Rare lung disease CT image automatic auxiliary diagnosis system based on deep learning

Abstract

The invention provides a lung rare disease CT image automatic auxiliary diagnosis system based on deep learning, and belongs to the technical field of deep learning and auxiliary diagnosis, and the system comprises a data collection and preprocessing module which is used for carrying out the privacy processing, format conversion, standardization processing and data enhancement of CT image data, and outputting a three-dimensional input block meeting the model input requirements; the three-dimensional feature modeling and training module is used for taking the three-dimensional compact connection neural network as a core, learning morphological and distribution features of cystic lesions through a training strategy, and constructing and training to obtain a disease classification model; the online reasoning and interpretable display module is used for performing probability calibration on the trained model output to generate a visual result; and the clinical application module is used for receiving the current preprocessed CT image, inputting the CT image into the calibrated model for processing and analysis, and outputting a classification result and a visualization basis. High-dimensional structure modeling and credible reasoning of the rare diffuse lung disease image are realized, and AI auxiliary diagnosis is facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and auxiliary diagnosis, and particularly relates to a deep learning-based automatic auxiliary diagnosis system for CT images of rare lung diseases. BACKGROUND

[0002] Lymphangioleiomyomatosis (LAM) is a rare, progressive cystic lung disease that mainly occurs in women of childbearing age. The disease is characterized by diffuse infiltration of abnormal smooth muscle-like cells in the lung parenchyma and lymphatic system, leading to progressive cystic destruction, decreased lung function, and respiratory failure. Imaging shows diffuse distribution of thin-walled cysts, often accompanied by increased lung volume, vascular compression, and interstitial destruction. Due to the extremely low incidence (estimated at about 3-8 per million), most primary or comprehensive hospitals lack experienced imaging specialists to identify LAM, so it is often confused with other diffuse cystic lung diseases (DCLD) such as lymphangioma, Langerhans cell histiocytosis, Birt-Hogg-Dubé syndrome, and idiopathic pulmonary emphysema, resulting in a high misdiagnosis rate and significant delay in diagnosis.

[0003] Currently, clinical diagnosis mainly relies on chest CT, and experienced imaging physicians make a comprehensive judgment based on the number, size, distribution, and lung density changes of the cysts. Diagnosis needs to be combined with patient gender, age, hormone status, and clinical symptoms, but in actual work, most image interpretation is still manually operated. Doctors need to read the film layer by layer and subjectively judge the shape and distribution of the cysts, which is extremely time-consuming and depends on personal experience. Multi-center studies show that the consistency (κ value) of image interpretation of LAM and non-LAM cystic lesions among different physicians is generally insufficient, especially in early or atypical cases.

[0004] In recent years, artificial intelligence (AI) image analysis technology has made progress in common diseases such as lung nodule detection, chronic obstructive pulmonary disease (COPD) classification, and tumor segmentation, but it is still almost blank in the field of rare diffuse cystic lung diseases. Most existing AI models are based on two-dimensional slice-level network structures, and the training samples are focused on common diseases or public data sets, lacking specialized feature learning ability for small sample, complex structure, and rare diseases such as LAM.

[0005] In addition, the existing AI systems generally have the following defects and deficiencies: Subjectivity and consistency problems: LAM and DCLD image features are highly overlapping, different physicians have different judgments on cyst shape and density, and there are significant differences in diagnosis results.

[0006] Low efficiency: manual layer-by-layer film reading takes tens of minutes to hours, which is not suitable for large-scale screening or follow-up monitoring.

[0007] Algorithm lacks pertinence: mainstream AI image systems focus on common diseases such as pulmonary nodules and pneumonia, and have few samples of rare disease categories, and the network fails to capture key pathological structure characteristics.

[0008] Lack of model interpretability: most deep learning algorithms are black box structures that cannot output visualized lesion basis, and doctors have difficulty trusting the results.

[0009] Therefore, the present application provides a deep learning-based automatic auxiliary diagnosis system for rare lung disease CT images. SUMMARY

[0010] The present application provides a deep learning-based automatic auxiliary diagnosis system for rare lung disease CT images to solve the above technical problems.

[0011] The present application provides a deep learning-based automatic auxiliary diagnosis system for rare lung disease CT images, comprising: A data acquisition and preprocessing module for privacy processing, format conversion, standardization processing and data enhancement of CT image data, and outputting three-dimensional input blocks meeting the model input requirements; A three-dimensional feature modeling and training module for learning the morphological and distributional characteristics of cystic lesions through training strategies using a three-dimensional dense connection neural network as the core, and constructing and training a disease classification model; An online inference and interpretable display module for probability calibration of the trained model output and generation of interpretable visual results; A clinical application module for receiving the current preprocessed CT images and inputting them to the calibrated model for processing and analysis, and outputting classification results and visual basis to assist clinicians in diagnosis.

[0012] Preferably, the data acquisition and preprocessing module comprises: An anonymization processing unit for removing personal identity information and institutional labels in the original DICOM file through an automated privacy cleaning script; A format conversion unit for converting the de-identified DICOM file to NIfTI format; A standardization processing unit for standardizing the converted image in three-dimensional voxel space, resampling to an isovoxel resolution of 1.0mm x 1.0mm x 5.0mm, mapping the gray value to the [-1000, 400] HU interval through linear normalization and standardizing to the [0, 1] floating point distribution, and cropping to a three-dimensional input block of 1 x 64 x 128 x 128; The data augmentation unit employs a 3D data augmentation strategy involving random rotation, flipping, random pruning, and mild perturbation, while suppressing class shift caused by long-tailed distribution through hierarchical sampling.

[0013] Preferably, the three-dimensional densely connected neural network is 3D-DenseNet-121, which consists of an initial convolutional layer, four DenseBlocks and corresponding TransitionLayers connected sequentially. A dense connection strategy is adopted to achieve cross-layer feature reuse and efficient gradient propagation. The kernel size of the initial convolutional layer is 7×7×7, the stride is 2, and a max pooling layer is connected after the initial convolutional layer.

[0014] Preferably, each DenseBlock consists of multiple layers of three-dimensional convolutional units stacked sequentially. Each convolutional unit adopts a combination structure of BatchNormalization + ReLU activation + 3D convolution. The output feature map of each layer is connected to the feature maps of all previous layers through channel-level concatenation to form a dense feature stream.

