Four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method

Abstract

The invention relates to the technical field of computer vision, in particular to a four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method, which comprises the following steps of: firstly, carrying out initial segmentation on an input medical image, and positioning a heart ROI (Region of Interest); then, performing blood vessel coarse segmentation in the heart ROI to obtain a coarse segmentation result; based on the coarse segmentation result, performing fine segmentation on the blood vessel to restore the fine structure of the blood vessel and obtain a fine blood vessel contour result; then, based on the fine blood vessel contour result, plaques on the blood vessel wall are segmented, the positions, boundaries, forms and attached blood vessel segment information of the plaques are determined, and a final segmentation result is obtained; and finally, converting the final segmentation result to generate a three-dimensional model for clinical analysis. According to the reconstruction method provided by the invention, the problems of dependence on manual intervention, low efficiency, poor result consistency, weak modal adaptability, non-universal output format and the like in the existing cardiac coronary vessel and plaque reconstruction method are solved.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a four-stage progressive medical image segmentation and STL three-dimensional reconstruction method for vascular plaques. Background Technology

[0002] The core pathological feature of coronary artery disease (CAD) is the formation of atherosclerotic plaques in the coronary artery walls. The location, size, shape, and nature of these plaques (such as the degree of calcification and stability) directly determine the disease risk and treatment plan. Clinically, it is necessary to obtain information about the coronary arteries and plaques through medical imaging, and use 3D reconstruction technology to transform 2D images into intuitive 3D models to assist doctors in qualitative analysis and quantitative measurement.

[0003] Currently, the mainstream coronary imaging technologies include: CTA (Computed Tomography Angiography), IVUS (Intravascular Ultrasound), and OCT (Optical Coherence Tomography). Among them, CTA has the advantages of being non-invasive, covering the entire coronary tree, and being able to initially show the location of plaques and the degree of vascular stenosis, but its accuracy in identifying the boundaries of soft plaques (mainly lipid cores) is relatively low. IVUS has the advantages of being invasive, having high resolution (about 100 μm), and being able to clearly distinguish plaque components (such as fibrous caps and lipid cores), but it can only acquire local images within the vascular lumen and requires catheter movement to achieve "scanning" data acquisition. OCT has even higher resolution (about 10-20 μm) and can accurately identify the microstructure of plaques (such as thin fibrous caps), but its imaging depth is shallow (only 1-3 mm) and it is also limited to the local intraluminal field of view.

[0004] With the increasing demand for precision medicine, the integrated process of "image acquisition - information extraction - 3D reconstruction" has become a development trend. However, existing reconstruction methods still cannot balance automation, accuracy and universality, and cannot fully meet the needs of efficient clinical diagnosis. Summary of the Invention

[0005] This application provides a four-stage progressive medical image segmentation and STL three-dimensional reconstruction method for vascular plaques, which solves the problems of existing cardiac coronary artery and plaque reconstruction methods, such as reliance on manual intervention, low efficiency, poor result consistency, weak modal adaptability, and non-universal output format.

[0006] To address the aforementioned technical problems, this application provides a four-stage progressive medical image segmentation and STL 3D reconstruction method for vascular plaques, comprising the following steps: First, initial segmentation is performed on the input medical image to locate the region of interest (ROI) of the heart; then, coarse segmentation of the blood vessels is performed within the cardiac ROI to obtain a coarse segmentation result; next, based on the coarse segmentation result, fine segmentation of the blood vessels is performed to restore the fine structure of the blood vessels, obtaining a fine blood vessel contour result; then, based on the fine blood vessel contour result, plaques on the blood vessel wall are segmented, and the location, boundary, morphology, and attached blood vessel segment information of the plaques are determined to obtain a final segmentation result; finally, the final segmentation result is converted to generate a 3D model for clinical analysis.

[0007] In some exemplary embodiments, the initial segmentation of the input medical image to locate the cardiac ROI includes: using the nnU-Net 3D full-resolution method to perform initial segmentation of the input medical image to delineate the cardiac ROI for subsequent vascular and plaque analysis, so as to eliminate interference from non-cardiac regions, reduce the computational load of subsequent segmentation tasks, and improve the focus of the target region.

[0008] In some exemplary embodiments, coarse segmentation of blood vessels is performed within the cardiac ROI to obtain coarse segmentation results. This includes: within the cardiac ROI, preliminary segmentation of the aorta and coronary arteries is completed based on the nnU-Net 3D full-resolution method to obtain coarse segmentation results, thereby providing a basic outline for subsequent fine segmentation, while providing differentiated optimization strategies for different clinical needs.

