An artificial intelligence model for quantitative evaluation of lung inflammatory lesions based on CT images

Abstract

The application discloses a lung inflammation lesion quantitative evaluation artificial intelligence model based on CT images and relates to the field of artificial intelligence models.The model comprises CT image preprocessing, self-supervised lesion segmentation and quantification, multi-modal joint pre-training coding and multi-task diagnosis and prognosis prediction module composition which are sequentially and communicatively connected, and realizes feature interaction through a unified feature embedding space.The model realizes self-supervised lesion segmentation and multi-dimensional quantification by adopting a three-dimensional generative reconstruction network, completes image-text feature fine-grained alignment through cross-modal contrast learning of an adversarial enhancement, and fuses multi-dimensional features.The model can complete multiple tasks such as ARDS diagnosis, non-invasive estimation of P / F ratio, severity grading and prognosis prediction in parallel, reduces dependence on artificial labeling, improves cross-center generalization capability and diagnosis precision, and provides a standardized intelligent tool for clinical evaluation of severe lung inflammation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence models, specifically to an artificial intelligence model for quantitative assessment of lung inflammatory lesions based on CT images. Background Technology

[0002] Severe pneumonia, especially ARDS, is a serious lung disease, usually caused by multiple factors, including infection, trauma, and inhalation of harmful substances. Besides general physiological testing, existing assessment methods mainly rely on clinical symptoms and traditional imaging examinations; however, these methods often lack objectivity and accuracy. In recent years, with the development of computer vision and deep learning technologies, image-based disease assessment has gradually become a research hotspot. Some studies have attempted to use CT images to assess pulmonary inflammatory lesions, but most methods still suffer from insufficient accuracy and poor clinical applicability. Existing methods struggle to provide reproducible quantitative biomarkers and cannot effectively integrate imaging, linguistic, and metadata features to achieve a comprehensive assessment of pulmonary inflammatory lesions, including diagnosis, severity stratification, prognostic prediction, and complication analysis. Currently, CT image-based assessment techniques mainly rely on fully supervised learning models, which require large amounts of labeled data for training and perform poorly when dealing with complex lesions. Traditional image analysis methods often rely on manual feature extraction, making it difficult to capture subtle pathological changes. Furthermore, existing assessment indicators, such as the PaO2 / FiO2 ratio, often rely on indirect monitoring methods, lacking real-time performance and accuracy. While some studies have attempted to combine multimodal data for analysis, numerous challenges remain in feature fusion and model training. With the deepening understanding of the pathological mechanisms of severe pneumonia, there is an urgent need for a novel assessment tool that can comprehensively consider imaging features, clinical data, and pathological information to improve diagnostic accuracy and the reliability of prognostic assessment. Summary of the Invention

[0003] The purpose of this invention is to provide an artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence model for quantitative assessment of lung inflammatory lesions based on CT images, comprising a CT image preprocessing module, a self-supervised lesion segmentation and quantification module, a multimodal joint pre-training coding module, and a multi-task diagnosis and prognosis prediction module connected in sequence. Each module achieves feature interaction and data encapsulation through a unified feature embedding space.

[0005] The CT image preprocessing module is used to sequentially perform spatial resampling, HU value intensity normalization and automatic lung parenchyma region segmentation on the input chest CT three-dimensional volume data, and output standardized lung parenchyma CT volume data.

[0006] The self-supervised lesion segmentation and quantification module is used to synthesize and reconstruct healthy lung structures from the standardized lung parenchyma CT volume data using a three-dimensional generative reconstruction network, generating healthy reference CT volume data that matches the individual anatomical structure of the input CT. It then calculates the voxel-level HU value difference between the standardized lung parenchyma CT volume data and the healthy reference CT volume data, generating an abnormal enhancement map that retains only high-density abnormal regions. Based on a variational constraint optimization model, iterative optimization of lesion boundaries and noise suppression are performed on the abnormal enhancement map to generate a three-dimensional lesion segmentation mask. Based on the three-dimensional lesion segmentation mask, multi-dimensional structured quantification indicators are extracted, and a lesion quantification feature vector is constructed. These multi-dimensional structured quantification indicators include at least the proportion of lesion volume in different HU density intervals to the total lung volume, and the spatial distribution parameters of lesions in each HU density interval relative to the geometric center of the lung.

[0007] The multimodal joint pre-training coding module incorporates an image coding unit, a text coding unit, an adversarial perturbation enhancement unit, a lesion soft label distillation unit, and a metadata prediction unit. The image coding unit extracts features from standardized lung parenchyma CT data to generate image embedding features. The text coding unit extracts features from the standardized radiology report text corresponding to the input CT scan to generate text embedding features. A cross-modal contrastive learning mechanism maps the image embedding features and text embedding features to a unified feature embedding space, achieving bidirectional fine-grained alignment between image features and radiological semantic features. The adversarial perturbation enhancement unit generates adversarial variants of key semantics related to lesions in the radiology report, strengthening the model's ability to identify fine-grained pathological features through contrastive constraints. The lesion soft label distillation unit uses the abnormal enhancement image as a lesion soft label, guiding the image coding unit to learn features of lung parenchyma lesion regions. The metadata prediction unit integrates patient demographic metadata for joint modeling, improving the model's domain adaptability across devices and institutions. Finally, it outputs a unified multimodal embedding feature that fuses image embedding features, lesion quantification features, text semantic features, and patient metadata features.

[0008] The multi-task diagnosis and prognosis prediction module is used to complete downstream multi-task parallel inference through task adaptive fine-tuning based on the unified multimodal embedding features. The downstream tasks include at least acute respiratory failure identification, acute respiratory distress syndrome diagnosis, non-invasive estimation of P / F ratio, Berlin standard severity classification, identification of right ventricular dysfunction related to acute respiratory distress syndrome, and prediction of short-term survival prognosis of patients. Finally, it outputs structured quantitative assessment results of pulmonary inflammatory lesions and clinical risk warning information.

