Artificial intelligence-based lesion analysis device and method using characteristics of blood vessels adjacent to lesion

Abstract

An artificial intelligence-based lesion analysis device and method using characteristics of blood vessels adjacent to a lesion is disclosed. The artificial intelligence-based lesion analysis method using characteristics of blood vessels adjacent to a lesion, according to one embodiment of the present application, may comprise the steps of: (a) acquiring a medical image of a subject; (b) inputting the medical image into a pre-constructed segmentation model so as to segment a lesion region included in the medical image and a blood vessel region in an adjacent portion of the lesion region; (c) extracting blood vessel feature information from the segmented blood vessel region; and (d) inputting the blood vessel feature information into a pre-constructed analysis model so as to generate analysis information associated with a target lesion of the subject.

Description

Artificial intelligence-based lesion analysis device and method utilizing characteristics of blood vessels surrounding the lesion

[0001] The present invention relates to an artificial intelligence-based lesion analysis device and method utilizing the characteristics of blood vessels surrounding a lesion. For example, the present invention relates to an artificial intelligence-based tumor-vascular biomarker for predicting the response to radiation therapy.

[0002] Lung cancer is one of the most common cancers worldwide, and in particular, non-small cell lung cancer (NSCLC) accounts for the majority of all lung cancer cases. The abnormal vascular structures observed in NSCLC hinder the supply of oxygen and nutrients and induce hypoxia, thereby promoting tumor growth and metastasis, which worsens the patient's prognosis and reduces the effectiveness of radiation therapy. Specifically, while normal blood vessels exhibit a hierarchical structure with uniform spacing between arteries, capillaries, and veins, tumor blood vessels display structural heterogeneity, irregular distribution, and tortuous shapes.

[0003] In this regard, while vascular density is the most commonly used parameter for quantifying vascular abnormalities, tumor vessels are often more abundant at the tumor-host interface than in the central tumor region, and vascular density tends to decrease as the tumor grows. Although previous studies have attempted to predict vascular abnormalities through the mathematical measurement of morphological features such as curvature and torsion, these methods have had limitations in their ability to capture the heterogeneity and complex distribution of tumor vessels.

[0004] Meanwhile, Stereotactic Body Radiation Therapy (SBRT) is an advanced technique that delivers precise, high-dose radiation to tumors and surrounding tissues while minimizing damage to normal tissues and blood vessels. This method is primarily used for inoperable lung cancer patients, and through such high-dose radiation therapy, it is possible to improve the tumor microenvironment by selectively destroying abnormal tumor blood vessels; this can be considered a key mechanism supporting the efficacy of SBRT. In particular, the death of tumor vascular endothelial cells significantly influences the tumor's response to radiation therapy and plays an important role in inhibiting tumor growth.

[0005] However, conventional vascular density measurements alone have limitations in effectively analyzing the relationship between the complex characteristics of tumor blood vessels and the response to radiation therapy. Simple vascular density measurements fail to adequately distinguish between normal and abnormal tumor vessels and do not sufficiently reflect the heterogeneous characteristics that negatively affect prognosis.

[0006] Furthermore, due to the complex characteristics of tumor blood vessels, there is a problem in accurately predicting the response to radiation therapy using existing methods. Accordingly, there is a need for a new method that can effectively analyze the complex characteristics of tumor blood vessels and accurately predict the response to radiation therapy.

[0007] The technology forming the background of the present invention is disclosed in Korean Published Patent Application No. 10-2024-0129697.

[0008] The present invention aims to solve the problems of the aforementioned conventional technology by providing an AI-based lesion analysis device and method utilizing the characteristics of blood vessels surrounding a tumor, which can quantitatively evaluate the difference between tumor blood vessels and normal blood vessels by utilizing AI technology to effectively analyze the complex characteristics of blood vessels surrounding a tumor.

[0009] The present invention aims to solve the problems of the aforementioned conventional technology by utilizing artificial intelligence to quantify the degree of abnormality in peritumoral blood vessels in contrast-enhanced CT images of non-small cell lung cancer patients, and thereby to provide an AI-based lesion analysis device and method utilizing the characteristics of peritumoral blood vessels that can accurately predict the response to stereotactic radiation therapy and the prognosis.

[0010] The present invention aims to solve the problems of the aforementioned conventional technology by developing a biomarker that can be utilized to select patients suitable for radiation therapy and establish personalized treatment plans through an AI-based Vessel Risk Score (VRS).

[0011] However, the technical problems that the embodiments of the present invention aim to solve are not limited to the technical problems described above, and other technical problems may exist.

[0012]

[0013] As a technical means for achieving the above-mentioned technical problem, an artificial intelligence-based lesion analysis method using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention may include: (a) acquiring a medical image of a subject; (b) inputting the medical image into a pre-established segmentation model to segment a lesion region included in the medical image and a blood vessel region surrounding the lesion region; (c) extracting blood vessel feature information from the segmented blood vessel region; and (d) inputting the blood vessel feature information into a pre-established analysis model to generate analysis information associated with a target lesion of the subject.

[0014] In addition, step (d) above can calculate a quantified risk for the target lesion by comparing the vascular characteristic information and reference information reflecting the characteristics of normal blood vessels using the analysis model.

[0015] In addition, an artificial intelligence-based lesion analysis method using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention may include (e) a step of predicting the prognosis when applying targeted treatment to the subject using the analysis information including the risk level.

[0016] In addition, step (d) above can quantify the difference between the vascular feature information and the reference information based on the Mahalanobis Distance.

[0017] In addition, the above-mentioned segmentation model can be pre-constructed as a structure combining a Convolutional Neural Network (CNN) and a U-Net model.

[0018] In addition, the above analysis model can be pre-built with a structure combining a Vision Transformer (ViT) and Long Short-Term Memory (LSTM).

