Lung adenocarcinoma preoperative prediction system, medical device, method, and medium

Abstract

This application relates to the field of preoperative prediction technology for lung adenocarcinoma, and specifically to a preoperative prediction system, medical device, method, and medium for lung adenocarcinoma. The system includes an interactive interface and a processor. The interactive interface responds to user interactions and inputs multiple data from a lung adenocarcinoma patient, including inflammatory markers, tumor markers, CT images, and clinical data. The processor, connected to the interactive interface, is used to extract clinical features from the clinical data and morphological features from the CT images, and then calculate visual indicators, radiomics scores, and clinical scores for lung adenocarcinoma. An input vector is constructed based on the visual indicators, radiomics scores, and clinical scores, and input into a pre-trained prediction model, outputting the preoperative prediction result for lung adenocarcinoma. This solves the problems of related technologies relying on physicians' subjective experience, limited features, and insufficient interpretability of the prediction model.

Description

Technical Field

[0001] This application relates to the field of preoperative prediction technology for lung adenocarcinoma, and in particular to a preoperative prediction system, medical device, method and medium for lung adenocarcinoma. Background Technology

[0002] Preoperative accurate staging of lung adenocarcinoma, particularly in differentiating its invasiveness and lymph node metastasis, is crucial for developing individualized surgical plans and improving patient prognosis. Current techniques often rely on physicians' subjective judgment of images or analysis of single features, which not only leads to inconsistent assessments due to individual differences in experience but also fails to comprehensively capture the pathological characteristics of the tumor. Furthermore, many predictive models suffer from a "black box" problem—opaque decision-making logic—resulting in insufficient interpretability and limited clinicians' trust in the predictions, thus hindering the clinical translation and practical application of these technologies. Summary of the Invention

[0003] This application provides a preoperative prediction system, medical device, method, and medium for lung adenocarcinoma to address the problems of related technologies relying on doctors' subjective experience, having limited features, and insufficient interpretability of prediction models.

[0004] The first aspect of this application provides a preoperative prediction system for lung adenocarcinoma, comprising: an interactive interface for responding to user interaction actions, and inputting multiple data from a lung adenocarcinoma patient, including inflammatory indicators, tumor markers, CT images, and clinical data; and a processor connected to the interactive interface, for extracting clinical features from clinical data, extracting image morphology features from CT images, calculating visual indicators of lung adenocarcinoma based on image morphology features, calculating a radiomics score based on radiomics feature mining results of image morphology features, calculating a clinical score based on inflammatory indicators, tumor markers, clinical features, and image morphology features, constructing an input vector based on visual indicators, radiomics score, and clinical score, inputting the input vector into a pre-trained prediction model, and outputting preoperative prediction results for lung adenocarcinoma through the input model.

[0005] Based on the aforementioned technical means, this application embodiment constructs a standardized and automated preoperative prediction system for lung adenocarcinoma through the collaboration of an interactive interface and a processor. The interactive interface enables unified input of multiple data sources, including inflammatory indicators, tumor markers, and CT images, breaking the limitations of single data dimensions. The processor handles data processing, automatically extracting clinical and imaging morphological features and converting them into three types of quantitative results: visual indicators, radiomics scores, and clinical scores. Finally, the model outputs the prediction results. This system not only replaces subjective judgment and single-feature analysis by physicians, avoiding assessment biases caused by experience differences, but also captures tumor characteristics through multi-dimensional information, improving the objectivity and accuracy of predictions. This provides reliable support for the development of individualized surgical plans and facilitates the clinical application of the technology.

[0006] Optionally, the processor is also configured to: perform preprocessing on the CT images before extracting image morphological features from the CT images, the preprocessing including at least one of image normalization, image enhancement and annotation, image normalization being used to normalize the density values ​​of the CT images to a preset range, image enhancement employing an adaptive algorithm, and annotation being manually determined for the target region.

[0007] Based on the aforementioned technical means, this application embodiment effectively solves the problems of uneven quality, blurred details, and unclear target areas in original image data by setting optional preprocessing procedures for CT images. Image standardization can eliminate density value differences caused by different CT equipment and scanning parameters, improving data consistency; image enhancement strengthens the detailed features of the tumor and surrounding areas through adaptive algorithms, facilitating accurate extraction of morphological information in subsequent processing; manual annotation ensures accurate positioning of the target area and reduces interference from irrelevant areas. This approach retains processing flexibility while optimizing image data quality from the source, providing reliable input for subsequent feature extraction and index calculation, and further improving the accuracy and stability of prediction results.

[0008] Optionally, the processor is also configured to: calculate visual metrics based on the geometric parameters of the segmented target region, wherein the visual metrics include at least one of the following: the ratio of the diameter of the solid region to the tumor, the ratio of the tumor volume under different image windows, and the ratio of the solid region to the total tumor volume; and the geometric parameters include the maximum diameter and volume of the target region.

[0009] Based on the aforementioned technical means, this application embodiment achieves quantitative characterization of tumor-related features through computational visual indicators. This quantification process uses objective geometric parameters as the basis for calculation, replacing the assessment method that relies on doctors' subjective observation. It accurately captures the relative morphological characteristics of the tumor and the solid region while avoiding assessment bias caused by individual experience differences. Simultaneously, the optional settings of various ratio-type indicators can cover the analytical needs of different clinical scenarios, providing quantitative input for subsequent clinical scoring construction and model prediction, further improving the scientific rigor and reliability of the prediction results.

[0010] Optionally, the processor is also used to: determine multiple regions based on image morphological features, extract radiomics features from the multiple regions and, after screening, construct radiomics scores for each of the multiple regions.

[0011] Based on the above technical means, the embodiments of this application can comprehensively capture the heterogeneous characteristics of tumors and related regions; by screening the extracted radiomics features, redundant information is effectively removed, high-value features are retained, and irrelevant data interference is avoided; at the same time, scores are constructed for multiple regions, which can accurately reflect the unique pathological association characteristics of each region, providing more targeted quantitative support for subsequent input vector construction and model prediction, and further improving the comprehensiveness and accuracy of the prediction results.

[0012] Optionally, the processor is also used to: obtain the respective weights of inflammatory markers, tumor markers, clinical features, and imaging morphological features; and calculate a clinical score based on the respective weights of inflammatory markers, tumor markers, clinical features, and imaging morphology.

[0013] Based on the aforementioned technical means, this application embodiment assigns weights to inflammatory indicators, tumor markers, clinical features, and imaging morphological features and calculates clinical scores, fully reflecting the different contributions of various features to the preoperative assessment of lung adenocarcinoma. This avoids the problem of information redundancy or weakening of key information caused by indiscriminate feature fusion. At the same time, the weight-based quantitative calculation method makes the clinical scores more consistent with the actual logic of tumor pathological assessment, improving the objectivity and pertinence of the scores. It provides accurate quantitative basis for subsequent input vector construction and model prediction, further enhancing the reliability and scientific nature of the prediction results.

