System and computer-readable medium for rib fracture detection and evaluation

Abstract

A system for rib fracture detection and assessment includes a data receiving module for receiving CT images of a patient's chest cavity obtained by a computed tomography scanner; a bone segmentation module for performing three-dimensional analysis of the CT images to generate an image mask; a three-dimensional volume rendering module for retaining voxels corresponding to the ribs as target rib voxels based on the image mask and performing volume rendering to generate a two-dimensional stereoscopic rendered image based on multiple preset angle combinations; a fracture detection module for detecting suspected rib fracture areas and outputting corresponding coordinates; and a result integration and display module for overlaying the detection results of suspected fracture areas onto the two-dimensional stereoscopic rendered image and outputting image data for clinical physician review.

Description

System for rib fracture detection and assessment System and Computer-readable medium for Rib Fracture Detection and Evaluation Technical Field

[0001] This invention relates to a system for the detection and assessment of rib fractures, particularly a system for the detection and assessment of rib fractures based on thoracic computed tomography (CT) images. Prior Technology

[0002] Rib fractures are a common diagnosis in chest trauma, and accurately assessing the number, location, and type of fractures is crucial for clinical treatment decisions (such as whether to perform surgical fixation). However, current diagnostic techniques have the following limitations: 1. Cognitive burden of interpreting 2D slides: The most common method involves interpreting a series of 2D computed tomography (CT) slides (axial, coronal, or sagittal planes). Physicians need to scroll back and forth across multiple slides and mentally integrate them to reconstruct a three-dimensional spatial perception. This is not only time-consuming but also prone to misinterpretation or omissions due to errors in spatial judgment. 2. Distortion of distorted views: Some techniques use unfolded views, such as "fishbone diagrams," to flatten the ribs. However, these images are geometrically distorted and do not conform to the actual human anatomy. Literature indicates that such software may produce artifacts due to its inability to accurately detect the rib centerline, and physicians need to take additional steps to map them back to the original 3D structure, resulting in insufficient intuitiveness and reliability. 3. Lack of intuitive surgical planning information: Pure text reports or 2D diagrams cannot allow surgeons to have a clear understanding of the entire fracture (such as the direction of displacement and the degree of comminutedness), which reduces the efficiency of preoperative planning.

[0003] Therefore, there is an urgent clinical need for a 3D visualization solution for fracture detection that can provide an intuitive, comprehensive, and accurate view. Summary of the Invention

[0004] In view of the problems in the prior art, this invention provides a system for rib fracture detection and assessment, comprising: a data receiving module for receiving CT images of a patient's chest cavity obtained by a computed tomography scanner; a bone segmentation module electrically connected to the data receiving module, and including a first deep learning model configured to perform three-dimensional analysis on the CT images to generate image masks corresponding to the ribs, clavicle, sternum, thoracic vertebrae, and scapula; and a three-dimensional volumetric rendering module electrically connected to the bone segmentation module, configured to retain voxels corresponding to the ribs as a set of target rib voxels based on the image masks, and to render the data from the CT images. The system removes voxels corresponding to soft tissue and non-target structures, and uses a volumetric rendering engine to perform volumetric rendering on the target rib voxels to generate multiple two-dimensional stereoscopic rendering images based on a set of preset angle combinations; a fracture detection module, electrically connected to the three-dimensional volumetric rendering module, includes a second deep learning model, which is configured to receive the two-dimensional stereoscopic rendering images as input, detect suspected rib fracture areas, and output corresponding coordinates; and a result integration and display module, electrically connected to the fracture detection module, is configured to overlay the detection results of the suspected fracture areas onto the two-dimensional stereoscopic rendering images and output image data for clinical physicians to review.

[0005] In some specific embodiments, the result integration and display module is configured to: map the coordinates back to the original three-dimensional coordinate system through the projection and transformation relationship used by the three-dimensional volume rendering engine to obtain the position of each suspected fracture area in three-dimensional space; use a clustering algorithm to perform spatial cluster analysis on the position to merge multiple detection results falling within the same spatial proximity range into a single fracture candidate and exclude detection results that do not meet the clustering conditions; and overlay the markers corresponding to the fracture candidate areas onto one or more of the two-dimensional stereoscopic rendering images to form an output image indicating the rib fracture candidate areas.