[0015] Preferably, the training strategy includes: The loss function uses FocalLoss with a modulation factor, combined with class weights set according to the frequency of samples by class. The optimizer uses the Adam optimizer, combined with a piecewise cosine annealing learning rate scheduling strategy; Channel-level Dropout and L2 regularization are used to reduce the risk of overfitting. Model training follows a protocol of partitioned cross-validation and independent validation set evaluation.

[0016] Preferably, the probability calibration method is temperature scaling or Platt scaling; Temperature scaling is achieved by applying a scalar temperature parameter T>0 to the logit vector Z before softmax. Reparameterization; Platt scaling maps confidence scores using logistic regression. Probabilistic calibration quality is quantitatively assessed using expected calibration error and Brier score, and a reliability curve is generated for visual verification.

[0017] Preferably, the visualization results are generated as follows: a class-related attention heatmap is generated on the volume data based on Grad-CAM, and the heatmap is resampled and aligned to the original CT coordinate system according to the voxel space element information of the input CT, and then overlaid and displayed on the axial and MPR reconstructed coronal / sagittal planes.

[0018] Preferably, the classification results include probabilistic discrimination results, confidence intervals, and visualization heatmaps of the model's region of interest for LAM, DCLD, and healthy controls.

[0019] Preferably, the three-dimensional feature modeling and training module comprises: a memory unit for storing rare category sample features and reorganizing the features of the same rare category sample during model training.

[0020] Compared with the prior art, the application has the following beneficial effects: The high-dimensional structure modeling and reliable reasoning of rare diffuse lung disease images are realized at the algorithm level, and an efficient, interpretable and integrable AI diagnosis solution is formed at the clinical level. The comprehensive performance and clinical benefits show that the system has significant innovation and practical value, and can provide a demonstration for intelligent diagnosis of rare diseases and standardized application of medical AI.

[0021] Other features and advantages of the present application will be described in the following description, and some will become apparent from the description, or will be understood by those skilled in the art. The purpose and other advantages of the present application can be achieved and obtained by the structure specifically pointed out in the written description and the accompanying drawings.

[0022] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0023] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 is a structural diagram of a lung rare disease CT image automatic auxiliary diagnosis system based on deep learning in an embodiment of the present application; Figure 2 is an implementation flowchart of a lung rare disease CT image automatic auxiliary diagnosis method based on deep learning in an embodiment of the present application; Figure 3 is a structural diagram of an algorithm model in an embodiment of the present application. DETAILED DESCRIPTION

[0024] The preferred embodiments of the present application will be described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and do not limit the present application.

[0025] The present application provides a lung rare disease CT image automatic auxiliary diagnosis system based on deep learning, as shown in Figure 1 , comprising: a data acquisition and preprocessing module for privacy processing, format conversion, standardization processing and data enhancement of CT image data, and outputting three-dimensional input blocks meeting the model input requirements; The 3D feature modeling and training module is used to learn the morphological and distributional features of cystic lesions through training strategies, with a 3D densely connected neural network as the core, and to build and train a disease classification model. The online inference and interpretable visualization module is used to perform probability calibration on the output of the trained model and generate interpretable visualization results. The clinical application module is used to receive pre-processed CT images and input them into a calibrated model for processing and analysis, outputting classification results and visualization evidence to assist clinicians in diagnosis.

[0026] Preferably, the data acquisition and preprocessing module includes: The identifier processing unit is used to remove personally identifiable information and organizational tags from the original DICOM file using an automated privacy cleanup script; The format conversion unit is used to convert the de-identified DICOM file into NIfTI format; The standardization processing unit is used to standardize the converted image in three-dimensional voxel space, resample it to an equal voxel resolution of 1.0mm×1.0mm×5.0mm, map the gray values ​​to the [-1000,400]HU range through linear normalization and standardize them to a [0,1] floating-point distribution, and crop them into a 1×64×128×128 three-dimensional input block; The data augmentation unit employs a 3D data augmentation strategy involving random rotation, flipping, random pruning, and mild perturbation, while suppressing class shift caused by long-tailed distribution through hierarchical sampling.

[0027] Preferably, the three-dimensional densely connected neural network is 3D-DenseNet-121, which consists of an initial convolutional layer, four DenseBlocks and corresponding TransitionLayers connected sequentially. A dense connection strategy is adopted to achieve cross-layer feature reuse and efficient gradient propagation. The kernel size of the initial convolutional layer is 7×7×7, the stride is 2, and a max pooling layer is connected after the initial convolutional layer.

[0028] Preferably, each DenseBlock consists of multiple layers of three-dimensional convolutional units stacked sequentially. Each convolutional unit adopts a combination structure of BatchNormalization + ReLU activation + 3D convolution. The output feature map of each layer is connected to the feature maps of all previous layers through channel-level concatenation to form a dense feature stream.

[0029] Preferably, the training strategy includes: The loss function uses FocalLoss with a modulation factor, combined with class weights set according to the frequency of samples by class. The optimizer uses the Adam optimizer, combined with a piecewise cosine annealing learning rate scheduling strategy; Channel-level Dropout and L2 regularization are used to reduce the risk of overfitting. Model training follows a protocol of partitioned cross-validation and independent validation set evaluation.

[0030] Preferably, the probability calibration method is temperature scaling or Platt scaling; Temperature scaling is achieved by applying a scalar temperature parameter T>0 to the logit vector Z before softmax. Reparameterization; Platt scaling maps confidence scores using logistic regression. Probabilistic calibration quality is quantitatively assessed using expected calibration error and Brier score, and a reliability curve is generated for visual verification.

[0031] Preferably, the visualization results are generated as follows: a class-related attention heatmap is generated on the volume data based on Grad-CAM, and the heatmap is resampled and aligned to the original CT coordinate system according to the voxel space element information of the input CT, and then overlaid and displayed on the axial and MPR reconstructed coronal / sagittal planes.

[0032] Preferably, the classification results include probabilistic discrimination results, confidence intervals, and visualization heatmaps of the model's region of interest for LAM, DCLD, and healthy controls.

[0033] Preferably, the 3D feature modeling and training module includes: Memory units are used to store features of rare class samples and to reorganize them based on the features of samples of the same rare class during model training.