[0014] Optionally, the prediction model includes an embedding layer, a Transformer encoder, and a multi-task output head. The embedding layer maps the input vector to a pre-defined vector space to generate embedding vectors for visual indicators, radiomics scores, and clinical scores in the vector space. The Transformer encoder concatenates the embedding vectors of visual indicators, radiomics scores, and clinical scores in the vector space and performs a fusion feature representation on the concatenated vectors through a self-attention mechanism. The multi-task output head executes a first prediction task and a second prediction task in parallel based on the fusion feature representation. The first prediction task is an invasiveness prediction task, and the second prediction task is a lymph node metastasis prediction task.

[0015] Based on the aforementioned technical means, this application employs a prediction model composed of an embedding layer, a Transformer encoder, and a multi-task output head. The embedding layer maps the input vector to a preset vector space and generates corresponding embedding vectors, laying the foundation for deep feature fusion. Leveraging the self-attention mechanism of the Transformer encoder, the embedding vectors of visual indicators, radiomics scores, and clinical scores are efficiently spliced ​​and fused, accurately capturing the complex relationships between features of different dimensions and avoiding information fragmentation caused by simple feature superposition. The multi-task output head executes two prediction tasks—invasiveness and lymph node metastasis—in parallel, simultaneously mining the correlation information between the two prediction targets, improving prediction efficiency and collaborative accuracy. This model enhances the depth and specificity of feature fusion while achieving parallel advancement of prediction tasks, further improving the comprehensiveness, accuracy, and clinical applicability of preoperative prediction for lung adenocarcinoma.

[0016] Optionally, the preoperative prediction results for lung adenocarcinoma include the probability of invasiveness in four categories and the probability of lymph node metastasis in two categories.

[0017] Based on the aforementioned technical means, this application embodiment clearly defines the preoperative prediction results for lung adenocarcinoma as a four-category probability of invasiveness and a two-category probability of lymph node metastasis, achieving a detailed representation of the prediction results. Specifically, the four-category probability of invasiveness can accurately match the clinical needs for assessing tumor invasiveness grading, while the two-category probability of lymph node metastasis can clearly reflect the level of metastasis risk. The probabilistic output provides clinicians with an intuitive quantitative reference, replacing vague qualitative judgments, facilitating doctors to more accurately grasp the severity of the disease and formulate individualized surgical plans, further improving the clinical applicability and decision-making reference value of the prediction results.

[0018] A second aspect of this application provides a medical device, including the lung adenocarcinoma preoperative prediction system of the above embodiments.

[0019] A third aspect of this application provides a data processing method for a preoperative prediction system for lung adenocarcinoma. The method is applied to the processor of the preoperative prediction system for lung adenocarcinoma described above. The method includes the following steps: receiving multiple data from a lung adenocarcinoma patient, including inflammatory markers, tumor markers, CT images, and clinical data, through an interactive interface; extracting clinical features from the clinical data and extracting morphological features from the CT images; calculating visual indicators based on the morphological features; calculating a radiomics score based on the radiomics feature mining results of the morphological features; calculating a clinical score based on the inflammatory markers, tumor markers, clinical features, and morphological features; constructing an input vector based on the visual indicators, radiomics score, and clinical score; inputting the input vector into a pre-trained prediction model; and outputting the preoperative prediction result for lung adenocarcinoma through the input model.

[0020] Based on the aforementioned technical means, this application embodiment automates the data processing for preoperative prediction of lung adenocarcinoma through a standardized process of "receiving multimodal data, extracting features, calculating three types of quantitative scores, constructing input vectors, and outputting model results." By automating feature extraction and score calculation, it replaces the manual subjective analysis mode, avoiding assessment bias caused by individual experience differences. This ensures both the standardization and repeatability of the processing flow, and improves the objectivity and accuracy of the prediction results through the integration of multi-dimensional information, providing a standardized and feasible technical implementation path for clinical preoperative assessment.

[0021] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a data processing method for a lung adenocarcinoma preoperative prediction system as described in the above embodiments.

[0022] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0023] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is an architecture diagram of a preoperative prediction system for lung adenocarcinoma provided according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a medical device provided according to an embodiment of this application; Figure 3 This is a flowchart of a data processing method for a preoperative prediction system for lung adenocarcinoma provided according to an embodiment of this application. Detailed Implementation

[0024] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0025] Currently, accurate preoperative staging of lung adenocarcinoma, especially invasiveness, and differentiation between AIS (Adenocarcinoma in situ), MIA (Minimally Invasive Adenocarcinoma), and IAC (Invasive Adenocarcinoma), as well as lymph node metastasis, are crucial for developing individualized surgical plans and improving patient prognosis. Radiomics based on CT (Computed Tomography) images, with its high-throughput image feature extraction capabilities, has become an important non-invasive tool for assessing tumor heterogeneity. Current development mainly focuses on peritumoral region feature analysis, such as texture analysis in 2mm and 5mm ranges, and multimodal information fusion, such as combining CECT (Contrast-Enhanced Computed Tomography) radiomics features with morphological features. These two directions have improved the ability to differentiate sub-centimeter lung adenocarcinoma nodules to some extent, demonstrating potential to assist in clinical decision-making.

[0026] The relevant technologies have the following shortcomings in the field of preoperative prediction of lung adenocarcinoma: Predictive performance is limited and stability is poor; the validation set AUC (Area Under the ROC Curve) is mostly between 0.85 and 0.89, and its generalization and robustness are limited due to differences in study design, sample size, and equipment parameters; key technologies and definitions are not standardized, and there is a lack of unified standards for the scope of peritumoral analysis, CT acquisition parameters, and preprocessing methods, making the process difficult to reproduce; feature utilization is limited, relying on manual feature selection or subjective image evaluation, and failing to deeply integrate multi-dimensional information, making it impossible to uncover complex relationships between features; clinical applicability and interpretability are insufficient, with some models relying on enhanced CT, limiting their applicability, and many being "black box" structures, with opaque decision-making logic reducing clinical trust.

[0027] In summary, although the relevant technologies have shown potential to assist clinical decision-making, the aforementioned shortcomings make it difficult to achieve accurate and stable predictive effects, which restricts their in-depth transformation and widespread application in clinical practice.

[0028] The following description, with reference to the accompanying drawings, describes a preoperative prediction system, medical device, method, and medium for lung adenocarcinoma according to embodiments of this application. Addressing the issues mentioned in the background art, such as reliance on physician subjective experience, limited features, and insufficient interpretability of prediction models, this application provides a preoperative prediction system for lung adenocarcinoma invasiveness and lymph node metastasis based on multimodal artificial intelligence. The system integrates visual imaging indicators: C / T ratio (Consolidation / Tumor Ratio), TDR (Tumor Density Ratio), VE (Volume Ratio of Consolidation to Tumor), deep radiomics features (intratumoral GTV (Gross Tumor Volume) and peritumoral GPTV (Gross Peritumoral Volume) regions) and clinical indicator scores to construct a multi-task deep learning model based on the Transformer architecture. The aim is to simultaneously and accurately predict the invasiveness grade of lung adenocarcinoma (AAH (Atypical Adenomatous Hyperplasia) / AIS / MIA / IAC, 4 categories) and lymph node metastasis status (N0 / N+, 2 categories) without relying on intraoperative pathology, providing key basis for clinical surgical options, such as decisions on sublobar resection versus lobectomy.