[0009] In some exemplary embodiments, the optimization strategy is an improvement on the nnU-Net 3D full-resolution method.

[0010] In some exemplary embodiments, the operation of the optimization strategy includes replacing the loss function layer of the nnU-Net 3D full-resolution method with the CLRecall loss function; the CLRecall loss function improves the initial accuracy of blood vessel contours by enhancing the recall rate of true positive blood vessel regions, thereby reducing false positive detections in blood vessel segmentation.

[0011] In some exemplary embodiments, the operation of the optimization strategy includes: preserving the standard loss function of the nnU-Net 3D full-resolution method to enhance the connectivity of vascular structures; the standard loss function includes the Dice loss function and the cross-entropy loss function; and leveraging the modeling advantages of the continuous spatial structure through this optimization strategy to ensure the contour continuity of the aortic trunk and major coronary branches.

[0012] In some exemplary embodiments, fine segmentation of blood vessels is performed based on the coarse segmentation results to restore the fine structure of blood vessels and obtain fine blood vessel contour results. This includes: using the coarse segmentation results as an initial mask, inputting the coarse segmentation results into the high-resolution sub-network of the nnU-Net 3D full-resolution method to complete the fine segmentation of the aorta and coronary arteries, optimizing the vessel wall thickness and the minimum diameter of the stenotic segment, restoring the morphological changes of small branches with a diameter <2mm and the plaque attachment site, and providing a precise vascular structure benchmark for subsequent plaque segmentation.

[0013] In some exemplary embodiments, based on the detailed vascular contour results, plaques on the vascular wall are segmented, and the location, boundary, shape, and attached vascular segment information of the plaques are determined to obtain the final segmentation result. This includes: using the detailed vascular contour results as a region mask, enabling the segmentation function of the nnU-Net 3D full-resolution method in the region inside the vascular structure, and fine-tuning the feature extraction layer of the network by inputting the image features of the plaques, so that the network prioritizes the identification of lesion areas inside the blood vessels, avoids misjudging extravascular tissues as plaques, and achieves accurate separation of lesion areas inside the vascular structure.

[0014] In some exemplary embodiments, the final segmentation result is transformed to generate a three-dimensional model for clinical analysis, including: performing operations on the final segmentation result in a data preparation stage, a core transformation stage, and a post-processing output stage to generate a three-dimensional model for clinical analysis.

[0015] In some exemplary embodiments, the data preparation phase includes: starting from the NII file input, including operations such as loading, binarization, and data validation branches; the core transformation phase includes: independently demonstrating operations such as oversampling, Gaussian smoothing, Marching Cubes algorithm, and coordinate transformation; the post-processing output phase includes: demonstrating Laplacian smoothing, mesh simplification, STL model export, and finally generating a 3D model.

[0016] The technical solution provided in this application has at least the following advantages: This application provides a four-stage progressive medical image segmentation and STL 3D reconstruction method for vascular plaques, comprising the following steps: First, initial segmentation is performed on the input medical image to locate the cardiac region of interest (ROI); then, coarse segmentation of blood vessels is performed within the cardiac ROI to obtain coarse segmentation results; next, based on the coarse segmentation results, fine segmentation of blood vessels is performed to restore the fine structure of blood vessels and obtain fine blood vessel contour results; then, based on the fine blood vessel contour results, plaques on the blood vessel wall are segmented, and the location, boundary, morphology, and attached blood vessel segment information of the plaques are determined to obtain the final segmentation results; finally, the final segmentation results are converted to generate a 3D model for clinical analysis.

[0017] This application proposes a four-stage progressive medical image segmentation technique based on the nnU-Net 3D full-resolution method. Through a hierarchical constraint process of "cardiac localization → coarse vessel segmentation → fine vessel segmentation → plaque segmentation," it overcomes the accuracy bottleneck of traditional single-stage segmentation. Compared to the original nnU-Net single-stage segmentation method, this application achieves progressive accuracy improvement through a four-stage progressive architecture, effectively solving the problems of "non-target region interference" and "low accuracy in multi-structure collaborative segmentation." The innovative dual-strategy optimization (CLRecall loss / connectivity enhancement) adapts to different clinical scenarios, embedding anatomical hierarchical relationships into the algorithm to significantly improve lesion detection rate, and realizing an integrated process from image segmentation to 3D reconstruction, improving segmentation accuracy by more than 10%. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0019] Figure 1 This is a flowchart illustrating a four-stage progressive medical image segmentation and STL three-dimensional reconstruction method for vascular plaques provided in this application embodiment.

