Apparatus and method of brain lesion tracking based on habitat
The apparatus and method use deep learning to analyze multiparameter MRI data for voxel-level classification of brain lesions, enhancing the accuracy of distinguishing between radiation necrosis and tumor recurrence by integrating structural and physiological sequences.
Patent Information
- Application Number
- KR1020250190224
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2045-12-04
AI Technical Summary
Conventional methods struggle to accurately distinguish between radiation necrosis and tumor recurrence in brain lesions using MRI, as they rely on single-parameter image analysis and fail to quantify heterogeneous internal characteristics, limiting the prediction of treatment prognosis.
An apparatus and method utilizing deep learning technology for voxel-level analysis of multiparameter MRI data, integrating structural and physiological sequences, and classifying brain lesions into habitat types to provide quantitative prognosis.
Enhances the accuracy and objectivity of distinguishing between radiation necrosis and tumor recurrence, improving universality and reproducibility by leveraging metadata-based preprocessing and nnU-Net-based 3D segmentation.
Smart Images

Figure 112025136860005-PAT00004_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to medical image analysis technology, and more specifically, to an apparatus and method for detecting brain lesions and predicting treatment prognosis using magnetic resonance imaging (MRI). Background Technology
[0002] Stereotactic Radiosurgery (SRS) is widely utilized as an important treatment option in brain tumor therapy and is recognized as an effective treatment, particularly for patients with brain metastases and glioblastoma. However, accurately distinguishing between radiation necrosis and tumor recurrence during follow-up after radiation therapy is a critical task in determining the direction of treatment.
[0003] Conventional methods for tracking brain lesions have primarily relied on MRI imaging to observe changes in lesion size or contrast enhancement patterns. While this structural imaging-based approach is useful for identifying morphological changes, distinguishing between radiation necrosis and recurrence remains a difficult issue. This is because radiation necrosis and recurrence share common imaging findings, such as contrast enhancement, edema, and mass effects, making it difficult to clearly differentiate them; furthermore, radiation-treated tissues exhibit a complex interplay of viable tumors, capillary damage, ischemia, and toxicity to anticancer drugs.
[0004] Meanwhile, although medical image analysis utilizing deep learning technology has recently garnered attention, there are several difficulties when applying it to the analysis of brain lesions. Existing deep learning-based approaches are primarily focused on the segmentation of lesions, which limits their ability to predict prognosis by quantitatively analyzing the microenvironment (habitat) within the lesion at the voxel level. Furthermore, there is a lack of systematic methodologies to integrate and utilize multi-parameter information acquired from various MRI sequences.
[0005] Furthermore, in the case of longitudinal analyses tracking changes in lesions over time, existing methods primarily focus on changes in lesion size or contrast enhancement, making it difficult to quantitatively analyze dynamic changes in the lesion's internal microenvironment. In particular, analyzing the entire lesion as a single Region of Interest (ROI) risks overlooking heterogeneous internal characteristics. Consequently, there are limitations in predicting early recurrence or accurately evaluating treatment response. Due to these limitations, there is a need for a new technical approach capable of accurately distinguishing between radiation necrosis and tumor recurrence, integratively utilizing multiparameter MRI data, and quantitatively analyzing heterogeneous internal characteristics of the lesion. Prior art literature
[0006] Korean Registered Patent 10-2789232 (April 1, 2025) The problem to be solved
[0007] The embodiments disclosed in this disclosure disclose an apparatus and method for classifying Habitat types through voxel-level analysis and for detecting brain lesions and predicting treatment prognosis based on quantitative information for each Habitat type.
[0008] Another embodiment disclosed in the present disclosure discloses an apparatus and method for performing preprocessing on medical images by optimizing preprocessing tasks to maximize quantitative analysis accuracy.
[0009] The problems that this disclosure aims to solve are not limited to those mentioned above, and other unmentioned problems will be clearly understood by a person skilled in the art from the description below. means of solving the problem
[0010] An apparatus according to the present disclosure for achieving the technical problem described above comprises: a memory storing at least one process for tracking a brain lesion from a medical image; and a processor operating according to the at least one process. The processor acquires sequences of the medical image having multiple parameters, inputs the acquired sequences into a deep learning algorithm to segment a lesion region, classifies voxels within the segmented lesion region into a plurality of habitat types based on physiological variable values among the multiple parameters, and distinguishes the treatment prognosis of the brain lesion as radiation necrosis or tumor recurrence based on quantitative information of the classified plurality of habitat types.