[0009] Furthermore, the specific steps of the CT image preprocessing module in performing standardization processing are as follows: First, isotropic spatial resampling is performed on the input chest CT three-dimensional volume data to unify the voxel spacing to a preset standard value; then, HU value intensity normalization is performed on the resampled CT data, limiting the HU value to the effective lung range of [-1024, 400] and linearly mapping it to the range of [0, 1]; finally, automatic extraction and masking of the lung parenchyma region is completed through a pre-trained lung segmentation network to remove interference data from the chest wall, mediastinum, and background regions, and output standardized lung parenchyma CT volume data.

[0010] Furthermore, the self-supervised lesion segmentation and quantization module incorporates a three-dimensional CycleGAN generative reconstruction network to generate healthy reference CT volume data. The three-dimensional CycleGAN generative reconstruction network includes a first generator, a second generator, a first discriminator, and a second discriminator. The first generator maps abnormal lung CT domain data to a healthy lung CT domain, the second generator maps healthy lung CT domain data back to the abnormal lung CT domain, the first discriminator determines the authenticity of the healthy domain image, and the second discriminator determines the authenticity of the abnormal domain image. Both the first and second generators employ a three-dimensional convolutional residual structure, consisting of an input layer, a double downsampling layer, six three-dimensional ResNet residual blocks, a double upsampling layer, and an output layer. During training, adversarial loss and cycle consistency loss are simultaneously introduced for joint optimization.

[0011] Furthermore, the specific method by which the self-supervised lesion segmentation and quantification module generates abnormal enhancement maps and lesion segmentation masks is as follows: Voxel-by-voxel HU difference is calculated between standardized lung parenchyma CT volume data and healthy reference CT volume data, and the difference is expressed using the following formula:

[0012]

[0013] Only the difference in density between abnormal areas in the original CT scan and those in the healthy reference CT scan is retained. For raw CT body data, For healthy reference CT scan data, interval clipping and normalization are performed on the difference results to generate a three-dimensional abnormal enhancement map. Based on the Chan-Vese variational constraint optimization model, the region energy of the abnormal enhancement map is minimized, the lesion boundary is iteratively updated and morphologically constrained, noise interference is suppressed and boundary continuity is improved, and finally, a logical AND operation is performed with the lung parenchyma mask to generate a three-dimensional lesion segmentation mask.

[0014] Furthermore, the multi-dimensional structured quantitative indicators extracted by the self-supervised lesion segmentation and quantification module are eight-dimensional quantitative features divided according to HU value density intervals. Specifically, they include: the standardized proportion of lesion volume to total lung volume in four density intervals: less than -600HU, -600HU to -400HU, -400HU to -200HU, and greater than -200HU; and the average spatial distance between the centroid of the lesion and the geometric center of the lung in each density interval. The self-supervised lesion segmentation and quantification module also has a built-in principal component analysis unit, which is used to construct a unified disease burden characterization principal axis PC1 based on the eight-dimensional quantitative features and patient age and gender metadata, so as to realize the mapping of disease changes and progression assessment of CT images at different time points.

[0015] Furthermore, in the multimodal joint pre-trained encoding module, the image encoding unit adopts a hierarchical three-dimensional SwingTransformer backbone network to extract multi-scale features of CT images using window attention and hierarchical feature pyramid structure. The input is complete lung parenchyma CT volume data after uniform size resampling. The text encoding unit adopts a language Transformer network pre-trained on chest radiology report corpus to encode standardized radiology reports into fixed-length text embedding vectors. During training, the parameters of the text encoding unit are kept frozen to stabilize the semantic embedding space.

[0016] Furthermore, the adversarial perturbation enhancement unit of the multimodal joint pre-training coding module is used to generate adversarial variants for each standardized radiology report during the pre-training phase. The adversarial variants only rewrite and replace key fields related to lung parenchymal lesions. Key fields include lesion texture, distribution range, anatomical location, severity, and lesion type label, while other non-lesion content, stylistic structure, and sentence pattern remain unchanged. During training, the original report and the adversarial variant are input simultaneously to construct a contrast constraint that "the similarity between the original CT and the original report is higher than the similarity between the original CT and the adversarial variant," thereby enhancing the model's ability to identify fine-grained pathological features.

[0017] Furthermore, the lesion soft label distillation unit of the multimodal joint pre-trained coding module uses the abnormal enhancement map output by the self-supervised lesion segmentation and quantization module as the soft label supervision signal, and adds a lightweight decoding head to the multi-scale features output by the image coding unit to predict the lesion probability map consistent with the soft label. The consistency between the prediction result and the soft label is constrained by the soft supervision loss. The metadata prediction unit, based on the global features output by the image coding unit, simultaneously performs patient age regression and gender classification prediction tasks, realizes the model's encoding of patient demographic features and collection background information, and improves cross-domain generalization ability.

[0018] Furthermore, the multi-task diagnosis and prognosis prediction module incorporates multiple sets of parallel task heads, corresponding to downstream tasks such as acute respiratory failure identification, acute respiratory distress syndrome diagnosis, non-invasive P / F ratio estimation, Berlin standard severity grading, right ventricular dysfunction identification, and 28-day survival prognosis prediction. The multi-task diagnosis and prognosis prediction module also includes a threshold warning unit. When the output diagnostic probability, risk score, and physiological estimate exceed a preset threshold range, a graded warning signal is automatically triggered, and the prediction results and lesion heatmap visualization data are simultaneously pushed to the clinical interactive terminal and the hospital information system.