[0020] Figure 2 A flowchart illustrating the specific process of blood vessel segmentation provided in this application embodiment.

[0021] Figure 3 This is a schematic diagram of the network structure shared by all tasks provided in the embodiments of this application.

[0022] Figure 4 This is a schematic diagram of the process for reconstructing an STL file provided in an embodiment of this application. Detailed Implementation

[0023] As can be seen from the background technology, existing methods for cardiac coronary artery and plaque reconstruction suffer from problems such as reliance on manual intervention, low efficiency, poor consistency of results, weak modal adaptability, and non-universal output formats.

[0024] Currently, the technical solutions most similar to "three-dimensional reconstruction of coronary arteries and plaques" can be mainly divided into three categories, all of which have clear reliance on manual intervention or functional limitations, as detailed below: Option 1: A semi-automated coronary artery and plaque reconstruction method based on CTA images.

[0025] Technical principle: Using clinically common CTA images as input, reconstruction is achieved through "image preprocessing → manual intervention segmentation → 3D stitching". The typical process is as follows: Image preprocessing: Denoising (such as Gaussian filtering) and enhancing (such as improving vascular contrast) are performed on the raw CTA data.

[0026] Manual intervention segmentation: Physicians need to manually delineate the boundaries of coronary arteries on 2D CTA tomographic images (e.g., marking the vessel central axis once every 5-10 slices) and manually confirm the plaque region based on CT value thresholds (e.g., CT value of calcified plaques > 130 HU) (excluding vessel wall calcification artifacts); 3D reconstruction: The software generates a 3D model of the vessel based on the manually marked vessel axis and boundaries using interpolation algorithms (e.g., cubic spline interpolation). The plaque model needs to be stitched together separately based on the manually confirmed region; Output format: Mostly proprietary formats (e.g., DICOM 3D), which can only be viewed on the original imaging workstation (e.g., Siemens SOMATOM, GE Revolution CT workstation) and cannot be directly exported to general 3D software. Application scenarios: Mostly used for preliminary assessment of coronary artery stenosis, but requires 1-2 hours of manual operation per case by the physician, and is only suitable for simple coronary artery anatomy (e.g., without severe branch overlap).

[0027] Option 2: A semi-automated stitching and reconstruction method based on IVUS / OCT images.

[0028] Technical Principle: Reconstruction is achieved using IVUS / OCT local intravascular images through "catheter position calibration → manual frame matching → 3D stitching." The process is as follows: Data Synchronization: The IVUS / OCT tomographic images are synchronized with the catheter retraction speed (e.g., 0.5 mm / s) to obtain the axial position information of each slice; Manual Frame Matching: Due to potential offsets during catheter retraction, the physician needs to manually adjust the positions of adjacent slices (e.g., aligning them using vessel wall feature points) to avoid model distortion; Plaque Segmentation: Based on the grayscale features of IVUS / OCT (e.g., soft plaques appear as low signal and calcified plaques as high signal with shadows in OCT), the physician needs to manually mark the inner and outer boundaries of the plaques; 3D Reconstruction: The aligned tomographic images are stacked to generate a local 3D model of the vessel lumen and plaques. The output format is mostly internal binary format and only supports dedicated analysis software (e.g., Boston Scientific iLab). Application scenarios: It is mostly used for plaque characterization assessment before stent implantation, but it can only reconstruct local vascular segments of 3-5cm, cannot cover the entire coronary tree, and manual frame matching is prone to introducing errors.

[0029] Option 3: A single-modal automated reconstruction method during the research phase.

[0030] In recent years, some literature (such as the 2022 article "Automated Coronary Plaque Segmentation from CTA" published in *Medical Image Analysis*) has proposed automated methods based on deep learning. The core process is as follows: using a U-Net-like neural network to segment blood vessels in CTA images; identifying calcified plaque regions based on a pre-trained plaque classification model (such as ResNet); and directly converting the segmentation results into a 3D mesh model, with the output format being PLY (polygonal file format). Limitations: It only supports single-modal CTA images, and the accuracy rate for identifying soft plaques is less than 60% (due to the low contrast between soft plaques and the vessel wall in CTA); furthermore, it does not consider complex situations such as overlapping vessel branches and metal artifacts (e.g., after coronary stenting), limiting its practical clinical applicability.