[0011] Meanwhile, the processor may input the acquired sequences into the deep learning algorithm by executing at least one preprocessing task among Bias Field Correction (N4), Brain Extraction, Nonlinear Registration, Co-registration & Resample, Denoise, NAWM Intensity Normalization based on normal white matter regions, Motion Correction, Image Arithmetic, and Morphological Operations.
[0012] In addition, the processor may dynamically configure the execution order or parameters of the at least one preprocessing task based on the metadata of the acquired sequences, or optimize the at least one preprocessing task by combining known preprocessing tasks.
[0013] Additionally, the sequences of medical images having the above multiple parameters include a structural MRI sequence and a physiological MRI sequence, wherein the structural MRI sequence includes at least one of a T1 contrast-enhanced image (T1CE) and a T2-weighted image (T2WI), and the physiological MRI sequence may include at least one of an apparent diffusion coefficient (ADC) map and a cerebral blood volume (CBV) map.
[0014] In addition, the deep learning algorithm may include a 3D segmentation network based on nnU-Net.
[0015] In addition, the above physiological variable values include apparent diffusion coefficient (ADC) values and normalized cerebral blood flow (nCBV) values, and the processor can classify the voxels using a K-means clustering algorithm.
[0016] In addition, the above plurality of habitat types may include hypervascular cellular habitat, hypovascular cellular habitat, and nonviable tissue habitat.
[0017] Additionally, the processor further classifies the voxels within the segmented lesion region into structural habitat types based on the structural variable value among the multiple parameters, wherein the structural variable value includes the signal intensity of the T1 contrast-enhanced image and the signal intensity of the T2-weighted image, and the structural habitat types may include an enhancing tissue habitat, a solid low-enhancing habitat, and a nonviable tissue habitat.
[0018] In addition, the processor can align medical image sequences acquired at multiple time points for the same subject, calculate changes between classified Habitat types at each time point, and predict the risk of tumor recurrence based on said changes.
[0019] Meanwhile, the method according to the present disclosure is performed by a brain lesion tracking device that tracks a brain lesion from a medical image, comprising: a step of acquiring sequences of said medical images having multiple parameters; a step of inputting said acquired sequences into a deep learning algorithm to segment a lesion region; a step of classifying voxels within said segmented lesion region into a plurality of habitat types based on physiological variable values among said multiple parameters; and a step of distinguishing the treatment prognosis of said brain lesion as radiation necrosis or tumor recurrence based on quantitative information of said classified habitat types.
[0020] In addition to this, a computer program stored on a computer-readable recording medium for implementing the present disclosure may be further provided. Effects of the invention
[0021] According to the present invention, by quantitatively evaluating the microenvironmental characteristics of brain lesions through integrated analysis of multiparameter MRI sequences and Habitat-based voxel classification, it is possible to provide a technical effect that significantly improves the accuracy and objectivity of distinguishing between radiation necrosis and tumor recurrence compared to conventional single-parameter image analysis or qualitative reading methods.
[0022] In addition, the metadata-based dynamic preprocessing pipeline and nnU-Net-based 3D segmentation of the present invention ensure consistent analysis quality even under various image acquisition conditions, thereby significantly improving universality and reproducibility compared to existing fixed preprocessing methods.
[0023] The effects of the present disclosure are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below. Brief explanation of the drawing
[0024] FIG. 1 is a conceptual diagram of the entire system according to one embodiment of the present disclosure. Figure 2 is a control block diagram of the device shown in Figure 1. FIG. 3 is a flowchart of a method according to one embodiment of the present disclosure. FIGS. 4 and 5 are examples showing trace results according to the present disclosure. Specific details for implementing the invention
[0025] Specific structural or functional descriptions regarding embodiments according to the concept of the present invention disclosed herein are provided merely for the purpose of explaining embodiments according to the concept of the present invention, and embodiments according to the concept of the present invention may be implemented in various forms and are not limited to the embodiments described herein.
[0026] Additionally, terms such as "first" or "second" may be used to describe various components, but said components should not be limited by said terms. Such terms are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component. The term "and / or" may include a combination of a plurality of related described items or any of a plurality of related described items.
[0027] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.
[0028] The terms used in this specification are used merely to describe specific embodiments and are not intended to limit the invention. Singular expressions may include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “comprising” or “having” are intended to indicate the presence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.
[0029] The present invention relates to an image analysis method for follow-up observation after treatment of brain lesions. By integrally processing multiparameter MRI images using deep learning technology and a habitat-based analysis method, it is possible to perform microenvironment analysis at the voxel level and provide quantitative and objective prediction of treatment prognosis.
[0030] FIG. 1 is a conceptual diagram of a brain lesion tracking system according to one embodiment of the present disclosure.