[0019] Additionally, the above step (b) may include the step of setting a blood vessel area located within a preset radius from a reference position for the lesion area as a Region of Interest (ROI).

[0020] In addition, the target lesion may include non-small cell lung cancer (NSCLC).

[0021] In addition, the above-mentioned target therapy may include stereotactic body radiation therapy (SBRT).

[0022] In addition, the above prognosis may include at least one of progression-free survival (PFS) and overall survival (OS).

[0023] Meanwhile, an artificial intelligence-based lesion analysis device utilizing the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention may include an image acquisition unit for acquiring a medical image of a subject, an image segmentation unit for inputting the medical image into a pre-established segmentation model to segment a lesion region included in the medical image and a blood vessel region surrounding the lesion region, a feature extraction unit for extracting blood vessel feature information from the segmented blood vessel region, and an analysis execution unit for inputting the blood vessel feature information into a pre-established analysis model to generate analysis information linked to a target lesion of the subject.

[0024] In addition, the analysis performing unit can calculate a quantified risk for the target lesion by comparing the vascular characteristic information and reference information reflecting the characteristics of normal blood vessels using the analysis model.

[0025] In addition, the analysis performing unit can predict the prognosis when applying targeted treatment to the subject using the analysis information including the risk level.

[0026] In addition, the analysis unit can quantify the difference between the vascular characteristic information and the reference information based on the Mahalanobis Distance.

[0027] The means for solving the problem described above are merely exemplary and should not be interpreted as intended to limit the present invention. In addition to the exemplary embodiments described above, additional embodiments may exist in the drawings and the detailed description of the invention.

[0028] According to the solution to the problem of the present invention described above, an AI-based lesion analysis device and method utilizing the characteristics of blood vessels surrounding a lesion can be provided, which can quantitatively evaluate the difference between tumor blood vessels and normal blood vessels by utilizing AI technology to effectively analyze the complex characteristics of blood vessels surrounding a tumor.

[0029] According to the aforementioned means for solving the problem of the present invention, it is possible to provide an AI-based lesion analysis device and method utilizing the characteristics of blood vessels surrounding a lesion, which quantifies the degree of abnormality of blood vessels surrounding a tumor in contrast-enhanced CT images of non-small cell lung cancer patients and thereby accurately predicts the response to stereotactic radiation therapy and the prognosis.

[0030] According to the solution to the problem described above, the accuracy of predicting treatment response can be improved by utilizing artificial intelligence to analyze complex tumor-vascular relationships and the microenvironment.

[0031] According to the solution to the problem of the present invention described above, the calculated Vascular Risk Score (VRS) can be utilized to select patients suitable for SBRT and determine appropriate doses and frequencies, which can be used as a reference indicator for establishing a customized treatment plan.

[0032] According to the solution to the problem described above, there is an advantage in that it can be widely applied to other types of cancer, such as hepatocellular carcinoma (HCC), in addition to non-small cell lung cancer.

[0033] However, the effects obtainable from this invention are not limited to those described above, and other effects may exist.

[0034] FIG. 1 is a schematic diagram of an artificial intelligence-based lesion analysis system according to one embodiment of the present invention.

[0035] Figure 2 is a conceptual diagram illustrating an artificial intelligence-based analysis framework that performs learning on vascular feature information and derives analysis information on target lesions reflecting the degree of vascular risk.

[0036] Figure 3 is a conceptual diagram illustrating, in an exemplary manner, the cohort collection process for collecting training data used to build an artificial intelligence-based segmentation model and an analysis model.

[0037] FIGS. 4a and 4b are experimental examples linked with an artificial intelligence-based lesion analysis device utilizing the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention, and are graphs showing the prediction performance of the treatment response by comparing the quantified risk derived by the analysis model with the blood vessel density.

[0038] FIGS. 5a and 5b are graphs illustrating Kaplan-Meier survival curves comparing VRS and blood vessel density for PFS and OS, as experimental examples linked with an artificial intelligence-based lesion analysis device utilizing the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0039] FIG. 6 is a schematic diagram of an artificial intelligence-based lesion analysis device using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0040] FIG. 7 is a schematic diagram of a learning device for an artificial intelligence-based lesion analysis model using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0041] FIG. 8 is a flowchart of an operation for an artificial intelligence-based lesion analysis method using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0042] FIG. 9 is a flowchart of the operation of a learning method for an artificial intelligence-based lesion analysis model using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0043]

[0044] Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement them. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0045] Throughout this specification, when a part is described as being "connected" to another part, this includes not only cases where they are "directly connected," but also cases where they are "electrically connected" or "indirectly connected" with other elements interposed between them.

[0046] Throughout the entire specification, when a component is described as being located "on," "on top," "on top," "under," "on bottom," or "on bottom" of another component, this includes not only cases where the component is in contact with the other component but also cases where another component exists between the two components.

[0047] Throughout this specification, when a part is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0048] The present invention relates to an artificial intelligence-based lesion analysis device and method utilizing the characteristics of blood vessels surrounding a lesion. For example, the present invention relates to an artificial intelligence-based tumor-vascular biomarker for predicting the response to radiation therapy.

[0049] FIG. 1 is a schematic diagram of an artificial intelligence-based lesion analysis system according to one embodiment of the present invention.

[0050] Referring to FIG. 1, the artificial intelligence-based lesion analysis system (10) may include an artificial intelligence-based lesion analysis device (100) utilizing the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention (hereinafter referred to as the 'lesion analysis device (100)'), a learning device (200) for an artificial intelligence-based lesion analysis model utilizing the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention (hereinafter referred to as the 'model learning device (200)'), a medical imaging device (300), a user terminal (400), and a database (500).