[0029] Specifically, Figure 1 This is an architectural diagram of a preoperative prediction system for lung adenocarcinoma provided in an embodiment of this application.

[0030] like Figure 1 As shown, the preoperative prediction system 10 for lung adenocarcinoma includes: an interactive interface 110 and a processor 120.

[0031] The interactive interface 110 is used to respond to user interaction actions, and input multiple data from the lung adenocarcinoma patient's inflammatory indicators, tumor markers, CT images, and clinical data according to the interaction actions. The processor 120 is connected to the interactive interface 110 and is used to extract clinical features from clinical data, extract image morphology features from CT images, calculate visual indicators of lung adenocarcinoma based on image morphology features, calculate radiomics scores based on radiomics feature mining results of image morphology features, calculate clinical scores based on inflammatory indicators, tumor markers, clinical features, and image morphology features, construct an input vector based on visual indicators, radiomics scores, and clinical scores, input the input vector into a pre-trained prediction model, and output the preoperative prediction results of lung adenocarcinoma through the input model.

[0032] It is understood that this application embodiment constructs a standardized and automated preoperative prediction system for lung adenocarcinoma through the collaboration of the interactive interface 110 and the processor 120. The interactive interface enables unified input of multiple data such as inflammatory indicators, tumor markers, and CT images, breaking the limitations of single data dimensions; the processor is responsible for data processing, automatically extracting clinical features and imaging morphological features, and converting them into three types of quantitative results: visual indicators, radiomics scores, and clinical scores, and finally outputting prediction results through the model. This not only replaces the doctor's subjective judgment and single feature analysis, avoiding assessment bias caused by experience differences, but also captures tumor characteristics through multi-dimensional information, improving the objectivity and accuracy of prediction, providing reliable support for the formulation of individualized surgical plans, and facilitating the clinical application of the technology.

[0033] It should be noted that inflammatory markers include NLR (Neutrophil-to-Lymphocyte Ratio), tumor markers include CEA (Carcinoembryonic Antigen), CT images refer to preoperative thin-slice (≤1mm) scan images, and clinical data covers basic information such as age, smoking index, and medical history. The clinical features extracted by the processor are key information with predictive value, such as smoking history and comorbidities. Imaging morphological features include nodule size, marginal spiculation, and vascular convergence, which are visually identifiable. The calculated visual indicators are parameters that quantify tumor consolidation and density characteristics. The radiomics score is generated after screening based on intratumoral and peritumoral regional features. The clinical score is a weighted comprehensive result of multiple indicators. The prediction model used is built based on the Transformer architecture, and the final output of the preoperative prediction results for lung adenocarcinoma specifically includes the probability of four-category invasiveness and the probability of two-category lymph node metastasis.

[0034] In this embodiment of the application, the processor 120 is further configured to: perform preprocessing on the CT image before extracting image morphological features from the CT image, the preprocessing including at least one of image normalization, image enhancement and annotation, image normalization being used to normalize the density value of the CT image to a preset range, image enhancement employing an adaptive algorithm, and annotation being manually determined for the target region.

[0035] It is understood that the embodiments of this application effectively solve the problems of uneven quality, blurred details, and unclear target areas in the original image data by setting optional preprocessing procedures for CT images. Image standardization can eliminate density value differences caused by different CT equipment and scanning parameters, improving data consistency; image enhancement strengthens the detailed features of the tumor and surrounding areas through adaptive algorithms, facilitating accurate extraction of morphological information in the subsequent process; manual annotation ensures accurate positioning of the target area and reduces interference from irrelevant areas. This approach retains processing flexibility while optimizing image data quality from the source, providing reliable input for subsequent feature extraction and index calculation, and further improving the accuracy and stability of prediction results.

[0036] It should be noted that the "preset range" of CT image preprocessing in this application refers to the normalized range of CT density values ​​(HU (Hounsfield Unit) values). In the embodiment, -1000 to +1000HU can be selected. The specific value can be adjusted according to the actual application scenario. This application does not make a unique limitation. The "target area" is the tumor boundary and the consolidation area, which needs to be marked by a professional radiologist. This application does not limit the specific marking details, as long as it can meet the needs of subsequent feature extraction and model training.

[0037] Specifically, the system first performs standardization processing on the input preoperative thin-section (≤1mm) CT images, i.e., data preprocessing, to ensure data quality and consistency. This includes image standardization, image enhancement, and expert annotation and quality control: image standardization unifies CT scan parameters and normalizes HU values ​​to -1000 to +1000; image enhancement uses the CLAHE (Contrast-Limited Adaptive Histogram Equalization) algorithm to enhance image details; and annotation is performed by radiologists on the CT images to mark tumor boundaries and consolidation areas, which serve as standards for subsequent automatic segmentation model training and validation.

[0038] The preprocessing workflow is as follows: all CT images are uniformly converted to DICOM (Digital Imaging and Communications in Medicine) format and quality assessment is performed. After grayscale normalization, isotropic resampling is applied to ensure consistent spatial resolution. An adaptive Gaussian filter is used to dynamically adjust parameters based on the noise level to reduce interference. The annotation process is completed independently by two thoracic radiologists with more than 5 years of experience using ITK-SNAP software, combining a semi-automatic segmentation algorithm with manual correction to ensure accuracy.

[0039] In this embodiment of the application, the processor 120 is further configured to: calculate visual indicators based on the geometric parameters of the segmented target region, wherein the visual indicators include at least one of the following: the ratio of the diameter of the solid region to the tumor, the ratio of the tumor volume under different image window positions, and the ratio of the solid region to the total tumor volume; and the geometric parameters include the maximum diameter and volume of the target region.

[0040] It is understood that the embodiments of this application can achieve quantitative characterization of tumor-related features through computational visual indicators. This quantification process is based on objective geometric parameters, replacing the assessment method that relies on the subjective observation of doctors. It accurately captures the relative morphological characteristics of tumors and consolidation areas, while avoiding assessment bias caused by individual experience differences. At the same time, the optional settings of various ratio-type indicators can cover the analysis needs of different clinical scenarios, providing quantitative input for subsequent clinical scoring construction and model prediction, further improving the scientificity and reliability of the prediction results.

[0041] Specifically, the calculation of visual metrics is based on the accurate segmentation of tumor and solid regions. Segmentation can be achieved using zero-shot / few-shot segmentation techniques based on large medical imaging models (such as SAM-Med2D / 3D) or through nnUNet training.

[0042] Furthermore, the segmentation of the tumor region and solid region is preferably achieved using a 3D medical image segmentation network based on the nnU-Net framework. The nnU-Net network adopts an encoder-decoder U-shaped structure, with spatial detail information transmitted between the encoder and decoder through same-scale skip connections. The encoder part includes multiple progressively downsampled resolution levels, each level including at least two convolutional units. Each convolutional unit consists of a 3D convolutional layer, an instance normalization layer, and a nonlinear activation function, wherein the nonlinear activation function is preferably Leaky ReLU; downsampling is achieved between adjacent levels through stride convolution. The decoder part includes multiple progressively upsampled levels corresponding to the encoder levels. Each level first restores the spatial resolution through transposed convolution or upsampled convolution, then concatenates it with the feature map of the corresponding layer of the encoder, and completes feature fusion through at least two convolutional units to output the voxel-level segmentation result of the target region.