[0019] Furthermore, the model incorporates a domain adaptation module, which achieves consistent processing of multi-center heterogeneous data across devices and institutions through feature standardization and distribution alignment mechanisms, reducing domain shifts caused by differences in scanning parameters and slice thicknesses of different CT devices. The model also supports temporal analysis of multi-phase CT images, enabling tracking of patient disease progression and estimation of image-derived biological age, providing dynamic monitoring and risk stratification for patients with acute respiratory distress syndrome.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] The model described in this invention enables continuous and quantitative monitoring of pulmonary inflammatory lesions and supports automatic identification of -ARF and ARDS. It also achieves P / F ratio estimation, Berlin standard severity grading, ARDS-related right ventricular dysfunction (RVD) assessment, and 28-day survival prognosis prediction. By constructing a unified image-physiology joint representation system, diagnostic consistency and risk stratification accuracy are improved. The model employs a self-supervised lesion segmentation mechanism to generate stable lesion masks, achieving reproducible quantification of pulmonary lesion burden. Abnormal tissues are separated from normal lung structures through synthetic image reconstruction and voxel-level difference mapping, thereby reducing reliance on manual annotation and subjective experience. Combined with an adversarial enhancement visual-linguistic alignment strategy, image features and radiological semantic information are finely matched within the same embedding space, enhancing the model's sensitivity to key pathological attributes and cross-center generalization ability. The system integrates a multimodal pre-training framework and a task-adaptive fine-tuning module, maintaining stable performance in multi-center heterogeneous data environments. Experimental results show that in a multi-center cohort, the Dice coefficient for lesion segmentation, the AUC for ARDS diagnosis, and the Pearson correlation coefficient for P / F ratio estimation are all significant. The model supports real-time inference and result output, and can synchronously transmit predicted scores and risk indicators to clinical terminals, realizing a closed-loop data flow from CT acquisition to clinical decision-making. Through the above technical solution, this invention achieves structured quantitative expression and multi-task integrated analysis of pulmonary inflammatory lesions, providing a standardized and scalable intelligent support tool for the diagnosis, grading, and prognostic assessment of severe pneumonia and ARDS. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the automated disease workflow and clinical integration in this invention;

[0023] In the figure: a represents ARDS data preparation, quantization, and application; b represents the AutoARDS process and pre-training strategy.

[0024] Figure 2 A schematic diagram of ARDS lung injury and disease progression assessment based on quantitative CT.

[0025] In the figure: a) Dataset information for self-supervised and fully supervised training; “LLD” = Mild lung abnormalities without respiratory failure; b) Quantitative evaluation of segmentation performance; “DSC” = Dice Score Coefficient, “Sen” = Sensitivity, “Spe” = Specificity; Best results are marked in bold; c) Visualization of the self-supervised segmentation workflow and segmentation results; Image magnification to focus on the lung region for clearer visualization; d) CT volume t-SNE clustering based on patient metadata information and lesion quantification, which naturally indicates disease type and severity even without supervision; e) PCA analysis can indicate the progression of ARDS; we use changes in the P / F ratio to represent recovery or deterioration; the x-axis represents the change in PCA component values ​​between two consecutive measurements, with the red line 0 indicating no change; the y-axis represents the probability density of cases; in ARDS recovery cases, PCA component values ​​typically decrease; conversely, in deterioration cases, the value typically increases; the upper half shows studies using internal datasets, and the lower half shows studies using external datasets.

[0026] Figure 3 This is a schematic diagram illustrating the comprehensive evaluation of ARDS diagnosis, P / F ratio estimation, and severity stratification in the internal dataset of this invention.

[0027] In the figure: a) is the ROC curve for ARF diagnosis based on all CT scan results under different models; b) is the ROC curve for ARDS detection of all ARF cases, and is compared with manual interpretation (four resident physicians and three junior ICU doctors); c) is a scatter plot of the P / F ratio of predicted value to true value, and the correlation index of each method is marked; d) is the confusion matrix of severity classification (mild, moderate, severe), showing the overall classification accuracy (ACC) of each method.

[0028] Figure 4 This is a schematic diagram illustrating the comprehensive evaluation of ARDS diagnosis, P / F ratio estimation, and severity stratification in an external dataset in this invention.

[0029] In the figure: a) is the ROC curve for ARF diagnosis based on all CT scan results under different models; b) is the ROC curve for ARDS detection of all ARF cases, and is compared with manual interpretation (four resident physicians and three junior ICU doctors); c) is a scatter plot of the P / F ratio of predicted value to true value, and the correlation index of each method is marked; d) is the confusion matrix of severity classification (mild, moderate, severe), showing the overall classification accuracy (ACC) of each method.

[0030] Figure 5 For a comprehensive evaluation of ARDS prognosis, RVD prediction, and imaging-derived biological age;

[0031] In the figure: a is the area under the curve of the time dependence of the 28-day survival rate prediction for the internal dataset, with the shaded area representing the 95% confidence interval; b is the area under the curve of the time dependence of the 28-day survival rate prediction for the external dataset, with the shaded area representing the 95% confidence interval; c is the ROC curve for the prediction of ARDS-related right ventricular dysfunction (RVD), with the shaded area representing the 95% confidence interval; d is the prediction error distribution density map of image-derived biological age, which marks the prediction bias statistics and model fitting accuracy indices for different health status groups. Detailed Implementation

[0032] The following will refer to the appendices in the embodiments of the present invention. Figure 1-5 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figure 1-5 The present invention provides an artificial intelligence model for quantitative assessment of lung inflammatory lesions based on CT images, comprising a CT image preprocessing module, a self-supervised lesion segmentation and quantification module, a multimodal joint pre-training coding module, and a multi-task diagnosis and prognosis prediction module connected in sequence. Each module achieves feature interaction and data encapsulation through a unified feature embedding space.

[0034] The CT image preprocessing module is used to sequentially perform spatial resampling, HU value intensity normalization, and automatic lung parenchyma region segmentation on the input chest CT three-dimensional volume data, and output standardized lung parenchyma CT volume data.

