Measuring semiconductor devices in electron micrographs using fast marching level sets

Abstract

Measuring semiconductor devices in electron micrographs using fast marching level sets. Disclosed herein are apparatuses and methods for measuring devices using fast marching level sets. An example method includes at least creating a fast marching level set seed on an image, propagating a fast marching level set curve from the fast marching level set seed to locate boundaries of a plurality of regions of interest within the image, and performing a measurement on the regions of interest based in part on the boundaries.

Description

Technical Field

[0001] This invention generally relates to image processing and automatic measurement, and more particularly to deep learning-based image processing and measurement of charged particle-based micrographs using fast-moving level sets. Background Technology

[0002] Processing images obtained from electron microscopy is challenging due to their typical image quality (potentially poor contrast and signal-to-noise ratio). For example, such poor image quality can make structural analysis (e.g., measuring structures within an image) difficult due to the lack of defined structural boundaries. While various image processing techniques can overcome some of these difficulties, these techniques typically require manual control by highly skilled technicians. In most use cases, automated and efficient analysis is desired. Summary of the Invention

[0003] This document discloses an apparatus and method for measuring a device using a fast-moving level set. One example method includes at least: creating a fast-moving level set seed on an image; propagating fast-moving level set curves from the seed to locate boundaries of a plurality of regions of interest within the image; and performing a measurement on the regions of interest, in part based on the boundaries.

[0004] An example method includes at least electron microscopy to acquire an image of a sample comprising multiple regions of interest, coupled to or included in the electron microscope, and a kernel. The kernel includes or is coupled to code that, when executed by the kernel, causes the kernel to create a fast-moving level set seed within the image, propagate fast-moving level set curves from the fast-moving level set seed to locate the boundaries of each of the multiple regions of interest within the image, and perform a metric on the regions of interest in part based on the boundaries. Attached Figure Description

[0005] Figure 1 This is an example of a charged particle microscope system according to embodiments of the present disclosure.

[0006] Figure 2 This is a representative system for segmenting and measuring ROIs (Regions of Interest) in images and other applications according to embodiments of this disclosure.

[0007] Figure 3 This is an example method according to embodiments of the present disclosure for segmenting an image and performing a metric on one or more ROIs therein.

[0008] Figure 4 This is a sequence of example images illustrating the initialization and propagation of a fast-moving level set contour according to embodiments of the present disclosure.

[0009] Figure 5 This is another example image sequence illustrating a high-level diagram of an image being processed using the disclosed method according to embodiments of this disclosure.

[0010] Figure 6 This is another example image sequence illustrating the optimization of a fast-moving level set around multiple ROIs according to embodiments of the present disclosure.

[0011] Figure 7 Two images showing multiple layers with positioning using fast-moving level sets according to an embodiment of the present disclosure are illustrated.

[0012] Figure 8 This includes example images showing different shapes on non-circular ROIs that implement the techniques disclosed herein.

[0013] Figure 9A It is an instance deep learning filter for preprocessing images according to the disclosed embodiments.

[0014] Figure 9B The diagram illustrates an example of reconstructing the output image of an input / output image pair used for deep model training.

[0015] Figure 10 This is an example functional block diagram based on an embodiment of the present disclosure.

[0016] In the various views of the accompanying drawings, the same reference numerals denote corresponding parts. Detailed Implementation

[0017] Embodiments of the invention are described below in the context of electron microscopy, using a hybrid approach of conventional and deep learning-based algorithms and models to perform segmentation and measurement techniques. For example, an electron image can be smoothed and filtered using a trained model, which then incorporates one or more fast marching level set algorithms to find one or more boundaries within the image. The delineated boundaries are then used to anchor measurement algorithms to measure various aspects of the region of interest (ROI) in the image. The various embodiments disclosed herein provide examples of implementing the disclosed techniques and should not be construed as limiting the scope of the disclosed techniques.

[0018] As used in this application and claims, the singular forms “a” (a, an) and “the” include the plural forms unless the context clearly specifies otherwise. Additionally, the term “comprising” means “including”. Furthermore, the term “coupled” does not exclude the existence of intermediate elements between coupled items.

[0019] In some instances, values, processes, or devices are referred to as “lowest,” “best,” “smallest,” etc. It should be understood that such descriptions are intended to indicate that a choice can be made among many functional alternatives used, and that such a choice is not necessarily better, smaller, or otherwise superior to other choices.

[0020] The systems, apparatuses, and methods described herein should not be construed as limiting in any way. In fact, this disclosure applies to all novel and non-obvious features and aspects of the various disclosed embodiments, whether individually or in various combinations and sub-combinations formed with each other. The disclosed systems, methods, and apparatuses are not limited to any particular aspect or feature or combination thereof, nor are they required to have any one or more particular advantages or solve any one or more particular problems. Any operational theory is provided for ease of interpretation, but the systems, methods, and apparatuses of this disclosure are not limited to such operational theories.

[0021] Although some of the methods disclosed are described in a specific sequential order for ease of presentation, it should be understood that this descriptive approach encompasses rearrangement unless specific language requirements dictate a particular order. For example, in some cases, the operations described sequentially may be rearranged or performed simultaneously. Furthermore, for simplicity, the accompanying drawings may not show the various ways in which the disclosed systems, methods, and apparatus can be used in conjunction with other systems, methods, and apparatuses. Additionally, this specification sometimes uses terms such as “produce” and “provide” to describe the disclosed methods. These terms are high-level abstractions of the actual operations performed. The actual operations corresponding to these terms will vary depending on the specific implementation and can be readily identified by one of ordinary skill in the art.