[0043] In a preferred embodiment, the network may employ a multi-layer resolution hierarchical structure, such as a six-layer resolution hierarchical structure. The input is a three-dimensional volumetric data block of lung CT. The number of channels in each layer of the encoder is set to 32, 64, 128, 256, 320, and 320 in a layer-by-layer incremental manner, and the decoder recovers the data layer by layer in a symmetrical manner. The size of each convolution kernel is preferably 3×3×3, and some low-resolution directions can be adaptively adjusted to 1×3×3 or 3×3×1 according to the voxel spacing to adapt to the anisotropy of thin-slice CT in the axial and slice thickness directions. The network ends with a 1×1×1 convolution to output a segmentation category probability map, wherein the category includes at least one or more of the following: background, nodule / tumor region, and solid region.

[0044] The nnU-Net framework can further include a deep supervision mechanism and, in some implementations, introduce residual connections or other feature enhancement structures to enhance multi-scale feature representation capabilities. The deep supervision mechanism refers to outputting auxiliary segmentation maps at multiple scales of the decoder, calculating auxiliary losses separately, and then weighted summing them to enhance the collaborative training capability of shallow and deep features, improving the recognition accuracy for ground-glass components, partially real components, and regions with blurred boundaries. Through this structure, without relying on manual design of the network topology, the convolution kernel, stride, normalization method, and patch size can be adaptively determined based on the voxel spacing, field of view size, and target region scale of the training data, thereby improving adaptability to different CT equipment and scanning protocols.

[0045] Furthermore, the training data for the segmentation model preferably includes publicly available medical imaging datasets and retrospectively annotated data from within hospitals. The publicly available datasets are preferably selected from at least one of the following: the LIDC-IDRI lung nodule dataset, the lung lesion-related dataset from Medical Segmentation Decathlon, or the publicly available datasets from The Cancer Imaging Archive (TCIA) that can be used for learning the contours of lung tumors or nodules. When within-hospital data is available, desensitized preoperative thin-slice chest CT scans can be combined with the publicly available datasets for transfer training or incremental training. The publicly available datasets are used to improve the model's initial representation ability and cross-center generalization ability, while the within-hospital annotated data are used to enhance the adaptability to the target task of this application, namely, lung adenocarcinoma nodules, consolidation components, and peritumoral boundary features.

[0046] In a preferred embodiment, the following preprocessing steps are performed on the original CT images before training: (1) Reconstruct the original DICOM sequence into three-dimensional volume data and convert it to NIfTI format; (2) Perform window width and window level clipping and normalization on the CT gray value. It is preferable to cut the HU value into the range of [-1000, 400] or [-1000, 1000] and then perform linear normalization. (3) Perform isotropic or quasi-isotropic resampling on the image, preferably resampling to 1.0mm×1.0mm×1.0mm or voxel spacing consistent with the statistical median spacing of the training set; (4) A coarse lung field extraction method based on lung window threshold and connected component analysis was adopted to remove irrelevant areas such as chest wall and bed board to narrow the search range; (5) Perform consistency checks on the annotation masks. For samples with inconsistent annotations by multiple physicians, generate reference labels through review by senior physicians, consensus confirmation, or fusion of multiple annotation results. (6) Training patches are cropped based on the target region location, and the ratio of positive and negative samples is balanced to avoid training bias caused by an excessively high proportion of background voxels.

[0047] Furthermore, data augmentation includes at least one or more of the following: random rotation, random scaling, random flipping, Gamma intensity transformation, Gaussian noise perturbation, Gaussian blur, low-resolution simulation, and random cropping. For ground-glass nodules and some solid nodules, random contrast perturbation and local elastic deformation enhancement can be added to enhance boundary learning capabilities. Through preprocessing and enhancement workflows, the robustness of the model to images from different manufacturers' CT equipment, with different slice thicknesses, and with different noise levels can be improved.

[0048] Furthermore, the nnU-Net segmentation model is trained using supervised learning. Training labels are assigned by at least two physicians with experience in chest imaging, who annotate the tumor contour and consolidation region. If discrepancies exist, a senior physician reviews and determines the final labels. A composite loss function is used during training, comprising at least Dice loss and cross-entropy loss. Dice loss enhances adaptability to small-volume targets and class imbalance, while cross-entropy loss improves class discrimination stability. Preferably, the total loss function is expressed as: .in, and The loss weighting coefficient is preferably set within the ranges of 0.3–0.7 and 0.3–0.7, respectively. In one embodiment, it can be set to... =0.5, =0.5. For samples with a low proportion of real variation regions or ambiguous boundaries, boundary loss or Top-k loss can be introduced as additional terms to improve boundary consistency. In other implementations, Dice loss, cross-entropy loss, Focal loss, boundary loss, or combinations thereof can also be used.

[0049] The optimization algorithm preferably employs either the Stochastic Gradient Descent (SGD) optimizer or the Adam / AdamW optimizer. In a preferred embodiment, the SGD optimizer is used with an initial learning rate of 0.01, a momentum of 0.99, and a weight decay coefficient of [missing value]. A multinomial decay learning rate strategy is employed, with the preferred number of training iterations being 500 to 1000 epochs. For memory-constrained scenarios, a mixed-precision training and gradient accumulation strategy can be used. The batch size is adaptively set according to GPU memory, preferably 2 to 8. In some implementations, an early stopping strategy can be set to reduce the risk of overfitting. Training is terminated and the optimal weights are saved when the Dice coefficients on the validation set do not improve within a preset number of epochs.

[0050] Furthermore, a five-fold cross-validation process is employed for training. All training samples are divided into five subsets based on the patient dimension. Each time, one subset is selected as the validation set, and the remaining subsets are used as the training set. This process is repeated five times, and the optimal model for each fold is saved. During the inference phase, either a single model output or an ensemble of five-fold models can be used; it is preferable to use probability averaging or majority voting to form the final segmentation result to improve model stability.

[0051] Furthermore, the key hyperparameters of the segmentation model include input patch size, batch size, number of base channels, learning rate, loss weights, sampling strategy, and inference threshold. The parameter settings are based on factors such as the voxel spacing statistical distribution of the training data, the range of nodule target sizes, GPU memory capacity, and validation set performance. Preferably: (1) The input patch size is determined by the maximum diameter of the target nodule, the peritumoral analysis range and the video memory capacity. The three-dimensional patch is preferably one of 128×128×128, 96×160×160 or 64×192×192. (2) The batch size is preferably set to 2 to 4 to balance gradient stability and memory usage; (3) The initial learning rate is preferably searched within the range of 0.001 to 0.01; (4) The weights of Dice loss and cross-entropy loss are preferably adjusted within the range of [0.3, 0.7]; (5) The sampling probability of positive sample patches is preferably set to 0.33 to 0.66 to ensure that the lesion-containing areas are fully learned; (6) During inference, a sliding window prediction method is adopted, with the window overlap rate preferably being 0.5, and the prediction results of the overlapping area are fused by Gaussian weighting; (7) The output probability map is thresholded and processed by three-dimensional connected domain post-processing to obtain the final segmentation mask. The threshold is preferably 0.4 to 0.6, and the largest connected lesion or candidate region that meets the volume constraint is retained.