[0034] In this embodiment, considering the differences in scanner manufacturers and protocols, intensity clipping and linear normalization based on the HU (Henry's unit) range are adopted. Subsequently, lung field mask extraction is performed to remove mediastinal and chest wall interference. Then, resampling is performed with a fixed voxel spacing (e.g., 1.0 mm³) to ensure that data from different sources fall on the same geometric reference. For the spatial dimension of volume data, a regular three-dimensional window covering the lung field is selected, and each volume block can be seamlessly stitched together during inference. During training, three-dimensional data augmentation strategies such as random rotation, flipping, random clipping, and mild intensity perturbation are introduced to improve the network's robustness to the variability of vesicle size, morphology, and density distribution.

[0035] In this embodiment, such as Figure 2 , Figure 3As shown, the model input is a standardized volume tensor. During forward propagation, the shallow channels focus on capturing the local texture of vesicle boundaries and thin-walled structures, while the deep channels accumulate the global vesicle load and distribution features. At the end of the network, a dense three-dimensional semantic embedding is obtained through global average pooling, which is then connected to a fully connected classification head and outputs the three probabilities using Softmax.

[0036] In this embodiment, to mitigate the erosion of decision boundaries by class imbalance, Focal Loss combined with class weights is employed during the training phase. Simultaneously, the Adam optimizer and piecewise cosine annealing learning rate scheduling are used to jointly control the convergence process. To reduce the risk of overfitting, channel-level Dropout and L2 regularization are introduced. Model training follows a protocol of partitioned cross-validation and independent validation set evaluation to ensure the robustness of the results to sample partitioning and center differences. In this embodiment, the feature learning module uses a 3D DenseNet-121 network as the backbone. This network is structurally modified for 3D medical image volume data to adapt to application scenarios involving small samples of rare diseases, high-dimensional voxel input, and fine-grained modeling of structural features. The network consists of an initial convolutional layer, four Dense Blocks, and corresponding Transition Layers connected sequentially. A dense connectivity strategy is used to achieve cross-layer feature reuse and efficient gradient propagation, thereby maintaining stable training and good generalization ability under limited sample conditions. Specifically, the network input is a 3D volume tensor with dimensions 1×D×H×W, processed by unified spatial resampling and grayscale normalization. The input is first downsampled by a 3D convolutional layer (kernel size=7×7×7, stride=2) and a max-pooling layer to extract low-level spatial structure features and reduce computational complexity. It then enters the first to fourth Dense Blocks. Each Dense Block consists of multiple layers of sequentially stacked 3D convolutional units. Each convolutional unit uses a combination of "Batch Normalization + ReLU activation + 3D convolution," where the output feature map of each layer is concatenated with the feature maps of all preceding layers through channel-level concatenation, forming a dense feature stream.

[0037] At the feature representation level, the shallow Dense Blocks primarily focus on local structural information, capturing fine-grained texture features such as vesicle boundaries, thin-walled structures, and focal low-density regions through 3D convolutional operations with a small receptive field. The intermediate Dense Blocks gradually expand the receptive field, accumulating multi-scale spatial contextual information to characterize the continuity and regional clustering features of cystic distribution. The deep Dense Blocks, on the other hand, focus on integrating global structural information, performing high-level semantic modeling of the cystic burden, spatial distribution patterns, and overall degree of destruction in the lung parenchyma. A Transition Layer is set between each Dense Block, using 1×1×1 3D convolution to achieve channel compression and combining it with average pooling to achieve a smooth transition in spatial scale, thereby controlling parameter scale and suppressing redundant features.

[0038] At the network's end, global average pooling is used to aggregate the spatial dimensions of the final feature map, resulting in a fixed-length three-dimensional semantic embedding vector to represent the overall image features of each sample. This embedding vector is then fed into a fully connected classification head, which outputs the predicted probability distributions for LAM, DCLD, and healthy controls through a softmax function.

[0039] In terms of training strategy, to mitigate the impact of class imbalance on decision boundary learning, a Focal Loss with a modulation factor is adopted as the loss function, combined with weight parameters set according to the frequency of samples of each class. This reduces the dominant role of easily classified samples in the gradient and increases the learning contribution of minority class samples. The optimization process uses the Adam optimizer, combined with a piecewise cosine annealing learning rate scheduling strategy to dynamically control the training process, balancing convergence speed and stability. To further reduce the risk of overfitting, channel-level Dropout is introduced within the Dense Block, and L2 regularization constraints are applied to the network parameters.

[0040] Model training and evaluation follow an experimental protocol that combines hierarchical cross-validation and independent validation sets. The model is trained repeatedly under different data partitions, and the model weights with the best performance are selected to ensure that the feature learning module still has good robustness and repeatability under the conditions of sample partition changes and potential center differences.

[0041] After training the three-class classification network and freezing its parameters, to improve the usability of the probability output in clinical decision support, this invention trains and then parameterizes the confidence level of the classification head output. Specifically, temperature scaling is applied to the logit vector before softmax on the independent validation set: the logit vector Z is scaled with a scalar temperature parameter T>0. The model is reparameterized to ensure that the output probability distribution is more consistent with the true hit rate within each threshold interval without changing the class ranking. In another optional implementation, Platt scaling is used to map the confidence level through logistic regression to obtain the calibrated probability output. The calibration quality is quantitatively evaluated using expected calibration error (ECE) and Brier score, and a reliability curve is generated for visualization verification, thereby ensuring that the model output probability can serve as a stable reference indicator for clinical risk stratification, threshold decision-making, and follow-up management. To enhance interpretability and verifiability, a class-related Grad-CAM volumetric data-level attention heatmap is calculated on the three-dimensional feature layer during the inference stage. Based on the voxel spatial metadata (spacing / origin) of the input CT, the heatmap is resampled and aligned to the original CT coordinate system. Subsequently, it is overlaid on the axial and MPR reconstructed coronal / sagittal planes, so that the model's area of ​​interest is strictly registered with the original image in space, making it easier to trace the network's discrimination criteria back to the image semantics familiar to physicians (such as cystic clusters). To enhance interpretability, during inference, a class-related attention heatmap is generated on the in-situ data based on Grad-CAM and spatially registered with the original CT slices and MPR reconstructed planes to present the cystic clusters and boundary structures that the model focuses on. This allows the network's discrimination criteria to be traced back to the image semantics familiar to physicians, thereby enhancing the verifiability of human-machine collaboration.