[0006] In some specific embodiments, the three-dimensional volume rendering module is further configured to calculate a midline section of the thoracic cavity using the image mask of the sternum and thoracic vertebrae, and to use the midline section to distinguish and render the left and right ribs separately to avoid visual overlap.

[0007] In some specific embodiments, the first deep learning model employs a three-dimensional fully convolutional neural network architecture, and the second deep learning model employs an object detection network architecture. Other objects and advantages of this invention are partly described in the following description, or may be understood through embodiments of this invention. It should be understood that the foregoing description and the following implementation methods are merely illustrative and explanatory, and not intended to limit the invention as the scope of the claims suggests. Simple Explanation of the Diagram

[0008] Figure 1 is a structural diagram of a system for rib fracture detection and assessment according to a specific embodiment of this invention. Implementation

[0009] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this work pertains.

[0010] As used herein, the singular forms “a,” “an,” and “the” include plural indicators unless the context clearly specifies otherwise. Thus, for example, reference to “a sample” includes a plurality of such samples and their equivalents known to those skilled in the art.

[0011] The “thoracic computed tomography images” used in this article refer to multiple fine-section axial images obtained after scanning the chest of an examinee with a computed tomography scanner (CT Scanner), which are usually transmitted and stored in the Digital Imaging and Communications in Medicine (DICOM) format.

[0012] The term "image mask" used in this article refers to the labeled data generated after semantic segmentation of an image. It is a two-dimensional or three-dimensional matrix corresponding to the size of the original image, used to indicate the anatomical structure classification (such as ribs, sternum, etc.) to which a specific voxel belongs.

[0013] The term "volume rendering" used in this article refers to a three-dimensional visualization technique that directly processes three-dimensional voxel data. By defining transformation functions, it maps voxel scalar values ​​(such as CT values) to colors and opacities, thereby generating two-dimensional projected images with a sense of depth.

[0014] The "transformation functions" used in this article include "opacity transformation functions" and "color transformation functions," which are used to determine the visual presentation characteristics of different density tissues in an image.

[0015] The "clustering algorithm" used in this paper refers to an unsupervised learning algorithm (e.g., Density-based spatial clustering of applications with noise, DBSCAN) used to group spatially adjacent point cloud data into the same set and to identify and remove outliers that do not belong to any set.

[0016] This invention provides a computer-based method, system, and computer-readable medium for rib fracture detection and assessment. Its core technology lies in using deep learning for bone segmentation, then converting the three-dimensional information into multi-angle two-dimensional stereoscopic images that facilitate AI detection and physician interpretation through specific volumetric rendering techniques, and combining this with spatial integration algorithms to improve diagnostic accuracy.

[0017] The methodology of this invention first includes a data receiving step, which involves receiving chest computed tomography (CT) images of the patient obtained by a computed tomography scanner. For example, the patient's chest CT images are received via a medical digital imaging protocol (e.g., DICOM).

[0018] The subsequent steps involve generating image masks and voxel selection. A first deep learning model is used to perform 3D analysis (voxel-level analysis) on the 3D CT images, identifying and generating image masks corresponding to structures such as the ribs, clavicle, sternum, thoracic vertebrae, and scapula. In the voxel selection step, based on the image masks, voxels are selected from the 3D CT image data to retain voxels corresponding to the ribs as a target rib voxel group, while voxels corresponding to soft tissue and non-target structures are removed. The purpose of this step is to separate the target bones from surrounding soft tissue, the heart, lungs, and potential external noise (such as physiological monitoring leads).

[0019] The next step involves generating a stereoscopic rendered image. A volumetric rendering engine is used to perform volumetric rendering on the target rib voxels. Specifically, this volumetric rendering is based on a set of preset combinations of multiple angles. For example, the angle combinations include three dimensions: eye level, a 20-degree downward angle, and a 20-degree upward angle. The image is then rotated horizontally around the center of the thorax, capturing one image every 10 degrees to generate a total of 108 two-dimensional stereoscopic rendered images. Based on the aforementioned image masking, the system retains only the target rib voxels, ensuring that the generated two-dimensional stereoscopic rendered image accurately represents the spatial features of the rib structure.