[0052] Furthermore, the hyperparameters are not fixed but are obtained through a combination of nnU-Net's default adaptive configuration, validation set parameter tuning, grid search, or manual verification. The optimal parameter combination is chosen as the final training parameters, based on the combined optimal values ​​of the validation set Dice coefficients, 95% Hausdorff distance, and average surface distance.

[0053] Based on the tumor region mask and solid region mask output by the above segmentation model, visual indicators can be further automatically calculated. Specifically, the maximum tumor diameter is obtained from the maximum Euclidean distance or the longest diameter of the maximum cross-section of the 3D tumor region mask; the maximum solid diameter is obtained from the solid region mask in the same way; the C / T ratio is defined as the ratio of the maximum solid diameter to the maximum tumor diameter. The total tumor volume is obtained by multiplying the number of voxels in the tumor mask by the voxel volume, and the solid volume is obtained by multiplying the number of voxels in the solid mask by the voxel volume; VE is defined as the ratio of the solid volume to the total tumor volume. The volume ratio under different window levels can be obtained by voxel screening of the same segmented region under different HU threshold conditions. For example, the solid component volume can be counted under the mediastinal window threshold condition, and the total nodule volume can be counted under the lung window threshold condition, and a TDR or similar volume density ratio can be constructed. All the above indicators are automatically generated by the program, thereby reducing human measurement errors and improving repeatability.

[0054] To verify the technical effectiveness of the nnU-Net-based segmentation scheme, experimental evaluation can be conducted using a public dataset and an independent test set. Preferably, the public dataset is divided into a training set, a validation set, and a test set, and an independent test set outside the hospital is set up when conditions permit to evaluate cross-center generalization performance. Segmentation performance metrics include at least the Dice similarity coefficient, intersection-over-union ratio, 95% Hausdorff distance, and mean surface distance; visual consistency metrics include at least the intraclass correlation coefficient (ICC) or Pearson correlation coefficient between automatic measurements and physician measurements; downstream predictive value is evaluated by inputting the C / T ratio, VE, and Rad-score obtained from automatic segmentation into the subsequent prediction model, and using AUC, accuracy, sensitivity, specificity, and calibration curves.

[0055] Furthermore, in another embodiment, the segmentation of the tumor region and the solid region can be achieved using a cue-based segmentation method based on the MedSAM medical imaging fundamental segmentation model. The MedSAM model preferably comprises three parts: an image encoder, a cue encoder, and a mask decoder. The image encoder extracts high-dimensional visual features from the input medical image, the cue encoder encodes external cue information into cue features, and the mask decoder fuses the image features and cue features to output a segmentation mask for the target region. This structure inherits the overall framework of the Segment Anything Model and is adapted for medical imaging scenarios, making it suitable for chest CT target region segmentation in this application.

[0056] In a preferred embodiment, the input image is a two-dimensional slice of a preoperative chest CT scan or a key-level image selected from three-dimensional volume data. After preprocessing, it is sent to the image encoder. The cue information is preferably a bounding box cue surrounding the target nodule region, but can also be one or more of positive / negative point cue, coarse segmentation mask cue. The cue encoder converts the bounding box or point cue into a cue embedding representation aligned with image features. The mask decoder outputs a probability mask of the target nodule region based on the image features and the cue embedding, and obtains the final segmentation result after thresholding and connected component filtering. Public information shows that MedSAM is commonly used to guide segmentation through cueing such as bounding boxes, rather than directly outputting end-to-end without any cueing. When using two-dimensional cue segmentation, reasoning can be performed layer by layer or layer by layer on the three-dimensional CT volume data, and the segmentation results of each layer can be stacked, interpolated, or fused into a three-dimensional target region mask for subsequent volume and scale index calculations.

[0057] Furthermore, the MedSAM model can be directly inferred using pre-trained weights, or it can be adapted using few-sample data on the task data described in this application. For the implementation method that directly calls the pre-trained model, the focus of this application is not on retraining the base model itself, but on inputting the pre-processed lung CT image and target region prompts into the pre-trained MedSAM model to obtain segmentation results for the tumor region and the solid region. For the implementation method that uses pre-trained weights for direct inference, the focus of this application is on the preprocessing of the input image, the construction of the target region prompts, the model inference process, and the post-processing of the segmentation results.

[0058] In another preferred embodiment, to improve the adaptability of MedSAM in scenarios involving lung adenocarcinoma nodules, ground-glass nodules, or partially solid nodules, the pre-trained MedSAM can be fine-tuned on a small number of labeled samples. Fine-tuning can involve updating only the mask decoder parameters, or updating some of the later-layer parameters of the image encoder while freezing the remaining parameters; the cue encoder is preferably kept frozen. Publicly available code examples also include implementations that train only the image encoder and mask decoder and freeze the cue encoder.

[0059] Furthermore, when using MedSAM for target segmentation, the input data is preferably thin-slice chest CT images. The preprocessing steps include at least: (1) Convert the original DICOM sequence into two-dimensional slices or three-dimensional volume data; (2) The grayscale values ​​are cropped to the HU window width and window level, preferably to the range corresponding to the lung window or mediastinal window; (3) Resample or scale the image size to meet the model input requirements; (4) Normalize the image; (5) Generate initial bounding box prompts for the target region through thresholding, lung field segmentation, candidate nodule detection or manual interaction.

[0060] In this approach, if 2D MedSAM inference is used, bounding box cues can be generated layer by layer for CT slices, and 2D masks can be output. These masks are then stacked to form a 3D nodule region. If key-level inference is used, the largest nodule layer and adjacent layers can be segmented separately and combined to form the target contour. Bounding box cues can originate from manual selection, automatic output from the detection network, or automatic generation from the bounding rectangle of the coarse segmentation result. Using these methods, even without retraining the base model, those skilled in the art can achieve the target region extraction required in this application based on the existing pre-trained MedSAM. MedSAM papers and official repositories both demonstrate its implementation path based on medical image preprocessing followed by cue segmentation.

[0061] Based on the segmentation results, three key visual indicators were calculated: the ratio of the diameter of the solid region to that of the tumor (C / T ratio), the ratio of the maximum diameter of the solid region to that of the maximum diameter of the tumor, the ratio of the tumor volume under different imaging windows (TDR), the ratio of the tumor volume in the mediastinal window (threshold ≥ -50HU) to that in the lung window (threshold ≤ -600HU), and the ratio of the solid region to the total tumor volume (VE).

[0062] In this embodiment of the application, the processor 120 is further configured to: determine multiple regions based on image morphological features, extract radiomics features from the multiple regions and, after screening, construct radiomics scores for each of the multiple regions.