[0031] In summary, existing technical solutions cannot meet the clinical needs of "fully automated, multimodal, high precision, and universality." Their main shortcomings are reflected in the following five aspects: (1) There is a lot of manual intervention and the efficiency is extremely low: Option 1 (CTA semi-automation) requires doctors to manually outline the blood vessel boundaries and confirm the plaque area, which takes 1 to 2 hours per case; Option 2 (IVUS / OCT semi-automation) requires manual frame matching, which takes more than 2 hours for complex cases, and cannot meet the needs of clinical batch case processing.

[0032] (2) Poor consistency of results and strong subjectivity: The marking of vascular boundaries and plaque areas depends on the doctor's experience. The difference in operation results between different doctors (or even the same doctor at different times) can reach 15%~20% (such as the error in measuring the degree of stenosis), which affects the accuracy of diagnosis and the consistency of treatment plan.

[0033] (3) Limited modal adaptability and incomplete information extraction: Scheme 1 is only applicable to CTA and cannot utilize the high-resolution plaque component information of IVUS / OCT; Scheme 2 is only applicable to IVUS / OCT and cannot cover the entire coronary tree; Scheme 3 only supports CTA and has low accuracy in soft plaque recognition. None of them can achieve "multimodal fusion + complete information extraction".

[0034] (4) The output format is not universal and the scalability is poor: the output of the existing solution is mostly in a special format (such as DICOM 3D, internal binary format), which cannot be directly imported into general three-dimensional software (such as AutoCAD, ANSYS, 3D printing slicing software), which makes it impossible for doctors to perform further mechanical analysis (such as plaque rupture risk assessment), surgical simulation or 3D printing solid model.

[0035] (5) Weak adaptability to complex cases and insufficient reconstruction accuracy: When faced with complex situations such as severe vascular stenosis (stenosis rate > 70%), branch overlap, and metal artifacts (such as stents and pacemakers), the existing scheme is prone to discontinuous vascular segmentation, missed / false detection of plaques, and deviation of the reconstruction model from the actual anatomical structure by more than 10%.

[0036] Therefore, addressing the problems of existing methods for reconstructing coronary arteries and plaques, such as reliance on manual intervention, low efficiency, poor result consistency, weak modal adaptability, and non-universal output formats, this application provides a four-stage progressive medical image segmentation and STL three-dimensional reconstruction method for vascular plaques. This method accurately extracts the geometric information of coronary arteries (including trunks and branches) and plaques (including calcified plaques, soft plaques, etc.) without manual intervention and automatically reconstructs them into a universal STL (stereolithography) format three-dimensional model. This meets the high-efficiency and accurate requirements of clinical diagnosis (such as plaque nature assessment and stenosis measurement) and surgical planning (such as stent size matching and implantation path simulation). The purpose of the four-stage progressive medical image segmentation and STL three-dimensional reconstruction method for vascular plaques provided in this application is: (1) Achieve fully automated processing: Without any manual intervention, the entire process from input of multimodal images (CTA+IVUS / OCT) to extraction of coronary artery and plaque information and three-dimensional reconstruction is automated, reducing the processing time of a single case to less than 10 minutes.

[0037] (2) Ensure consistency and accuracy of results: eliminate subjective human error through standardized algorithms (such as multimodal fusion segmentation and adaptive artifact removal) to keep the difference in reconstruction results of different cases and different equipment within 5% and meet the precise needs of clinical diagnosis.

[0038] (3) Compatible with multimodal images and complete information extraction: It integrates the advantages of CTA in full coronary tree coverage and IVUS / OCT in high-resolution plaque composition to achieve the extraction of "full structure of coronary trunk and branches + full information on plaque location, size and properties".

[0039] (4) Output general STL format to expand application scenarios: automatically export the three-dimensional model to STL format, which is compatible with general three-dimensional analysis, simulation and 3D printing software, and supports subsequent mechanical analysis, surgical planning, 3D printed solid model and other extended applications.

[0040] (5) Improve adaptability to complex cases: Through dedicated complex case processing modules (such as narrow area interpolation completion and artifact suppression algorithms), ensure the reconstruction accuracy in scenarios such as severe stenosis, branch overlap, and metal artifacts. The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0041] See Figure 1 This application provides a four-stage progressive medical image segmentation and STL three-dimensional reconstruction method for vascular plaques, including the following steps: Step S1: Perform initial segmentation on the input medical image and locate the cardiac ROI.

[0042] Step S2: Perform coarse segmentation of blood vessels within the cardiac ROI to obtain the coarse segmentation results.

[0043] Step S3: Based on the coarse segmentation results, perform fine segmentation of blood vessels to restore the fine structure of blood vessels and obtain fine blood vessel contour results.

