Quantitative assessment method and system for chest x-ray images

Abstract

The application discloses a method and system for quantitative evaluation of chest X-ray images, which includes providing an input image to a pre-trained classification model to obtain prediction probabilities of multiple disease categories, generating a corresponding heat map for each disease category based on the classification model to locate the lesion area related to the disease category in the input image, and labeling the spatial semantic label of the lesion area according to its relative position in the preset lung field region mask, and finally performing quantitative evaluation on the lesion area of each disease category and outputting the quantitative evaluation result with the spatial semantic label. In this way, the scheme can provide accurate and readable quantitative evaluation report for chest X-ray images.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging artificial intelligence and medical information processing technology, specifically relating to a method and system for quantitative evaluation of chest X-ray images. Background Technology

[0002] The statements in this section are merely to provide background information related to the technical solutions of this application to aid understanding, and do not necessarily constitute prior art for the technical solutions of this application.

[0003] Chest X-ray (CXR) is a commonly used and fundamental imaging examination in clinical medicine, widely used for the screening and diagnosis of various chest diseases such as pneumonia, tuberculosis, lung cancer, pleural effusion, and pneumothorax. With the rapid increase in medical imaging data, traditional image analysis methods in clinical practice suffer from problems such as heavy workload of image interpretation, strong subjectivity, and insufficient consistency. Utilizing artificial intelligence technologies such as deep learning to achieve automatic classification and intelligent assisted diagnosis of chest X-ray images is gradually becoming a mainstream research direction. Intelligent assisted diagnosis based on deep learning can improve diagnostic efficiency and stability to a certain extent.

[0004] Deep convolutional neural networks have achieved remarkable performance in chest X-ray image classification tasks. Among them, the CheXNet model has achieved groundbreaking results in pneumonia detection. For example, the CheXNet model based on DenseNet121 can perform multi-label disease probability prediction on chest X-rays and identify multiple types of chest abnormalities, demonstrating potential that surpasses that of human doctors.

[0005] However, the clinical application of deep learning models in medical image analysis still faces significant challenges. The most critical issue lies in the "black box" nature of deep learning models. Although deep neural networks can achieve classification accuracy comparable to or even surpassing that of human experts, their decision-making process lacks transparency and interpretability. Clinicians and patients struggle to understand "why" and "on what basis" the model makes specific diagnostic judgments. This lack of interpretability severely restricts the implementation and trust-building of AI-assisted diagnostic systems in clinical practice. Summary of the Invention

[0006] During the process of researching the engineering implementation of the chest X-ray assisted analysis system, the inventors discovered that the existing image report results are all based on disease classification probabilities, which are provided by deep learning models with black-box characteristics. Therefore, clinicians cannot fully trust the results of image reports and still need to make judgments based on their own knowledge and experience, resulting in reports serving only as references without actual clinical significance.

[0007] Therefore, this application aims to provide a quantitative assessment report on chest X-ray images that can be trusted by clinicians, is highly accurate, and is readable.

[0008] The above objective is achieved through the following technical solution:

[0009] According to a first aspect of the embodiments of this application, a quantitative evaluation method for chest X-ray images is provided, comprising: providing an input image to a pre-trained classification model to obtain a classification result for the image, the classification result including predicted probabilities of multiple disease categories; generating a heatmap for each disease category based on the classification model, and locating lesion regions related to the disease category in the input image based on the heatmap of each disease category; labeling the lesion region with corresponding spatial semantic labels according to the relative position of the lesion region of each disease category in a pre-set lung field region mask, the lung field region mask dividing the lung area into multiple anatomical regions, each region being labeled with a different spatial semantic label; performing a quantitative evaluation on the lesion region of each disease category and outputting a quantitative evaluation result, the quantitative evaluation result being associated with the spatial semantic label of the lesion region.

[0010] In some embodiments, the method may further include preprocessing the input image, including color adjustment and image transformation; segmenting the preprocessed input image to obtain a segmentation mask including the background, heart region, and thoracic region; locating the leftmost and rightmost points of the heart shadow in the heart region mask; locating the leftmost and rightmost points of the thoracic cavity in the thoracic cavity mask; calculating the horizontal distance between the left and right points of the heart shadow as the maximum transverse diameter of the heart, calculating the horizontal distance between the left and right points of the thoracic cavity as the maximum transverse diameter of the thoracic cavity, and dividing the two to obtain the CTR parameter, which is used as one of the results of quantitative evaluation.

[0011] In some embodiments, the method may further include statistically analyzing the area of ​​the lesion region for each disease category and calculating the area ratio of the lesion region for that disease category in the input image, generating a severity score or grade for the disease category based on the predicted probability corresponding to the disease category and the area ratio of the lesion region, and using it as one of the results of quantitative evaluation.

[0012] In some embodiments, the method may further include converting all results of the quantitative assessment into clinical terminology and populating them into a predefined report template to generate a structured diagnostic report; the predefined report template is selected from at least one of a single lesion template, a multi-lesion template, or a template containing a cardiothoracic ratio parameter, based on differences in disease category combinations, number of lesions, or analysis conditions; the generated structured diagnostic report is presented in any of the following forms: text string, key-value pair data, JSON structure, database record, or exportable file.

[0013] In some embodiments, the input image can be a DICOM format image.

[0014] In some embodiments, locating lesion regions related to a disease category in an input image based on a heatmap of each disease category may include: selecting a preset response threshold to apply the heatmap... Figure 2 Value-based analysis yields one or more lesion regions corresponding to disease categories.