[0020] To enhance the features, the following steps can be performed: (1) Separation of left and right sides: Define a midline section of the thoracic cavity based on the relative position of the sternum and thoracic vertebrae to distinguish the left and right ribs and render them separately to avoid visual overlap of the left and right ribs during 3D projection; and / or (2) Visual enhancement: Mapping is performed through the "opacity conversion function" and the "color conversion function". The opacity function makes low-density tissue transparent and high-density bone tends to be opaque; the color function gives different density values ​​specific color contrasts to highlight the rib outline and fracture line.

[0021] After the image is generated, a fracture detection step is performed. A second deep learning model receives the two-dimensional stereoscopic rendered image and uses it as input to automatically detect the suspected rib fracture area and output the corresponding two-dimensional coordinates.

[0022] Finally, the process of overlaying and outputting images is performed. The detection results of the suspected fracture area are overlaid on the two-dimensional stereoscopic rendered image, and output image data is provided for clinical review. Since the same fracture may appear in rendered images from different angles, this method utilizes projection and transformation relationships to back-calculate the 2D coordinates into 3D space, and uses clustering algorithms for spatial grouping analysis. The technical effect is to merge multiple detection results of the same lesion, identify and remove outliers (false positives) that do not meet the clustering conditions, and finally overlay the accurate results onto the image and transmit them to the image server.

[0023] This invention also provides a system (10) for rib fracture detection and assessment, which includes multiple hardware modules that can be implemented on one or more computing devices: (1) Data receiving module (110): Connects to CT scanner or hospital database. (2) Skeletal segmentation module (120): Load the first deep learning model to perform three-dimensional image segmentation. (3) 3D volume rendering module (130): contains a rendering engine, responsible for performing voxel selection, midline segmentation and multi-angle projection. (4) Fracture detection module (140): Load the second deep learning model for lesion identification. (5) Results integration and display module (150): Responsible for coordinate transformation, spatial cluster analysis and final image overlay and transmission.

[0024] Referring to Figure 1, in one specific embodiment, the system 10 for rib fracture detection and assessment includes: a data receiving module 110; a bone segmentation module 120 electrically connected to the data receiving module 110; a three-dimensional volume rendering module 130 electrically connected to the bone segmentation module 120; a fracture detection module 140 electrically connected to the three-dimensional volume rendering module 130; and a result integration and display module 150 electrically connected to the fracture detection module 140.

[0025] The data receiving module 110 is used to receive chest computed tomography images of the subject obtained by a computed tomography scanner.

[0026] The skeletal segmentation module 120 includes or is equipped with a first deep learning model, which is configured to perform three-dimensional analysis on the thoracic computed tomography image data to generate image masks corresponding to the ribs, clavicle, sternum, thoracic vertebrae and scapula respectively.

[0027] The 3D volumetric rendering module 130 is configured to retain voxels corresponding to the ribs as a set of target rib voxels based on the image mask, and remove voxels corresponding to soft tissue and non-target structures from the thoracic computed tomography image data. Furthermore, the 3D volumetric rendering module 130 is configured to perform volumetric rendering on the target rib voxels using a volumetric rendering engine to generate multiple 2D stereoscopic rendering images based on a set of preset angle combinations.

[0028] The fracture detection module 140 includes or is equipped with a second deep learning model, which is configured to receive the two-dimensional stereo rendering image as input, detect suspected rib fracture areas in the image, and output the corresponding coordinates.

[0029] The result integration and display module 150 is configured to overlay the detection results of the suspected fracture area onto the two-dimensional stereoscopic rendering image and output image data for clinicians to review.

[0030] The result integration and display module 150 can be further configured to: map the coordinates back to the original three-dimensional coordinate system through the projection and transformation relationships used by the three-dimensional volume rendering engine to obtain the position of each suspected fracture area in three-dimensional space; use a clustering algorithm to perform spatial cluster analysis on the positions to group multiple detection results falling within the same spatial proximity range into a single fracture candidate and exclude detection results that do not meet the clustering conditions; and overlay the markers corresponding to the fracture candidate areas onto one or more of the two-dimensional stereoscopic rendering images to form an output image indicating the rib fracture candidate areas. The output image can be displayed on a display device and electrically or signalally connected to the system 10 used for rib fracture detection and assessment.

[0031] In addition, the 3D volume rendering module can be further configured to: calculate a midline section of the thoracic cavity using the image mask of the sternum and thoracic vertebrae, and use the midline section to distinguish and render the left and right ribs separately to avoid visual overlap.