[0042] In multi-center internal validation, this invention achieved an average discrimination AUC of 0.950 and an overall accuracy of 0.932 for a three-class classification task, demonstrating good calibration at the probabilistic level. Sensitivity, specificity, and macro-average F1 remained stable across data subsets from different scanning protocols and equipment manufacturers, indicating the model's adaptability to imaging differences and case heterogeneity. Combined with heatmap quality control, errors were concentrated in boundary cases with sparse vesicles or those combined with other lung lesions. Physicians can quickly identify potential sources of misclassification through visualization and make manual corrections, further reducing the risk of clinical use.

[0043] In this embodiment, the core logic of the automated privacy cleanup script is as follows: For the original DICOM file, field location technology is used to identify and extract the privacy information that needs to be removed, specifically including key fields such as PatientName, PatientID, PatientBirthDate, InstitutionName, and InstitutionAddress. Regular expression matching is then used to further validate and remove hidden privacy information. The regular expression is as follows: (Matching patient IDs that are 10-18 alphanumeric characters) (Matches 2-8 Chinese or English names).

[0044] The script execution flow is as follows: read DICOM file metadata → locate target fields → clean up by overwriting fields (replacing them with placeholders "XXX") or deleting them directly → verify the cleanup results → output the de-identified DICOM file.

[0045] A lung field mask extraction algorithm based on grayscale thresholding combined with morphological processing is adopted: Based on the grayscale distribution characteristics of CT images, a threshold of -300HU is set, and areas with grayscale values ​​greater than -300HU (such as the mediastinum, chest wall, and bones) are initially marked as non-lung field areas. Morphological opening operations were performed on the preliminary segmentation results (corrosion kernel size 3×3×3, expansion kernel size 3×3×3) to remove minor noise. Through connected component analysis, the two connected components with the largest volume (corresponding to the left and right lung parenchyma) are retained. The boundaries of connected regions are smoothed to generate a complete lung field mask, which is used to remove non-lung field interference in subsequent processing.

[0046] In this embodiment, the modulation factor in the Focal Loss is 0.25, which is determined by 5-fold cross-validation and can effectively balance the gradient contributions of positive and negative samples. The method for calculating the category weight based on the sample frequency of each category is as follows: weight = total number of samples / (number of categories × number of samples in that category). In this invention, the number of categories is 3 (LAM, DCLD, and healthy controls). If the total number of samples is 370 (20 LAM, 50 DCLD, and 300 healthy controls), then the weights of the three categories are calculated according to the above formula.

[0047] The dropout rate of channel-level Dropout is set to 0.2, which only applies to the convolutional output channels within the DenseBlock during the training phase; The penalty coefficient for L2 regularization is set to... It is used to constrain the weight parameters of fully connected layers and convolutional layers to suppress overfitting interference.

[0048] The scalar temperature parameter T for temperature scaling is determined by a grid search method with a search range of 0.5-2.0 and a step size of 0.1. The goal is to minimize the expected calibration error (ECE) on the independent validation set, and finally select the optimal T value (the optimal T value after verification in this invention is 1.2). In Platt-scaled logistic regression mapping, the input features are the logit vectors output by the model, and the mapping function is... Where z is the logit vector, , These are the parameters obtained by fitting the data to an independent validation set; In the quantitative assessment of probabilistic calibration quality, the expected calibration error (ECE) is calculated using 10 equally probable intervals, and the Brier score is calculated using the standard formula. , To predict probabilities for the model, (where N is the number of samples in the validation set, representing the true labels).

[0049] The target layer for Grad-CAM visualization is set as the output layer of the last DenseBlock in a 3D densely connected neural network, which can effectively characterize the high-level semantic features of lung lesions. The normalization threshold for the attention heatmap is set to 0.3, meaning that only regions with response values ​​greater than 0.3 in the heatmap are retained as regions of interest for the model, while regions with response values ​​below this threshold are set to transparent. The thermal map resampling uses a trilinear interpolation algorithm to ensure that the spatial resolution of the resampled thermal map is consistent with that of the original CT image. Coordinate alignment method: Based on the voxel spatial element information (spacing / origin) of the input CT, the spatial coordinates of the heat map are rigidly registered with the world coordinates of the original CT to ensure that the position of the heat map in the axial, coronal and sagittal planes completely coincides with the position of the lesion in the original CT. Overlay display method: On the coronal / sagittal images reconstructed by axial and MPR, a heat map is overlaid using a red gradient transparency method (the higher the response value, the lower the transparency), while retaining the grayscale information of the original CT, which is convenient for doctors to compare and observe.

[0050] Interface Protocol: It adopts the DICOM 3.0 standard interface to interface with the clinical PACS system, and also supports the RESTful API interface for data interaction. The interface address format is http: / / [server IP]:[port number] / api / ct-analysis; Data transmission format: CT image data is transmitted in DICOM format via TCP / IP protocol, and AES-256 encryption is used during transmission; model inference results are returned in JSON format, and the JSON fields include PatientID, AnalysisTime, LAM_Probability, DCLD_Probability, Healthy Control_Probability, ConfidenceInterval, Heatmap_Path, and Model_Version. Report feedback process: After the model generates the inference results, it automatically associates them with the patient's examination number in the PACS system and pushes them to the PACS system's report module through the interface. Doctors can directly retrieve and view them in the image reading interface. Audit trail implementation: The system automatically records key information for each analysis, including patient ID, examination number, data upload time, inference start time, inference end time, analyst (if any), and result modification records (if any). The log file is named "YYYYMMDD_HHMMSS_PatientID.log", stored in the specified server directory, and retained for 5 years.

[0051] Compared with traditional manual image reading and general AI image analysis systems, this invention has achieved significant improvements and advantages in terms of technical performance, clinical application value, reliability and scalability.

[0052] Firstly, regarding diagnostic performance, this invention overcomes the limitation of two-dimensional slicing methods in capturing spatial continuity by introducing 3D DenseNet-121 to model the spatial texture and vesicle distribution patterns of lung parenchyma at the voxel level. Validation results on multi-center real-world data show that for the three-class classification task involving LAM, DCLD, and normal samples, the AUC reaches 0.950, and the overall accuracy is 0.932. After temperature calibration, the confidence distribution of the model output is highly consistent with the true positive rate. This result demonstrates that the system still possesses stable and reliable recognition capabilities in rare disease imaging scenarios. Through the introduction of class-imbalanced training strategies and FocalLoss, the model maintains high sensitivity and specificity even under conditions of scarce samples and skewed label distribution, fundamentally solving the problem of the difficulty in generalizing rare disease AI models.