[0035] The self-supervised lesion segmentation and quantification module is used to synthesize and reconstruct healthy lung structures from standardized lung parenchyma CT volume data using a 3D generative reconstruction network, generating healthy reference CT volume data that matches the individual anatomical structure of the input CT. It calculates the voxel-level HU value difference between the standardized lung parenchyma CT volume data and the healthy reference CT volume data, generating an abnormal enhancement map that retains only high-density abnormal regions. Based on a variational constraint optimization model, iterative optimization of lesion boundaries and noise suppression are performed on the abnormal enhancement map to generate a 3D lesion segmentation mask. Based on the 3D lesion segmentation mask, multi-dimensional structured quantification indicators are extracted, and a lesion quantification feature vector is constructed. These multi-dimensional structured quantification indicators include at least the proportion of lesion volume to the total lung volume in different HU density intervals and the spatial distribution parameters of lesions in each HU density interval relative to the geometric center of the lung.

[0036] The multimodal joint pre-trained encoding module incorporates image encoding units, text encoding units, adversarial perturbation enhancement units, lesion soft label distillation units, and metadata prediction units. The image encoding unit extracts features from standardized lung parenchyma CT data to generate image embedding features. The text encoding unit extracts features from the standardized radiology report text corresponding to the input CT scan to generate text embedding features. A cross-modal contrastive learning mechanism maps image embedding features and text embedding features to a unified feature embedding space, completing the image feature mapping.

[0037] The system features bidirectional fine-grained alignment with radiological semantic features; an adversarial perturbation enhancement unit to generate adversarial variants of key semantics related to lesions in radiological reports, enhancing the model's ability to identify fine-grained pathological features through contrastive constraints; a lesion soft-label distillation unit to use abnormal enhancement maps as lesion soft labels, guiding the image coding unit to learn features of lung parenchymal lesion regions; a metadata prediction unit to integrate patient demographic metadata for joint modeling, improving the model's domain adaptability across devices and institutions; and finally, a unified multimodal embedding feature that integrates image embedding features, lesion quantification features, textual semantic features, and patient metadata features.

[0038] The multi-task diagnosis and prognosis prediction module is used to complete downstream multi-task parallel inference based on unified multimodal embedding features and through task adaptive fine-tuning. The downstream tasks include at least acute respiratory failure identification, acute respiratory distress syndrome diagnosis, non-invasive estimation of P / F ratio, Berlin standard severity classification, identification of right ventricular dysfunction related to acute respiratory distress syndrome, and prediction of short-term survival prognosis of patients. Finally, it outputs structured quantitative assessment results of pulmonary inflammatory lesions and clinical risk warning information.

[0039] The specific steps of the CT image preprocessing module for standardization are as follows: First, the input chest CT three-dimensional volume data is resampled isotropically to unify the voxel spacing to a preset standard value; then, the HU value intensity is normalized for the resampled CT data, limiting the HU value to the effective lung range of [-1024, 400] and linearly mapping it to the range of [0, 1]; finally, the pre-trained lung segmentation network completes the automatic extraction and masking of the lung parenchyma region, removes interference data from the chest wall, mediastinum, and background regions, and outputs standardized lung parenchyma CT volume data.

[0040] The self-supervised lesion segmentation and quantization module incorporates a 3D CycleGAN generative reconstruction network to generate healthy reference CT volume data. The 3D CycleGAN generative reconstruction network includes a first generator, a second generator, a first discriminator, and a second discriminator. The first generator maps abnormal lung CT domain data to healthy lung CT domain data, the second generator maps healthy lung CT domain data back to abnormal lung CT domain data, the first discriminator determines the authenticity of healthy domain images, and the second discriminator determines the authenticity of abnormal domain images. Both the first and second generators employ a 3D convolutional residual structure, consisting of an input layer, a double downsampling layer, six 3D ResNet residual blocks, a double upsampling layer, and an output layer. During training, adversarial loss and cycle consistency loss are introduced for joint optimization.

[0041] The self-supervised lesion segmentation and quantification module generates abnormal enhancement maps and lesion segmentation masks in the following way: It calculates the voxel-by-voxel HU difference between standardized lung parenchyma CT volume data and healthy reference CT volume data, and then uses the difference formula...

[0042]

[0043] Only the difference in density between abnormal areas in the original CT scan and those in the healthy reference CT scan is retained. For raw CT body data, For healthy reference CT scan data, interval clipping and normalization are performed on the difference results to generate a three-dimensional abnormal enhancement map. Based on the Chan-Vese variational constraint optimization model, the region energy of the abnormal enhancement map is minimized, the lesion boundary is iteratively updated and morphologically constrained, noise interference is suppressed and boundary continuity is improved, and finally, a logical AND operation is performed with the lung parenchyma mask to generate a three-dimensional lesion segmentation mask.

[0044] The self-supervised lesion segmentation and quantification module extracts multi-dimensional structured quantitative indicators, which are eight-dimensional quantitative features divided according to HU value density intervals. Specifically, these include: the standardized proportion of lesion volume to total lung volume in four density intervals: less than -600HU, -600HU to -400HU, -400HU to -200HU, and greater than -200HU; and the average spatial distance between the centroid of the lesion and the geometric center of the lung in each density interval. The self-supervised lesion segmentation and quantification module also has a built-in principal component analysis unit, which is used to construct a unified disease burden characterization principal axis PC1 based on the eight-dimensional quantitative features and patient age and gender metadata, so as to realize the mapping of disease changes and progression assessment of CT images at different time points.

[0045] In the multimodal joint pre-trained encoding module, the image encoding unit adopts a hierarchical three-dimensional SwinTransformer backbone network to extract multi-scale features of CT images using window attention and hierarchical feature pyramid structure. The input is complete lung parenchyma CT volume data after uniform size resampling. The text encoding unit adopts a language Transformer network pre-trained on chest radiology report corpus to encode standardized radiology reports into fixed-length text embedding vectors. During training, the parameters of the text encoding unit are kept frozen to stabilize the semantic embedding space.