[0031] Referring to FIG. 1, a system (1) according to one embodiment of the present disclosure may be configured to include a device (100) and a user terminal (200), and these components may be organically connected to perform a medical image-based brain lesion analysis and treatment prognosis prediction function.
[0032] The device (100) can acquire multi-parameter medical images and analyze them using a deep learning algorithm to perform Habitat-based classification. To this end, the device (100) may internally include a plurality of specialized processing modules, each of which performs a specific function while forming an integrated analysis pipeline overall. A detailed explanation regarding this will be provided later.
[0033] In one embodiment of the present invention, the device (100) may be implemented in the form of a workstation within a hospital, but is not limited thereto and may also be implemented in the form of a cloud-based server system or a distributed processing system.
[0034] The user terminal (200) can serve as an interface that allows medical personnel to interact with the device (100) and check the analysis results.
[0035] In one embodiment of the present invention, the user terminal (200) may be implemented in the form of a workstation equipped with a high-resolution medical monitor, but is not limited thereto and may be implemented in various forms such as a tablet, a smartphone, or a web browser-based client.
[0036] The user terminal (200) can visualize and provide a spatial habitat map generated from the device (100), thereby allowing medical personnel to intuitively understand the distribution of the microenvironment inside the lesion. According to some other embodiments, the user terminal (200) may also provide a three-dimensional habitat map using augmented reality (AR) or virtual reality (VR) technology, thereby providing a more immersive analysis environment.
[0037] Figure 2 is a conceptual diagram of the device shown in Figure 1.
[0038] Referring to FIG. 2, the device (100) may be configured to include a processor (110), a communication module (120), a memory (130), a display (140), and an input module (150), and these components may be organically connected through an internal bus and operate cooperatively.
[0039] The processor (110) is a computational processing unit of the device (100) and can execute the brain lesion tracking and treatment prognosis prediction algorithm according to the present invention.
[0040] In one embodiment of the present invention, the processor (110) may be configured as a heterogeneous computing architecture combining a multi-core CPU and a GPU.
[0041] The processor (110) may internally include an image preprocessing unit, a deep learning inference unit, a Habitat classification unit, a longitudinal analysis unit, and a prognosis prediction unit. A detailed explanation regarding this will be provided later.
[0042] For example, the processor (110) acquires sequences of medical images having multiple parameters, inputs the acquired sequences into a deep learning algorithm to segment the lesion area, classifies the voxels within the segmented lesion area into multiple habitat types based on the physiological variable values among the multiple parameters, and can distinguish the treatment prognosis of the brain lesion as radiation necrosis or tumor recurrence based on the quantitative information of the classified multiple habitat types.
[0043] Additionally, the processor (110) can input the acquired sequences into the deep learning algorithm by executing at least one preprocessing task among Bias Field Correction (N4), Brain Extraction, Nonlinear Registration, Co-registration & Resample, Denoise, NAWM Intensity Normalization based on normal white matter regions, Motion Correction, Image Arithmetic, and Morphological Operations.
[0044] Additionally, the processor (110) can dynamically configure the execution order or parameters of the at least one preprocessing task based on the metadata of the acquired sequences, or optimize the at least one preprocessing task by combining a known preprocessing task.
[0045] Additionally, the sequences of medical images having the above multiple parameters include a structural MRI sequence and a physiological MRI sequence, wherein the structural MRI sequence includes at least one of a T1 contrast-enhanced image (T1CE) and a T2-weighted image (T2WI), and the physiological MRI sequence may include at least one of an apparent diffusion coefficient (ADC) map and a cerebral blood flow (CBV) map.
[0046] In addition, the deep learning algorithm may include a 3D segmentation network based on nnU-Net.
[0047] Additionally, the above physiological variable values include apparent diffusion coefficient (ADC) values and normalized cerebral blood flow (nCBV) values, and the processor (110) can classify the voxels using a K-means clustering algorithm.
[0048] In addition, the plurality of habitat types may include a hypervascular cellular habitat having low ADC values and high nCBV values, a hypovascular cellular habitat having low ADC values and low nCBV values, and a nonviable tissue habitat having high ADC values and low nCBV values.
[0049] Additionally, the processor (110) can classify the voxels within the divided lesion area into additional structural Habitat types based on the structural variable value among the multiple parameters, and can distinguish between radiation necrosis and tumor recurrence by combining the structural Habitat type and the plurality of Habitat types.
[0050] Additionally, the structural variable values include signal intensity of the T1 contrast-enhanced image and signal intensity of the T2-weighted image, and the structural habitat types may include an enhancing tissue habitat with a high T1 contrast-enhancing signal, a solid low-enhancing habitat with a low T2 signal and a low T1 contrast-enhancing signal, and a nonviable tissue habitat with a high T2 signal and a low T1 contrast-enhancing signal.