[0044] Step S4: Based on the detailed vascular contour results, the plaques on the vascular wall are segmented, and the location, boundary, shape and attached vascular segment information of the plaques are determined to obtain the final segmentation result.

[0045] Step S5: Convert the final segmentation results to generate a three-dimensional model for clinical analysis.

[0046] This application utilizes the nnU-Net 3D full-resolution method, employing a five-stage progressive workflow from coarse localization to fine segmentation to achieve accurate segmentation and 3D reconstruction of blood vessels and plaques. The method first uses coarse segmentation of the cardiac region to locate the Region of Interest (ROI) and eliminate non-target interference. Second, coarse segmentation of blood vessels is performed within the cardiac ROI, employing two strategies: CLRecall loss function and standard loss function, to optimize false positive control and connectivity preservation, respectively. Third, a high-resolution subnetwork is used to perform fine segmentation of blood vessels, restoring subtle structures. Fourth, plaque lesions are segmented under precise blood vessel contour constraints. Finally, the segmentation results are converted into an STL model, which undergoes three stages: data preparation, Marching Cubes core transformation, and post-processing optimization, generating a 3D model suitable for clinical analysis. The overall workflow, through hierarchical constraints and progressive optimization, achieves accurate extraction from global localization to local lesion extraction.

[0047] In some embodiments, step S1 involves initial segmentation of the input medical image to locate the cardiac ROI, including: using the nnU-Net 3D full-resolution method to perform initial segmentation of the input medical image to delineate the cardiac ROI for subsequent vascular and plaque analysis, thereby eliminating interference from non-cardiac regions, reducing the computational load of subsequent segmentation tasks, and improving the focus of the target region.

[0048] Specifically, step S1 is mainly used for initial localization. The vessel and plaque segmentation in this application uses the nnU-Net (a U-Net-based adaptive medical image segmentation framework) 3D full-resolution method (reference: Isensee F, JaegerPF, Kohl SAA, et al. nnU-Net: Self-adapting Framework for U-Net-Based MedicalImage Segmentation. arXiv:1809.10486, 2018) as the base segmentation network. It is achieved through a four-stage progressive process "from coarse localization to fine segmentation." The specific vessel segmentation process is as follows: Figure 2 As shown.

[0049] In step S1, during the initial segmentation of the input medical images, this application uses the nnU-Net 3D full-resolution method to perform coarse segmentation of the heart region, defining the region of interest (ROI) for subsequent vascular and plaque analysis. This effectively eliminates interference from non-heart regions such as the lungs and thoracic bones, reducing the computational load of subsequent segmentation tasks and improving the focus of the target region. Specifically, this application uses the nnU-Net 3D full-resolution method to perform initial segmentation of the input medical images (such as CTA, IVUS / OCT). This method, through adaptive adjustment of network layer parameters and optimization of resolution matching strategies, can quickly adapt to the anatomical scale features of the heart images, achieving efficient coarse localization of the entire heart region and providing "target range constraints" for subsequent subdivision tasks.

[0050] In some embodiments, step S2 involves performing coarse segmentation of blood vessels within the cardiac ROI to obtain coarse segmentation results. This includes: within the cardiac ROI, performing preliminary segmentation of the aorta and coronary arteries based on the nnU-Net 3D full-resolution method to obtain coarse segmentation results, thereby providing a basic outline for subsequent fine segmentation, while also providing differentiated optimization strategies for different clinical needs.

[0051] Specifically, step S2 is mainly used for coarse segmentation of the aorta and coronary arteries. Within the cardiac ROI defined in Task 1, the aorta and coronary arteries are initially segmented based on the nnU-Net 3D full-resolution method, providing a basic outline for subsequent fine segmentation, while also providing differentiated optimization strategies for different clinical needs.

[0052] In some embodiments, the optimization strategy is an improvement on the nnU-Net 3D full-resolution method. Specifically, step S2 can be implemented using either strategy A or strategy B (either of two strategies, both based on the improvement of the nnU-Net 3D full-resolution method).

[0053] In some embodiments, the optimization strategy of Strategy A includes replacing the loss function layer of the nnU-Net 3D full-resolution method with the CLRecall loss function. The core focus is on reducing false positive detections in vessel segmentation (i.e., reducing the misidentification of perivascular fat and myocardial tissue as vessels). The CLRecall loss function enhances the recall rate of true positive vessel regions to reduce false positive detections in vessel segmentation and improve the initial accuracy of vessel contours. It is suitable for images with low contrast between vessels and surrounding tissues (such as CTA images with a high proportion of soft plaques).