[0046] The adversarial perturbation enhancement unit of the multimodal joint pre-training coding module is used to generate adversarial variants for each standardized radiology report during the pre-training phase. The adversarial variants only rewrite and replace key fields related to lung parenchymal lesions, including lesion texture, distribution range, anatomical location, severity, and lesion type label. Other non-lesion content, stylistic structure, and sentence pattern remain unchanged. During training, the original report and the adversarial variant are input simultaneously to construct a contrast constraint that "the similarity between the original CT and the original report is higher than the similarity between the original CT and the adversarial variant", thereby enhancing the model's ability to identify fine-grained pathological features.

[0047] The lesion soft-label distillation unit of the multimodal joint pre-trained coding module uses the abnormal enhancement map output by the self-supervised lesion segmentation and quantization module as the soft-label supervision signal. It adds a lightweight decoding head to the multi-scale features output by the image coding unit to predict the lesion probability map consistent with the soft label. The consistency between the prediction result and the soft label is constrained by the soft supervision loss. The metadata prediction unit, based on the global features output by the image coding unit, simultaneously performs patient age regression and gender classification prediction tasks, realizes the model's encoding of patient demographic features and collection background information, and improves cross-domain generalization ability.

[0048] The multi-task diagnosis and prognosis prediction module has multiple sets of parallel task heads, which correspond to downstream tasks such as acute respiratory failure identification, acute respiratory distress syndrome diagnosis, non-invasive P / F ratio estimation, Berlin standard severity classification, right heart function abnormality identification, and 28-day survival prognosis prediction. The multi-task diagnosis and prognosis prediction module also has a threshold warning unit. When the output diagnostic probability, risk score, and physiological estimate exceed the preset threshold range, a graded warning signal is automatically triggered, and the prediction results and lesion heat map visualization data are pushed to the clinical interactive terminal and the hospital information system simultaneously.

[0049] The model incorporates a domain adaptation module, which achieves consistent processing of multi-center heterogeneous data across devices and institutions through feature standardization and distribution alignment mechanisms, reducing domain shifts caused by differences in scanning parameters and slice thicknesses of different CT devices. The model also supports temporal analysis of multi-phase CT images, enabling tracking of patient disease progression and estimation of image-derived biological age, providing dynamic monitoring and risk stratification for patients with acute respiratory distress syndrome.

[0050] A lesion quantification method for an artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images includes the following steps: First, preprocessing and standardizing the CT three-dimensional volume data to automatically segment and obtain the lung parenchyma region; constructing a differentially enhanced image within a defined lung area and generating an initial lesion region using a variational constraint model; optimizing the boundary and suppressing noise in the lesion region to obtain a stable lesion segmentation mask. Based on this, multi-dimensional quantitative indicators are extracted, including the lesion volume ratio in each density interval and the spatial distribution parameters of the lesion relative to the lung geometric center, to form a structured quantitative feature vector; further, a unified disease representation axis is constructed through principal component analysis (PCA) to realize the mapping of disease changes and progression assessment between different time points.

[0051] An overall framework for the AI ​​model for quantitative assessment of pulmonary inflammatory lesions based on CT images includes the following components:

[0052] (1) Adversarial enhancement vision-language joint pre-training module, which is used to align CT image features with corresponding radiological text information across modalities, and to build an enhancement training mechanism by perturbating key semantic elements to improve the model’s ability to identify fine-grained pathological features.

[0053] (2) A multimodal coding and fusion module is used to receive structured indicators and original image features from the lesion quantification module and perform unified embedding expression; the encoder adopts parameter optimization and computation compression strategies to achieve low-power inference and efficient data output;

[0054] (3) An extended interface module is used to access multi-center data sources and external clinical information systems to achieve standardized integration and cross-platform deployment of heterogeneous data. The framework also includes a clinical interactive interface for real-time display of model running status and prediction results; when the diagnostic probability, risk score or physiological estimate exceeds the preset threshold range, the system automatically triggers prompts or warning signals to support clinical decision-making and intervention.

[0055] A system based on an artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions using CT images can adapt to various clinical environments, including heterogeneity of multicenter datasets and dynamic changes in pulmonary inflammatory conditions.

[0056] To perform self-supervised lesion quantification, we propose a synthesis-contrast framework (…). Figure 1 b). First, lung regions are extracted from the entire CT volume to focus on the lung parenchyma, and then a model is generated to reconstruct the corresponding healthy lung structures. Lesions are presented as voxel-by-voxel Huntsfield unit (HU) deviations between the original and synthesized volumes, forming a difference map. Then, we employ a variational model to refine the difference map into precise lesion masks. From these masks, we extract eight quantitative indicators to provide a detailed characterization of lesion burden. Finally, the refined segmentation results and derived indicators are integrated into the downstream model as information priors tailored for ARDS-specific applications. To capture fine-grained and diagnostically relevant information for ARDS, a task-unified multi-model pre-training framework is provided ( Figure 1 b). The framework first incorporates a contrastive language-image pre-training (CLIP) paradigm, aligning CT volumes with their corresponding radiology reports to learn joint representations. In addition to this baseline alignment, we introduce a textual adversarial perturbation strategy to enhance sensitivity to subtle but clinically significant cues. Specifically, for each structured radiology report, we generate adversarial variants by selectively altering key descriptors (e.g., disease label, anatomical location, lesion distribution, severity, or texture) while keeping the rest of the report unchanged. These targeted perturbations force the model to precisely focus on linguistic attributes that can differentiate the condition, thereby improving its ability to capture ARDS-specific diagnostic signals. AutoARDS was first pre-trained on the CT-RATE dataset and then fine-tuned on our multicenter cohort (containing 6,835 chest CT scans from 6,153 individuals, along with corresponding demographic information and radiology reports) (Table 1).