[0015] In some embodiments, labeling a lesion region with a corresponding spatial semantic label based on the relative position of the lesion region of each disease category in a pre-defined lung field region mask may include: identifying the corresponding region in the pre-defined lung field region mask based on the center point location, coverage area, or center of gravity of the lesion region, and using the spatial semantic label of the corresponding region as the corresponding spatial semantic label of the lesion region to indicate the anatomical location of the lesion region.

[0016] In some embodiments, the quantitative assessment results may include at least the severity of each disease category, which is determined based on the predicted probability of the disease category and the area ratio of the lesion region of the disease category in the input image.

[0017] According to a second aspect of the embodiments of this application, a computer program product is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described method.

[0018] According to a third aspect of the embodiments of this application, a quantitative evaluation system for chest X-ray images is provided, comprising: a classification module, which provides an input image to a pre-trained classification model to obtain a classification result for the image, the classification result including predicted probabilities of multiple disease categories; a heatmap generation module, which generates a heatmap for each disease category based on the classification model, and locates lesion regions related to the disease category in the input image based on the heatmap of each disease category; a spatial semantic description module, which labels the lesion region with corresponding spatial semantic labels according to the relative position of the lesion region of each disease category in a pre-set lung field region mask, the lung field region mask dividing the lung area into multiple anatomical regions, each region being labeled with a different spatial semantic label; and a quantitative result output module, which performs quantitative evaluation on the lesion region of each disease category and outputs the quantitative evaluation result, the quantitative evaluation result being associated with the spatial semantic label of the lesion region.

[0019] Compared with existing technologies, the advantages of this application are mainly as follows: Standardized preprocessing ensures consistency of images input to the deep learning model; heatmaps are introduced to provide visual and interpretable quantification of the classification probabilities derived by the deep learning model; based on the heatmap of each disease category, lesion regions related to that disease category are located in the input image, and corresponding spatial semantic labels are assigned; finally, quantification evaluation of the lesion regions for each disease category is performed, and quantification evaluation results with spatial semantic labels are output. This provides a highly accurate and readable quantification evaluation method that can be trusted by clinicians. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0021] Figure 1 This is a flowchart illustrating a method for quantitative evaluation of chest X-ray images according to an embodiment of this application.

[0022] Figure 2 This is a schematic diagram illustrating the automatic calculation process of the cardiothoracic ratio in a quantitative evaluation method for chest X-ray images, according to an embodiment of this application.

[0023] Figure 3 This is a schematic diagram of the report generation process of a quantitative evaluation method for chest X-ray images, according to an embodiment of this application.

[0024] Figure 4 This is a schematic diagram of the structure of a quantitative evaluation system for chest X-ray images according to an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided through specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0026] As described in the background section, due to the "black box" nature of deep learning models, existing technologies for intelligent analysis of chest X-rays only output a single disease classification probability indicator. Without any explanation of this indicator, clinicians cannot easily trust it, and even if they do, differences in skill and experience may lead to different clinical diagnoses, resulting in medical risks. Therefore, although artificial intelligence demonstrates the potential to surpass human doctors, many problems still exist in clinical practice.

[0027] To address the aforementioned issues, this application introduces a quantitative evaluation method and system for chest X-ray images. Figure 1 A flowchart illustrating a method for quantitative evaluation of chest X-ray images according to an embodiment of this application is shown. Figure 1 As shown, the quantitative evaluation method mainly includes the following steps: Step S1) The input image is provided to a pre-trained classification model to obtain the classification result for the image, the classification result including the predicted probability of multiple disease categories; Step S2) Based on the classification model, a heat map of the corresponding disease category is generated for each disease category, and the lesion area related to the disease category is located in the input image based on the heat map of each disease category; Step S3) According to the relative position of the lesion area of ​​each disease category in a pre-set lung field area mask, the lesion area is labeled with the corresponding spatial semantic label, the lung field area mask divides the lung area into multiple anatomical regions, and different regions are labeled with different spatial semantic labels; Step S4) The lesion area of ​​each disease category is quantitatively evaluated and the quantitative evaluation result is output, the quantitative evaluation result is associated with the spatial semantic label of the lesion area (the above steps will be described in detail below).

[0028] The aforementioned quantitative assessment method achieves a fully automated processing flow from inputting raw chest X-ray images to outputting quantitative assessment results. It eliminates the manual measurement and subjective interpretation steps in the traditional process and provides visual evidence for the uninterpretable parts of the deep learning model, making the quantitative assessment results more accurate, objective, and consistent. This enables the AI-assisted diagnostic system to be effectively integrated into the clinical workflow.

[0029] although Figure 1 Not shown, but in some embodiments, this quantitative evaluation method may also include a preprocessing step for the input chest X-ray image. That is, before inputting the chest X-ray image into the deep learning model (also known as a classification model), a series of preprocessing steps are performed to convert the raw image from any source into a uniform digital format that the deep learning model can recognize. This preprocessing step can be divided into two stages: a color adjustment stage and an image transformation stage.

[0030] During the color adjustment stage, the PIL library is used to read and perform basic processing on image files. The PIL library (Python Imaging Library) is a third-party image processing library in the Python ecosystem. Simply put, it can perform various functions on image files, such as opening, saving, and format conversion; image geometric transformations; and color and channel operations. Because images generated by different devices or in different formats may use different color encoding methods, such as grayscale, RGBA (A being the alpha channel, representing the transparency of each pixel from 0 to 255), four-channel encoding, or even CMYK (the four-color printing model, where C represents cyan, M represents magenta, Y represents yellow, and K represents black), the PIL library needs to force all images to be uniformly converted to the standard RGB three-color channel mode. This ensures that all images have the same mathematical structure for color representation, thus eliminating the need for additional adaptation for different input formats.