[0051] A lesion analysis device (100), a model learning device (200), a medical imaging device (300), a user terminal (400), and a database (500) can communicate with each other through a network (20). The network (20) refers to a connection structure that enables information exchange between each node, such as terminals and servers. Examples of such a network (20) include, but are not limited to, a 3GPP (3rd Generation Partnership Project) network, an LTE (Long Term Evolution) network, a 5G network, a WIMAX (World Interoperability for Microwave Access) network, the Internet, a LAN (Local Area Network), a Wireless LAN (Wireless Local Area Network), a WAN (Wide Area Network), a PAN (Personal Area Network), a Wi-Fi network, a Bluetooth network, a satellite broadcasting network, an analog broadcasting network, and a DMB (Digital Multimedia Broadcasting) network.

[0052] The user terminal (400) can be any type of wireless communication device, such as a smartphone, smartpad, tablet PC, PCS (Personal Communication System), GSM (Global System for Mobile communication), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), Wibro (Wireless Broadband Internet) terminal.

[0053] Meanwhile, in the description of the embodiments of the present invention, the medical imaging device (300) may be a computerized tomography (CT) scanner, but is not limited thereto. As another example, the medical imaging device (300) may be a magnetic resonance imaging (MRI) scanner, an ultrasound imaging device, an X-ray imaging device, etc. Additionally, depending on the type of medical imaging device (200), the medical image provided to the lesion analysis device (100) may correspond to a CT image, an X-ray image, a magnetic resonance imaging (MRI) image, an ultrasound image, etc. Hereinafter, the embodiments of the present invention will be described assuming that the medical image is a contrast-enhanced CT image.

[0054] Additionally, in the description of the embodiments of the present invention, the database (500) may be a device or server for storing learning data including a plurality of medical images, normal blood vessel sample data, medical image data of a subject, and analysis information regarding target lesions derived for a subject.

[0055] Specifically, the database (500) can store contrast-enhanced computerized tomography (CT) image data obtained from multiple patients including non-small cell lung cancer (NSCLC) and corresponding clinical information of the patients.

[0056] Additionally, the database (500) can store parameter values ​​related to the segmentation model and analysis model built by the model learning device (200), and can store the mean vector and covariance matrix of normal blood vessel features as reference information reflecting the characteristics of normal blood vessels.

[0057] Additionally, the database (500) can store analysis information including the Vessel Risk Score (VRS) calculated by the lesion analysis device (100) and the prognosis prediction information associated therewith, in conjunction with patient data.

[0058] Figure 2 is a conceptual diagram illustrating an artificial intelligence-based analysis framework that performs learning on vascular feature information and derives analysis information on target lesions reflecting the degree of vascular risk.

[0059] Referring to FIG. 2, the artificial intelligence-based analysis framework disclosed herein can be broadly divided into two stages. Specifically, the first row of FIG. 2 illustrates the first stage (Step 1) of learning vascular characteristics, and the second row illustrates the second stage (Step 2) of calculating vascular risk.

[0060] In the first step, the tumor region and the vascular region are first segmented from the contrast-enhanced CT image using a tumor-vascular segmentation model. As described in detail below, this segmentation model combines the ConvNeXt-Small and U-Net structures to identify the tumor and surrounding blood vessels with high accuracy. The segmented vascular images are input into a vascular feature extractor to extract vascular feature information. This vascular feature extractor can be composed of a model combining a Vision Transformer (ViT) and Long Short-Term Memory (LSTM), and can generate a 768-dimensional tumor-vascular feature vector from each CT slice.

[0061] In the second step, the Vessel Risk Score (VRS) is calculated by comparing Normal Vessel Features and Tumor Vessel Features extracted from a normal vessel cohort. Specifically, the VRS can be quantified by calculating the mean vector and covariance matrix from the normal vessel features and measuring the Mahalanobis Distance with the tumor vessel features. The calculated VRS is normalized to a value between 0 and 1, and a value closer to 1 indicates a higher degree of abnormality in the tumor vessels.

[0062] Below, we will first explain in detail the process by which the model learning device (200) constructs the segmentation model and the analysis model.

[0063] The model learning device (200) can collect learning data including multiple medical images and normal blood vessel samples associated with the target lesion.

[0064] In this regard, Figure 3 is a conceptual diagram illustrating, exemplarily, the cohort collection process for collecting training data used to build an artificial intelligence-based segmentation model and an analysis model.

[0065] Referring to FIG. 3, the model learning device (200) can perform a cohort collection process to build a large dataset containing medical images of lung cancer patients collected from major medical institutions (e.g., large hospitals, etc.). Next, the model learning device (200) can exclude patients who do not have non-small cell lung cancer (NSCLC) and patients who have not received radiation therapy from the built dataset, and select data of NSCLC patients who have received radiation therapy.

[0066] Additionally, the model learning device (200) can classify selected patients into three main cohorts. First, the learning cohort consists of patients who received hypofractionated radiotherapy (number of fractions ≤ 10) and can be used for the initial learning of the fractionation model and the analysis model. Next, the external validation cohort consists of early-stage NSCLC patients who received stereotactic body radiotherapy (SBRT) and can be used to independently evaluate the performance of the developed model. Finally, the normal blood vessel cohort consists of normal blood vessel images extracted from the opposite lung (the side without tumor) of early-stage NSCLC patients and can be used to obtain the distribution of normal blood vessel features. This classification of the three cohorts allows the learning and validation of the model to be performed in different patient groups, thereby improving the generalization ability of the developed analysis model and enabling objective evaluation of its performance.

[0067] Additionally, the model learning device (200) can generate reference information reflecting the characteristics of normal blood vessels from a pre-prepared normal blood vessel sample.

[0068] More specifically, reference information reflecting the characteristics of normal blood vessels may include the statistical characteristics of blood vessel feature vectors extracted from a normal blood vessel cohort. For example, a model learning device (200) can extract a 768-dimensional feature vector (Normal Vessel Feature) from a normal blood vessel sample using a Vision Transformer (ViT) of the analysis model and a Long Short-Term Memory (LSTM)-based blood vessel feature extractor.