[0053] Secondly, in clinical applications, this invention significantly improves diagnostic efficiency and consistency. Traditional LAM image interpretation often requires senior physicians to meticulously review multiple CT slices layer by layer, while the system's inference can automatically analyze the entire lung volume and output structured results within ten seconds. The system can automatically generate a visual heatmap containing disease probability, confidence intervals, and the model's region of interest. Doctors can directly verify the model's judgment criteria on the image interpretation interface, thus achieving a dual confirmation mechanism of AI assistance and doctor-led analysis, effectively avoiding clinical risks caused by AI misjudgments.

[0054] Furthermore, this invention achieves breakthroughs in interpretability and credibility. By introducing the Grad-CAM visualization algorithm, the model can automatically generate high-response regions consistent with pathological features during the inference phase. This allows doctors to clearly see the overlap between the AI's focus and the clinical lesion location, enhancing the understandability and clinical trustworthiness of the results. The model output is not limited to single-class labels but also provides probability-calibrated continuous risk values, which can be directly used for clinical follow-up, disease grading, and research statistical analysis.

[0055] Furthermore, this invention exhibits excellent system compatibility and scalability. The system can be used on computers with varying configurations to achieve automatic data retrieval, inference, report feedback, and audit trail recording. Its modular design supports hybrid deployment on local area servers and in the cloud, and can be expanded as needed to automatically identify other diffuse lung diseases, such as Birt-Hogg-Dubé syndrome or Langerhans cell histiocytosis, providing a general algorithmic framework for the construction of intelligent diagnostic platforms for rare diseases.

[0056] Finally, in terms of social and scientific research value, the application of this invention helps improve the early identification capabilities of primary care and non-specialized hospitals for rare lung diseases, shortens the time to diagnosis for patients, and reduces the occurrence of misdiagnosis and delayed treatment. Through standardized and quantitative image analysis processes, the system can also provide objective and consistent data characteristics and reproducible diagnostic evidence for multi-center clinical studies, promoting integrated research on LAM and other diffuse cystic lung diseases.

[0057] In summary, this invention not only achieves high-dimensional structural modeling and reliable inference of rare diffuse lung disease images at the algorithmic level, but also forms an efficient, interpretable, and integrable AI diagnostic solution at the clinical level. Its comprehensive performance and clinical benefits demonstrate that the system possesses significant innovation and practical value, and can serve as a model for the standardized application of intelligent diagnosis of rare diseases and medical AI.

[0058] This invention provides an automated assisted diagnostic system for CT images of rare lung diseases based on deep learning. The data enhancement unit employs a three-dimensional data enhancement strategy involving random rotation, flipping, random cropping, and mild perturbation to enhance the images, including: Using a pre-trained lung multi-structure segmentation network, the standardized input 3D CT block V is segmented to obtain a set of binary masks M of stable anatomical structures including the thoracic cage, left and right lung parenchyma, major blood vessels, and tracheobronchial tree. A lung CT template library containing healthy lungs and various known lesion patterns is constructed. For the input block V and its corresponding anatomical mask, a non-rigid alignment is performed with the most similar template in the library using three-dimensional deformable registration technology to obtain a dense three-dimensional deformation field. ; Deformation field The grayscale information of the input block V is input into a conditional deformation field generation network. This network uses an anatomical mask M as an adversarial constraint. Its training objective, guided by adversarial loss, is to learn to generate an incremental deformation field that applies realistic deformation to lung tissue (especially the lung parenchyma) while maintaining the topological continuity and relative spatial relationships within all anatomical masks. ; The adversarial loss is provided by a discriminator that simultaneously receives the real CT block and the result of the deformation field. The distorted CT block is analyzed to determine its authenticity, while a constraint is imposed: each anatomical mask distorted by the same deformation field should have a high degree of consistency with the original mask. For each data augmentation request, an incremental deformation field satisfying the anatomical constraints is obtained by sampling from the conditional prior distribution of the conditional deformation field generation network. ; This incremental deformation field is superimposed on a basic, small-scale stochastic elastic deformation field, and an energy minimization problem based on thin plate splines is solved to obtain a globally smooth and locally anatomically reasonable final deformation field. ; use Three-dimensional bilinear interpolation sampling is performed on the original input block V to obtain the spatially transformed image. ; Subsequently, within the lung parenchyma region defined by the segmentation mask, a nonlinear intensity modulation related to the Jacobian determinant of local deformation (reflecting the degree of tissue compression / expansion) is applied. It simulates the changes in CT values ​​caused by variations in tissue density, generating the final enhanced image. .

[0059] The pre-trained lung multi-structure segmentation network uses the 3DU-Net network and is trained on a public dataset containing 500 chest CT scans. The input is a 1×64×128×128 three-dimensional CT block, and the output is a multi-channel mask of the corresponding size. Each channel corresponds to the thoracic cage, left lung parenchyma, right lung parenchyma, major blood vessels, and tracheobronchial tree, respectively. The standardized input three-dimensional CT block V is a three-dimensional chest CT data block that has undergone format conversion, grayscale normalization, and size cropping. The anatomical mask set M is a set of binary data representing different anatomical structures of the lung, obtained through a segmentation network. Each mask corresponds to a specific structure. The segmentation network outputs masks with five channels: channel 1 corresponds to the thoracic cavity, channel 2 to the left lung parenchyma, channel 3 to the right lung parenchyma, channel 4 to the large blood vessels, and channel 5 to the tracheobronchial tree. These five channels together constitute M. The lung CT template library is a collection of chest CT data containing healthy individuals and various known lung lesions (such as LAM and common pneumonia). It provides reference templates for deformable registration. Data from 200 cases were collected, including 50 healthy lung CTs, 50 LAM patient CTs, 50 other DCLD patient CTs, and 50 common pneumonia CTs. All data were processed into standardized 3D blocks of 1×64×128×128 pixels and stored according to lesion type, forming the template library. Three-dimensional deformable registration technology is an algorithm that enables spatial alignment of two three-dimensional images, achieving non-rigid position matching. It employs a B-spline free deformation registration algorithm, using the anatomical mask set M of the input block V as a reference, to calculate the deformation field between V and a CT template of a LAM patient in the template library, ensuring that the overlap of the lung parenchyma regions of the two images reaches more than 95%, resulting in a dense three-dimensional deformation field. .