[0022] Precise segmentation and measurement of regions of interest (ROIs) in electron microscope micrographs can be difficult due to undesirable image characteristics such as low contrast and low signal-to-noise ratio (SNR). Furthermore, the resolution at which the image is acquired can also affect image quality. For example, the segmentation of 3D-NAND memory structures and the material interfaces associated with these structures is fundamental to reliable measurement in 3D-NAND memory manufacturing. To address the measurement challenges of 3D-NAND, device manufacturers are experimenting with various techniques. However, currently, there is no fully automated solution for measuring critical dimensions (CDs) in 3D-NAND data collected using electron microscopy. Traditionally, recipe builders included in a given electron microscope (whether scanning electron, transmission electron, or scanning transmission electron) are typically used to establish a measurement process to identify possible edges within the ROI. However, in most cases, due to the manual nature of the work, only a small area of ​​the ROI is used. Establishing a recipe for finding edges (e.g., pixels associated with the edges) assumes good contrast between different materials, consistency of image quality within and across the image, and the user's ability to place the edge finder appropriately. Fitting the geometry assumes that each material layer has been rigorously adhered to the geometry during manufacturing and that there is sufficient contrast between the materials. In both cases, human interaction is necessary. These methods and their associated measurement limitations reduce the effectiveness of accurate automated detection of multiple CDs, which is essential for automated monitoring of process variations. It should be noted that while much of the discussion and illustration in this paper uses 3D-NAND as an example, the techniques disclosed herein are not limited to this, and any object or ROI within an image (even an optical image) can be analyzed using the disclosed techniques.

[0023] One solution to the aforementioned segmentation and measurement problem is to use Fast Moving Level Sets (FMLS), which can optionally be aided by machine learning-based image preprocessing, to efficiently delineate boundaries within a Region of Interest (ROI). For example, the disclosed techniques can be used to identify and extensively measure various boundaries within 3D-NAND vias. Optional preprocessing can use deep learning algorithms to at least enhance the image's SNR and contrast, resulting in a smoother image. With or without preprocessing, FMLS is initialized and allows optimization (e.g., localization or identification) at or around boundaries within or around the ROI. In some cases, a large number of boundaries can be localized within a single image using only a single initialized FMLS curve or point. Furthermore, this technique works in conjunction with single or multiple initializations propagating towards region / object boundaries, independent of topology merging and segmentation. In some cases, additional evolution of the FMLS curves at the boundaries is allowed to further optimize the curves to the boundary locations. Once the boundaries are identified and the FMLS curves are optimized, deep learning-based metrics can be performed to measure various aspects of the ROI, such as area, shape factor, eccentricity, realism, convexity, best-fit ellipse features, total area within the closed boundary, features of pixel grayscale clusters within the closed object region, boundary length, etc. Each stage of boundary delineation and measurement can be performed automatically without user assistance.

[0024] The disclosed solution provides an automated technique for descriptive metrics of ROIs, such as 3D-NAND memory cells. This technique overcomes some limitations of edge finder-based methods and can work on low-contrast images, depicting the interfaces between different material layers constituting the desired ROI. Furthermore, the technique allows for descriptive measurements of the shape characteristics of each material layer, the spatial distribution of material thickness, etc., which is a significant advantage for the statistical analysis of structural variations. The technique can also enable additional unsupervised machine learning models to pick out outliers, implicit biases, and manufacturing constraints during the manufacturing process.

[0025] Figure 1This is an example of a charged particle microscopy (CPM) system 100 according to embodiments of the present disclosure. The charged particle microscopy (CPM) system 100, or simply system 100, includes at least a CPM environment 102, a network 104, one or more servers 106, and an artificial neural network 114. The CPM system 100 can be used to study and analyze samples of various sizes and compositions. In one example, the CPM system 100 may be implemented at least partially at an integrated circuit manufacturing site and used to analyze and measure various aspects of wafers and circuits manufactured at the site. In some embodiments, the CPM system 100 may be distributed across various locations. For example, the CPM environment 102 may be located at a manufacturing or development site, the network 104 may be distributed locally, regionally, or nationally, and the server 106 may be located at a server farm and coupled to the CPM environment 100 via the network 104. Regardless of the organization of the CPM system 100, the system 100 may be used at least to implement one or more deep learning modules 114 and one or more analysis algorithms to perform various other image processing tasks, such as image processing algorithm 116.

[0026] CPM environment 102 includes any type of charged particle microscope, but the applications of neural networks and analyses disclosed herein are not limited to charged particle microscopes and are for illustrative purposes only. Examples of CPMs include scanning electron microscopes (SEM), transmission electron microscopes (TEM), scanning transmission electron microscopes (STEM), focused ion beam (FIB), and dual-beam (DB) systems that include both electron and ion beam capabilities, to name just a few. CPM environment 102 can be used to obtain electron or ion images of samples, some of which may be thin slices, such as thin layers, taken from larger samples or wafers. CPM environment 102 may include aspects that can be contained in a single tool or located within a separate tool. For example, CPM environment 102 may include an imaging platform 108 (e.g., SEM, TEM, or STEM), a sample preparation platform 110, and one or more controllers 112. Of course, each platform 108 and 110 may also include more than one microscope / sample preparation tool.