[0031] After reading images in common formats such as JPEG and PNG, if the input image itself is a three-channel RGB format, it is further converted to the standard color space sRGB to ensure consistent color visual performance on different display devices. If the input image is a single-channel grayscale format, PIL will copy the brightness value of this channel three times and fill it into the red, green, and blue channels respectively, forming an equivalent color image that is exactly the same in all three channels. If the input image has a transparency channel, the transparency information will be discarded first, and the image will be mixed with a white background before generating an opaque three-channel image.

[0032] After completing the image format unification and color adjustment, a PIL image object is obtained. Subsequent preprocessing operations in the image transformation stage are executed via a pipeline built using the `transform` module from the `torchvision` library. `torchvision` is the official toolkit in the PyTorch ecosystem specifically designed for computer vision tasks, while the `transform` module is an image processing toolkit containing tools such as `Resize`, `ToTensor`, and `Normalize`. These tools allow for multiple preprocessing combinations:

[0033] Because deep learning models require input tensors to have exactly the same height and width in order to perform batch parallel computations using fixed-dimensional matrices, the Resize tool is first used to scale all images to a fixed size of 224×224 pixels, consistent with the model training size.

[0034] The ToTensor tool is then used to convert the PIL image object into a tensor data structure specific to the PyTorch framework. A tensor is essentially a multi-dimensional array of values, serving as the basic unit for all mathematical operations and gradient calculations in PyTorch. This step automatically reduces the integer range of pixel values ​​from 0 to 255 to a floating-point range of 0.0 to 1.0, making the numerical magnitude more suitable for subsequent gradient descent optimization.

[0035] Finally, the Normalize tool is used to standardize the tensors using the mean and standard deviation of the ImageNet dataset. Each pixel value is subtracted from the mean used by the pre-trained model and then divided by the corresponding standard deviation, making the distribution of the entire dataset approximate a standard normal distribution with a mean of zero and a variance of one. Since most backbone networks (such as ResNet) are pre-trained on the ImageNet dataset, their layer parameters have already adapted to the statistical distribution of this dataset. Even when transferring to medical imaging tasks, maintaining the input distribution consistent with pre-training allows for more efficient utilization of the feature extraction capabilities learned during pre-training. This standardization operation ensures numerical benchmark alignment between different images, avoiding systematic biases caused by overall image brightness or darkness affecting model judgments.

[0036] Through the steps executed in the above sequence, the original chest X-ray image is finally transformed into a tensor with uniform size, standardized numerical range, standardized color space, and aligned distribution reference, thus possessing all the conditions to be input into a deep learning model for training or inference.

[0037] In another embodiment of this application, the chest X-ray image input to the deep learning model can be DICOM or other medical image formats. DICOM files not only contain image pixel data, but also embed a large amount of metadata such as patient information, equipment parameters, and scanning position. The pixel values ​​themselves represent measurements with clear physical meaning (such as CT values ​​or MRI signal intensity), rather than simple brightness or color values; and these pixel values ​​are usually stored with 12-bit or 16-bit precision, with a value range far exceeding the 0 to 255 grayscale range that the human eye can distinguish.

[0038] These characteristics mean that DICOM files cannot be directly opened by general image libraries such as PIL; their binary structure must first be parsed using a specialized pydicom library. This involves extracting key parameters from the metadata (such as window width and window level) and the original pixel data, correctly processing the photometric interpretation parameters, and ensuring the correct display of the grayscale image.

[0039] Secondly, window width and window level settings are applied to optimize image contrast. Since the complete grayscale range covered by a DICOM image encompasses an extremely wide density range from air to bone, directly compressing these tens of thousands of grayscale levels proportionally to a display range of 0 to 255 would result in extremely low contrast in critical diagnostic areas such as soft tissue, with lesion features submerged in a large amount of irrelevant information. Therefore, in the window width and window level settings, the window width defines the width of the grayscale range that needs to be highlighted, while the window level determines the center position of this range. The system only extracts pixel values ​​within the window width range, setting all values ​​below the lower limit of the window level to the lower limit value and all values ​​above the upper limit to the upper limit value. This extracted grayscale range is then linearly stretched to the 0 to 255 range. This targeted highlighting of the most relevant tissue details greatly enhances the identifiability of lesions.

[0040] Finally, the pixel values ​​are normalized as follows:

[0041]

[0042] This converts the original pixel values ​​from 16-bit physical measurements to 8-bit grayscale values ​​in the range of 0 to 255. At this point, the data format of the medical image is completely consistent with the conventional grayscale images that the PIL library can process, while maintaining the integrity of the diagnostic information.

[0043] By standardizing the preprocessing of input images of different formats, we ensure that all input images have a consistent data format and eliminate interference from non-pathological factors such as device differences and image noise. This ensures the accuracy and consistency of input information in subsequent processing.

[0044] Continue to refer to Figure 1 In step S1, the input chest X-ray image or the preprocessed chest X-ray image described above can be provided to a pre-trained classification model to obtain a classification result for the input image. A significant milestone in classification models for intelligent chest X-ray analysis is the CheXNet model proposed by Rajpurkar et al. at Stanford University in 2017. This model uses DenseNet121 as its backbone network structure, achieving efficient feature reuse and smooth gradient propagation through a dense connection mechanism, maintaining a low parameter count and good training stability even at deep network depths. The model input is a 224×224×3 RGB image, and its output layer corresponds to multiple disease labels, each label outputting a probability value belonging to that abnormal category. In the embodiments of this application, the CheXNet model is used as an example for classifying chest X-ray images; however, it should be understood that the scheme of this application is also applicable to other classification models.

[0045] In one embodiment of this application, after obtaining classification results including predicted probabilities of multiple disease categories for an input image using a CheXNet-based classification model, target categories that need to be included in subsequent interpretable region analysis can be selected or determined from these probabilities according to preset thresholds or ranking rules. For example, all categories with predicted probabilities exceeding a preset threshold can be selected, the top k categories with the highest predicted probabilities can be selected, or positive categories belonging to a pre-defined set of key diseases of concern can be selected.

