to acquire the original chest X-ray image and decompose it into a low-frequency structural information stream that represents the overall structure and a high-frequency texture information stream that represents the texture details using discrete wavelet transform. A dual-information-flow evidence extraction module is used to input the low-frequency structural information flow into a first deep learning model to generate positive structural evidence and negative structural evidence, and to input the high-frequency texture information flow into a second deep learning model to generate positive texture evidence and negative texture evidence; wherein, the output terminals of the first and second deep learning models are respectively connected to an evidence extraction multilayer perceptron with ELU as the activation function; wherein, the first and second deep learning models are convolutional neural networks with the same or different structures, and the output terminals of the first and second deep learning models are respectively connected to the evidence extraction multilayer perceptron with ELU as the activation function, for mapping the extracted features into scalar form evidence greater than 1; The probability synthesis module constructs a first Beta distribution based on the structural positive evidence and structural negative evidence for each disease, obtaining a first positive probability and a first confidence level for that disease; and constructs a second Beta distribution based on the texture positive evidence and texture negative evidence for each disease, obtaining a second positive probability and a second confidence level for that disease; wherein, the first Beta distribution is derived from the structural positive evidence α for each disease. sk Negative evidence of structure β sk Parameterization, the first positive probability is y sk =α sk / (α sk +β sk The first confidence level is b. sk =(α sk +β sk -2) / (α sk +β sk The second Beta distribution is composed of positive evidence α for the texture of each disease. tk Negative evidence of texture β tk Parameterization, the second positive probability is y tk =α tk / (α tk +β tk The second confidence level is b. tk =(α tk +β tk -2) / (α tk +β tk );α sk β represents positive structural evidence for the k-th disease. sk α represents the structural negative evidence for the k-th disease. tk β represents positive evidence of texture for the k-th disease. tk The texture represents negative evidence for the k-th disease, k∈(1,K), where K is the total number of disease categories included in the training or original image; The classification module is used to concatenate the first positive probability, the first confidence level, the second positive probability, and the second confidence level into a classification multilayer perceptron, generate the final classification label and final confidence level for each disease, and output an interpretable classification result. The interpretable classification results include the final classification label for each disease and its corresponding final confidence level.

8. The classification system as described in claim 7, characterized in that... Also includes: The consistency verification module is used to set a threshold, compare the first positive probability with the threshold to generate a structural flow intermediate prediction label, compare the second positive probability with the threshold to generate a texture flow intermediate prediction label, and calculate the first Cohen's Kappa coefficient between the structural flow intermediate prediction label and the final classification label, and the second Cohen's Kappa coefficient between the texture flow intermediate prediction label and the final classification label. The first Cohen's Kappa coefficient is used to quantify the consistency between the evidence collected by the low-frequency structural information stream and the final classification result, while the second Cohen's Kappa coefficient is used to quantify the consistency between the evidence collected by the high-frequency texture information stream and the final classification result. The closer the coefficient value is to 1, the more reliable the evidence of the corresponding information stream.

Images