[0060] The Conditional Deformation Field Generation Network is a deep learning network that generates spatial deformation fields that conform to anatomical structures, using anatomical masks as constraints. It adopts a conditional generative adversarial network with a U-Net structure as the generator. The generator's input is the deformation field φ and the grayscale information of the input block V. At the same time, the set of anatomical masks M is used as the conditional input. During training, the generated deformation field is designed so that the trachea and blood vessels in the mask still maintain their accompanying relationship after transformation.

[0061] Adversarial loss is a loss function used to train generative adversarial networks and distinguish between real and generated data. The discriminator is set as a 3D convolutional network, receiving both real CT blocks and data generated by deformation fields. For the distorted CT block, the adversarial loss uses binary cross-entropy loss. When the discriminator cannot distinguish between the real and distorted CT blocks, the loss value is less than 0.1.

[0062] Incremental deformation field In the basic deformation field An additional deformation field, conforming to anatomical constraints, is added to further expand the deformation diversity of the sample. This involves sampling from the conditional prior distribution of the conditional deformation field generation network to obtain a deformation field that causes slight local expansion of the lung parenchyma, with the expansion not exceeding 5% of the original lung parenchyma volume. This deformation field is then used as the additional deformation field. .

[0063] The basic, small-range random elastic deformation field is a pre-defined elastic deformation field with a small deformation range, used to be superimposed on the incremental deformation field to increase diversity. It adopts a randomly generated elastic deformation field, and the deformation displacement range is limited to ±2 voxels to ensure that the anatomical structure will not break after deformation.

[0064] The energy minimization problem based on thin plate splines is a mathematical optimization problem that uses thin plate spline algorithms to ensure the global smoothness of the superimposed deformation field. It uses the bending energy of the thin plate spline as the objective function to minimize the incremental deformation field. Using the field superimposed on the basic elastic deformation field as the initial value, the deformation field that minimizes bending energy is solved to obtain the globally smooth final deformation field. Three-dimensional bilinear interpolation sampling is an interpolation algorithm used to apply the deformation field to the original CT block to obtain a spatially transformed image; it utilizes the final deformation field... For each voxel in the original input block V, a three-dimensional bilinear interpolation method is used to calculate the gray value corresponding to its transformed position, thus obtaining the spatially transformed image. .

[0065] The Jacobian determinant of local deformation is a mathematical index that reflects the degree of compression or expansion of tissue by the deformation field in a local region. It is used to calculate the final deformation field. The determinant value of the Jacobian matrix at each voxel location: a value of 0.8 indicates that the tissue in that region has been compressed by 20%; a value of 1.2 indicates that it has been expanded by 20%. Nonlinear intensity modulation. This is a CT value adjustment value that simulates changes in tissue density and is related to the degree of local deformation. The value is 100. When the Jacobian determinant of a voxel is 0.8, That is, the CT value of the voxel decreases by 20; when the determinant value is 1.2, That is, the CT value increases by 20.

[0066] In this embodiment, the network structure of the pre-trained 3DU-Net segmentation network is as follows: the encoder contains 4 layers, the decoder contains 4 layers, and the convolutional kernel size of each layer is 3×3×3; the number of encoder channels is 64→128→256→512 in sequence, the number of decoder channels is 512→256→128→64 in sequence, and the final output layer has 5 channels (corresponding to 5 types of anatomical structures). The training parameters are set as follows: learning rate = 1e-4, iterations = 10000, batch size = 4; loss function is a weighted sum of Dice loss and cross-entropy loss, with a weight ratio of 1:1, where the Dice loss formula is... Where A is the prediction mask and B is the real mask; the training dataset for this network is a public dataset containing 500 chest CT scans (sourced from a subset of the LIDC-IDRI dataset). All data have been annotated by clinicians, and the annotations include the boundaries of the thoracic cage, left lung parenchyma, right lung parenchyma, major blood vessels, and tracheobronchial tree.

[0067] In this embodiment, the B-spline free deformation registration algorithm is used, and the key parameters are as follows: The control point spacing was set to 10 mm, the number of iterations was 50, the similarity metric used was mutual information, and the convergence threshold was set to 1e-6.

[0068] The registration process is as follows: using the anatomical mask set M of the input block V as a reference, the most similar template (with the highest mutual information value and greater than 0.85) is selected from the template library by calculating the mutual information value; B-spline mesh control points are initialized; the positions of the control points are adjusted through iterative optimization to ensure that the overlap between the lung parenchyma region of the input block V and the template reaches more than 95%; and a dense three-dimensional deformation field is output. .

[0069] In this embodiment, the generator (U-Net structure) has 64→128→256 encoder channels and 256→128→64→3 decoder channels (outputting a three-dimensional deformation field). Each layer is activated using BatchNormalization+LeakyReLU (negative slope 0.2). Discriminator (3D convolutional network): It contains 3 layers of 3D convolution, with a kernel size of 3×3×3 and the number of channels from 64 to 128 to 256. Finally, it is connected to a fully connected layer to output a binary classification result (real / generated). The activation function is LeakyReLU (negative slope 0.2). Adversarial loss formula: Where x is the real CT block, z is random noise, M is the anatomical mask, G is the generator, and D is the discriminator; Simultaneously add anatomical constraint loss ,in, For the original anatomical mask, The anatomical mask is distorted by the deformation field, and N is the number of mask categories; The total loss is: .

[0070] Training objective: When the discriminator cannot distinguish between real and distorted CT blocks, the adversarial loss value should be less than 0.1 and the anatomical constraint loss value should be less than 0.05.

[0071] In this embodiment, the final deformation field is calculated. The Jacobian matrix at each voxel location is obtained through the following steps: For deformation field The spatial partial derivatives of the three components (x, y, and z directions) are calculated respectively to obtain a 3×3 partial derivative matrix. ; Calculate the determinant value det(J) of the matrix. A determinant value less than 1 indicates that the tissue in the region is compressed, greater than 1 indicates that it is expanded, and equal to 1 indicates that there is no volume change.