[0046] Through the aforementioned flexible selection mechanism, the analysis scope can be dynamically determined according to the actual needs of different clinical scenarios, ensuring sufficient regional localization and visual assessment of high-risk or high-probability diseases. Furthermore, for each selected target category, its corresponding probability value can be compared with a preset dynamic judgment threshold to determine whether the disease category is positive, and simultaneously generate the severity level of that category and its spatial distribution information in various anatomical regions of the lung.

[0047] Continue to refer to Figure 1 In step S2, after obtaining the classification results of each disease category with probability values, a heat map of the corresponding category can be generated for each disease category based on the classification model, and the lesion area related to the disease category can be located in the input image based on the heat map of each disease category to enhance the interpretability of the model output.

[0048] For example, the Grad-weighted Class Activation Mapping (Grad-CAM) algorithm can be used to generate a heatmap that highlights the image regions in the input image that make the main contribution to the classification model's judgment of a specific disease category by utilizing the gradient information of each target category on the convolutional feature map, thereby providing a basis for lesion localization and subsequent spatial semantic analysis.

[0049] In one embodiment of this application, the process of obtaining a heatmap using the Grad-CAM algorithm is as follows:

[0050] We weight and combine the feature maps of the target convolutional layers of the classification model to obtain the activation heatmap corresponding to each target category. Let... This represents the nth feature map of the last convolutional layer in the network. If the output score represents category 𝑐, then the weight of the corresponding channel is defined as:

[0051]

[0052] in, This represents the total number of spatial locations in the feature map. This reflects the importance of the nth feature map in class determination.

[0053] After obtaining the weights of each channel, the Grad-CAM heatmap can be represented as:

[0054]

[0055] ReLU(⋅) is used to preserve response regions that positively contribute to the target category and suppress negative responses, thus highlighting the image regions that the model focuses on when determining a disease category. In the above formula, This represents the nth feature map of the last convolutional layer. The output score represents category A. This represents the importance weight of category A to channel B. The final generated heatmap is upsampled and mapped back to the original chest X-ray space for visualization and subsequent post-processing analysis.

[0056] Regarding the selection of convolutional layers, the last convolutional layer is used as the feature extraction layer for Grad-CAM in the embodiments of this application. This is because the features of this layer have both strong high-level semantic information and a certain spatial resolution, which can achieve a good balance between category discrimination ability and localization ability. If shallow features are selected, the semantic expression ability will be insufficient; if the output of a fully connected layer is selected, spatial location information cannot be preserved.

[0057] Furthermore, in the embodiments of this application, a ReLU operation is introduced in the heatmap generation to retain only regions that have a positive promoting effect on the target category, thereby avoiding interference from negative gradient regions with the lesion visualization results. In practical applications, heatmaps are typically generated only for categories predicted to be positive or for high-probability categories with the highest probability (r), rather than visualizing all categories one by one. This design reduces computational overhead and also helps to reduce noise introduced by irrelevant categories, improving the clinical readability of the results.

[0058] Based on the heatmap obtained for each disease category, further post-processing can be performed on the heatmap to extract the lesion region corresponding to that disease category. In the embodiments of this application, post-processing of the heatmap may include the following three steps: normalizing the heatmaps of multiple diseases, performing threshold segmentation on the heatmap based on a preset response threshold, and obtaining the lesion region corresponding to the disease category based on contour extraction or connected component analysis.

[0059] The heatmaps obtained using the Grad-CAM algorithm are essentially a set of activation values, typically floating-point numbers. Their original numerical range can vary significantly depending on the input image or target category. For example, the original activation values ​​of a pneumonia heatmap might be between 0 and 5, while those of a cardiomegaly heatmap might be between 0 and 0.5. Therefore, by first normalizing the heatmaps for multiple diseases, the activation value range of each heatmap is uniformly compressed to the interval of 0 to 1. This ensures that when the heatmaps for each disease category are drawn using the same color mapping rules (e.g., a gradient from red, representing the highest level of attention, through yellow, to blue, representing the lowest level of attention), they have a consistent visual reference. This also ensures that the same preset response threshold can be used to segment the heatmaps for each disease category during subsequent threshold segmentation.

[0060] After normalizing the heatmaps of various diseases, a preset response threshold is selected to perform binarization segmentation of the heatmaps. Specifically, pixels in the normalized heatmap above the threshold are considered high-response regions, and pixels below the threshold are considered background. Let the normalized heatmap be denoted as... Thus, a binarized region map is obtained. :

[0061]

[0062] Here, 𝜏 is the preset response threshold of the heatmap. This threshold can be a fixed value or dynamically set according to the target category, image statistics, or empirical rules.

[0063] Since a preset response threshold that is too low will introduce a lot of background noise, while one that is too high may result in incomplete high-response regions, to avoid the risks associated with an inappropriate preset threshold, contour-based extraction or connected component analysis is then performed on the binarized region map to obtain one or more lesion regions corresponding to the disease category. Preferably, connected component analysis is performed on multiple high-response regions. Regions with an area smaller than the preset minimum area threshold of the heatmap are considered noise regions and deleted, thus filtering out local responses that are too small and lack clinical significance. The remaining high-response regions represent the lesion regions corresponding to the disease category.

[0064] When multiple connected regions exist in an image, this application can adopt two strategies depending on actual needs: one is to retain only the connected region with the largest area to highlight the most important region; the other is to retain the high-response regions with the largest area to accommodate situations where multiple lesions coexist. The retention strategy can be flexibly adjusted according to the specific disease type and clinical display needs.