[0069] Additionally, the model learning device (200) can define a normal blood vessel distribution by calculating a mean vector and a covariance matrix from a plurality of normal blood vessel feature vectors extracted in this way. These mean vector and covariance matrix represent the characteristics of normal blood vessels statistically in a multidimensional space and can be used as reference information to quantify the degree of abnormality of tumor blood vessels. This method allows for a comprehensive evaluation of the degree of deviation from the normal blood vessel distribution rather than simply measuring individual blood vessel characteristics, thereby providing a clinically more meaningful indicator.

[0070] Additionally, the model learning device (200) can use the collected learning data to train a segmentation model that operates to segment a lesion area included in a medical image and a blood vessel area surrounding the lesion area when a medical image is input.

[0071] For example, the model learning device (200) can construct a segmented model having a structure that combines a Convolutional Neural Network (CNN) and a U-Net model.

[0072] In this regard, the segmentation model can be constructed to segment tumor and vascular regions with high accuracy by using the ConvNeXt-Small architecture as the encoder and combining it with a U-Net structured decoder. More specifically, ConvNeXt-Small provides enhanced feature extraction capabilities compared to existing CNN models and is a lightweight model capable of effective training with a small number of parameters.

[0073] This model can extract features at various scales through a hierarchical structure, making it effective for accurately recognizing tumors and vascular structures with complex shapes. The U-Net model generates segmentation masks with the same resolution as the original image by progressively expanding features extracted from the encoder. Through U-Net's characteristic Skip Connection, low-level features from the encoder are directly transferred to the decoder, enabling precise segmentation with clear boundaries. This combination of ConvNeXt-Small and U-Net can effectively segment the complex structures of tumor regions and surrounding blood vessels, and offers the particular advantage of accurately identifying even small vessels that appear faint in contrast-enhanced CT images.

[0074] Additionally, the model learning device (200) can input each of the multiple medical images included in the learning data into a segmentation model and extract blood vessel feature information from the segmented blood vessel regions using the segmentation model.

[0075] Specifically, in the description of the embodiments of the present invention, vascular feature information may refer to a multidimensional feature vector extracted from a segmented tumor-peripheral vascular region. This feature vector is generated through an analysis model based on Vision Transformer (ViT) and Long Short-Term Memory (LSTM), and represents the morphological and structural characteristics of blood vessels in a high-dimensional space of 768 dimensions. Vascular feature information may include not only various morphological characteristics such as blood vessel thickness, density, branching patterns, curvature, and tortuousness, but also complex structural characteristics such as connectivity and distribution patterns of the vascular network.

[0076] In particular, features extracted from each CT slice are integrated into sequence information through an LSTM, allowing for the capture of interactions between vascular structures in three-dimensional space. This extracted vascular feature information can be used as an important indicator to quantify abnormalities in tumor blood vessels and predict the response to radiation therapy.

[0077] Additionally, the model learning device (200) can use vascular feature information, reference information, and clinical result information of a subject (subject) related to the vascular feature information to train an analysis model that operates to generate analysis information associated with a target lesion by comparing the vascular feature information with the reference information when vascular feature information is input.

[0078] In the description of the embodiments of the present invention, clinical outcome information may refer to various medical data related to the course of a patient's treatment. Specifically, clinical outcome information may include the results of tumor response evaluation after radiation therapy, which may be classified into Complete Response (CR), Partial Response (PR), Stable Disease (SD), and Progressive Disease (PD) according to the Response Evaluation Criteria in Solid Tumors (RECIST).

[0079] In addition, clinical outcome information may include survival data such as progression-free survival (PFS) and overall survival (OS), and may include various clinical variables such as the patient's comorbidities, general status (Eastern Cooperative Oncology Group performance status, ECOG PS), smoking status, histological classification of the tumor (squamous cell carcinoma, adenocarcinoma, etc.), radiation therapy dose, and fractionation information.

[0080] This clinical outcome information can be utilized as a target variable during analysis model training, playing a key role in constructing a model that predicts treatment response and prognosis from vascular characteristic information.

[0081] In particular, by learning the relationship between vascular characteristic information and survival data through survival analysis based on the Cox Proportional Hazard Model, an analysis model capable of accurately predicting the prognosis of new patients can be developed.

[0082] In addition, logistic regression analysis can be used to model the relationship between vascular characteristic information and treatment response, which can then be utilized to predict the response to radiation therapy. As such, training an analysis model using clinical outcome information can identify the complex relationship between tumor vascular characteristics and treatment outcomes, thereby contributing to the determination of personalized treatment strategies.

[0083] For example, the model learning device (200) can build an analysis model having a structure that combines a Vision Transformer (ViT) and a Long Short-Term Memory (LSTM).

[0084] In this regard, the Vision Transformer is a model that applies a transformer structure to computer vision tasks, dividing images into patches and learning the relationships between each patch through a self-attention mechanism. This structure is effective in capturing relationships between distant vascular structures in vascular images and enables feature extraction that considers global context without being biased toward local features. ViT generates a 768-dimensional feature vector for each CT slice, allowing it to represent the morphological and structural characteristics of blood vessels in a high-dimensional space.

[0085] Long Short-Term Memory (LSTM) is a type of Recurrent Neural Network (RNN) that is a structure capable of effectively learning long-term dependencies from sequence data. LSTM receives feature vectors extracted by ViT from each CT slice as a sequence input, enabling it to capture not only the relationships between adjacent slices but also the relationships between distant slices. Through this, it is possible to model the interactions and continuity between vascular structures in three-dimensional space.