[0051] Additionally, the processor (110) can align medical image sequences acquired at multiple points in time for the same subject, calculate changes between the classified Habitat types at each point in time, and predict the risk of tumor recurrence based on the changes.
[0052] Additionally, the change between the above Habitat types may include at least one of the volume change value (ΔV) of each Habitat type, the centroid drift of each Habitat type, and the change in spatial distribution of each Habitat type.
[0053] In addition, the processor (110) can analyze the longitudinal increase of hypovascular cell habit and predict that the risk of recurrence is high if the hypovascular cell habit increases above a reference value.
[0054] In addition, the quantitative information of the plurality of Habitat types may include the volume of each Habitat type, the volume fraction of each Habitat type, and the number of voxels of each Habitat type.
[0055] The above processor (110) can determine that it is a tumor recurrence if the volume fraction of solid low-contrast enhanced habitat is above a reference value, and determine that it is radiation necrosis if the volume fraction of non-living tissue habitat is above a reference value.
[0056] The processor (110) can generate a spatial habitat map by mapping the classified habitat types into the original image space and provide the spatial habitat map by visualizing it through a user interface.
[0057] The processor (110) can highlight areas with a high probability of recurrence on the spatial habitat map when tumor recurrence is predicted, and present them as additional treatment target areas.
[0058] The processor (110) can analyze the association between the quantitative information of the Habitat type and clinical results using logistic regression or a mixed-effects model.
[0059] The above processor (110) can analyze medical images of a patient with brain metastasis or a patient with a brain tumor who is being followed up after stereotactic radiosurgery (SRS).
[0060] The above processor (110) can analyze medical images of a patient with IDH-wild type glioblastoma and identify pathological cellular tumor regions and infiltrating tumor regions.
[0061] The processor (110) can generate a report including at least one of a recurrence risk score, a treatment recommendation category, and prediction reliability information, along with the result of distinguishing between radiation necrosis and tumor recurrence.
[0062] The processor (110) includes a model prediction task for clustering voxels within a lesion area of a brain tumor, the model prediction task includes a parameter for defining the number of clusters for clustering, and the processor (110) can execute the model prediction task in which the number of clusters parameter is set to 3.
[0063] The communication module (120) can serve as an interface that enables the device (100) to transmit and receive data to and from an external system.
[0064] In one embodiment of the present invention, the communication module (120) supports the DICOM communication protocol and can be linked with a PACS (Picture Archiving and Communication System) or a RIS (Radiology Information System), thereby automatically receiving medical image data and transmitting the analysis results. The communication module (120) also supports the HL7 (Health Level Seven) standard and can be linked with an EMR (Electronic Medical Record) system, allowing for the integrated use of the patient's clinical information.
[0065] According to some other embodiments, the communication module (120) may be connected to a remote medical institution via a cloud-based API to provide remote diagnostic services.
[0066] The memory (130) can store at least one process for tracking brain lesions, and may include all algorithms and data according to the present invention.
[0067] In one embodiment of the present invention, the memory (130) may be configured as a hierarchical storage structure combining a high-speed SSD and a large-capacity HDD, and frequently used deep learning models and preprocessing algorithms may be stored in the SSD and large-capacity image data may be stored in the HDD.
[0068] The memory (130) may internally include a model storage area, an image data storage area, an analysis result storage area, and a temporary work area. The model storage area may store pre-trained nnU-Net models, Habitat classification models, prognosis prediction models, etc., and the image data storage area may store original DICOM files and preprocessed image data. The analysis result storage area may store segmentation masks, Habitat maps, quantitative analysis results, prediction reports, etc., and the temporary work area may temporarily store intermediate results generated during the processing.
[0069] The display (140) may be an output device capable of visually representing the analysis results and providing them to medical staff.
[0070] The display (140) can visualize the spatial habitat map using various colors and transparencys, and hypervascular cell habitat can be displayed in, for example, red, hypovascular cell habitat in blue, and non-viable tissue habitat in gray, but this color scheme can be adjusted according to user preference or clinical practice. The display (140) can also display longitudinal changes in an animated form along the time axis so that the pattern of habitat change can be intuitively grasped.
[0071] The input module (150) can serve as an input interface that allows medical personnel to interact with the device (100).
[0072] The input module (150) may include a keyboard, a mouse, and a touchscreen in one embodiment of the present invention, and may also include a voice recognition interface or a gesture recognition interface according to some other embodiments.