[0065] For each high-response region, multiple parameters can be calculated for quantitative evaluation, such as region bounding box coordinates, region area, region center point, region proportion of the total image area, and region relative coverage within the lung field. These parameters serve as inputs for subsequent spatial semantic analysis and severity assessment. In step S3, after obtaining lesion localization information suitable for clinical expression, the lesion region can be labeled with corresponding spatial semantic tags based on its relative position within a pre-defined lung field region mask for each disease category, indicating its corresponding anatomical location. The lung field region mask divides the lung area into multiple anatomical regions, each labeled with a different spatial semantic tag to indicate its corresponding anatomical location.

[0066] The lung field mask can be a pre-defined standard chest X-ray image. Based on the anatomical partitioning rules of the standard chest X-ray image, the lung area of ​​the lung field mask can be pre-divided into multiple anatomical regions, for example, horizontally divided into left and right sides, and vertically divided into upper, middle, and lower lung fields. Therefore, the anatomical location of a lesion region can be determined by matching the lesion region of a disease category with the pre-defined lung field mask. For example, based on the center point, coverage area, or centroid of the lesion region, its corresponding region in the pre-defined lung field mask can be identified. The spatial semantic label of this corresponding region is then used as the corresponding spatial semantic label for the lesion region, indicating its anatomical location and outputting a corresponding spatial semantic description.

[0067] For example, when the lesion center is located in the right half of the lung field mask and mainly covers the lower third of the area, the lesion area can be labeled with the spatial semantic tag "right lower lung field"; when the lesion area covers the left half of the lung field mask and spans both the upper and middle lung fields, the lesion area can be labeled with the spatial semantic tag "left upper lung field and middle lung field"; when there are multiple significant lesion areas, multiple spatial semantic tags can be labeled for each lesion area, and the spatial semantic tags can be sorted or merged according to area proportion, response intensity, or disease probability. In addition, not only can a single spatial semantic tag be labeled for a single lesion area, but more granular site descriptions can also be generated, such as "diffuse distribution in both lungs" or "mainly in a certain area and involving other areas".

[0068] For each lesion area, by calculating the relative position of its centroid coordinates with the lung region to which it belongs, it is assigned a corresponding spatial semantic label such as "left upper lung field" or "right middle lung field". This transforms the description of the lesion area into a spatial semantic label that is more in line with clinical expression habits, which can be used in the "imaging description", "positive findings" or "impression summary" sections of the subsequent report.

[0069] Continue to refer to Figure 1 In step S4, after obtaining the corresponding spatial semantic description of the lesion region, a quantitative assessment of the lesion region for each disease category can be performed, and a quantitative assessment result with the spatial semantic description can be output. The quantitative assessment result may include multiple quantitative assessment parameters calculated for the lesion region of each disease category, such as the coordinates of the region bounding box, the region area, the region center point, the proportion of the region to the total image area, and the relative coverage ratio of the region within the lung field. The quantitative assessment result may also include an assessment of the disease severity.

[0070] In real clinical settings, the assessment of disease severity relies solely on disease classification probabilities, mapping a continuous probability of 0-1 to three severity levels: 0-0.4 for mild, 0.4-0.7 for moderate, and 0.7-1 for severe. This linear mapping is simple and intuitive, but it may not fully reflect the actual impact of the lesion.

[0071] Therefore, in the embodiments of this application, in order to further quantify the lesion area of ​​each disease category more accurately, based on assessing the severity of the disease using the predicted probability of the target disease category, a severity score or level is generated by combining the area information of the lesion area. Specifically, the predicted probability of the target disease category and the area ratio of the lesion area corresponding to that disease category are used as inputs and jointly input into the severity calculation function. This severity calculation function comprehensively considers two dimensions: disease probability and lesion area ratio, and calculates the severity score using the formula shown below. :

[0072]

[0073] in, The predicted probability value representing the target disease category. The severity score, calculated by representing the area of ​​the lesion region that represents the target disease category, can be a comprehensive score within a continuous numerical range (usually 0 to 1).

[0074] By analyzing disease probability, lesion area proportion, and severity score, not only can the above-mentioned numerical severity score be obtained, but also discrete grading results or predefined text labels can be generated. Specifically, for the obtained severity score, the continuous interval from 0 to 1 is divided into several non-overlapping sub-intervals, each sub-interval is mapped to a corresponding severity level, similar to the basic assessment of disease classification probability. The severity score is then output as a discrete grading result, such as mild, moderate, or severe, based on the sub-interval it falls into. Furthermore, based on this discrete grading result or by inputting disease probability and lesion area proportion, corresponding fields in a predefined report template can be selected to generate text descriptions such as mild involvement, extensive involvement, or obvious lesions.

[0075] Furthermore, the areas of multiple lesion regions can be summed to obtain the overall lesion region area ratio, which is then combined with disease probability to form a basis for severity assessment. This avoids relying solely on classification probability while ignoring the spatial extent of lesions, or relying solely on region area while ignoring category confidence, thereby improving the stability and clinical understandability of severity expression. The obtained severity results can be used to refine disease descriptions, indicate the degree of abnormality, or generate imaging impressions in the final report.

[0076] In another embodiment of this application, the quantitative evaluation result may also include the cardiothoracic ratio (CTR), and the above-described quantitative evaluation method for chest X-ray images may further include a step of automatically calculating the cardiothoracic ratio (CTR). Figure 2 The flowchart shown is for the automatic calculation of the cardiothoracic ratio (CTR) in a quantitative evaluation method for chest X-ray images. After preprocessing the input image, including color adjustment and image transformation, the preprocessed input image is segmented to obtain a segmentation mask including the background, heart region, and thoracic cavity region. The leftmost and rightmost points of the heart shadow are located in the heart region mask; the leftmost and rightmost points of the thoracic cavity are located in the thoracic cavity region mask; the horizontal distance between the left and right points of the heart shadow is calculated as the maximum transverse diameter of the heart, and the horizontal distance between the left and right points of the thoracic cavity is calculated as the maximum transverse diameter of the thoracic cavity, and the two are divided to obtain the CTR parameter.