[0072] The beneficial effects of the above technical solution are as follows: the introduction of an adversarial generative network with anatomical structure as a hard constraint condition, the deformation space learned by the network excludes invalid deformations that would lead to anatomical structure breakage, crossing, or topological errors, and the generated deformation strictly respects the anatomical constraints of the lung (such as the accompanying relationship between the bronchi and blood vessels, and the continuity of the pleura), ensuring the anatomical rationality of the enhanced samples, the correlation between intensity modulation and local voxel deformation, simulating the morphology-density correlation changes commonly seen in lung lesions (such as cystic changes and solid changes), enhancing the pathological realism, and by sampling fully in a reasonable anatomical deformation space, greatly expanding the sample distribution that conforms to the variation law of the real world seen by the model.

[0073] This invention provides an automated assisted diagnostic system for rare lung diseases based on CT images, which suppresses class shift caused by long-tailed distribution through hierarchical sampling, including: The output feature map of the last DenseBlock of the 3D densely connected neural network is transformed through two parallel and structurally different projection head networks: General Feature Projection Head Output general feature vector This feature is designed to capture the basic anatomical and imaging patterns of the lungs common to all categories (common diseases, rare diseases, and healthy individuals); Rare and unique feature projection head Output specific feature vectors This feature is designed to capture the subtle, specific patterns needed to distinguish various rare lesions; During training, an orthogonal loss function is used. This forces the representations of the two types of features in the vector space to be independent of each other; For each tail (rare) category c in the training set, maintain a dynamically updated feature prototype memory. Used to store samples of this category extracted by the current model. and Vector pairs; In each training batch, for the head (common) category samples, their features are used for training in the usual way; For each tail category sample, perform cross-sample feature remixing: From the memory bank of the same category Feature pairs of randomly selected other samples ; Perform feature-level remixing: synthesize feature pairs as ,in ,and Remain unchanged; To retrieve from memory banks of the same category The general feature vector of another sample selected from the sample; General characteristics after remixing Rare specific features of the original The parts are reassembled and then fed into the classifier through a lightweight fusion module. The classifier receives the final features output by the fusion module and makes a class prediction: Design a gradient isolation mechanism: During backpropagation, calculate the gradient of the loss function with respect to the general feature branch. and gradient of specific feature branches By using a gradient projection layer, we can ensure... It does not contain components that can be explained by general features, thus enabling rare-specific feature branches to uncover marginal information that is ignored by general features and is important for distinguishing rare categories.

[0074] In this embodiment, the feature map output by the last DenseBlock of 3DDenseNet-121 has a size of 1024×8×16×16 (number of channels × depth × height × width). This feature map contains high-level features such as the morphology and distribution of lung tissue in the CT block. (Universal feature projection head) This is a network module used to extract basic lung features common to all categories (common diseases, rare diseases, and healthy individuals). It adopts a network structure containing one 3D convolutional layer (3×3×3 kernel, 256 output channels) and one fully connected layer (128 output dimensions). The input is the output feature map of the last DenseBlock, and the output is a general feature vector with a dimension of 128. General feature vectors It is a vector output by the general feature projection head that represents the basic features common to all categories, such as the vector corresponding to a CT scan of a healthy lung. The elements in its vector primarily represent common features shared by all lung samples, such as the normal texture of the lung parenchyma and the basic morphology of the trachea. Rare specific feature projection head. This is a network module used to extract subtle features needed to distinguish various rare diseases. It adopts a network structure containing two 3D convolutional layers (the first layer has a 3×3×3 kernel and 128 output channels; the second layer has a 3×3×3 kernel and 64 output channels) and one fully connected layer (output dimension 64). The input is the output feature map of the last DenseBlock, and the output is a specific feature vector with a dimension of 64. Specific feature vectors It is a vector representing the unique pathological features of rare diseases, output by a rare-specific feature projection head, such as the vector corresponding to a CT block of a certain LAM patient. The elements in its vector mainly represent the characteristics of LAM, such as the distribution of diffuse cystic changes and the size of vesicles, which are unique to rare diseases.

[0075] Orthogonality loss It is a loss function used to force the general feature vector and the specific feature vector to be independent of each other in the vector space, such as calculating... (Dimension 128) and The squared L2 norm of the transpose product of (dimension 64) is used to ensure that the loss value is less than 0.01 during training, thus ensuring that there is no redundant correlation between the two types of features.

[0076] The tail (rare) category c is a category with a very small number of samples in the training data. It corresponds to rare lung diseases, such as the LAM category, which has only 20 training samples, far fewer than the healthy category (200 cases) and the common disease category (150 cases). It belongs to the tail category.

[0077] Feature Prototype Memory It is a dynamic collection maintained for each rare class, storing feature vector pairs of samples from that class, such as the memory maintained for the LAM class (c=LAM). The LAM samples that have been trained are stored in ( , There are 20 pairs of vectors, and each time a new LAM sample is trained, its corresponding vector pair is added to the memory.

[0078] Cross-sample feature remixing is an operation that selects features from other samples in the same rare class memory and mixes them with the features of the current sample to increase diversity, such as for the current LAM sample. ,from Select another LAM sample The two are mixed according to a certain weight to obtain . The distribution is a probability distribution used to generate mixed weights, when When they are equal, they are symmetrically distributed, such as when set =2, obtained by sampling from the Beta(2,2) distribution. =0.4, this That is, the current sample With memory samples Mixed weights.

[0079] The lightweight fusion module is a small network module used to concatenate and integrate the remixed general features with the original specific features. It employs a module containing one fully connected layer (input dimension 128 + 64 = 192, output dimension 128) and a ReLU activation function. The input is... and The concatenated vectors are used to output the integrated feature vector.

[0080] The gradient isolation mechanism is a mechanism that ensures that the gradient of the specific feature branch does not contain interpretable components of the general feature during backpropagation. During training, the gradient of the specific feature branch is processed to prevent it from containing information from the gradient of the general feature branch, thereby strengthening the independent learning of the specific feature.

[0081] gradient and The gradient is obtained by taking the derivative of the loss function with respect to the parameters of the general feature branch and the specific feature branch, respectively. The gradient of the classification loss with respect to the parameters of the general feature projection head is calculated as follows. ;Calculate the gradient of the classification loss with respect to the projection head parameters of rare specific features. .