[0073] Through the input module (150), medical personnel can select an image to be analyzed, adjust preprocessing parameters, set a Habitat classification threshold, and specify a report format. The input module (150) also provides a Region of Interest (ROI) setting function, which allows medical personnel to perform detailed analysis only on a specific area.
[0074] A method for tracking brain lesions performed by the device (100) shown in FIG. 1 will be described below.
[0075] FIG. 3 is a flowchart of a method for tracking brain lesions according to one embodiment of the present disclosure.
[0076] Referring to FIG. 3, a brain lesion tracking method according to one embodiment of the present disclosure may sequentially perform steps of acquiring sequences of medical images (S100), executing a preprocessing task (S300), segmenting a lesion region (S500), classifying a Habitat type (S700), and predicting a treatment prognosis based on Habitat analysis (S900), and each step may be executed accurately and efficiently through a highly optimized algorithm.
[0077] In the step of acquiring sequences of medical images (S100), the processor (110) can systematically collect and verify sequences of medical images having multiple parameters.
[0078] The processor (110) can perform multi-parameter image acquisition and integrated processing functions through the image acquisition unit.
[0079] For example, the image acquisition unit can receive various MRI sequences compliant with the DICOM (Digital Imaging and Communications in Medicine) standard and manage them in an integrated manner.
[0080] According to one embodiment of the present invention, an image acquisition unit can process T1 contrast-enhanced images (T1CE) and T2-weighted images (T2WI) as structural MRI sequences, and can simultaneously process apparent diffusion coefficient (ADC) maps and cerebral blood flow (CBV) maps as physiological MRI sequences.
[0081] According to some other embodiments, additional sequences such as FLAIR (Fluid Attenuated Inversion Recovery) images, SWI (Susceptibility Weighted Imaging), or ASL (Arterial Spin Labeling) images may also be included to further improve the accuracy of the analysis.
[0082] The image acquisition unit may further perform a metadata analysis substep to extract and analyze information such as the shooting conditions, resolution, slice thickness, and magnetic field strength of each image. This information can be utilized for parameter optimization in a subsequent preprocessing step.
[0083] In addition, the image acquisition unit may further perform a sub-step of quality verification by evaluating the image's signal-to-noise ratio, motion artifacts, magnetic susceptibility artifacts, etc., to confirm whether it has quality suitable for analysis.
[0084] In the preprocessing task execution step (S300), the processor (110) can dynamically configure and execute an optimized preprocessing pipeline based on the metadata of the acquired sequences.
[0085] The processor (110) can perform metadata-based adaptive preprocessing on multi-parameter images through an image preprocessing unit. In this process, optimal preprocessing parameters suitable for the characteristics of each image can be automatically selected.
[0086] For example, the image preprocessing unit may perform a sub-step of pipeline configuration that determines the execution order and parameters of at least one preprocessing task stored according to the characteristics of the image. In one embodiment of the present invention, the preprocessing task may include Bias Field Correction (N4), Brain Extraction, Nonlinear Registration, Co-registration & Resampling, Denoise, NAWM Intensity Normalization based on normal-looking white matter regions, Motion Correction, Image Arithmetic, and Morphological Operations. Alternatively, the image preprocessing unit may optimize the pipeline configuration by combining known preprocessing tasks with at least one preprocessing task stored.
[0087] In addition, the image preprocessing unit may perform additional individual preprocessing execution sub-steps by utilizing parallel processing techniques to efficiently execute each preprocessing task constituting the preprocessing pipeline.
[0088] In addition, the image preprocessing unit may further perform a sub-step of quality evaluation that evaluates the quality of the preprocessing results through quantitative indicators.
[0089] In the lesion region segmentation step (S500), the processor (110) can input the preprocessed sequences into a 3D segmentation network based on nnU-Net to precisely segment the lesion region.
[0090] The processor (110) can divide the lesion region in the preprocessed sequences through the deep learning inference unit.
[0091] For example, the deep learning inference unit can segment lesion regions by executing a 3D segmentation network based on nnU-Net, and to do this, it may internally include a model loading subunit, an inference execution subunit, and a post-processing subunit.
[0092] The model loading subunit can load a pre-trained deep learning model into memory and deploy it in a form optimized for GPU memory, and the inference execution subunit can input preprocessed image data into the network to generate voxel-level segmentation results. The post-processing subunit can generate a final lesion region mask by applying morphological operations, connectivity analysis, size-based filtering, etc., to the network output.
[0093] According to some other embodiments, the deep learning inference unit can further improve partitioning accuracy by utilizing ensemble techniques to combine the results of multiple different network structures.
[0094] In the Habitat type classification step (S700), the processor (110) can classify the voxels within the divided lesion area into multiple Habitat types based on physiological and structural variable values.