[0057] Table 1. Detailed information on the AutoARF development dataset.

[0058]

[0059] For subsequent analysis, all CT scans were divided into four groups: healthy controls, limited lung disease (LLD, representing mild lung abnormalities without respiratory failure), non-ARDS ARF (hereinafter referred to as ARF), and ARDS. Based on our dataset, AutoARDS supports a wide range of downstream tasks, from ARDS identification to ABG assessment, RVD estimation, and outcome prediction. Figure 1 a) By unifying these tasks within a single framework, AutoARDS not only improves diagnostic accuracy but also enhances clinical management efficiency, thus providing a foundational and scalable tool for timely analysis and decision-making in ARDS care.

[0060] Self-supervised ARDS lesion segmentation

[0061] The proposed segmentation method was evaluated by comparing it with several cutting-edge pneumonia segmentation methods. A fully supervised benchmark model was trained from scratch using nn-UNet, with datasets including two public pneumonia datasets, CLISD and MosMedData, totaling 70 CT scan images. A Chan-vese (CV) model and direct thresholding were also used for ablation studies. For ARDS patients, 20 CT images were selected as an independent test set, and these were strictly excluded from training to prevent data leakage. Segmentation performance was quantitatively evaluated using Dice similarity coefficient (DSC), sensitivity (Sen), and specificity (Spe). The results are presented in quantitative comparison charts. Figure 2 b) and qualitative visualization analysis ( Figure 2 c).

[0062] Experimental results show that models trained solely on pneumonia datasets cannot be directly applied to ARDS lesion segmentation, resulting in poor performance and inadequate segmentation. Specifically, these models exhibit significantly low sensitivity (below 60%), and a large number of lesions fail to be identified. This limitation can be attributed to the significant morphological differences between COVID-19 and ARDS lesions. Figure 2 a) In contrast, the self-supervised learning method, trained on a diverse dataset of real-world ARDS cases, successfully identified the vast majority of lesions and demonstrated superior performance. This method achieved a DSC value of 75.62±9.32% and a Sen value of 80.13±12.13%, significantly outperforming all other methods while maintaining competitive specificity (6.26±6.90%). Among the two optimization strategies evaluated, the cross-validation-based method achieved better DSC and Sen values ​​compared to the thresholding method.

[0063] Lesion quantification enables clinically interpretable quantification and tracking of ARDS.

[0064] Lung lesions are a core feature of ARDS lung pathology, and their quantitative assessment provides objective imaging biomarkers, which are crucial for ARDS diagnosis. Eight quantitative indicators were extracted from lesions and lung tissue segmented from CT scans. Specifically, lesions were divided into four groups based on HU density: below -600 HU, -600 HU to -400 HU, -400 HU to -200 HU, and above -200 HU. For each density category, the total lesion volume was calculated and standardized based on the total lung volume to obtain the volume ratio.

[0065] like Figure 2 As shown in Figure .c, the clustering results revealed four natural groups: the healthy control group clustered tightly on the right; LLD cases were distributed in the intermediate transitional region; non-ARDS-related acute respiratory distress syndrome cases were sandwiched between the LLD and ARDS groups, reflecting an intermediate level of lesion burden; while ARDS cases formed an independent cluster on the left, characterized by a significantly higher lesion involvement rate. This result indicates that the diagnosis of ARDS can be directly determined through lesion characteristics, and lesion quantification analysis can also provide important reference for the diagnosis and differentiation of ARDS.

[0066] This study analyzed paired CT scans and P / F ratios acquired at different time points in the same group of patients with acute respiratory distress syndrome (ARDS). An increase in the P / F ratio at the second time point was considered an improvement in the patient's condition; conversely, a decrease in the P / F ratio was considered a deterioration. In the graph, the x-axis represents the range of PC1 value variation, with 0 indicating no change, and the y-axis displays the probability density distribution of the paired CT scans, visually presenting the probability distribution within different ranges of PC1 value variation. Figure 2 As shown in the .d file, the analysis results for the internal and external datasets revealed significant differences: in ARDS recovery cases (increased OI values), PC1 values ​​mainly showed a decreasing trend (88.43% in the internal dataset, 94.00% in the external dataset); while in cases of worsening condition (decreased P / F ratio), PC1 values ​​mainly showed an increasing trend (84.62% in the internal dataset, 90.38% in the external dataset). These results indicate that Principal Component 1 (PC1) can serve as a stable and reliable quantitative indicator and biomarker for monitoring the progression or improvement of acute respiratory distress syndrome (ARDS), with an accuracy exceeding 85%. Logistic regression analysis further confirmed a significant correlation between PC1 values ​​and changes in the P / F ratio, a lung function indicator. This longitudinal reproducibility not only highlights the value of quantitative CT detection as a diagnostic parameter but also demonstrates its potential as a dynamic monitoring tool, providing a scientific basis for adjusting treatment plans and intervention timing.

[0067] When blood gas analysis results are unavailable, oxygen saturation (SpO2) is currently the most commonly used surrogate indicator. However, recent research shows that the SpO2 / FiO2 (SF ratio) only correctly identifies the degree of disease change between two consecutive data points in 19.6% of cases. In contrast, the AutoARDS system achieved over 85% consistency in tracking clinical trajectories.