[0086] This combination of ViT and LSTM can simultaneously consider spatial features in 2D CT slices and time-series features in 3D volumes, so the analysis model can be effective for comprehensively analyzing the characteristics of complex tumor vascular structures.

[0087] Meanwhile, in the description of the embodiments of the present invention, the target lesion may include non-small cell lung cancer (NSCLC). In this regard, NSCLC is characterized by abnormal vascular structures within and around the tumor, and these abnormal vessels impede oxygen supply, causing hypoxia and reducing the effectiveness of radiation therapy. Taking this into consideration, the analysis model disclosed herein learns the characteristic vascular structures of NSCLC to quantitatively evaluate tumor vascular abnormalities in individual patients and predict the response to radiation therapy.

[0088] However, in the description of the embodiments of the present invention, the target lesion may be broadly applied to various types of cancer rich in blood vessels, in addition to non-small cell lung cancer, depending on the embodiment of the present invention. For example, hepatocellular carcinoma (HCC) is a cancer that exhibits high angiogenesis characteristics, and the tumor vascular structure directly influences the patient's prognosis and treatment response. Furthermore, in various solid tumors such as renal cell carcinoma (RCC), brain tumors, pancreatic cancer, and breast cancer, the characteristics of tumor blood vessels and treatment response can be analyzed and the lesion analysis technique utilizing AI-based vascular features disclosed herein can be applied.

[0089] Below, a process for deriving analysis information for a target lesion using a segmented model and an analysis model constructed by a lesion analysis device (100) will be described in detail.

[0090] First, the lesion analysis device (100) can acquire a medical image of the subject. For example, the lesion analysis device (100) can receive a medical image of the subject taken from a medical image capturing device (300).

[0091] In addition, the lesion analysis device (100) can input a medical image into a pre-established segmentation model to segment the lesion area included in the medical image and the vascular area surrounding the lesion area.

[0092] Specifically, the lesion analysis device (100) can set a blood vessel area located within a preset radius from a reference position for the lesion area as a Region of Interest (ROI).

[0093] In this regard, the lesion analysis device (100) can set the center point of the divided tumor area as a reference position and designate a blood vessel area located within a predetermined radius (e.g., 20 mm) from this center point as a region of interest (ROI).

[0094] Meanwhile, such setting of the region of interest may be applied to comprehensively analyze the tumor and its surrounding microenvironment. According to one embodiment of the present invention, the radius of the region of interest may be optimized through preliminary experiments, and the radius may be adjusted according to the tumor size. For example, methods such as setting it to 1.5 times the longest diameter of the tumor or setting it in proportion to the square root of the tumor volume may be applied.

[0095] In addition, according to one embodiment of the present invention, the lesion analysis device (100) may operate to extract blood vessel features by assigning weights based on distance from the center, taking into account that the density distribution of blood vessels within the region of interest may be non-uniform, thereby reflecting biological characteristics in which blood vessels closer to the center of the tumor have a direct influence on the tumor microenvironment.

[0096] In addition, the lesion analysis device (100) can extract vascular feature information from the divided vascular region.

[0097] In addition, the lesion analysis device (100) can generate analysis information linked to the target lesion of the subject by inputting vascular feature information into a pre-established analysis model.

[0098] Specifically, the lesion analysis device (100) can calculate a quantified risk of a target lesion by using an analysis model to compare vascular feature information extracted from a medical image of a subject with reference information reflecting the characteristics of a normal blood vessel.

[0099] Here, the quantified risk can be defined as the Vessel Risk Score (VRS), which is an indicator that quantifies the degree of abnormality in blood vessels surrounding a tumor into a value between 0 and 1. The VRS is a value obtained by measuring how far tumor vessels deviate from the normal blood vessel distribution using the Mahalanobis Distance, and then normalizing this value to a range of 0–1 through min-max scaling. A VRS value closer to 0 indicates characteristics similar to normal blood vessels, while a value closer to 1 indicates an abnormal vascular structure.

[0100] Unlike simple vascular density measurements, VRS evaluates abnormalities by comprehensively considering the morphological and structural characteristics of blood vessels; therefore, it can function as a more effective biomarker for predicting radiation therapy response and prognosis. In particular, high VRS values ​​reflect abnormal vascular structures associated with hypoxia, and thus can be utilized as an indicator to predict resistance to radiation therapy. This implies that VRS is an indicator that evaluates the qualitative characteristics, rather than merely the quantitative characteristics, of blood vessels.

[0101] According to one embodiment of the present invention, a lesion analysis device (100) can quantify the difference between vascular characteristic information and reference information based on the Mahalanobis Distance.

[0102] More specifically, the lesion analysis device (100) can quantify the difference between vascular feature information (x) and reference information (mean vector μ and covariance matrix Σ) using the Mahalanobis distance. Specifically, the Mahalanobis distance can be calculated through the following Equation 1.

[0103] [Equation 1]

[0104]

[0105] Here, x is the feature vector of the tumor blood vessel being evaluated, μ is the mean vector of normal blood vessel features, and Σ is the covariance matrix of normal blood vessel features. Mahalanobis distance is an effective statistical method for detecting outliers in multivariate data by considering the correlation between each feature, and can accurately measure the degree of deviation from normal blood vessel characteristics in a 768-dimensional feature space.

[0106] As another example, the lesion analysis device (100) can quantify the difference between the vascular feature information and the reference information by measuring the angle-based similarity between two feature vectors using cosine similarity according to an embodiment of the present invention, or by calculating the straight-line distance using Euclidean distance.

[0107] As another example, the lesion analysis device (100) can also quantify the difference between vascular feature information and reference information by using a method of measuring distance between distributions such as Maximum Mean Discrepancy (MMD), a method of utilizing an autoencoder to use the reconstruction error as an outlier score, or a method of using an outlier detection algorithm such as a One-Class Support Vector Machine (One-Class SVM) or an Isolation Forest. The various methods described above each have different characteristics, and the lesion analysis device (100) can operate to select and apply an appropriate method according to the characteristics of the data to be analyzed.