[0095] The processor (110) can cluster the voxels within the lesion, which are divided through the Habitat classification unit, according to physiological or structural characteristics.
[0096] For example, the Habitat classification unit can classify voxels within a segmented lesion region into multiple Habitat types based on the K-means clustering algorithm.
[0097] In one embodiment of the present invention, the habitat classification unit may utilize the apparent diffusion coefficient (ADC) value and the normalized cerebral blood flow (nCBV) value, which are physiological variable values, as mediators, and in this case, the number of clusters may be set to 3 to classify into hypervascular cellular habitat, hypovascular cellular habitat, and nonviable tissue habitat.
[0098] Hypervascular cell habitat can be defined, for example, as voxels with low ADC values (less than approx. 0.8 × 10³ mm² / s) and high nCBV values (1.5 or higher), but is not limited thereto and may be appropriately adjusted depending on the characteristics of the patient group or scanner characteristics. Hypovascular cell habitat can be defined, for example, as voxels with low ADC values (less than approx. 0.8 × 10³ mm² / s) and low nCBV values (less than approx. 1.0), and non-viable tissue habitat can be defined, for example, as voxels with high ADC values (more than approx. 1.2 × 10³ mm² / s) and low nCBV values (less than approx. 1.0).
[0099] In another embodiment of the present invention, the habitat classification unit may further classify structural habitat types by utilizing the signal intensity of the T1 contrast-enhanced image and the signal intensity of the T2-weighted image, which are structural variable values. Structural habitat types may include enhancing tissue habitats, solid low-enhancing habitats, and nonviable tissue habitats, which can be combined with physiological habitat types to enable more precise tissue characteristic analysis.
[0100] Contrast-enhancing tissue habitats may be defined, for example, as voxels with high T1 contrast enhancement signals (signal intensity is about 1.5 times or more compared to normal white matter), low contrast enhancement habitats may be defined as voxels with low T2 intensity signals and low T1 contrast enhancement signals, and non-living tissue habitats may be defined as voxels with high T2 intensity signals and low T1 contrast enhancement signals.
[0101] Meanwhile, the Habitat classification unit may internally include a feature extraction subunit, a clustering subunit, and a classification verification subunit.
[0102] The feature extraction subunit can extract multidimensional feature vectors, such as ADC values, nCBV values, T1CE signal strength, and T2WI signal strength, from each voxel, and the clustering subunit can cluster the voxels based on these feature vectors by applying a K-means algorithm. In one embodiment of the present invention, a weighting technique may be applied during the clustering process to assign relatively high weights to ADC and nCBV values; however, this is exemplary and may be adjusted according to the characteristics of the patient group or clinical requirements. The classification verification subunit can evaluate the quality of the clustering results and, if necessary, adjust the number of clusters or initial values to derive optimal classification results.
[0103] In the treatment prognosis prediction step (S900) based on the Habitat analysis, the processor (110) can comprehensively analyze the quantitative information of the classified Habitat types to distinguish between radiation necrosis and tumor recurrence and predict the risk of recurrence.
[0104] The processor (110) can quantitatively analyze the change in Habitat over time through a longitudinal analysis unit.
[0105] For example, the longitudinal analysis unit may internally include a temporal alignment subunit, a change quantification subunit, and a statistical analysis subunit.
[0106] The temporal alignment subunit enables the precise alignment of images captured at different time points to track changes at the same anatomical location. In this process, rigid and non-rigid transformations can be applied sequentially, and anatomical structural deformations caused by brain edema or postoperative changes can also be considered. The change quantification subunit can calculate volume change (ΔV), centroid drift, and spatial distribution changes for each Habitat type, while the statistical analysis subunit can evaluate whether these changes are statistically significant.
[0107] Additionally, the processor (110) can distinguish between radiation necrosis and tumor recurrence based on the results of the Habitat analysis through the prognosis prediction unit and predict the risk of recurrence.
[0108] For example, the prognosis prediction unit may internally include a feature selection subunit, a classification model subunit, and a risk calculation subunit.
[0109] The feature selection subunit can select the features most useful for prognostic prediction among quantitative indicators such as volume by type of habitat, volume fraction, number of voxels, and rate of change. The classification model subunit can distinguish between radiation necrosis and tumor recurrence by utilizing logistic regression analysis or a mixed-effects model. In one embodiment of the present invention, if the volume fraction of solid hypocontrast-enhancing habitat relative to the total lesion volume is greater than or equal to a threshold value (e.g., 73.6%), it is determined that there is a very high probability of tumor recurrence, and if the volume fraction of non-living tissue habitat is greater than or equal to a threshold value, it is determined to be radiation necrosis. In another embodiment of the present invention, if the volume fraction of hypervascular cell habitat exceeds a threshold value (e.g., 3.9%) or the volume fraction of non-living tissue habitat is less than or equal to a threshold value (e.g., 10.1%), it suggests that there are many living tumor cells, so it can be determined to be tumor recurrence. In another embodiment of the present invention, the longitudinal increase of hypovascular cell habit is analyzed, and if it increases above a reference value, it can be predicted that there is a high risk of recurrence.