[0027] Imaging platform 108 is used to acquire images of the sample, some of which may have been prepared by sample preparation platform 110, but this is not necessary. Images are acquired by irradiating the sample with their respective charged particle beams using electron and / or ion sources. In some instances, charged particle beam imaging is obtained by scanning the beam, for example, by moving it over the sample, while in other instances, the charged particle beam is not scanned. Electrons, such as those backscattered, secondary, or transmitted, are then detected, and a grayscale image is formed based on them. The image includes grayscale contrast that depends on the sample material, where variations in grayscale levels indicate changes in material type or crystal orientation. Imaging platform 108 may be controlled by internal controls (not shown), controller 112, or a combination thereof.

[0028] Sample preparation platform 110 forms some of the samples imaged by imaging platform 108. Of course, imaging samples can also be formed using other tools (not shown). For example, sample preparation 110 can be a DB system that uses FIB to prepare and assist in removing thin samples from larger samples, such as by ion milling, ion-induced etching, or a combination thereof, and other techniques to process the sample for imaging. Other techniques may include, but are not limited to, planarization milling / etching, reference production, cross-section formation, top-down sheet preparation, flattening, etc. Sample preparation platform 110 may also include an electronic imaging component that allows monitoring of the sample preparation process, but is not required. In some embodiments, sample preparation platform 110 may include other physical preparation aspects (lasers, cutting tools, resin encapsulation tools, cryogenic tools, etc.). Other physical preparation aspects for preparing samples for imaging platform 108 include lasers, cutting tools, resin encapsulation tools, cryogenic tools, etc. Sample preparation platform 110 may be controlled by internal controls (not shown), controller 112, or a combination thereof.

[0029] Network 104 can be any kind of network used to transmit signals between CPM environment 102 and one or more servers 106. For example, network 104 can be a local area network, a wide area network, or a distributed network, such as the Internet, a telephone backbone, or a combination thereof.

[0030] Server 106 may include one or more virtual and / or physical computing platforms capable of running code for various algorithms, deep learning / neural networks, and analytics suites. Although not shown, users of CPM environment 102 can access server 106 to retrieve data, update software code, perform data analysis tasks, etc., with access occurring via network 104 from the user's local computing environment (not shown). In some embodiments, users access and / or retrieve image data stored on server 106, perform segmentation and / or measurement using deep learning module 114 and image processing algorithm 116 (which may execute on server 106 or CPM environment 102), and perform segmentation and / or measurement at their local computing environment.

[0031] In operation, one or more images of the sample are acquired by imaging platform 108. At least one of the images, including, for example, one or more structures of interest, can be segmented, for example, by a combination of modules 114 and 116, to locate boundaries. The segmented images can then be used to locate / identify desired features of each or at least one imaging structure in the one or more acquired images. Initially, the images can be preprocessed, for example, by deep learning module 114, to improve the intensity gradients of the images, which will aid in segmentation. In this way, the images are preprocessed to at least smooth them, thereby reducing noise and enhancing object boundaries. In some embodiments, image preprocessing is optional and unnecessary for implementing further processing, such as seeding and evolution of FMLS. Using the smoothed images, FMLS is performed by image processing algorithm module 116, which seeds initial points or curves on the smoothed images and propagates the seeds to locate one or more boundaries. Propagation uses strong gradient magnitudes of intensity within the image to move and is slowed down or stopped at locations with low gradients. These locations are typically where there are boundaries between different layers in the image.

[0032] In some embodiments, additional optimizations to the propagation / evolution curve can be performed to improve the accuracy of the evolution profile. In such embodiments, level set algorithms or active profile (e.g., snake-like) algorithms can be applied. This optimization allows for more accurate measurements to be performed subsequently.

[0033] For metrics, deep learning module 114 is used to measure various aspects of the now segmented image. In some embodiments, unsupervised machine learning processes are used to perform the metrics. Any number of machine learning techniques, such as Bayesian learning, classification trees, random forests, neural networks, K-nearest neighbors (KNN), etc., can be used to classify and measure the segmented features in the image.

[0034] Although the images provided to the deep learning module 114 are described as being obtained by the imaging platform 108, in other embodiments, the images may be provided by different imaging platforms and provided to the ANN 114 via the network 104.

[0035] Figure 2This is a representative system 200 for segmenting and measuring Regions of Interest (ROIs) in an image according to embodiments of this disclosure. System 200 includes at least an imaging system 202 (such as a SEM, STEM, or TEM) that provides images of one or more ROIs to an image processor 204 for segmentation and measurement. TEMs or other electron microscopes typically include an electron beam source, electron optics, and an electronic detection system that captures or generates electron beam-based images. As mentioned above, segmentation of images obtained with an electron beam system such as system 202 can be challenging because the quality of the images produced by imaging system 202 is typically limited. Image processor 204 can be implemented in a processor associated with imaging system 202, either with a separate local processing unit as discussed, or with a remote processor coupled to imaging system 202 via a local area network (LAN) or wide area network (WAN) 206. Figure 2 As shown, a local deep learning (DL) library 212 is stored in a computer-readable medium coupled to an image processor 204 and contains processor-executable instructions and storage defining values ​​for one or more deep learning modules. Additionally, computer-readable media 210 and 214 are provided for storing supplemental training data for one or more deep learning modules stored respectively in the DL library 212 and the image processing algorithm 214. A single memory or storage device may be used for all of these, and some or all may be remotely stored and available via network 206. For example, a remote DL library 208 and a remote image processor 220 communicate with the image processor 204 via network 206. The image processor 204 may also be remotely located relative to the imaging system 202. In some instances, remote DL library specifications are transferred to local storage such as the local DL library 212.