[0068] ARDS Diagnosis

[0069] Based on a reproducible quantitative CT analysis system, a two-stage supervised binary classification problem was implemented: the first stage used CT scans for preliminary screening to distinguish acute respiratory failure (ARF) cases from all scan results; the second stage detected acute respiratory distress syndrome (ARDS) cases from the identified ARF cohort for accurate diagnosis. Given the limited number of studies related to ARDS diagnosis, AutoARDS was compared with two representative pre-trained CT baseline models—CT-CLIP and 3DMAE. CT-CLIP employs an image-language contrastive learning strategy, a coarse-grained pre-training method; while 3DMAE is based on a masked autoencoder architecture, a technique proven to more effectively capture high-frequency information. Furthermore, a baseline model with the same architecture as AutoARDS was constructed to evaluate the specific contribution of pre-training to performance improvement. The baseline model was trained on a novel dataset. Five-fold cross-validation was used for the internal dataset, and the average results of all folds were reported; for the external dataset, the average performance metrics of five independently trained models were calculated. Relevant experimental results and receiver operating characteristic (ROC) curves are shown in [reference to relevant data]. Figure 3 and Figure 4 .

[0070] In the ARF classification task on the internal dataset ( Figure 3 a) The AutoARDS model achieved a high area under the curve (AUC) of 0.9748, an overall accuracy of 0.9275, and an F1 score of 0.9282. This was specifically designed for the ARDS detection task in ARF cases. Figure 3 (b) AutoARDS also showed an AUC of 0.8874, an accuracy of 0.8820, and an F1 score of 0.9297.

[0071] In external dataset testing, AutoARDS continues to demonstrate strong performance, achieving an AUC of 0. In ARF detection, the AutoARDS model performs exceptionally well: an overall accuracy of 0.9181 and an F1 score of 0.9267. Figure 4a) In ARDS detection, the AUC value reached 0.8240, the accuracy reached 0.8203, and the F1 score was 0.8907. Compared with the baseline model trained from scratch, the AutoARDS model showed less performance degradation, fully demonstrating the effectiveness of pre-training techniques in improving cross-dataset generalization ability. We also organized an expert review team, consisting of three junior ICU physicians and four interns, to evaluate ARDS detection in ARF cases. Each reviewer was required to review 25 CT images and obtain patient age and gender information. The results showed that the diagnostic performance of all human reviewers was lower than the ROC curve of AutoARDS ( Figure 3 a, Figure 4 a) In ARDS diagnosis, AutoARDS significantly outperformed human review, with an average F1 score 0.4463 higher than that of interns and 0.2618 higher than that of junior ICU physicians. Notably, the worst-performing trainee had an F1 score as low as 0.2857, highlighting the difficulty of consistently interpreting borderline acute respiratory distress syndrome cases in the absence of additional arterial blood gas analysis or prior imaging data.

Claims

1. An artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images, characterized in that, It includes a CT image preprocessing module, a self-supervised lesion segmentation and quantization module, a multimodal joint pre-training coding module, and a multi-task diagnosis and prognosis prediction module that are connected in sequence. Each module realizes feature interaction and data encapsulation through a unified feature embedding space. The CT image preprocessing module is used to sequentially perform spatial resampling, HU value intensity normalization and automatic lung parenchyma region segmentation on the input chest CT three-dimensional volume data, and output standardized lung parenchyma CT volume data. The self-supervised lesion segmentation and quantification module is used to synthesize and reconstruct healthy lung structures from the standardized lung parenchyma CT volume data through a three-dimensional generative reconstruction network, generating healthy reference CT volume data that matches the individual anatomical structure of the input CT; and to calculate the difference in voxel-level HU values ​​between the standardized lung parenchyma CT volume data and the healthy reference CT volume data, generating an abnormal enhancement map that retains only the high-density abnormal areas. Based on the variational constraint optimization model, the abnormal enhancement map is iteratively optimized for lesion boundaries and noise suppression is performed to generate a three-dimensional lesion segmentation mask. Based on the three-dimensional lesion segmentation mask, multi-dimensional structured quantitative indicators are extracted and lesion quantitative feature vectors are constructed. The multi-dimensional structured quantitative indicators include at least the proportion of lesion volume in different HU density intervals to the total lung volume and the spatial distribution parameters of lesions in each HU density interval relative to the geometric center of the lung. The multimodal joint pre-trained coding module includes an image coding unit, a text coding unit, an adversarial perturbation enhancement unit, a lesion soft label distillation unit, and a metadata prediction unit. The image coding unit is used to extract features from standardized lung parenchyma CT data and generate image embedding features. The text encoding unit is used to extract features from the standardized radiology report text corresponding to the input CT and generate text embedding features; The image embedding features and text embedding features are mapped to a unified feature embedding space through a cross-modal contrastive learning mechanism, achieving bidirectional fine-grained alignment between image features and radiological semantic features; the adversarial perturbation enhancement unit is used to generate adversarial variants of key semantics related to lesions in the radiological report, and enhances the model's ability to identify fine-grained pathological features through contrastive constraints; the lesion soft label distillation unit is used to use the abnormal enhancement map as a lesion soft label to guide the image coding unit to learn the features of the lung parenchymal lesion region; The metadata prediction unit is used to integrate patient demographic metadata for joint modeling, improving the model's domain adaptability across devices and institutions; the final output is a unified multimodal embedding feature that integrates image embedding features, lesion quantification features, text semantic features and patient metadata features. The multi-task diagnosis and prognosis prediction module is used to complete downstream multi-task parallel inference through task adaptive fine-tuning based on the unified multimodal embedding features. The downstream tasks include at least acute respiratory failure identification, acute respiratory distress syndrome diagnosis, non-invasive estimation of P / F ratio, Berlin standard severity classification, identification of right ventricular dysfunction related to acute respiratory distress syndrome, and prediction of short-term survival prognosis of patients. Finally, it outputs structured quantitative assessment results of pulmonary inflammatory lesions and clinical risk warning information.

2. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 1, characterized in that, The specific steps of the CT image preprocessing module for standardization are as follows: First, isotropic spatial resampling is performed on the input chest CT three-dimensional volume data to unify the voxel spacing to a preset standard value; then, HU value intensity normalization is performed on the resampled CT data, limiting the HU value to the effective lung range of [-1024, 400] and linearly mapping it to the range of [0, 1]; finally, automatic extraction and masking of the lung parenchyma region is completed through a pre-trained lung segmentation network to remove interference data from the chest wall, mediastinum, and background regions, and output standardized lung parenchyma CT volume data.

3. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 2, characterized in that, The self-supervised lesion segmentation and quantization module incorporates a three-dimensional CycleGAN generative reconstruction network to generate healthy reference CT volume data. The three-dimensional CycleGAN generative reconstruction network includes a first generator, a second generator, a first discriminator, and a second discriminator. The first generator maps abnormal lung CT domain data to a healthy lung CT domain, the second generator maps healthy lung CT domain data back to the abnormal lung CT domain, the first discriminator determines the authenticity of the healthy domain image, and the second discriminator determines the authenticity of the abnormal domain image. Both the first and second generators employ a three-dimensional convolutional residual structure, consisting of an input layer, a double downsampling layer, six three-dimensional ResNet residual blocks, a double upsampling layer, and an output layer. During training, adversarial loss and cycle consistency loss are introduced for joint optimization.

4. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 3, characterized in that, The self-supervised lesion segmentation and quantification module generates abnormal enhancement maps and lesion segmentation masks in the following manner: It calculates the voxel-by-voxel HU difference between standardized lung parenchyma CT volume data and healthy reference CT volume data, and then uses the difference formula... Only the difference in density between abnormal areas in the original CT scan and those in the healthy reference CT scan is retained. For raw CT body data, For healthy reference CT scan data, interval clipping and normalization are performed on the difference results to generate a three-dimensional abnormal enhancement map. Based on the Chan-Vese variational constraint optimization model, the region energy of the abnormal enhancement map is minimized, the lesion boundary is iteratively updated and morphologically constrained, noise interference is suppressed and boundary continuity is improved, and finally, a logical AND operation is performed with the lung parenchyma mask to generate a three-dimensional lesion segmentation mask.

5. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 1, characterized in that, The self-supervised lesion segmentation and quantification module extracts multi-dimensional structured quantitative indicators, which are eight-dimensional quantitative features divided according to HU value density intervals. Specifically, these include: the standardized proportion of lesion volume to total lung volume in four density intervals: less than -600HU, -600HU to -400HU, -400HU to -200HU, and greater than -200HU; and the average spatial distance between the centroid of the lesion and the geometric center of the lung in each density interval. The self-supervised lesion segmentation and quantification module also has a built-in principal component analysis unit, which is used to construct a unified disease burden characterization principal axis PC1 based on the eight-dimensional quantitative features and patient age and gender metadata, so as to realize the mapping of disease changes and progression assessment of CT images at different time points.

6. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 1, characterized in that, In the multimodal joint pre-trained encoding module, the image encoding unit adopts a hierarchical three-dimensional SwinTransformer backbone network to extract multi-scale features of CT images using window attention and hierarchical feature pyramid structure. The input is complete lung parenchyma CT volume data after uniform size resampling. The text encoding unit adopts a language Transformer network pre-trained on chest radiology report corpus to encode standardized radiology reports into fixed-length text embedding vectors. During training, the parameters of the text encoding unit are kept frozen to stabilize the semantic embedding space.

7. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 6, characterized in that, The adversarial perturbation enhancement unit of the multimodal joint pre-training coding module is used to generate adversarial variants for each standardized radiology report during the pre-training phase. The adversarial variants only rewrite and replace key fields related to lung parenchymal lesions. Key fields include lesion texture, distribution range, anatomical location, severity, and lesion type label. Other non-lesion content, stylistic structure, and sentence pattern remain unchanged. During training, both the original report and the adversarial variant are input simultaneously. A contrast constraint is constructed that "the similarity between the original CT and the original report pair is higher than the similarity between the original CT and the adversarial variant pair" to enhance the model's ability to identify fine-grained pathological features.

8. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 1, characterized in that, The lesion soft-label distillation unit of the multimodal joint pre-trained coding module uses the abnormal enhancement map output by the self-supervised lesion segmentation and quantization module as the soft-label supervision signal. It adds a lightweight decoding head to the multi-scale features output by the image coding unit to predict the lesion probability map consistent with the soft label. The consistency between the prediction result and the soft label is constrained by the soft supervision loss. The metadata prediction unit, based on the global features output by the image coding unit, simultaneously performs patient age regression and gender classification prediction tasks, realizing the model's encoding of patient demographic features and collection background information, and improving cross-domain generalization ability.

9. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 1, characterized in that, The multi-task diagnosis and prognosis prediction module has multiple sets of parallel task heads, which correspond to downstream tasks such as acute respiratory failure identification, acute respiratory distress syndrome diagnosis, non-invasive P / F ratio estimation, Berlin standard severity classification, right ventricular dysfunction identification, and 28-day survival prognosis prediction. The multi-task diagnosis and prognosis prediction module also has a threshold warning unit. When the output diagnostic probability, risk score, and physiological estimate exceed the preset threshold range, a graded warning signal is automatically triggered, and the prediction results and lesion heat map visualization data are pushed to the clinical interactive terminal and the hospital information system simultaneously.

10. The artificial intelligence model for quantitative assessment of pulmonary inflammatory lesions based on CT images according to claim 1, characterized in that, The model incorporates a domain adaptation module, which achieves consistent processing of multi-center heterogeneous data across devices and institutions through feature standardization and distribution alignment mechanisms, reducing domain shifts caused by differences in scanning parameters and slice thicknesses of different CT devices. The model also supports temporal analysis of multi-phase CT images, enabling tracking of patient disease progression and estimation of image-derived biological age, providing dynamic monitoring and risk stratification for patients with acute respiratory distress syndrome.

Images