[0108] In addition, the lesion analysis device (100) can predict the prognosis when applying targeted treatment to a subject using analysis information including risk levels.

[0109] For example, if the target lesion is non-small cell lung cancer (NSCLC), the lesion analysis device (100) can predict at least one of progression-free survival (PFS) and overall survival (OS) as a prognosis of targeted therapy including stereotactic body radiation therapy (SBRT).

[0110] In this regard, the specific process by which the analysis model predicts the prognosis can proceed as follows. First, the lesion analysis device (100) can input clinical information of the subject (patient's age, gender, general condition, smoking status, histological classification of the tumor, radiation therapy dose, etc.) along with the vascular risk score (VRS) derived from the subject's medical images into the analysis model. Based on patterns acquired during the learning process, the analysis model can estimate the probability of survival by applying statistical methods, such as the Cox Proportional Hazard Model or Random Forest survival analysis, to the data of a new patient.

[0111] Specifically, the lesion analysis device (100) can calculate the probability of disease-free survival or overall survival at a specific point in time by referring to the pattern of existing patient data having similar characteristics centered on the VRS value.

[0112] In addition, the lesion analysis device (100) can classify patients into high-risk and low-risk groups by comparing the VRS value with a predetermined threshold value (e.g., median value 0.5119), and can generate and visually represent a predicted survival curve for each risk group. This prognostic prediction information can be used as an important indicator for clinicians to establish a personalized treatment plan for each patient.

[0113] For example, for patients classified as high-risk, more aggressive treatment approaches may be considered, such as increasing the radiation therapy dose or combining adjuvant therapies like immunotherapy. Additionally, the predicted prognostic information can be used to determine the follow-up period after treatment, and in the case of high-risk patients, disease progression can be detected early through more frequent follow-up examinations. In this way, the lesion analysis device (100) supports clinical decision-making through prognostic prediction based on VRS, which is a quantified risk level, and contributes to the realization of personalized medicine.

[0114] Below, an experimental example linked with a lesion analysis device (100) will be described with reference to FIGS. 4a, 4b, 5a, and 5b.

[0115] FIGS. 4a and 4b are experimental examples linked with an artificial intelligence-based lesion analysis device utilizing the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention, and are graphs showing the prediction performance of the treatment response by comparing the quantified risk derived by the analysis model with the blood vessel density.

[0116] Referring to Figures 4a and 4b, Figures 4a and 4b show the results of a comparative analysis of VRS and vascular density according to treatment response. Specifically, Figure 4a is a box plot showing the difference in VRS and vascular density between the responder and non-responder groups for SBRT treatment. The mean VRS of the responder group (complete response [CR], partial response [PR], stable disease [SD]) was 0.494 (95% confidence interval: 0.47-0.52), which was statistically significantly lower than the mean VRS of the non-responder group (progressive disease [PD]) at 0.578 (95% confidence interval: 0.54-0.62) (p < 0.0001). On the other hand, vascular density did not show a significant difference between the responder group (mean: 0.404, 95% confidence interval: 0.36-0.45) and the non-responder group (mean: 0.428, 95% confidence interval: 0.36-0.50) (p=0.3907).

[0117] Figure 4b shows the Receiver Operating Characteristic (ROC) curves of VRS and vascular density for predicting non-response to SBRT. The Area Under ROC Curve (AUROC) of VRS was 0.69 (95% confidence interval: 0.61-0.77, p=0.0001), indicating statistically significant predictive power, whereas vascular density did not show a significant AUROC value. These results demonstrate that VRS is a more effective biomarker for predicting SBRT treatment response than simple vascular density measurement.

[0118] FIGS. 5a and 5b are graphs illustrating Kaplan-Meier survival curves comparing VRS and blood vessel density for PFS and OS, as experimental examples linked with an artificial intelligence-based lesion analysis device utilizing the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0119] Referring to Figures 5a and 5b, which show Kaplan-Meier survival curves for progression-free survival (PFS) and overall survival (OS) according to VRS and vascular density, Figure 5a specifically shows a comparison of PFS and OS between the high-risk and low-risk VRS groups based on the median (0.5119). The high-risk group showed a significantly shorter PFS compared to the low-risk group (median: 20 months vs. >60 months, p = 0.007), suggesting that high VRS is associated with an increased risk of disease progression. Regarding OS, neither group reached the median survival time (>60 months), but the high-risk group tended to have a lower survival rate than the low-risk group (p = 0.077).

[0120] Figure 5b shows a comparison of PFS and OS between high-risk and low-risk groups based on the median vascular density. No significant differences were observed in PFS (p=0.574) and OS (p=0.913) between the groups according to vascular density. These results demonstrate that VRS is a more effective indicator than vascular density for predicting the prognosis of NSCLC patients after SBRT. In particular, the low-risk VRS group (low VRS) showed excellent long-term prognosis after SBRT, suggesting that VRS can be utilized to select patients who can maximize the benefits of SBRT.

[0121] FIG. 6 is a schematic diagram of an artificial intelligence-based lesion analysis device using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0122] Referring to FIG. 6, the lesion analysis device (100) may include an image acquisition unit (110), an image segmentation unit (120), a feature extraction unit (130), and an analysis execution unit (140).

[0123] The image acquisition unit (110) can acquire a medical image of the subject.

[0124] The image segmentation unit (120) can input a medical image into a pre-established segmentation model to segment the lesion area included in the medical image and the vascular area surrounding the lesion area.

[0125] Specifically, the image segmentation unit (120) can set a blood vessel area located within a preset radius from a reference position for the lesion area as a Region of Interest (ROI).

[0126] The feature extraction unit (130) can extract blood vessel feature information from the divided blood vessel region.