[0110] Hereinafter, with reference to FIGS. 4 and FIGS. 5, the prognosis prediction results based on a brain lesion tracking method according to one embodiment of the present disclosure will be described.
[0111] Figure 4 is a diagram showing an example of the results of predicting the treatment prognosis of a primary brain tumor.
[0112] Referring to FIG. 4, T1CE, FLAIR, DSC, and DWI sequences were used as input data (I), which may correspond to a combination of structural MRI sequences and physiological MRI sequences defined in the present invention.
[0113] For example, T1CE is a contrast-enhanced T1-weighted image that can indicate the degree of disruption of the blood-brain barrier, and FLAIR is a type of T2-weighted image that suppresses cerebrospinal fluid signals to show edema or infiltration around the lesion more clearly. Dynamic Susceptibility Contrast (DSC) provides information on cerebral blood flow and can be used to generate CBV maps, while Diffusion Weighted Imaging (DWI) can provide information on tissue cell density through the generation of ADC maps.
[0114] The HABITAT SEG image may be an example of a spatial HABITAT map that visually represents the HABITAT classification results according to the present invention. In this map, hypervascular cell HABITAT, hypovascular cell HABITAT, non-viable tissue HABITAT, etc., may be distinguished and displayed through different colors, allowing medical professionals to intuitively understand the distribution of the microenvironment within the lesion. In one embodiment of the present invention, hypervascular cell HABITAT may refer to an area of high tumor activity exhibiting active angiogenesis and high cell density, while hypovascular cell HABITAT may refer to an area where tumor cells exist that are still viable despite limited blood supply. Non-viable tissue HABITAT may represent an area where necrosis or cystic changes have occurred.
[0115] The output report(R) includes quantitative indicators such as all volume and volume change, which may correspond to quantitative information such as volume, volume fraction, and number of voxels for each Habitat type as defined in the present invention. Longitudinal tracking and change analysis can be performed by comparing images captured on different dates, through which the volume change value (ΔV), center point shift, and spatial distribution change of each Habitat type can be quantified.
[0116] For example, the response to treatment can be evaluated through changes in HABITAT SEG and volume change values, and the risk of recurrence can be predicted based on the increase or decrease patterns of specific HABITAT types. In one embodiment of the present invention, if a longitudinal increase in hypovascular cell HABITAT is observed, this may indicate the proliferation of treatment-resistant tumor cells, and thus a high risk of recurrence may be predicted.
[0117] In addition, a spatial Habitat map generated by mapping classified Habitat types into the original image space can be visualized through a user interface. This visualization can provide medical professionals with important information for understanding the spatial heterogeneity of lesions and establishing treatment plans. If tumor recurrence is predicted, areas with a high probability of recurrence can be highlighted on the Habitat map and presented as additional treatment target areas, which can be utilized in planning stereotactic radiosurgery or local treatment.
[0118] Figure 5 is an example of the results of meta-detection and tracking of brain metastasis.
[0119] Referring to FIG. 5, T1CE (GRE / TSE) images were used as input data (I), which can represent different sequence types of T1-weighted images after contrast enhancement. GRE (Gradient Recalled Echo) has the characteristics of enabling rapid imaging and clearly showing vascular structures, while TSE (Turbo Spin Echo) can provide high resolution and excellent contrast. In one embodiment of the present invention, these different sequence types can be automatically recognized through metadata analysis in the preprocessing step, and parameters optimized for each can be applied.
[0120] The output report(R) displays the results of a longitudinal analysis using multiple time points, and allows for the quantification of the appearance of new lesions and changes in the volume of existing lesions through the calculation of changes over time. The "New" area may refer to new metastatic lesions not observed in previous examinations, while the "Increase" area may indicate parts of existing lesions that have increased in size. This analysis can be performed in conjunction with the analysis of changes between Habitat types defined in this invention, enabling a more precise evaluation of the progression of lesions through changes in volume, center point shift, and spatial distribution of each Habitat type.
[0121] Meanwhile, the disclosed embodiments may be implemented in the form of a recording medium that stores instructions executable by a computer. The instructions may be stored in the form of program code and, when executed by a processor, may generate a program module to perform the operation of the disclosed embodiments. The recording medium may be implemented as a computer-readable recording medium.