[0082] Feature Prototype Memory The dynamic update mechanism is as follows: the update frequency is once after every epoch of training; the upper limit of storage capacity is set to twice the number of samples in this category. When the storage capacity reaches the upper limit, the earliest stored feature vector pair is eliminated using the first-in-first-out (FIFO) rule; each time it is updated, the cosine similarity between the new sample feature vector and the existing feature vector in the memory is calculated. If the similarity is greater than 0.95, it is determined to be a redundant feature and is not added to the memory to ensure the diversity of features in the memory.

[0083] In this embodiment, the specific calculation process of the gradient projection layer is as follows: Calculate the gradient of the general feature branch gradient with specific feature branches dot product ; Calculate the L2 norm squared of the gradient of the general feature branch. ; Calculate the projection component of the specific feature gradient onto the general feature gradient direction. ; The processed specific feature gradient is This ensures that the processed gradient does not contain components that can be explained by general features.

[0084] In this embodiment, the lightweight fusion module adopts a structure of fully connected layer + BatchNormalization + ReLU activation: the input dimension of the fully connected layer is 128 (general feature dimension) + 64 (specific feature dimension) = 192, and the output dimension is 128; the BatchNormalization layer is set with momentum of 0.99 and epsilon of 1e-5 to accelerate training convergence and suppress overfitting; the ReLU activation function is used to introduce nonlinear feature transformation to enhance the feature integration capability of the module.

[0085] The beneficial effects of the above technical solution are: preventing the discriminative features of the tail category from being submerged by the common features of the head category, ensuring that the subtle features of rare diseases are effectively extracted and enhanced, increasing the performance diversity of the tail category through remixing in the feature space and protecting the specific features from being contaminated or diluted, and enhancing the network's ability to mine edge-specific information to distinguish rare disease patterns that rarely appear in the training set with higher feature sensitivity.

[0086] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A deep learning-based automated assisted diagnosis system for rare lung diseases using CT images, characterized in that, include: The data acquisition and preprocessing module is used to perform privacy processing, format conversion, standardization processing and data augmentation on CT image data, and output three-dimensional input blocks that meet the model input requirements; The 3D feature modeling and training module is used to learn the morphological and distributional features of cystic lesions through training strategies, with a 3D densely connected neural network as the core, and to build and train a disease classification model. The online inference and interpretable visualization module is used to perform probability calibration on the output of the trained model and generate interpretable visualization results. The clinical application module is used to receive pre-processed CT images and input them into a calibrated model for processing and analysis, outputting classification results and visualization evidence to assist clinicians in diagnosis.

2. The deep learning-based automated assisted diagnosis system for rare lung diseases using CT images, as described in claim 1, is characterized in that... The data acquisition and preprocessing module includes: The identifier processing unit is used to remove personally identifiable information and organizational tags from the original DICOM file using an automated privacy cleanup script; The format conversion unit is used to convert the de-identified DICOM file into NIfTI format; The standardization processing unit is used to standardize the converted image in three-dimensional voxel space, resample it to an equal voxel resolution of 1.0mm×1.0mm×5.0mm, map the gray values ​​to the [-1000,400]HU range through linear normalization and standardize them to a [0,1] floating-point distribution, and crop them into a 1×64×128×128 three-dimensional input block; The data augmentation unit employs a 3D data augmentation strategy involving random rotation, flipping, random pruning, and mild perturbation, while suppressing class shift caused by long-tailed distribution through hierarchical sampling.

3. The deep learning-based automated assisted diagnosis system for rare lung diseases using CT images, as described in claim 1, is characterized in that... The three-dimensional densely connected neural network is 3D-DenseNet-121, which consists of an initial convolutional layer, four DenseBlocks and corresponding TransitionLayers connected sequentially. It adopts a dense connection strategy to achieve cross-layer feature reuse and efficient gradient propagation. The kernel size of the initial convolutional layer is 7×7×7, the stride is 2, and the initial convolutional layer is followed by a max pooling layer.

4. The deep learning-based automated assisted diagnosis system for rare lung diseases using CT images, as described in claim 3, is characterized in that... Each DenseBlock consists of multiple layers of three-dimensional convolutional units stacked sequentially. Each convolutional unit adopts a combination structure of BatchNormalization + ReLU activation + 3D convolution. The output feature map of each layer is connected to the feature maps of all previous layers through channel-level concatenation to form a dense feature stream.

5. The deep learning-based automated assisted diagnosis system for rare lung diseases using CT images according to claim 1, characterized in that, The training strategy includes: The loss function uses FocalLoss with a modulation factor, combined with class weights set according to the frequency of samples by class. The optimizer uses the Adam optimizer, combined with a piecewise cosine annealing learning rate scheduling strategy; Channel-level Dropout and L2 regularization are used to reduce the risk of overfitting. Model training follows a protocol of partitioned cross-validation and independent validation set evaluation.

6. The deep learning-based automated assisted diagnosis system for rare lung diseases using CT images according to claim 1, characterized in that, The probability calibration method is either temperature scaling or Platt scaling; Temperature scaling is achieved by applying a scalar temperature parameter T>0 to the logit vector Z before softmax. The reparameterization; Platt scaling maps confidence scores using logistic regression; Probabilistic calibration quality is quantitatively assessed using expected calibration error and Brier score, and a reliability curve is generated for visual verification.

7. The deep learning-based automated assisted diagnosis system for rare lung diseases using CT images according to claim 1, characterized in that, The visualization results are generated as follows: based on Grad-CAM in-volumetric data, a class-related attention heatmap is generated, and the heatmap is resampled and aligned to the original CT coordinate system according to the voxel space element information of the input CT. The heatmap is then overlaid and displayed on the axial and MPR reconstructed coronal / sagittal planes.

8. The deep learning-based automated assisted diagnosis system for rare lung diseases using CT images, as described in claim 1, is characterized in that... The classification results include probabilistic discrimination results, confidence intervals, and visualization heatmaps of the model's areas of interest for LAM, DCLD, and healthy controls.

9. The deep learning-based automated assisted diagnosis system for rare lung diseases using CT images, as described in claim 1, is characterized in that... The 3D feature modeling and training module includes: Memory units are used to store features of rare class samples and to reorganize them based on the features of samples of the same rare class during model training.

Images