[0036] In operation, system 200 acquires an image using imaging system 202 and provides the image to image processor 204 or 220. The image may contain one or more object features to be measured. Furthermore, image processor 204 (or image processor 220) can access deep learning-based image filters from a remote DL library 208 or a local DL library 212 for image preprocessing. For example, image processor 208 can access a deep learning-based nonlinear reaction diffusion filter to reduce any noise in the image and enhance contrast. The noise- and contrast-enhanced image provides a suitable image for the seed evolution of FMLS. This preprocessed image provides sharp intensity gradients and a smooth, uniform background between different material layers and can also be used as the output image for input / output image pairs used for training deep neural networks. In other embodiments, image processor 204 or 220 can access conventional reaction diffusion filters from image processing algorithm 214 to reduce noise and enhance object features in the image. While conventional filters or machine learning filters can be used, machine learning filters offer faster preprocessing (without calculating any training time). However, it should be noted that image preprocessing is optional and can only be implemented when the image quality makes the boundaries between different layers in the image not sufficiently obvious for FMLS evolution.

[0037] Once the image has been preprocessed, it can be segmented to delineate regions of object features for measurement purposes. In some embodiments, the segmentation process can be performed in two steps, for example, including creating FMLS to propagate an initial seed to label the actual shape of the object features before creating a classic level set method to improve segmentation accuracy. In other embodiments, only one step may be used, which is to create FMLS to propagate the initial seed to label the actual shape of the object features. Regardless of the implementation, the FMLS curve propagates along the normal direction at each point in the curve, see [link to relevant documentation]. Figure 4 As an example, in some embodiments, the image intensity gradient function can be used as a velocity function, which provides a simple behavior of slowing down the movement of the curve when it approaches a high intensity gradient (e.g., a high force) and accelerating the propagation of the curve in regions of low intensity gradient (e.g., regions of low force or uniformity).

[0038] As described above, to initiate a rapid traversal process, an initial seed or closed curve is provided on the image. In some embodiments, image boundary pixels are used to create a curve that wraps around the entire image, which can then propagate inward in the normal direction. The evolution of the FMLS curve can terminate when the change in curve shape is negligible over consecutive time steps. Such negligible changes can indicate that no further evolution is possible given a velocity function. See also Figure 4As an example, if a two-step process is implemented, the evolution curve can be viewed as an initial curve for evolution using a classical level set.

[0039] Once the curve has fully evolved through a one- or two-step process, it becomes the basis for measurement. However, prior to measurement, machine learning processes can be used to classify and eliminate artifacts in the image. For example, for a circular object of interest, a linear shape feature classifier based on support vector machines can be used to classify the segmented object into a 3D-NAND class and an artifact class. The segmented image is an image with an evolved FMLS curve that at least locates the outer boundary of the object in the image. For simplicity, the object used for classification can be the same object required for the measurement used in the metric analysis. The metric can determine many features of the object in the image, such as shape factor, eccentricity, solidity, convexity, best-fit ellipse features, total area within the closed boundary, features of pixel gray-level clusters within the closed object region, boundary length, etc. The only unconventional feature measured here is the gray-level cluster characteristics of the pixels within the closed boundary of the object.

[0040] Metrics can be performed using machine learning-based modules stored in the remote DL library 208 and / or the local DL library 212. System 200 can implement one of any number of machine learning techniques that can be used, such as Bayesian learning, classification trees, random forests, neural networks, K-nearest neighbors (KNN), etc., to classify and measure segmented features in an image. Once metrics have been performed, one or more objects (e.g., ROIs) in an image can be characterized by several measurements. Such measurements include one or more of the following: area, shape factor, eccentricity, realism, convexity, best-fit ellipse features, total area within a closed boundary, features of pixel gray-level clusters within a closed object region, boundary length, etc.

[0041] Figure 3 This is an example method 301 for segmenting an image and performing measurements on one or more ROIs therein, according to embodiments of this disclosure. For example, method 301 can be performed on system 100 or 200. Typically, method 301 is performed using a hybrid of deep learning-based modules and standard image processing algorithms to efficiently and accurately delineate boundaries and provide measurements of features contained within the delineated boundaries.

[0042] Method 301 may begin with an optional processing block 303, which includes preprocessing the image to at least reduce noise and enhance boundaries. For preprocessing, a deep learning-based trainable nonlinear reaction diffusion filter or other types of trainable filters may be used to enhance the image. The preprocessed image then provides a smooth, noise-free surface for seed / curve evolution used in FMLS.