[0127] The analysis execution unit (140) can generate analysis information linked to the target lesion of the subject by inputting vascular feature information into a pre-established analysis model.

[0128] Specifically, the analysis execution unit (140) can calculate a quantified risk level for a target lesion by using an analysis model to compare vascular feature information extracted from a medical image of a subject with reference information reflecting the characteristics of a normal blood vessel.

[0129] According to one embodiment of the present invention, the analysis performing unit (140) can quantify the difference between the vascular feature information and the reference information based on the Mahalanobis Distance.

[0130] In addition, the analysis unit (140) can predict the prognosis when applying targeted treatment to a subject using analysis information including risk levels.

[0131] For example, if the target lesion is non-small cell lung cancer (NSCLC), the analysis performing unit (140) can predict at least one of progression-free survival (PFS) and overall survival (OS) as a prognosis of targeted therapy including stereotactic body radiation therapy (SBRT).

[0132] FIG. 7 is a schematic diagram of a learning device for an artificial intelligence-based lesion analysis model using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0133] Referring to FIG. 7, the model learning device (200) may include a data collection unit (210), a segmentation model building unit (220), a feature collection unit (230), and an analysis model building unit (240).

[0134] The data collection unit (210) can collect training data including multiple medical images and normal blood vessel samples associated with the target lesion.

[0135] In addition, the data collection unit (210) can generate reference information reflecting the characteristics of normal blood vessels from a pre-prepared normal blood vessel sample.

[0136] The segmentation model building unit (220) can use collected training data to train a segmentation model that operates to segment a lesion area included in a medical image and a blood vessel area surrounding the lesion area when a medical image is input.

[0137] The feature collection unit (230) inputs each of the multiple medical images included in the training data into a segmentation model and can extract blood vessel feature information from the segmented blood vessel regions using the segmentation model.

[0138] The analysis model building unit (240) can train an analysis model that operates to generate analysis information linked to a target lesion by comparing the vascular characteristic information with the reference information when the vascular characteristic information is input, using vascular characteristic information, reference information, and clinical result information of a subject (subject) related to the vascular characteristic information.

[0139] Below, based on the details described above, we will briefly examine the operation flow of the present invention.

[0140] FIG. 8 is a flowchart of an operation for an artificial intelligence-based lesion analysis method using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0141] The artificial intelligence-based lesion analysis method using the characteristics of blood vessels surrounding the lesion illustrated in FIG. 8 can be performed by the lesion analysis device (100) described above. Therefore, even if the content is omitted below, the description of the lesion analysis device (100) can be equally applied to the description of the artificial intelligence-based lesion analysis method using the characteristics of blood vessels surrounding the lesion.

[0142] Referring to FIG. 8, in step S11, the image acquisition unit (110) can acquire (a) a medical image of the subject.

[0143] Next, in step S12, the image segmentation unit (120) can (b) input the medical image into a pre-established segmentation model to segment the lesion area included in the medical image and the vascular area surrounding the lesion area.

[0144] Specifically, in step S12, the image segmentation unit (120) can set a blood vessel area located within a preset radius from a reference position for the lesion area as a Region of Interest (ROI).

[0145] Next, in step S13, the feature extraction unit (130) can extract blood vessel feature information from (c) the divided blood vessel region.

[0146] Next, in step S14, the analysis execution unit (140) can (d) input vascular feature information into a pre-established analysis model to generate analysis information associated with the subject's target lesion.

[0147] Specifically, in step S14, the analysis performing unit (140) can calculate a quantified risk level for a target lesion by using an analysis model to compare vascular feature information extracted from a medical image of a subject with reference information reflecting the characteristics of a normal blood vessel.

[0148] According to one embodiment of the present invention, in step S14, the analysis performing unit (140) can quantify the difference between the vascular feature information and the reference information based on the Mahalanobis Distance.

[0149] Next, in step S15, the analysis performing unit (140) can predict the prognosis when applying targeted treatment to the subject using analysis information including (e) risk level.

[0150] For example, in step S15, the analysis performing unit (140) can predict at least one of progression-free survival (PFS) and overall survival (OS) as a prognosis of targeted therapy including stereotactic body radiation therapy (SBRT) if the target lesion is non-small cell lung cancer (NSCLC).

[0151] In the description above, steps S11 through S15 may be further divided into additional steps or combined into fewer steps according to an embodiment of the present invention. Additionally, some steps may be omitted as necessary, and the order between steps may be changed.

[0152] FIG. 9 is a flowchart of the operation of a learning method for an artificial intelligence-based lesion analysis model using the characteristics of blood vessels surrounding a lesion according to one embodiment of the present invention.

[0153] The method of learning an artificial intelligence-based lesion analysis model using the characteristics of blood vessels surrounding the lesion illustrated in FIG. 9 can be performed by the model learning device (200) described above. Therefore, even if the content is omitted below, the description of the model learning device (200) can be equally applied to the description of the method of learning an artificial intelligence-based lesion analysis model using the characteristics of blood vessels surrounding the lesion.

[0154] Referring to FIG. 9, in step S21, the data collection unit (210) can collect training data including a plurality of medical images associated with the target lesion and normal blood vessel samples.

[0155] Next, in step S22, the data collection unit (210) can generate reference information reflecting the characteristics of a normal blood vessel from a pre-prepared normal blood vessel sample.

[0156] Next, in step S23, the segmentation model building unit (220) can use the collected training data to train a segmentation model that operates to segment the lesion area included in the medical image and the vascular area surrounding the lesion area when a medical image is input.

[0157] Next, in step S24, the feature collection unit (230) inputs each of the multiple medical images included in the training data into a segmentation model and can extract blood vessel feature information from the segmented blood vessel regions using the segmentation model.