[0032] According to this invention, the first deep learning model preferably adopts a three-dimensional fully convolutional neural network architecture, and the second deep learning model preferably adopts an object detection network architecture.

[0033] This invention also provides a computer-readable medium that stores a computer program containing multiple instruction codes. When loaded and executed by a computer, the program enables the computer to perform the aforementioned detection and evaluation methods. Specific implementations of this medium may include: non-transitory recording media, such as hard disk drives (HDDs), solid-state drives (SSDs), read-only optical discs (CD-ROMs / DVD-ROMs), flash memory, or USB flash drives; network storage media, such as cloud server storage space, where the computer can download and execute the program via the network; or random access memory (RAM), for example, where the program is temporarily stored after loading for execution by the central processing unit (CPU) or graphics processing unit (GPU).

[0034] [Example]

[0035] Example 1: Rib fracture detection and assessment based on thoracic computed tomography (CT) images

[0036] 1.1 Hardware Environment

[0037] This example uses a single-user computer with an Ubuntu Linux operating system, an AMD 5800x CPU, 64GB of RAM, and a graphics accelerator (such as an NVIDIA 3090 GPU) with at least 12GB of video memory (VRAM).

[0038] 1.2 First Stage: Thoracic Skeletal Segmentation

[0039] The system receives raw, fine-section axial images of the thoracic cavity obtained from a computed tomography (CT) scan and transmits them to the host system via a DICOM interface. The host system uses a first deep learning model to analyze the image data, generating multiple image masks corresponding to the thoracic skeleton structures such as the ribs, clavicle, scapula, sternum, and thoracic vertebrae. The host system associates and stores the raw, fine-section axial images with the multiple image masks, using them as input for subsequent processing stages. The first deep learning model employs the nnU-Net architecture and is trained using at least a number of (e.g., 100) training images of thoracic skeleton regions annotated by a radiologist, enabling the model to perform three-dimensional voxel-level segmentation of the thoracic skeleton.

[0040] The goals of this stage are: (1) to automatically and precisely segment the ribs and other thoracic skeletons in 3D CT images, reducing the workload of manually drawing the skeleton outline layer by layer, and improving the consistency and reproducibility of skeleton segmentation; (2) to separate the thoracic skeleton from soft tissue, metal wires and other non-target structures, providing a clean and standardized skeleton area for subsequent 3D volume rendering and fracture detection, reducing noise and occlusion; and to maintain stable segmentation performance under different subjects and different scanning conditions by using an nnU-Net model trained with labeled data, which is helpful for the practical application of the overall system in the clinical environment.

[0041] 1.3 Second Stage: Target Voxel Selection and 3D Volumetric Rendering

[0042] Using the image mask coordinates of the ribs, clavicle, sternum, thoracic vertebrae and scapula obtained in the first stage, voxel screening is performed on the original fine-cut CT images: (1) Remove all non-rib pixels outside the target rib, including soft tissue, metal wires and other skeletal structures, so that the input only contains rib pixels; (2) Calculate the midline section of the thoracic cavity based on the image mask of the sternum and thoracic vertebrae, use the midline section to distinguish the left and right ribs, and remove the left rib pixels when generating the right rib image, and vice versa.

[0043] For the pixels of a single rib, a volumetric rendering engine (e.g., a 3D volumetric rendering module based on VTK (The Visualization Toolkit)) is used to generate multiple 2D stereoscopic rendering images based on a combination of preset rotation and pitch angles. Multiple images are generated for the right and left ribs respectively (e.g., 108 images each). Furthermore, using the mask coordinates of the ribs, clavicle, sternum, and scapula, non-skeletal pixels such as soft tissue and metallic wires are removed. 3D volumetric rendering is then performed on the pixels of the entire thoracic skeleton, and multiple stereoscopic rendering images of the entire thoracic skeleton are generated based on multiple horizontal rotation angles (e.g., 36 images).

[0044] During the volume rendering process, a piecewise linear transformation function is defined based on multiple preset scalar values ​​and corresponding opacity / color control points to map voxel scalar values ​​to opacity and RGB color values, making low-density tissues tend to be transparent and high-density bone gradually appear bright white, so as to emphasize the rib contours and subtle fracture lines. The parameters used are as follows.