[0043] If processing block 303 is implemented, it can be followed by processing block 305, which includes a velocity function used to construct the image. In other embodiments, method 301 may begin with processing block 305. The velocity function can be calculated as follows: F(x)=e^(-k|∇I(x)| ) (1) Where I(x) is the image intensity at point x, ∇I(x) is the image intensity gradient, and k is the curvature value at point x. This velocity function provides a simple behavior to slow down the movement of the seeded / initialized curve when it approaches a high image intensity gradient and to accelerate curve propagation in uniform regions (low intensity gradient regions). The velocity function image can then be used to seed an FMLS profile and evolve that profile to place / locate / delineate boundaries within the image.

[0044] Processing block 305 can be followed by processing block 307, which includes creating a fast-moving level set seed on the velocity function graph. The seed can be a closed contour around the image edges, such as at peripheral pixels of the image, or as one or more points within the image.

[0045] Processing block 307 can be followed by processing block 309, which includes propagating a fast-moving level set seed to determine boundaries within the velocity function image. When the seed is a closed contour, the propagation (e.g., evolution) of the seed propagates inward in the normal direction of the seed contour. The evolution of the curve can be terminated when the change in curve shape is negligible over consecutive time steps, meaning that no further evolution is possible under a given velocity function; see [link to relevant documentation]. Figure 4 For use in instances.

[0046] Processing block 309 can be followed by processing block 311, which includes optimizing the determined boundaries, such as the evolution curve. Processing block 311 is optional, as shown by the dashed line. In some instances, the evolution curve is considered to be the initial curve, which can be optimized using a standard level set or other active contour algorithms. Because the fast-moving level set makes the initial contour very close to the desired shape, only a few iterations of standard level set optimization may be needed before terminating processing. See also Figure 5 As an example, it shows the results of level set optimization with the final fast travel result as the initial contour.

[0047] Processing block 309 or 311 can be followed by processing block 313, which includes performing a metric on the ROI within the image based on the determined boundaries. The metric can be performed using a simple K-means clustering (unsupervised machine learning) process with multiple clusters as input parameters. The size of the clusters (e.g., the number of pixels in each cluster, the distance from the cluster center to the object center) can be used as a feature in the larger object / artifact classification process during the metric. In some embodiments, and specifically concerning 3D-NAND ROIs, grayscale analysis using unsupervised machine learning can be useful in eliminating artifacts that have a similar size and shape to the 3D-NAND ROI but are distinguishable by the pixel grayscale distribution within the structure. For example, a small number of 3D-NAND structures can be used to measure various features and a support vector machine can be used to construct a trained model for identifying targets of larger objects. Any of many machine learning techniques, such as Bayesian learning, classification trees, random forests, neural networks, K-nearest neighbors (KNN), etc., can be used to classify shapes into 3D-NAND ROIs and artifacts.

[0048] Figure 4 Image sequence 401 illustrates the initialization and propagation of a fast-moving level set contour according to embodiments of the present disclosure. Image sequence 401 is an example of image processing performed by system 100 and / or 200, and may be, for example, the result of at least a portion of method 301. Typically, image sequence 401 illustrates the initialization and optimization of a fast-moving level set toward a single ROI. As shown in the image labeled t=30, the result is the stabilization of contour 403 on the boundary 405 of the ROI.

[0049] The image labeled T=0 illustrates the seeding of a single contour 403. This contour is seeded on the outermost perimeter pixels of the image. Although the seed is shown as a closed contour, it can also be one or more pixels located in individual pixels. Either seeding strategy will result in optimization of the same boundary 405.

[0050] Images labeled T=2 to T=30 illustrate the progress of the fast-moving level set contour 407, where the number of labels indicates the time the fast-moving level set algorithm has been executing to find the boundary. It can be seen that only 30 iterations of the fast-moving level set algorithm are needed to optimize boundary 405. By determining where boundary 405 is located, the ROI shown in the image is segmented from the surrounding image. In some embodiments, the optimized contour, which can be further refined, can serve as the basis for ROI metrics such as determining diameter, perimeter, shape factor, eccentricity, realism, convexity, best-fit ellipse features, total area within the closed boundary, features of pixel gray-level clusters within the closed object region, boundary length, etc.

[0051] Figure 5 This is another example image sequence 501, a high-level illustration of an image being processed using method 301 according to an embodiment of this disclosure. Image 503 is an example of a raw image of a single 3D-NAND via, while image 505 is the raw image after processing block 305 (e.g., an applied velocity function) and processing block 307 (e.g., initialization of FMLS contours). The initialized contours are located at the edge pixels of the image. In subsequent processing block 309, the contours have evolved to locate the outer boundaries of the ROI, which takes 30 (arbitrary units) of time. In this example, processing block 311 is also performed to optimize the boundary contours. In such instances, the following level set equation is implemented: (2) In this context, the first term {g(I)‖∇φ‖ div(∇φ / ‖∇φ‖ )} is a smoothing term that moves the curve toward its curvature direction. The second term {g(I)‖∇φ‖∙v} is a ballooning term that controls the speed at which the curve propagates with the parameter v. Finally, the third term {∇g(I)∙∇φ} is an image attachment term that helps the curve converge. In many respects, at this point, the level set belongs to the same category as the active contour, where the smoothing term is controlled by the curvature of the contour and the image attachment term is the external force that attracts the contour toward the peak of the image intensity gradient. Alternatively, the “snake” method can be used to replace the level set formula in equation (2) to improve accuracy. In the case of high-resolution TEM images with several material layers and their approximate thicknesses known a priori, the snake algorithm can be used to depict the internal structure after detecting the external boundary using FMLS, see [link to relevant documentation]. Figure 7 As an example.