[0077] CTR, as an important geometric indicator reflecting the relative size of the heart to the transverse diameter of the thoracic cavity, is not only of practical value in judging whether the heart is enlarged and indicating abnormal cardiac function, but can also be used as one of the results of quantitative evaluation methods for chest X-ray images.

[0078] Specifically, the original chest X-ray image or the preprocessed image is input into a semantic segmentation model, such as DeepLabV3. DeepLabV3, as the semantic segmentation model, can perform pixel-level predictions in three semantic spaces (background / chest cavity / heart), outputting a multi-class segmentation mask where each pixel is assigned a class label: background, heart region, or chest cavity region. This mask has the same spatial resolution as the original image, providing a pixel-by-pixel anatomical localization basis for subsequent geometric measurements.

[0079] After obtaining the masks for the heart and thoracic regions, morphological and contour methods are used to extract their boundaries to calculate the transverse diameters of the heart and thoracic cavities. For the heart region, a contour tracing algorithm (such as the Suzuki algorithm) is used to extract the outer contour point set of the heart region. Then, all pixels on the contour are traversed to find the minimum and maximum values ​​of the horizontal coordinate (X-axis), corresponding to the leftmost and rightmost points of the heart shadow, respectively. The horizontal distance between these two points is the maximum transverse diameter of the heart. For the thoracic region, the same method is used to extract the outer contour of the thoracic cavity and locate the leftmost and rightmost points of the thoracic cavity. The horizontal distance between these points is the maximum transverse diameter of the thoracic cavity. In this process, the system typically takes the farthest points on the outer sides of the costophrenic angles as the thoracic cavity reference points to reflect the maximum internal diameter of the thoracic cavity.

[0080] After obtaining the two transverse diameter values ​​mentioned above, the calculation is performed according to the classic cardiothoracic ratio formula:

[0081]

[0082] Where represents the horizontal distance between the leftmost and rightmost points of the heart shadow, and represents the horizontal distance between the leftmost and rightmost points of the thoracic cavity. Both are absolute values, and the result is a floating-point number between 0 and 1 (typically within the range of approximately 0.42 to 0.50).

[0083] After obtaining the CTR parameter, it can be directly output as a numerical field in a structured report, that is, the calculated CTR floating-point number is directly filled into the corresponding quantitative field in the structured report, such as "cardiothoracic ratio: 0.48". Alternatively, a qualitative conclusion can be generated based on a preset threshold, that is, the CTR value is compared with a preset threshold (usually 0.50 or a threshold adjusted according to race, gender, and age group). If CTR > 0.50, a descriptive conclusion such as "increased cardiothoracic ratio" or "indicating a tendency for enlarged cardiac silhouette" is automatically generated; if CTR ≤ 0.50, "cardiothoracic ratio is within the normal range" is output.

[0084] In another embodiment of this application, to improve measurement stability, a series of post-processing operations can be performed on the original segmentation mask. First, hole filling is performed using morphological closing operations or connected component analysis to fill the hollow regions (i.e., internal pixels incorrectly labeled as background) within the heart or thoracic cavity region caused by insufficient model prediction confidence with the corresponding category, avoiding boundary extraction failure due to internal breaks. Second, smoothing is performed using Gaussian filtering or median filtering to denoise the segmentation boundaries, eliminating burrs and jagged edges caused by image noise or model output oscillations, making the heart and thoracic contours smoother and closer to the real anatomical boundaries. Finally, abnormal region removal is performed by calculating the area and shape characteristics of each connected component to remove mis-segmented regions with excessively small areas or abnormal eccentricities (e.g., narrow strips), preventing these abnormal regions from interfering with subsequent extreme point localization.

[0085] CTR is obtained through the steps described above and is completely independent of the logic of classification models (such as CheXNet). When both quantitative CTR results and disease probabilities output by the classification model are presented simultaneously, doctors can independently verify the diagnostic conclusions in the report based on CTR, an objective morphological parameter. For example, if the model outputs a probability of "enlarged cardiac silhouette" of 0.85, and the CTR is automatically calculated to be 0.55 (exceeding the normal threshold of 0.50), and the two are consistent, the credibility of the reported conclusion is significantly improved.

[0086] To address the challenge that while the raw probability values ​​and pixel-level heatmaps output by assisted diagnostic systems possess some diagnostic reference value, they are difficult to directly integrate with standard clinical workflows during clinical deployment, this paper proposes a method. Based on disease classification results and interpretability analysis, a rule engine and template constraint mechanism are used to transform the heterogeneous computational results generated by deep learning models into highly readable, professional diagnostic documents that conform to radiology writing conventions. Unlike directly outputting disease probabilities or heatmaps, this application further integrates geometric quantification indicators, spatial distribution semantics, and negative confirmation logic, thereby enabling standardized text format to serve the clinical diagnostic process.

[0087] In one embodiment of this application, such as Figure 3 The flowchart of the report generation method shown includes the following steps after completing disease probability prediction, lesion spatial localization, severity grading, and cardiothoracic ratio parameter measurement: converting all the results of the quantitative assessment into clinical terms and filling them into a predefined report template to generate a structured diagnostic report for multi-source information fusion.