[0063] It is understood that the embodiments of this application can comprehensively capture the heterogeneous characteristics of tumors and related regions; by screening the extracted radiomics features, redundant information is effectively removed, high-value features are retained, and irrelevant data interference is avoided; at the same time, scores are constructed for multiple regions, which can accurately reflect the unique pathological association characteristics of each region, providing more targeted quantitative support for subsequent input vector construction and model prediction, and further improving the comprehensiveness and accuracy of the prediction results.

[0064] It should be noted that in this application, "multiple regions" refers to regions at different distances within and around the tumor. In the embodiments, the intratumoral region GTV and the peritumoral region GPTV3 / 6 / 9 can be selected. The specific region division can be adjusted according to actual needs, and this application does not impose a unique limitation. Radiomic features are the image quantitative features extracted from each region. The "screening process" can adopt processes such as stability verification and correlation elimination. This application does not limit the specific type of feature extraction or the detailed steps of screening. The final goal is to construct a radiomic score with predictive value for each region.

[0065] Specifically, the study included multiple regions, including the intratumoral region (GTV) and several peritumoral regions (such as GPTV3, GPTV6, and GPTV9). Morphological, first-order statistical, and textural features (such as GLCM (Gray-Level Co-occurrence Matrix) and GLRLM (Gray-Level Run-Length Matrix)) were extracted from each region using the PyRadiomics library, resulting in 1218 initial features. After LoG (Laplacian of Gaussian) filtering and wavelet transform to enhance heterogeneity information capture, a four-step screening method was used to refine the features: ① Reproducible features were screened based on the intraclass correlation coefficient (ICC > 0.75) calculated using a two-way random effects model; ② Highly correlated features with a correlation coefficient r > 0.90 were removed; ③ An improved mRMR (Minimum Redundancy Maximum Ratio) was introduced by incorporating weighted mutual information. The Relevance algorithm (minimum redundancy maximum relevance) balances feature relevance and redundancy; ④ Adaptive weight-adjusted LASSO (Least Absolute Shrinkage and Selection Operator) regression is used to achieve feature dimensionality reduction, and finally a radiomics score (Rad-score) is constructed for each region.

[0066] In this embodiment of the application, the processor 120 is further configured to: obtain the respective weights of inflammatory indicators, tumor markers, clinical features and imaging morphological features; and calculate a clinical score based on the respective weights of inflammatory indicators, tumor markers, clinical features and imaging morphology.

[0067] It is understood that the embodiments of this application, by assigning weights to inflammatory indicators, tumor markers, clinical features, and imaging morphological features and calculating clinical scores, fully reflect the different contributions of various features to the preoperative assessment of lung adenocarcinoma, avoiding the problem of information redundancy or weakening of key information caused by indiscriminate feature fusion; at the same time, the weight-based quantitative calculation method makes the clinical scores more in line with the actual logic of tumor pathological assessment, improving the objectivity and pertinence of the scores, providing accurate quantitative basis for subsequent input vector construction and model prediction, and further enhancing the reliability and scientific nature of the prediction results.

[0068] It should be noted that the weights of various features in this application can be determined by multivariate logistic regression or LASSO regression in the embodiments. The specific acquisition method can be adjusted according to the actual scenario, and this application does not impose a unique limitation. The clinical score is calculated based on the weights of the above four types of features. This application does not impose a strict limitation on the specific weight values ​​of each feature, as long as a score with clinical predictive value can be obtained.

[0069] Specifically, the Clinical-score system integrates indicators including inflammatory markers, clinical features, and imaging morphological features. The selection of indicators follows the dual criteria of clinical importance and statistical significance. The relative importance is determined by AHP (Analytic Hierarchy Process). The weights are determined using a 1000-times Bootstrap resampling technique to ensure stability. Finally, a comprehensive clinical score is constructed through multivariate logistic regression or LASSO regression. The score is validated through internal 10-fold cross-validation and external multicenter validation to ensure its practicality.

[0070] In this embodiment, the prediction model includes an embedding layer, a Transformer encoder, and a multi-task output head. The embedding layer maps the input vector to a pre-defined vector space to generate embedding vectors for visual indicators, radiomics scores, and clinical scores in the vector space. The Transformer encoder concatenates the embedding vectors of visual indicators, radiomics scores, and clinical scores in the vector space and performs a fusion feature representation on the concatenated vectors through a self-attention mechanism. The multi-task output head executes a first prediction task and a second prediction task in parallel based on the fusion feature representation. The first prediction task is an invasiveness prediction task, and the second prediction task is a lymph node metastasis prediction task.

[0071] It is understood that the prediction model in this application employs an embedding layer, a Transformer encoder, and a multi-task output head. The embedding layer maps the input vector to a preset vector space and generates corresponding embedding vectors, laying the foundation for deep feature fusion. Leveraging the self-attention mechanism of the Transformer encoder, the embedding vectors of visual indicators, radiomics scores, and clinical scores are efficiently concatenated and fused, accurately capturing the complex relationships between features of different dimensions and avoiding information fragmentation caused by simple feature superposition. The multi-task output head executes two prediction tasks—invasiveness and lymph node metastasis—in parallel, simultaneously mining the correlation information between the two prediction targets, improving prediction efficiency and collaborative accuracy. This model enhances the depth and specificity of feature fusion while achieving parallel advancement of prediction tasks, further improving the comprehensiveness, accuracy, and clinical applicability of preoperative prediction for lung adenocarcinoma.

[0072] It should be noted that the pre-set vector space is a high-dimensional feature carrier constructed to achieve multimodal feature fusion. Its function is to uniformly map the three types of data—visual indicators, radiomics scores, and clinical scores—to a standardized space of the same dimension to generate embedding vectors. The specific dimension and mapping rules can be adjusted according to the number of features and model requirements. This application does not impose a unique limitation on this task, as long as it can provide an adaptation basis for subsequent fusion. The first prediction task is the invasiveness prediction task. In this embodiment, it is specifically a four-category prediction of lung adenocarcinoma invasiveness, outputting the probabilities of AAH, AIS, MIA, and IAC. This application does not impose a limitation on the level of classification refinement, as long as it meets the needs of clinical preoperative assessment. The second prediction task is the lymph node metastasis prediction task. In this embodiment, it is specifically a binary prediction of lymph node positive (N+) and negative (N0). This application does not impose a unique limitation on the judgment threshold for prediction, as long as it can provide a reliable basis for clinical lymph node status assessment.

[0073] Specifically, the input embedding layer of the prediction model maps visual indicators, Rad-scores, and Clinical-scores for each region to a high-dimensional vector space. The embedding process uses learnable positional encoding based on image spatial coordinates to transform the two-dimensional image positional information into a sequence. The Transformer encoder has four layers with a hidden layer dimension of 256 and is configured with eight attention heads (each head learns different types of feature relationships). Adaptive layer normalization and residual connections ensure training stability. The multi-task output head performs an aggressive prediction task using the Softmax activation function, outputting four-class probabilities (AAH, AIS, MIA, IAC). It performs a lymph node metastasis prediction task using the Sigmoid activation function, outputting binary probabilities (N+) and negative (N0) lymph nodes. During training, shared feature representation and joint loss function optimization are used to achieve complementary enhancement of tasks.