[0052] Figure 6 Image sequence 601 is another example of the optimization of a fast-moving level set set around multiple ROIs according to embodiments of this disclosure. All images in image sequence 601 are identical, but illustrate the progression of the fast-moving level set contours as it is optimized at the boundaries around all ROIs within the image. Image sequence 601 is an example of image processing performed by system 100 and / or 200, and may be, for example, the result of at least a portion of method 301. Generally, when multiple objects are present, the FMLS algorithm has an inherent ability to change propagation contours according to topology, to break and merge them as they evolve, and further to form sharp corners, cusps, etc. The images in sequence 601 illustrate the application of FMLS on a large image with multiple 3D-NAND memory cells and demonstrate the ability to propagate / evolve contours to break and merge according to topology and to delineate the boundaries of all individual ROIs.

[0053] Image 603 shows the progress of the seed contour at time 30, where the contour begins to move back and forth around the periphery of the ROI. In image 605, at time 200, the contour has found the boundaries around the surrounding ROIs, and the seed's journey has begun to move around the ROIs, just inside them. Images 607 and 609 show the corresponding times 200 and 300 and how the contour continues to advance and find boundaries around additional ROIs within the image. At time 300, the boundaries of all ROIs within the image have been located. It should be noted that the fast-moving level set contour can break when it connects to itself around the boundaries, but can still continue to propagate until all boundaries are located.

[0054] Figure 7 Two images showing multiple layers localized using a fast-moving level set according to embodiments of the present disclosure are illustrated. In some embodiments, such as when processing high-resolution images, the fast-moving level set and measurements can identify individual materials / layers within a Region of Interest (ROI). For example, individual layers within a 3D-NAND structure can be identified and measured. By marking / localizing the interfaces between adjacent layers / materials and measuring features such as average layer thickness and layer thickness variation in the radial direction, all or most layers in the ROI can potentially be characterized. Of course, the image resolution will determine the number of layers that can be identified and measured. For example, high-resolution STEM or TEM images can provide details that are not identifiable in low-resolution SEM images.

[0055] For illustration, image 701 is a high-resolution SEM image that allows not only the identification and localization of the outer boundaries of multiple ROIs, but also the identification and localization of the two inner boundaries within each ROI. Image 703 is a TEM image of a single ROI, and the techniques disclosed herein have been used to localize seven boundaries. In some embodiments, images 701 and 703 have an optimized set of fast-moving levels optimized using an active contour algorithm such as a serpentine algorithm.

[0056] In some embodiments, different algorithms and methods can be used to delineate different boundaries around and within a Region of Interest (ROI). For example, in high-resolution images, the metric requirement is to identify individual materials within a 3D-NAND structure by marking interfaces of individual materials within the 3D-NAND structure with neighboring materials and measuring features such as the average thickness of the layers and the variation of layer thickness in the radial direction. By performing multiple such measurements, any consistent variations detrimental to the performance of the 3D-NAND cell can be identified. For example, in high-resolution SEM data, and utilizing prior information from design specifications regarding the number of material layers and their approximate thicknesses, multiple profiles can be initialized and allowed to evolve using the FMLS techniques disclosed herein. Image 701 illustrates the results of such initialization for marking multiple interfaces within a memory cell, acquired using an SEM imaging mode.

[0057] After marking the outer boundary of the 3D-NAND memory cell as accurately as possible, multiple internal contours can be initialized at appropriate distances from the outer boundary using prior information about the number of interfaces to be detected and their approximate distances from the outer boundary (depending on the material thickness). Figure 703 illustrates this process for depicting multiple material interfaces within a 3D-NAND memory cell, where a high-resolution image is acquired, for example, using S / TEM mode.

[0058] Figure 8 Examples of images show different shapes implementing the techniques disclosed herein on non-circular ROIs. While 3D-NAND structures have been used as the primary examples herein, the disclosed techniques are not necessarily limited to such structures. Images 801, 803, and 805 show the application of level sets to the segmentation and subsequent measurements of fin-like and other structures. For example, image 801 is a fin-like structure with boundaries marked using the techniques disclosed herein. For another example, images 903 and 805 show another type of fin-like structure marked with an outer boundary (803) and an inner boundary (805).

[0059] Figure 9A This is an instance deep learning filter 900 for preprocessing images according to the disclosed embodiments. Specifically, Figure 9A A technique for training a deep learning-based trainable nonlinear reaction diffusion filter 902 using associated input-output pairs is disclosed. Training can be performed by system 100 or 200, or by another system, and then the trained model is ported to one of system 100 or 200.

[0060] For model training filter 902, input / output pairs are prepared from a small number of frames / images. For example, images are collected using 32 image frames (which, after drift correction and averaging, form the final single image), and selective pixel integration is used to correct possible drift and reduce noise. The resulting image is then subjected to brightness correction, contrast stretching, and preprocessing using a reactive diffusion filter to improve the signal-to-noise ratio (SNR). This preprocessed image is characterized by sharp intensity gradients between different material layers and a smooth, uniform background; see [link to relevant documentation]. Figure 9B Image 915 is used as an example. This preprocessed image serves as the output image 906 of the input / output image pair used for training the deep neural network. Figure 9BAn example of the output image from the reconstructed input / output image pair used for deep model training is illustrated. The image reconstructed using only 8 frames serves as the input image 904 for the input / output image pair used to train the deep model filter 902; see image 913 as an example. Multiple input / output pairs can be constructed for each 32-frame set obtained for training the deep learning model 902 by bootsampling 8 frames from 32 frames.