[0158] Next, in step S25, the analysis model building unit (240) can train an analysis model that operates to generate analysis information associated with a target lesion by comparing the vascular characteristic information with the reference information when the vascular characteristic information is input, using vascular characteristic information, reference information, and clinical result information of a subject (subject) related to the vascular characteristic information.

[0159] In the description above, steps S21 through S25 may be further divided into additional steps or combined into fewer steps according to an embodiment of the present invention. Additionally, some steps may be omitted as necessary, and the order of the steps may be changed.

[0160] An artificial intelligence-based lesion analysis method utilizing the characteristics of blood vessels surrounding a lesion and a learning method for an artificial intelligence-based lesion analysis model utilizing the characteristics of blood vessels surrounding a lesion, according to one embodiment of the present invention, may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either individually or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the present invention, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The above-described hardware device may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

[0161] In addition, the artificial intelligence-based lesion analysis method using the characteristics of blood vessels surrounding the lesion and the artificial intelligence-based lesion analysis model learning method using the characteristics of blood vessels surrounding the lesion described above can also be implemented in the form of a computer program or application executed by a computer stored on a recording medium.

[0162] The foregoing description of the present invention is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical concept or essential features of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0163] The scope of the present invention is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and the concept of equivalents thereof should be interpreted as being included within the scope of the present invention.

[0164] [Explanation of the symbol]

[0165] 10: AI-based lesion analysis system

[0166] 100: AI-based lesion analysis device utilizing characteristics of blood vessels surrounding the lesion

[0167] 110: Image acquisition unit

[0168] 120: Image segmentation section

[0169] 130: Feature extraction unit

[0170] 140: Analysis Execution Unit

[0171] 200: Learning device for an AI-based lesion analysis model using characteristics of blood vessels surrounding the lesion

[0172] 210: Data Collection Unit

[0173] 220: Partition Model Construction Section

[0174] 230: Feature Collection Unit

[0175] 240: Analysis Model Construction Department

[0176] 300: Medical imaging device

[0177] 400: User terminal

[0178] 500: Database

[0179] 20: Network

Claims

1. In an artificial intelligence-based lesion analysis method utilizing the characteristics of blood vessels surrounding the lesion, (a) A step of acquiring a medical image of the subject; (b) inputting the medical image into a pre-established segmentation model to segment the lesion region included in the medical image and the vascular region surrounding the lesion region; (c) a step of extracting vascular feature information from the divided vascular region; and (d) a step of generating analysis information linked to the target lesion of the subject by inputting the vascular characteristic information into a pre-established analysis model, A method for analyzing lesions, including 2. In Paragraph 1, The above step (d) is, A lesion analysis method that calculates a quantified risk for the target lesion by comparing the vascular characteristic information and reference information reflecting the characteristics of normal blood vessels using the above analysis model.

3. In Paragraph 2, (e) A step of predicting the prognosis when applying targeted treatment to the subject using the analysis information including the above risk level, A lesion analysis method that further includes 4. In Paragraph 2, The above step (d) is, A lesion analysis method characterized by quantifying the difference between the above-mentioned vascular characteristic information and the above-mentioned reference information based on the Mahalanobis Distance.

5. In Paragraph 1, A lesion analysis method in which the above-mentioned segmentation model is pre-constructed as a structure combining a Convolutional Neural Network (CNN) and a U-Net model.

6. In Paragraph 1, A lesion analysis method in which the above analysis model is pre-constructed as a structure combining a Vision Transformer (ViT) and Long Short-Term Memory (LSTM).

7. In Paragraph 1, The above step (b) is, A step of setting a blood vessel area located within a preset radius from a reference location for the above-mentioned lesion area as a Region of Interest (ROI), A lesion analysis method that includes 8. In Paragraph 1, A method for analyzing lesions, wherein the target lesion includes non-small cell lung cancer (NSCLC).

9. In Paragraph 3, The above-mentioned targeted therapy includes Stereotactic Body Radiation Therapy (SBRT), and A method for analyzing a lesion, characterized in that the above prognosis includes at least one of progression-free survival (PFS) and overall survival (OS).

10. An artificial intelligence-based lesion analysis device utilizing the characteristics of blood vessels surrounding a lesion, An image acquisition unit for acquiring a medical image of a subject; An image segmentation unit that inputs the medical image into a pre-established segmentation model to segment a lesion region included in the medical image and a vascular region surrounding the lesion region; A feature extraction unit for extracting blood vessel feature information from the above-mentioned divided blood vessel region; and An analysis execution unit that inputs the above-mentioned vascular characteristic information into a pre-established analysis model to generate analysis information linked to the target lesion of the subject, A lesion analysis device including 11. In Paragraph 10, The above analysis performing unit, A lesion analysis device that calculates a quantified risk for the target lesion by comparing the vascular characteristic information and reference information reflecting the characteristics of normal blood vessels using the above analysis model.

12. In Paragraph 11, The above analysis performing unit, Predicting the prognosis when applying targeted treatment to the subject using the analysis information including the above risk level, The above-mentioned targeted therapy includes Stereotactic Body Radiation Therapy (SBRT), and A lesion analysis device characterized by the above prognosis including at least one of progression-free survival (PFS) and overall survival (OS).

13. In Paragraph 11, The above analysis performing unit, A lesion analysis device characterized by quantifying the difference between the above-mentioned vascular characteristic information and the above-mentioned reference information based on the Mahalanobis Distance.

14. In Paragraph 10, The above partitioning model is pre-constructed with a structure that combines a Convolutional Neural Network (CNN) and a U-Net model, and A lesion analysis device in which the above analysis model is pre-built with a structure combining a Vision Transformer (ViT) and Long Short-Term Memory (LSTM).

15. In Paragraph 10, A lesion analysis device in which the above-mentioned target lesion includes non-small cell lung cancer (NSCLC).

Images