[0074] In the embodiments of this application, the preoperative prediction results for lung adenocarcinoma include the probability of invasiveness in four categories and the probability of lymph node metastasis in two categories.

[0075] It is understood that the embodiments of this application explicitly define the preoperative prediction results for lung adenocarcinoma as the probability of four-category invasiveness and the probability of two-category lymph node metastasis, achieving a detailed representation of the prediction results. Among them, the four-category invasiveness probability can accurately match the clinical assessment needs for tumor invasiveness grading, and the two-category lymph node metastasis probability can clearly reflect the level of metastasis risk. The probabilistic output provides clinicians with an intuitive quantitative reference, replacing vague qualitative judgments, making it easier for doctors to more accurately grasp the degree of disease and formulate individualized surgical plans, further improving the clinical adaptability and decision-making reference value of the prediction results.

[0076] Specifically, the preoperative prediction results for lung adenocarcinoma include the four-category probability of invasiveness and the two-category probability of lymph node metastasis: the four-category probability of invasiveness is the classification probability of AAH, AIS, MIA, and IAC, and the two-category probability of lymph node metastasis is the classification probability of lymph node positive (N+) and negative (N0). To improve the reliability of the results, the system uses SHAP (SHapley Additive explanations) value analysis (based on a tree model calculation and incorporating clinical knowledge constraints) to quantify the contribution of each feature, visualized in force maps or heatmaps; at the same time, an improved t-SNE (t-Distributed Stochastic Neighbor Embedding) algorithm (multi-scale similarity calculation) is used to reduce the dimensionality of high-dimensional fused features for intuitive verification of patient stratification. The prediction results can be combined with clinical scores to generate low, medium, and high risk levels, providing quantitative support for surgical decisions such as sublobar resection or lobectomy.

[0077] The preoperative prediction system for lung adenocarcinoma proposed in this application constructs a standardized and automated preoperative prediction system for lung adenocarcinoma through the collaboration of an interactive interface and a processor. The interactive interface enables unified input of multiple data sources, including inflammatory indicators, tumor markers, and CT images, breaking the limitations of single data dimensions. The processor handles data processing, automatically extracting clinical and imaging morphological features and converting them into three types of quantitative results: visual indicators, radiomics scores, and clinical scores. Finally, the model outputs the prediction results. This system not only replaces subjective judgment and single-feature analysis by physicians, avoiding assessment biases caused by experience differences, but also captures tumor characteristics through multi-dimensional information, improving the objectivity and accuracy of predictions. This provides reliable support for the development of individualized surgical plans and facilitates the clinical application of the technology.

[0078] Figure 2 This is a schematic diagram of the structure of a medical device provided in an embodiment of this application. It includes the lung adenocarcinoma preoperative prediction system 10 described in the above embodiment.

[0079] It should be noted that the foregoing explanation of the embodiment of the preoperative prediction system for lung adenocarcinoma also applies to the medical device of this embodiment, and will not be repeated here.

[0080] The following will combine Figure 3 The data processing method for the preoperative prediction system of lung adenocarcinoma is described. This method is applied to the processor of the preoperative prediction system of lung adenocarcinoma in the above embodiment, and includes the following steps: In step S301, multiple data from a lung adenocarcinoma patient, including inflammatory markers, tumor markers, CT images, and clinical data, are received through an interactive interface. Clinical features are extracted from the clinical data, and morphological features are extracted from the CT images.

[0081] It is understood that this application embodiment uses an interactive interface as the data entry point to acquire multimodal data such as inflammatory indicators, tumor markers, CT images, and clinical data, breaking the limitations of a single data dimension. The processor is responsible for feature extraction, extracting clinical features from clinical data and morphological features from CT images. This approach comprehensively covers tumor pathology, imaging, and clinical information through multimodal data fusion, avoiding information gaps. Furthermore, it replaces manual subjective screening with automated targeted extraction, reducing biases caused by individual experience differences. This provides accurate and consistent input for subsequent processes such as visual indicator calculation and score construction, laying a solid foundation for improving the reliability of prediction results.

[0082] In step S302, visual indicators are calculated based on image morphological features, radiomics scores are calculated based on the radiomics feature mining results of image morphological features, and clinical scores are calculated based on inflammatory indicators, tumor markers, clinical features, and image morphological features.

[0083] It is understood that the embodiments of this application calculate visual indicators based on image morphology features, obtain radiomics scores based on radiomics feature mining results, and then integrate inflammatory indicators, tumor markers, clinical features, and image morphology features to calculate clinical scores. Transforming fragmented features into quantitative scores avoids the bias of subjective judgment and comprehensively integrates multi-dimensional key information, providing accurate quantitative basis for subsequent input vector construction and model prediction.

[0084] In step S303, an input vector is constructed based on visual indicators, radiomics scores, and clinical scores. The input vector is then fed into a pre-trained prediction model, which outputs the preoperative prediction results for lung adenocarcinoma.

[0085] Understandably, the embodiments of this application integrate visual indicators, radiomics scores, and clinical scores to construct an input vector, which is then input into a pre-trained prediction model to ultimately output the preoperative prediction result for lung adenocarcinoma. This fully utilizes multi-dimensional quantitative information, enabling the model to comprehensively uncover key correlations, improve prediction accuracy, and provide a direct and reliable reference for clinical preoperative assessment.

[0086] It should be noted that the foregoing explanation of the embodiment of the preoperative prediction system for lung adenocarcinoma also applies to the data processing method of the preoperative prediction system for lung adenocarcinoma in this embodiment, and will not be repeated here.

[0087] The data processing method for the preoperative prediction system for lung adenocarcinoma proposed in this application automates the data processing for preoperative prediction of lung adenocarcinoma through a standardized process of "receiving multimodal data, extracting features, calculating three types of quantitative scores, constructing input vectors, and outputting results from the model." By automating feature extraction and score calculation, it replaces the manual subjective analysis mode, avoiding assessment bias caused by individual experience differences. This ensures both the standardization and repeatability of the processing flow, and improves the objectivity and accuracy of the prediction results through the integration of multi-dimensional information, providing a standardized and feasible technical implementation path for clinical preoperative assessment.

[0088] In summary, the embodiments of this application have at least the following beneficial effects: (1) Achieving high-precision preoperative dual-task prediction with significantly improved identification accuracy and robustness: Through deep integration of multimodal information, the Transformer multimodal fusion network (4-layer encoder structure, 8 multi-head attention units, 256-dimensional hidden layer configuration) to capture complex feature relationships and multi-scale feature fusion technology, the prediction accuracy of lung adenocarcinoma invasiveness grade (distinguishing between IAC and AIS / MIA) and lymph node metastasis status is excellent, with the test set AUC being improved by about 5%-8% compared to the single model. This advantage stems from the multi-region multi-scale feature extraction method (GTV and GPTV3 / 6 / 9 regions) and the four-step radiomics feature screening method (ICC>0.75, removal of features with r>0.90, mRMR, LASSO), ensuring feature stability and reliability, enabling the model to exhibit good generalization performance and stability under different hospitals and different CT equipment, while the shared feature representation and joint loss function optimization adopted by the dual-task output mechanism further enhance the prediction stability.