[0061] Figure 10 This is an example functional block diagram 1000 based on an embodiment of the present disclosure. Figure 10 This is a block diagram illustrating a computer system 1000 that may be included in embodiments of the present invention. The computing system 1000 may be an example of computing hardware included in systems 100 and / or 200. The computer system 1000 includes at least a bus 1040 or other communication mechanism for transmitting information, and a hardware processor 1042 coupled to the bus 1040 for processing information. The hardware processor 1042 may, for example, be a general-purpose microprocessor. The computing system 1000 may be used to implement the methods and techniques disclosed herein (such as methods 301 and 401), and may also be used to acquire images and segment said images using one or more classes.

[0062] Computer system 1000 also includes main memory 1044 (such as random access memory (RAM) or other dynamic storage device) coupled to bus 1040 for storing information and instructions to be executed by processor 1042. Main memory 1044 can also be used to store temporary variables or other intermediate information during the execution of instructions to be executed by processor 1042. When stored in non-transient storage media accessible to processor 1042, such instructions make computer system 1000 a special-purpose machine customized to perform the operations specified in the instructions.

[0063] The computer system 1000 further includes a read-only memory (ROM) 1046 or other static storage device coupled to the bus 1040 for storing static information and instructions for the processor 1042. A storage device 1048, such as a magnetic disk or optical disk, is provided and coupled to the bus 1040 for storing information and instructions.

[0064] Computer system 1000 can be coupled to display 1050 (such as a cathode ray tube (CRT)) via bus 1040 for displaying information to the computer user. Input device 1052, including alphanumeric keys and other keys, is coupled to bus 1040 for transmitting information and command selections to processor 1042. Another type of user input device is cursor controller 1054 (such as a mouse, trackball, or arrow keys) for transmitting directional information and command selections to processor 1042 and for controlling cursor movement on display 1050. This input device typically has two degrees of freedom on two axes (a first axis (e.g., x) and a second axis (e.g., y)), which allows the device to specify a position in a plane.

[0065] Computer system 1000 may implement the techniques described herein using custom hardwired logic, one or more ASICs or FPGAs, firmware, and / or program logic, which, in combination with the computer system, make computer system 1000 a special-purpose machine or program the computer system 1000 to be a special-purpose machine. According to one embodiment, the techniques herein are implemented by computer system 1000 in response to processor 1042 executing one or more sequences of one or more instructions contained in main memory 1044. Such instructions may be read into main memory 1044 from another storage medium, such as storage device 1048. Execution of the instruction sequence contained in main memory 1044 causes processor 1042 to perform the processing steps described herein. In alternative embodiments, hardwired circuitry may be used in place of or in combination with software instructions.

[0066] As used herein, the term "storage medium" refers to any non-transitory medium that stores data and / or instructions that enable a machine to operate in a particular manner. Such storage media can include non-volatile media and / or volatile media. Non-volatile media include, for example, optical discs or magnetic disks, such as storage device 1048. Volatile media include dynamic memory, such as main memory 1044. Common forms of storage media include, for example, floppy disks, flexible disks, hard disks, solid-state drives, magnetic tape or any other magnetic data storage media, CD-ROMs, any other optical data storage media, any physical media with a perforated pattern, RAM, PROMs and EPROMs, FLASH-EPROMs, NVRAMs, any other memory chips or cartridges, content-addressable memory (CAM), and ternary content-addressable memory (TCAM).

[0067] Storage media differ from transmission media, but can be used in conjunction with them. Transmission media participate in the transfer of information between storage media. For example, transmission media include coaxial cables, copper wires, and optical fibers, including conductors containing bus 1040. Transmission media can also take the form of sound waves or light waves, such as those generated during wireless and infrared data communication.

[0068] Various forms of media may be involved when carrying one or more sequences of one or more instructions for execution to processor 1042. For example, the instructions may initially be carried on a disk or solid-state drive of a remote computer. The remote computer may load the instructions into its dynamic memory and transmit the instructions over a telephone line using a modem. A local modem of computer system 1000 may receive data over the telephone line and convert the data into an infrared signal using an infrared transmitter. An infrared detector may receive the data carried in the infrared signal, and appropriate circuitry may place the data on bus 1040. Bus 1040 carries the data to main memory 1044, from which processor 1042 retrieves and executes the instructions. The instructions received by main memory 1044 may optionally be stored on storage device 1048 before or after execution by processor 1042.

[0069] Computer system 1000 also includes a communication interface 1056 coupled to bus 1040. Communication interface 1056 provides bidirectional data communication coupled to network link 1058 connected to local network 1060. For example, communication interface 1056 may be an Integrated Services Digital Network (ISDN) card, a cable modem, a satellite modem, or a modem providing data communication connectivity to a corresponding type of telephone line. As another example, communication interface 1056 may be a Local Area Network (LAN) card to provide data communication connectivity to a compatible LAN. A wireless link may also be implemented. In any such implementation, communication interface 1056 transmits and receives electrical, electromagnetic, or optical signals carrying digital data streams representing various types of information.