[0054] In some embodiments, the operation of strategy B optimization strategy includes: retaining the standard loss function of the nnU-Net 3D full-resolution method (such as Dice loss + cross-entropy loss), with the core focus on enhancing the connectivity of vascular structures (i.e. avoiding segmentation and breakage caused by small branches and stenotic segments of blood vessels). Through the "modeling advantage of continuous spatial structures", this method ensures the contour continuity of the main aortic trunk and major branches of the coronary arteries (such as the left anterior descending artery and circumflex artery), which is suitable for images with dense vascular branches and complex anatomical structures (such as CTA images of multivessel vascular lesions). In some embodiments, step S3 involves performing fine segmentation of blood vessels based on the coarse segmentation results to restore the fine structure of the blood vessels and obtain fine blood vessel contour results. This includes: using the coarse segmentation results as an initial mask, inputting the coarse segmentation results into the high-resolution sub-network of the nnU-Net 3D full-resolution method to complete the fine segmentation of the aorta and coronary arteries, optimizing the vessel wall thickness and the minimum diameter of the stenotic segment, restoring the morphological changes of small branches with a diameter <2mm and the plaque attachment site, and providing an accurate vascular structure benchmark for subsequent plaque segmentation.

[0055] Specifically, step S3 is mainly a high-resolution optimization process. Based on the coarse segmentation results obtained in step S2, step S3 further refines the segmentation of the aorta and coronary arteries through the high-resolution processing module of the nnU-Net 3D full-resolution method, optimizing the accuracy of vessel boundaries (such as vessel wall thickness and minimum diameter of stenotic segments), and restoring the fine structure of the vessels (such as small branches with a diameter <2mm and morphological changes of vessels at plaque attachment sites), providing a precise "vascular structure benchmark" for subsequent plaque segmentation. The specific implementation method includes: using the coarse segmentation results of the vessels output in step S2 as the "initial mask", inputting it into the high-resolution sub-network of the nnU-Net 3D full-resolution method (this sub-network improves the segmentation ability of fine structures by retaining more original image resolution information), and through the collaborative logic of "coarse segmentation contour constraint + high-resolution detail completion", resolving the contradiction between "inaccurate localization of large-scale structures" and "loss of small-scale details" in traditional single-stage segmentation, thereby achieving a progressive improvement in the accuracy of vessel segmentation.

[0056] In some embodiments, step S4 involves segmenting plaques on the vessel wall based on the detailed vessel contour results, and clarifying the location, boundaries, shape, and attached vessel segment information of the plaques to obtain the final segmentation result. This includes: using the detailed vessel contour results as a region mask, enabling the segmentation function of the nnU-Net 3D full-resolution method in the internal region of the vessel structure, and simultaneously fine-tuning the feature extraction layer of the network by inputting the image features of the plaques, so that the network prioritizes the identification of lesion areas within the vessel, avoids misjudging extravascular tissue as plaques, and achieves accurate separation of lesion areas within the vessel structure.

[0057] Specifically, step S4 mainly involves the precise extraction of the lesion area. Based on the precise vascular contour obtained in step S3, step S4 further segments plaques (such as calcified plaques and soft plaques) on the vessel wall using the nnU-Net 3D full-resolution method. This clarifies the location, boundaries, morphology, and attached vessel segment information of the plaques, providing complete lesion geometric data for subsequent collaborative 3D reconstruction of the "vessel + plaque". The specific implementation method includes: using the fine segmentation results of the vessels output from Task 3 as a "region mask," enabling the segmentation function of the nnU-Net 3D full-resolution method only in the internal regions of the vascular structure; and fine-tuning the network's feature extraction layer by inputting the image features of the plaques (such as CT values ​​>130HU for calcified plaques in CTA and low signal characteristics for soft plaques in OCT), so that the network prioritizes identifying lesion areas within the vessels, avoiding misidentification of extravascular tissue as plaques, thus achieving "precise separation of lesion areas within the vascular structure". The network structure shared by all the above tasks is as follows: Figure 3 As shown.

[0058] The four-stage progressive vascular plaque medical image segmentation and STL 3D reconstruction method proposed in this application differs from the nnU-Net 3D full-resolution method proposed by Isensee et al. in 2018 (paper "nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation") in the core difference of the segmentation process and its progressive design. Specifically, the substantial differences are as follows: The nnU-Net 3D full-resolution method in the paper takes "single-stage, single-target segmentation" as its core logic. That is, for single anatomical structures such as "heart" and "blood vessels", it directly completes the segmentation through an adaptive network without forming a progressive constraint of multiple targets and multiple scales. In contrast, the technical solution adopts "four-stage progressive segmentation", from "coarse localization of the heart region → coarse segmentation of blood vessels → fine segmentation of blood vessels → plaque segmentation". Each step is constrained by the result of the previous step, gradually narrowing the segmentation range and improving the segmentation accuracy, which solves the problems of "more interference from non-target regions and low accuracy of multi-structure collaborative segmentation" in single-stage segmentation.