[0088] Specifically, the process maps disease categories and probability values ​​to positive finding candidates, spatial semantic labels (such as "right lower lung field") to location phrases in image descriptions, severity grading (mild / moderate / severe) to corresponding severity modifiers, and CTR values ​​are directly used as quantitative indicators or, after thresholding, generated qualitative conclusions such as "tendency for cardiac enlargement." These diverse analytical results from different sources are organized into unified structured fields according to predefined mapping rules. Based on this, the above fields are further combined to generate "impression summary" and "suggestion fields." For example, when a case is pneumonia located in the right lower lung field, with moderate severity and a normal CTR, the above steps can generate "abnormally high-response lesion in the right lower lung field" as a positive finding, "moderate abnormal shadow visible in the right lower lung field, possibly indicating inflammatory lesion" as an image description, and form an impression summary of "possible inflammatory changes in the right lower lung field" and a follow-up suggestion of "re-examination recommended in conjunction with clinical findings and if necessary."

[0089] After fusion is completed, the report template generation engine is called to fill the structured fields into the predefined report template and output a structured diagnostic report containing chapters such as examination findings, positive findings, imaging description, quantitative indicators, impression summary and recommendations. Different templates such as single lesion, multiple lesion or CTR can be flexibly switched according to the disease combination and analysis conditions.

[0090] Compared with end-to-end free text generation methods, this invention adopts a rule- and template-driven approach, which can ensure stable output fields, standardized terminology, and compliance with clinical writing conventions, making it easy for doctors to quickly review and correct. The final structured report can be presented directly as a text string, or saved as key-value pairs, JSON structures, database records, or exportable files.

[0091] Figure 4 A schematic diagram of the structure of the quantitative evaluation system for chest X-ray images according to this application is shown. The quantitative evaluation system in this embodiment includes: an image preprocessing module, a classification module, a heatmap generation module, a spatial semantic description module, a CTR calculation module, a quantitative result output module, and a report generation module. The CTR calculation module can be selectively enabled according to system configuration or analysis requirements. The specific implementation of each module can be referred to the above-described combination. Figure 1 , 2 The steps related to the quantitative evaluation method for chest X-ray images are described in section 3.

[0092] Compared to systems used for chest X-ray images, the quantitative evaluation system in this embodiment can complete chest X-ray image preprocessing, disease classification, result generation, and visualization in a shorter time, helping to provide high-precision quantitative indicators and providing interpretable evidence for deep learning models. Meanwhile, the system design fully considers the standardization of medical image data processing and the integrity of business processes, ensuring that each functional module remains relatively independent while enabling collaborative work through a unified interface.

[0093] Specifically, the image preprocessing module receives an input image from, for example, an X-ray imaging device, and converts the original image from any source into a unified digital format that the model can recognize through a color adjustment stage and an image transformation stage. Specific implementation details can be found in the image preprocessing steps described above.

[0094] After inputting and preprocessing the image, the quantization evaluation system in this embodiment performs the following processing flow: First, in the classification module, the input image is provided to a classification model, such as the CheXNet model, for forward inference to obtain probability predictions for various diseases. Then, after obtaining the classification results for each disease category with probability values, the heatmap generation module generates a disease heatmap and locates the lesion region related to that disease category in the input image based on the heatmap of each disease category; the spatial semantic description module labels the lesion region with corresponding spatial semantic tags according to the relative position of the lesion region of each disease category in a pre-set lung field mask; the CTR calculation module includes a heart / thoracic segmentation module and a transverse diameter extraction module, and calculates the CTR using a segmentation model such as DeepLabV3. Finally, the quantization result output module performs quantization evaluation on the lesion region of each disease category and outputs the quantization evaluation results, which are returned to the front end for visualization in formats such as JSON. Specific implementation methods for each module can be found in the corresponding steps described above. To address the lack of clinical trust caused by the "black box" nature of deep learning models in the medical field, this application proposes a quantitative evaluation method and system for chest X-ray images. Firstly, the system generates heatmaps for specific disease categories. By backpropagating gradient information to the convolutional layer, it highlights the spatial regions that contribute most to the model's decision-making in a thermal format, thus intuitively revealing the model's judgment criteria. Furthermore, spatial constraints are applied to the lung segmentation mask and the heatmap, strictly limiting the highlighted areas to the lung parenchyma, effectively eliminating noise interference from non-anatomical structures and significantly improving the clinical credibility of the attribution results. Secondly, a semantic segmentation model is used to accurately extract masks of the heart and pleural cavity regions from chest X-ray images, and the cardiothoracic ratio parameter is automatically calculated accordingly. This process does not rely on the classification logic of deep learning but provides independent anatomical quantitative evidence based on traditional morphological measurements, forming cross-validation with the model's prediction results.

[0095] After obtaining the results of the multi-source analysis, the system uses semantic mapping to transform previously isolated and difficult-to-interpret technical parameters (such as disease probability values, heatmap activation areas, and CTR values) into anatomical locations, degree modifiers, and diagnostic terms familiar to doctors. Simultaneously, standardized templates ensure structural uniformity and terminology standardization in the reports, avoiding potential deviations or ambiguities in expression that might arise from freely generated models. This allows doctors to quickly grasp the logical relationship between diagnostic points and quantitative evidence. Doctors can comprehensively evaluate the rationality of the model's judgments from both quantitative parameters and spatial positioning dimensions, effectively realizing the transformation from "black-box prediction" to "white-box verification." Furthermore, it possesses practical value as a verifiable, traceable, and data-entryable electronic medical record system, providing doctors with clinically readable and trustworthy auxiliary diagnostic evidence.