[0045] 1.3.1 Opacity Conversion Function

[0046] For a scalar value (e.g., CT value) of the input voxel, a piecewise linear opacity transition function is defined, such that the opacity monotonically increases with the scalar value. This opacity transition function can be determined by several control points, each containing a scalar value and a corresponding opacity value, where: when the scalar value is 60, the opacity is 0; when the scalar value is 80, the opacity is approximately 0.1; when the scalar value is 150, the opacity is approximately 0.7; and when the scalar value is 200, the opacity is 1.

[0047] Between any two control points, the opacity of the intermediate scalar value is calculated by linear interpolation, making lower density tissue appear transparent or translucent, while high density bone structures gradually appear opaque, in order to emphasize the rib contour and fracture-related structures.

[0048] 1.3.2 Color Conversion Function (color)

[0049] Furthermore, a color conversion function is defined for the same range of scalar values ​​to map scalar values ​​to RGB colors, thereby enhancing the visual distinction between tissues of different densities. This color conversion function is also determined by multiple control points, each containing a scalar value and a corresponding RGB color value (each component ranging from 0 to 1). For example, when the scalar value is 0, the color is (0, 0, 0), i.e., black.

[0050] When the scalar value is 50, the color is approximately (1, 0.3, 0), used to indicate areas with lower density but close to the soft tissue boundary; when the scalar value is 130, the color is approximately (1, 0.75, 0.25), appearing as a brighter orange-yellow to highlight medium- to high-density structures; when the scalar value is 200, the color is approximately (1, 1, 0.9), which is close to white, used to represent the highest density bone.

[0051] Between adjacent control points, the RGB color of the intermediate scalar value is calculated using linear interpolation, so that the bone gradually transitions from a darker color to a bright white, improving the visibility of the rib outline and fine fracture lines.

[0052] Through this stage, the following technical effects can be achieved: (1) The three-dimensional curved rib structure is converted into a two-dimensional stereoscopic rendering image with multiple angles and significantly suppressed background interference, reducing the overlap and occlusion of ribs with soft tissues and non-target bones, and improving the visibility of fracture-related morphological features; (2) The left and right ribs are separated and rendered separately using the midline section of the thoracic cavity, avoiding visual overlap between the left and right ribs, so that each rib is clearly presented from different angles, which is beneficial for subsequent fracture detection or manual interpretation; (3) The original three-dimensional CT volume data is converted into hundreds of standardized two-dimensional rendering images, reducing the input dimension and computational cost of the deep learning model, while maintaining spatial and density information; and (4) The voxel scalar values ​​are mapped to opacity and RGB through a piecewise linear transformation function. Color makes low-density tissues appear transparent, while high-density bone gradually appears bright white, enhancing the contrast and visibility of rib contours and subtle fracture lines, and improving the recognition and display consistency of bone structure from different perspectives; the three-dimensional curved distribution of rib structure is converted into a set of multi-angle, two-dimensional stereoscopic rendering images with significantly suppressed background interference, reducing the overlap and occlusion of ribs with soft tissues and non-target bones, and improving the visibility of fracture-related morphological features.

[0053] 1.4 Phase 3: AI Fracture Detection and Integration of 3D Results

[0054] The multiple stereoscopic rendered images (e.g., 252 images) generated in the second stage are input one by one into a second deep learning model. This model employs an object detection architecture (training method described below) to detect and output multiple bounding boxes and their coordinates corresponding to the suspected rib fracture areas in each image. The center coordinates of each bounding box are then mapped back to the original 3D coordinate system using the projection and transformation relationships used by the second-stage 3D volumetric rendering engine to obtain the position of each suspected fracture area in 3D space.

[0055] Clustering algorithms (e.g., DBSCAN) are used to perform spatial clustering analysis on the coordinates of all suspected fracture centers. Multiple detection results falling within the same spatial proximity are grouped into a single fracture candidate, while isolated outlier detection results that do not meet the clustering criteria are excluded to reduce false positives. The bounding boxes corresponding to the clustered and filtered fracture candidate regions are then overlaid on the corresponding stereo rendering images generated in the second stage to form output images indicating the rib fracture candidate regions. Finally, the output images are transmitted to the hospital's Picture Archiving and Communication System (PACS) via the DICOM protocol for physicians to review in their existing imaging workstation environments.