[0059] The nnU-Net 3D full-resolution method in the paper focuses on "adaptive adjustment of network parameters" to adapt to different types of medical images, but does not design a dedicated process for the anatomical correlation between "heart-vessel-plaque". In contrast, this technical solution deeply integrates the network method with the anatomical logic of cardiac images (the heart is the carrier of blood vessels, and blood vessels are the carrier of plaques) through "progressive process improvement". It first excludes non-target areas and then gradually focuses on the lesion area, making the segmentation results more in line with the clinical needs for "structure-lesion co-analysis".

[0060] After completing the above four steps, a segmented .nii file can be obtained. Based on the intact segmentation labels and the original file, 3D reconstruction in STL format can be performed. The flowchart for reconstructing the STL file is as follows: Figure 4 As shown, rebuilding an STL file can be divided into three stages: data preparation stage (left), core conversion stage (middle), and post-processing output stage (right).

[0061] In some embodiments, step S5 involves converting the final segmentation result to generate a three-dimensional model for clinical analysis, including performing data preparation, core conversion, and post-processing output operations sequentially on the final segmentation result to generate a three-dimensional model for clinical analysis.

[0062] In some embodiments, the data preparation phase includes: starting from the NII file input, including operations such as loading, binarization, and data validation branches; the core transformation phase includes: independently demonstrating operations such as oversampling, Gaussian smoothing, Marching Cubes algorithm, and coordinate transformation; the post-processing output phase includes: demonstrating Laplacian smoothing, mesh simplification, STL model export, and finally generating a 3D model.

[0063] This application employs a complete four-stage progressive segmentation methodology and inter-stage constraint transfer mechanism, the application of the CLRecall loss function in vessel segmentation, high-resolution progressive optimization technology, an integrated workflow from NII segmentation to STL 3D reconstruction (including Marching Cubes core transformation and parameter optimization), and adaptive segmentation strategies for different lesion types and image modalities. These technical points constitute a complete intellectual property protection system from image input to 3D model output. Compared to the original nnU-Net single-stage segmentation method, this invention achieves progressive accuracy improvement through a four-stage progressive architecture, effectively solving the problems of "non-target region interference" and "low accuracy of multi-structure collaborative segmentation." Moreover, this application has been technically validated on publicly available CTA / OCT image datasets, achieving a Dice coefficient of 0.87 for vessel segmentation and a plaque detection sensitivity of 76%. Compared to the single-stage nnU-Net, the four-stage progressive workflow reduces the false positive rate by 30% and improves the recognition rate of small vessel branches by 35%. The generated STL 3D model exhibits good geometric accuracy, and the Marching Cubes algorithm reconstruction time is <50 seconds per case. Experimental results demonstrate that the technical approach of this method is feasible, can provide a reference for clinical diagnosis, and has the potential for further clinical application.

[0064] Based on the above technical solutions, this application provides a four-stage progressive medical image segmentation and STL three-dimensional reconstruction method for vascular plaques, including the following steps: First, the input medical image is initially segmented to locate the cardiac ROI; then, coarse segmentation of blood vessels is performed within the cardiac ROI to obtain coarse segmentation results; next, based on the coarse segmentation results, fine segmentation of blood vessels is performed to restore the fine structure of blood vessels and obtain fine blood vessel contour results; then, based on the fine blood vessel contour results, plaques on the blood vessel wall are segmented, and the location, boundary, shape, and attached blood vessel segment information of the plaques are clarified to obtain the final segmentation results; finally, the final segmentation results are converted to generate a three-dimensional model for clinical analysis.

[0065] This application proposes a four-stage progressive medical image segmentation technique based on the nnU-Net 3D full-resolution method. Through a hierarchical constraint process of "cardiac localization → coarse vessel segmentation → fine vessel segmentation → plaque segmentation," it overcomes the accuracy bottleneck of traditional single-stage segmentation. Compared to the original nnU-Net single-stage segmentation method, this application achieves progressive accuracy improvement through a four-stage progressive architecture, effectively solving the problems of "non-target region interference" and "low accuracy in multi-structure collaborative segmentation." The innovative dual-strategy optimization (CLRecall loss / connectivity enhancement) adapts to different clinical scenarios, embedding anatomical hierarchical relationships into the algorithm to significantly improve lesion detection rate, and realizing an integrated process from image segmentation to 3D reconstruction, improving segmentation accuracy by more than 10%.