[0096] In another embodiment of this application, a computer program product is also provided, which stores a computer program or instructions. When the computer program or executable instructions are executed by a processor, the technical solution described in the foregoing embodiments is implemented. The implementation principle is similar and will not be repeated here. This computer program product can be, but is not limited to, the following forms: a computer-readable storage medium, a downloadable program product, a program product embedded in a chip carrying program propagation signals, etc. The computer program product can be embodied in a computer-readable storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, a solid-state drive, a memory card, etc. When a computer program is stored on the storage medium, the processor reads and executes the computer program, thereby implementing the method according to the embodiments of this application. The computer program product can also be embodied in a program file that can be downloaded via a network, including installation packages, image files, compressed packages, executable files, script files, etc. After a user downloads the program product to a local computer, server, or terminal device via the Internet or a local area network, the processor executes the program through installation, loading, or running operations to achieve the corresponding function.

[0097] References to "various embodiments," "some embodiments," "one embodiment," or "embodiment," etc., in this specification refer to a specific feature, structure, or property described in connection with an embodiment that is included in at least one embodiment. Therefore, the appearance of the phrases "in various embodiments," "in some embodiments," "in one embodiment," or "in an embodiment," etc., throughout this specification does not necessarily refer to the same embodiment. Furthermore, specific features, structures, or properties can be combined in any suitable manner in one or more embodiments. Therefore, a specific feature, structure, or property shown or described in connection with one embodiment can be combined, in whole or in part, with features, structures, or properties of one or more other embodiments without limitation, provided that such combination is not illogical or inoperable.

[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments or implementation methods. Similar or identical parts between the various embodiments of the present invention can be referred to mutually. The implementation principles and technical effects of the inventive concept can be referred to mutually, and will not be repeated here. Where there is no conflict, the various embodiments or implementation methods in this invention can be combined with each other.

[0099] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used in the embodiments of the present invention is chosen to best explain the principles, practical applications, or technological improvements in the market, or to enable other those skilled in the art to understand the disclosed embodiments of the present invention.

Claims

1. A method for quantitative evaluation of chest X-ray images, comprising: The input image is fed into a pre-trained classification model to obtain a classification result for the image, the classification result including the predicted probabilities of multiple disease categories; Based on this classification model, heatmaps for each disease category are generated, and lesion regions related to that disease category are located in the input image based on the heatmaps for each disease category. Based on the relative position of the lesion area of ​​each disease category in the pre-set lung field area mask, the lesion area is labeled with a corresponding spatial semantic label. The lung field area mask divides the lung area into multiple anatomical regions, and different regions are labeled with different spatial semantic labels. For each disease category, the lesion area is quantitatively assessed and the quantitative assessment result is output. The quantitative assessment result is associated with the spatial semantic label of the lesion area.

2. The method according to claim 1, further comprising: The input image undergoes preprocessing including color adjustment and image transformation; The preprocessed input image is segmented to obtain a segmentation mask that includes the background, heart region, and chest cavity region. Locate the leftmost and rightmost points of the heart shadow within the heart region mask; Locate the leftmost and rightmost points of the thoracic cavity within the mask in the chest region; The horizontal distance between the left and right points of the heart shadow is calculated as the maximum transverse diameter of the heart, and the horizontal distance between the left and right points of the thorax is calculated as the maximum transverse diameter of the thorax. The two are then divided to obtain the CTR parameter, which is used as one of the results of quantitative evaluation.

3. The method according to claim 1 or 2, further comprising: The area of ​​the lesion region for each disease category is statistically analyzed, and the area ratio of the lesion region for that disease category in the input image is calculated. The severity score or grade of the disease category is generated based on the predicted probability and the proportion of lesion area, and is used as one of the results of quantitative assessment.

4. The method according to claim 1 or 2, further comprising: All results from the quantitative assessment are converted into clinical terminology and populated into a predefined report template to generate a structured diagnostic report; The predefined report template is selected from at least one of the following: single lesion template, multi-lesion template, or template containing cardiothoracic ratio parameter; The generated structured diagnostic report can be in any of the following formats: text string, key-value pair data, JSON structure, database record, or exportable file.

5. The method according to claim 1 or 2, wherein the input image is a DICOM format image.

6. The method according to claim 1 or 2, wherein, The process of locating lesion regions related to each disease category in the input image based on the heatmap for each disease category includes: The heatmap is binarized by selecting a preset response threshold to obtain one or more lesion regions corresponding to the disease category.

7. The method according to claim 1 or 2, wherein labeling the lesion region with a corresponding spatial semantic tag based on the relative position of the lesion region of each disease category in a pre-defined lung field region mask includes: Based on the location of the center point, coverage area, or center of gravity of the lesion area, the corresponding region in the pre-defined lung field area mask is identified, and the spatial semantic label of the corresponding region is used as the corresponding spatial semantic label of the lesion area to indicate the anatomical location of the lesion area.

8. The method according to claim 1 or 2, wherein, The quantitative assessment results include at least the severity of each disease category, which is determined based on the predicted probability of the disease category and the area ratio of the lesion region of the disease category in the input image.

9. A computer program product having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

10. A quantitative evaluation system for chest X-ray images, comprising: The classification module provides the input image to a pre-trained classification model to obtain a classification result for the image, the classification result including the predicted probabilities of multiple disease categories; The heatmap generation module generates heatmaps for each disease category based on the classification model, and locates lesion areas related to that disease category in the input image based on the heatmaps for each disease category. The spatial semantic description module labels the lesion area with corresponding spatial semantic tags according to the relative position of the lesion area of ​​each disease category in the pre-set lung field area mask. The lung field area mask divides the lung area into multiple anatomical regions, and different regions are labeled with different spatial semantic tags. The quantitative result output module performs quantitative evaluation on the lesion area for each disease category and outputs the quantitative evaluation result, which is associated with the spatial semantic label of the lesion area.

Images