[0122] Computer-readable recording media include all types of recording media that store instructions that can be decoded by a computer. Examples include ROM (Read Only Memory), RAM (Random Access Memory), magnetic tape, magnetic disk, flash memory, optical data storage devices, etc.
[0123] As described above, the disclosed embodiments have been explained with reference to the attached drawings. Those skilled in the art will understand that the present disclosure may be practiced in forms different from the disclosed embodiments without changing the technical spirit or essential features of the present disclosure. The disclosed embodiments are illustrative and should not be interpreted restrictively. Explanation of the symbols
[0124] 1: Brain lesion tracking system 100: Device 200: User terminal
Claims
Claim 1 A brain lesion tracking device comprising: a memory storing at least one process for tracking a brain lesion from a medical image; and a processor operating according to the at least one process; wherein the processor acquires sequences of the medical image having multiple parameters, inputs the acquired sequences into a deep learning algorithm to segment a lesion region, classifies voxels within the segmented lesion region into multiple habitat types based on physiological variable values among the multiple parameters, distinguishes the treatment prognosis of the brain lesion as radiation necrosis or tumor recurrence based on quantitative information of the classified multiple habitat types, sequentially applies rigid transformation and non-rigid transformation to medical image sequences acquired at multiple time points for the same subject to mutual alignment, calculates centroid drift between the habitat types classified at each time point as a change between the habitat types, and predicts the risk of tumor recurrence based on the change. Claim 2 A brain lesion tracking device according to claim 1, wherein the processor executes at least one preprocessing task among Bias Field Correction (N4), Brain Extraction, Nonlinear Registration, Co-registration & Resample, Denoise, NAWM Intensity Normalization based on a normal-looking white matter region, Motion Correction, Image Arithmetic, and Morphological Operations on the acquired sequences and inputs them to the deep learning algorithm. Claim 3 A brain lesion tracking device according to paragraph 2, wherein the processor dynamically configures the execution order or parameters of the at least one preprocessing task based on metadata of the acquired sequences, or optimizes the at least one preprocessing task by combining known preprocessing tasks. Claim 4 A brain lesion tracking device according to claim 1, wherein the sequences of medical images having multiple parameters include a structural MRI sequence and a physiological MRI sequence, wherein the structural MRI sequence includes at least one of a T1 contrast-enhanced image (T1CE) and a T2-weighted image (T2WI), and the physiological MRI sequence includes at least one of an apparent diffusion coefficient (ADC) map and a cerebral blood volume (CBV) map. Claim 5 In claim 4, the deep learning algorithm comprises a 3D segmentation network based on nnU-Net, a brain lesion tracking device. Claim 6 A brain lesion tracking device according to claim 4, wherein the physiological variable values include apparent diffusion coefficient (ADC) values and normalized cerebral blood flow (nCBV) values, and the processor classifies the voxels using a K-means clustering algorithm. Claim 7 A brain lesion tracking device according to claim 6, wherein the plurality of habitat types include a hypervascular cellular habitat, a hypovascular cellular habitat, and a nonviable tissue habitat. Claim 8 A brain lesion tracking device according to claim 7, wherein the processor further classifies voxels within the segmented lesion region into structural habitat types based on structural variable values among the multiple parameters, wherein the structural variable values include signal intensity of the T1 contrast-enhanced image and signal intensity of the T2-weighted image, and the structural habitat types include an enhancing tissue habitat, a solid low-enhancing habitat, and a nonviable tissue habitat. Claim 9 delete Claim 10 A method for tracking a brain lesion from medical images, performed by a brain lesion tracking device, comprising: acquiring sequences of medical images having multiple parameters; inputting the acquired sequences into a deep learning algorithm to segment a lesion region; classifying voxels within the segmented lesion region into multiple habitat types based on physiological variable values among the multiple parameters; and distinguishing the treatment prognosis of the brain lesion as radiation necrosis or tumor recurrence based on quantitative information of the classified multiple habitat types; wherein the distinguishing step comprises: sequentially applying rigid transformation and non-rigid transformation to medical image sequences acquired at multiple time points for the same subject to mutual alignment; calculating centroid drift between the habitat types classified at each time point as a change between the habitat types; and predicting the risk of tumor recurrence based on the change.
Citation Information
Patent Citations
Model construction method and device based on MRI habitat imaging, equipment and storage medium
CN119417749A
System and Methods Integrating Distributed Machine Learning Layers for Processing Multimodal Data Sets in Real Time to Optimize Outcomes
US20250364108A1