[0056] The second deep learning model uses the RF-DETR architecture (Roboflow) and is trained using training images of at least 250 cases and 63,000 images with at least 77,887 fracture areas labeled by radiologists, so that the model can detect fracture areas from stereoscopic rendered images from different angles.

[0057] Through this stage, the following technical effects can be achieved: (1) By using a second deep learning model to detect rib fracture candidate areas in multi-angle stereo rendering images, and integrating the detection results from each perspective into the same three-dimensional coordinate system, the system can simultaneously mark multiple fracture candidate locations in a single global 3D view, helping physicians to quickly grasp the number, distribution, and laterality of fractures, reducing the time spent repeatedly tracking and counting between multiple 2D slices; (2) Since the fracture candidate areas can be presented on an interactively rotatable 3D thoracic view, physicians (especially surgeons) can intuitively assess the specific morphology of each fracture from different perspectives, including the degree of comminutedness, the direction and angle of fracture displacement, the relationship between adjacent ribs and thoracic curvature, etc., providing key visual information for subsequent surgical fixation, incision design, and selection of fixation equipment; (3) Through clustering and outlier exclusion mechanisms, the system can merge duplicate detection results from multiple perspectives into a single fracture candidate, while reducing the impact of isolated false positives, so that the final 3D The fracture markers presented in the view are closer to the actual lesion distribution, which helps radiologists to quickly confirm and supplement the diagnosis, and provides a stable and repeatable visual basis for surgical planning; and (4) by outputting the images of the fracture candidate areas to PACS in DICOM format, the AI ​​detection results generated by this creation can be directly embedded into the existing radiology and surgery workflows, so that "AI detection", "radiology diagnosis" and "surgical planning" can be seamlessly connected in the same imaging system, improving the efficiency of clinical adoption and cross-disciplinary collaboration.

[0058] 10: Systems for rib fracture detection and assessment 110: Data receiving module 120: Skeletal Segmentation Module 130: 3D Volumetric Rendering Module 140: Fracture Detection Module 150: Results Integration and Display Module

Claims

1. A system for rib fracture detection and assessment, comprising: a data receiving module for receiving CT scan images of a subject's chest cavity obtained by a computed tomography scanner; a bone segmentation module electrically connected to the data receiving module, and including a first deep learning model configured to perform three-dimensional analysis on the CT scan images to generate image masks corresponding to the ribs, clavicle, sternum, thoracic vertebrae, and scapula; and a three-dimensional volumetric rendering module electrically connected to the bone segmentation module, configured to retain voxels corresponding to the ribs as a set of target rib voxels based on the image masks, remove voxels corresponding to soft tissue and non-target structures from the CT scan images, and perform volumetric rendering on the target rib voxels using a volumetric rendering engine to generate multiple two-dimensional stereoscopic rendering images based on a set of preset angle combinations. A fracture detection module, electrically connected to the 3D volumetric rendering module, includes a second deep learning model configured to receive the 2D stereoscopic rendering image as input, detect suspected rib fracture areas, and output corresponding coordinates; and a result integration and display module, electrically connected to the fracture detection module, configured to overlay the detection results of the suspected fracture areas onto the 2D stereoscopic rendering image, and output image data for clinical physician review.

2. The system as described in claim 1, wherein the result integration and display module is configured to: map the coordinates back to the original three-dimensional coordinate system through the projection and transformation relationships used by the three-dimensional volume rendering engine to obtain the position of each suspected fracture region in three-dimensional space; perform spatial clustering analysis on the position using a clustering algorithm to group multiple detection results falling within the same spatial proximity range into a single fracture candidate and exclude detection results that do not meet the clustering conditions; and overlay the markers corresponding to the fracture candidate regions onto one or more of the two-dimensional stereoscopic rendering images to form an output image indicating the rib fracture candidate regions.

3. The system as described in claim 1, wherein the three-dimensional volume rendering module is further configured to calculate a midline section of the thoracic cavity using image masks of the sternum and thoracic vertebrae, and to distinguish and render the left and right ribs separately using the midline section to avoid visual overlap.

4. The system as described in claim 1, wherein the first deep learning model employs a three-dimensional fully convolutional neural network architecture, and the second deep learning model employs an object detection network architecture.

Images