[0089] (2) Enhancing Model Interpretability and Clinical Acceptability: Through the model interpretability module, using SHAP value analysis (quantifying feature contribution), t-SNE visualization (intuitively presenting patient stratification), and heatmap display, key imaging areas on which the model's decisions depend are clearly presented, such as the peritumoral high-entropy area and vascular aggregation area, transforming the "black box" model into a "white box" tool. This interpretability technology allows clinicians to clearly understand the role of indicators such as the C / T ratio and the texture features of a certain peritumoral area, which can verify the model's judgment logic, significantly improve the trust in AI results, and promote the implementation of human-machine collaborative diagnosis and treatment models.

[0090] (3) Strong clinical applicability and outstanding decision support and integration capabilities: NCECT (Non-Contrast Enhanced Computed Tomography) can maintain high differentiation accuracy without relying on enhanced CT, making it suitable for patients with contraindications to contrast agents and broadening its application scope; the system can be seamlessly integrated into PACS (Picture Archiving and Communication Systems) workstations, relying on the automatic medical image segmentation and visual index calculation module, and automatically calculating indicators such as the HU value threshold segmentation strategy and C / T ratio, realizing full-process automation from nodule automatic segmentation and feature extraction to risk assessment, greatly improving diagnostic efficiency and reducing the burden on doctors. After combining the output probability results with clinical scores for risk stratification, the AI ​​risk stratification and surgical decision recommendation methods directly guide the decision on the timing and extent of surgery, such as sublobar or lobectomy, helping to downgrade treatment and avoid overtreatment or undertreatment.

[0091] (4) Promoting the implementation of precision medicine and improving patient prognosis, the technical architecture is flexible and scalable: It provides objective and quantitative assessment tools to support the formulation of individualized treatment strategies, such as priority surgery for high-risk IAC patients and active monitoring of AIS / MIA patients, which not only optimizes the treatment effect of patients but also saves medical resources. At the same time, the embodiments of this application have the ability to replace and extend. In terms of model architecture, GNN can be used to replace Transformer or XGBoost to adapt to scenarios with limited computing resources; in terms of feature extraction, features can be automatically learned through pre-trained CNNs such as ResNet and DenseNet, or extended to PET-CT multimodal fusion; in terms of application scenarios, it can be transferred to the prediction of invasiveness of other tumors such as breast cancer and liver cancer, and can also be extended to the fields of efficacy evaluation and prognosis prediction to form a complete diagnosis and treatment support system and further amplify medical value. This flexibility is based on the collaborative design of the system's five major modules (image segmentation and visual index calculation, radiomics feature processing, clinical score construction, Transformer fusion prediction, and interpretable decision support) to ensure stable performance during the expansion process.

[0092] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the data processing method of the preoperative prediction system for lung adenocarcinoma as described above.

[0093] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

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

[0095] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0096] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0097] Those skilled in the art will understand that all or part of the steps of the methods implementing the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0098] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A preoperative prediction system for lung adenocarcinoma, characterized in that, include: An interactive interface is used to respond to user interaction actions, and input multiple data from the lung adenocarcinoma patient's inflammatory indicators, tumor markers, CT images, and clinical data according to the interaction actions. The processor, connected to the interactive interface, is used to extract clinical features from the clinical data, extract morphological features from the CT images, calculate visual indicators for lung adenocarcinoma based on the morphological features, calculate a radiomics score based on the radiomics feature mining results of the morphological features, calculate a clinical score based on the inflammatory indicators, tumor markers, clinical features, and morphological features, construct an input vector based on the visual indicators, radiomics score, and clinical score, input the input vector into a pre-trained prediction model, and output a preoperative prediction result for lung adenocarcinoma through the input model.

2. The preoperative prediction system for lung adenocarcinoma according to claim 1, characterized in that, The processor is further configured to: perform preprocessing on the CT image before extracting image morphological features from the CT image, the preprocessing including at least one of image normalization, image enhancement and annotation, the image normalization being used to normalize the density values ​​of the CT image to a preset range, the image enhancement employing an adaptive algorithm, and the annotation being manually determined for the target region.

3. The preoperative prediction system for lung adenocarcinoma according to claim 2, characterized in that, The processor is further configured to: calculate visual indicators based on the geometric parameters of the segmented target region, wherein the visual indicators include at least one of the following: the ratio of the diameter of the solid region to the tumor, the ratio of the tumor volume under different image windows, and the ratio of the solid region to the total tumor volume; and the geometric parameters include the maximum diameter and volume of the target region.

4. The preoperative prediction system for lung adenocarcinoma according to claim 1, characterized in that, The processor is also used to: determine multiple regions based on image morphological features, extract radiomics features from the multiple regions and, after screening, construct radiomics scores for each of the multiple regions.

5. The preoperative prediction system for lung adenocarcinoma according to claim 1, characterized in that, The processor is also used for: Obtain the respective weights of the inflammatory markers, tumor markers, clinical features, and imaging morphological features; A clinical score is calculated based on the respective weights of the inflammatory markers, tumor markers, clinical features, and imaging morphology.

6. The preoperative prediction system for lung adenocarcinoma according to claim 1, characterized in that, The prediction model includes an embedding layer, a Transformer encoder, and a multi-task output head. The embedding layer maps the input vector to a pre-defined vector space to generate embedding vectors for the visual indicators, radiomics scores, and clinical scores in the vector space. The Transformer encoder concatenates the embedding vectors of the visual indicators, radiomics scores, and clinical scores in the vector space and performs a fusion feature representation on the concatenated vectors through a self-attention mechanism. The multi-task output head executes a first prediction task and a second prediction task in parallel based on the fusion feature representation. The first prediction task is an invasiveness prediction task, and the second prediction task is a lymph node metastasis prediction task.

7. The preoperative prediction system for lung adenocarcinoma according to claim 1 or 6, characterized in that, The preoperative prediction results for lung adenocarcinoma include the probability of invasiveness in four categories and the probability of lymph node metastasis in two categories.

8. A medical device, characterized in that, The lung adenocarcinoma preoperative prediction system includes any one of claims 1 to 6.

9. A data processing method for a preoperative prediction system for lung adenocarcinoma, characterized in that, The method is applied to the processor of the preoperative prediction system for lung adenocarcinoma according to claim 1, wherein the method includes the following steps: The system receives multiple data from lung adenocarcinoma patients, including inflammatory markers, tumor markers, CT images, and clinical data, through an interactive interface. Clinical features are extracted from the clinical data, and morphological features are extracted from the CT images. Visual indicators are calculated based on imaging morphological features; radiomics scores are calculated based on the radiomics feature mining results of imaging morphological features; and clinical scores are calculated based on inflammatory indicators, tumor markers, clinical features, and imaging morphological features. An input vector is constructed based on the visual indicators, the radiomics score, and the clinical score. The input vector is then fed into a pre-trained prediction model, which outputs the preoperative prediction result for lung adenocarcinoma.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed, they implement the data processing method of the preoperative prediction system for lung adenocarcinoma as described in claim 9.

Images