[0066] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A four-stage progressive medical image segmentation and STL three-dimensional reconstruction method for vascular plaques, characterized in that, Includes the following steps: The input medical images are initially segmented to locate the cardiac region of interest (ROI). Coarse segmentation of blood vessels was performed within the cardiac region of interest (ROI) to obtain the coarse segmentation results. Based on the coarse segmentation results, fine segmentation of blood vessels is performed to restore the fine structure of blood vessels and obtain fine blood vessel contour results. Based on the detailed vascular contour results, plaques on the vascular wall are segmented, and the location, boundaries, shape, and attached vascular segment information of the plaques are determined to obtain the final segmentation result. The final segmentation result is then transformed to generate a three-dimensional model for clinical analysis.

2. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 1, characterized in that, The input medical images are initially segmented to locate the cardiac region of interest (ROI), including: The nnU-Net 3D full-resolution method is used to perform initial segmentation on the input medical images, defining the cardiac ROI for subsequent vascular and plaque analysis to eliminate interference from non-cardiac regions, reduce the computational load of subsequent segmentation tasks, and improve the focus of the target region.

3. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 1, characterized in that, Coarse segmentation of blood vessels was performed within the cardiac region of interest (ROI) to obtain the coarse segmentation results, including: Within the cardiac ROI, the aorta and coronary arteries are initially segmented using the nnU-Net 3D full-resolution method, yielding coarse segmentation results that provide a basic outline for subsequent fine segmentation. At the same time, differentiated optimization strategies are provided for different clinical needs.

4. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 3, characterized in that, The optimization strategy is an improvement on the nnU-Net 3D full-resolution method.

5. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 4, characterized in that, The operation process of the optimization strategy includes: The loss function layer of the nnU-Net 3D full-resolution method is replaced with the CLRecall loss function. The CLRecall loss function improves the initial accuracy of blood vessel contours by enhancing the recall rate of true positive blood vessel regions, thereby reducing false positive detections in blood vessel segmentation.

6. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 4, characterized in that, The operation process of the optimization strategy includes: The standard loss function of the nnU-Net 3D full-resolution method is retained to enhance the connectivity of vascular structures; the standard loss function includes the Dice loss function and the cross-entropy loss function; the advantages of this optimization strategy in modeling continuous spatial structures are utilized to ensure the contour continuity of the aortic trunk and major coronary branches.

7. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 1, characterized in that, Based on the coarse segmentation results, fine segmentation of blood vessels is performed to restore the fine structure of the blood vessels, resulting in a fine blood vessel contour, including: Using the coarse segmentation result as the initial mask, the coarse segmentation result is input into the high-resolution sub-network of the nnU-Net 3D full-resolution method to complete the fine segmentation of the aorta and coronary arteries, optimize the vessel wall thickness and the minimum diameter of the stenotic segment, restore the morphological changes of small branches with a diameter <2mm and the plaque attachment site, and provide an accurate vascular structure benchmark for subsequent plaque segmentation.

8. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 1, characterized in that, Based on the detailed vascular contour results, plaques on the vessel wall are segmented, and the location, boundaries, morphology, and attached vessel segment information of the plaques are determined to obtain the final segmentation result, including: Using the refined vascular contour results as a region mask, the segmentation function of the nnU-Net 3D full-resolution method is enabled in the region inside the vascular structure. At the same time, by inputting the image features of the plaque, the feature extraction layer of the network is fine-tuned so that the network prioritizes the identification of lesion areas inside the blood vessel, avoids misidentifying extravascular tissue as plaque, and achieves accurate separation of lesion areas inside the vascular structure.

9. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 1, characterized in that, The final segmentation result is transformed to generate a three-dimensional model for clinical analysis, including: The final segmentation results are processed sequentially through data preparation, core transformation, and post-processing output stages to generate a three-dimensional model for clinical analysis.

10. The four-stage progressive vascular plaque medical image segmentation and STL three-dimensional reconstruction method according to claim 9, characterized in that, The data preparation phase includes the following steps: starting with the input of the NII file, including operations for loading, binarization, and data validation. The core transformation phase includes: independently demonstrating oversampling, Gaussian smoothing, Marching Cubes algorithm, and coordinate transformation operations; The post-processing output stage includes: displaying Laplacian smoothing, mesh simplification, STL model export, and finally generating a 3D model.

Images