[0070] Network link 1058 typically provides data communication to other data devices via one or more networks. For example, network link 1058 may provide a connection to host 1062 or to a data device operated by an Internet Service Provider (ISP) 1064 via local network 1060. ISP 1064, in turn, provides data communication services via a global packet data communication network now commonly referred to as the "Internet" 1066. Both local network 1060 and Internet 1066 use electrical, electromagnetic, or optical signals that carry digital data streams. Signals through various networks, signals on network link 1058, and signals through communication interface 1056 are instances of transmission media carrying digital data to and from computer system 1000.

[0071] Computer system 1000 can send messages and receive data (including program code) through (multiple) networks, network links 1058, and communication interfaces 1056. In an Internet instance, server 1068 can transmit application request codes through the Internet 1066, ISP 1064, local network 1060, and communication interface 1056.

[0072] The received code may be executed by processor 1042 upon receipt and / or stored in storage device 1048 or other non-volatile memory for later execution.

[0073] The embodiments discussed herein used to illustrate the disclosed techniques should not be considered limiting, but merely examples of implementation. For example, an X-ray shield as disclosed herein may include three elongated members formed in a Y-shape, or may have more than four members. For example, different numbers of members may affect the degree of torsion to achieve the desired amount of X-ray protection; generally, additional numbers of elongated members require less torsion. Those skilled in the art will understand that numerous other ways in which the disclosed techniques can be implemented are contemplated herein and are within the scope of this disclosure.

Claims

1. A method for image processing and measurement, comprising: Create a fast-moving level set seed on the image; Propagate fast-moving level set curves from the fast-moving level set seed to locate the boundaries of multiple regions of interest within the image; and The measurement of the region of interest is performed in part based on the boundary. Creating a fast-moving level set seed on the image includes creating a closed seed curve located around the perimeter of the image.

2. The method of claim 1, wherein creating a fast-moving level set seed on the image comprises creating one or more seed points within the image.

3. The method of claim 1, wherein propagating a fast-moving level set curve from the fast-moving level set seed to locate the boundaries of a plurality of regions of interest within the image comprises: The velocity function image of the image is constructed based on the velocity function algorithm; as well as The fast-moving level set seed is propagated to determine the boundaries of the plurality of regions of interest within the velocity function image.

4. The method of claim 1, further comprising: Preprocessing the image, wherein preprocessing the image includes: Improve the signal-to-noise ratio of the image; Correct brightness; Enhance contrast; as well as Enhance object boundaries.

5. The method of claim 4, wherein the preprocessing is performed using a deep learning-based model, and wherein the deep learning-based model is a trainable nonlinear reaction-diffusion filter.

6. The method of claim 4, wherein the preprocessed image includes sharp intensity gradients between different material layers and a smooth, uniform background.

7. The method of claim 1, further comprising: The boundaries of the multiple regions of interest are optimized based on active contour or level set algorithms.

8. The method of claim 1, wherein a machine learning-based model is used to perform the measurement to determine the measurement of one or more of the plurality of regions of interest.

9. A system for image processing and measurement, comprising: An electron microscope for acquiring images of a sample, the sample including multiple regions of interest; A nucleus coupled to or included in the electron microscope, the nucleus comprising or coupled to code, the code causing the nucleus to: Create a fast-moving level set seed within the image; Propagate the fast-moving level set curve from the fast-moving level set seed to locate the boundary of each of the plurality of regions of interest within the image; The boundaries of the plurality of regions of interest located via propagated fast-moving level set curves are optimized based on active contour or level set algorithms. as well as The measurement is performed on the region of interest in part based on the optimized boundary; The code that, when executed by the kernel, causes the kernel to create a fast-moving level set seed within the image, further includes: code that, when executed by the kernel, causes the kernel to create a closed seed curve within the image; The closed seed curve is located around the periphery of the image.

10. The system of claim 9, wherein the code that, when executed by the kernel, causes the kernel to create a fast-moving level set seed within the image, further comprises: The code that, when executed by the kernel, causes the kernel to create one or more seed points within the image.

11. The system of claim 9, wherein the code that, when executed by the kernel, causes the kernel to propagate the fast-travel level set curves from the fast-travel level set seed to locate the boundaries of the plurality of regions of interest within the image further comprises: When executed by the kernel, the following code is made to be executed by the kernel: The velocity function image of the image is constructed based on the velocity function algorithm; as well as The fast-moving level set seed is propagated to determine the boundaries of the plurality of regions of interest within the velocity function image.

12. The system of claim 9, wherein the code further includes code that, when executed by the kernel, causes the kernel to preprocess the image, wherein preprocessing the image includes one of the following: Improve the signal-to-noise ratio of the image; Correct brightness; Enhance contrast; as well as Enhance object boundaries.

13. The system of claim 12, wherein the kernel implements a deep learning-based model to preprocess the image.

14. The system of claim 13, wherein the deep learning-based model is a nonlinear reactive diffusion filter.

15. The system of claim 9, wherein the kernel implements a machine learning-based model to perform the metric.

16. The system of claim 15, wherein the machine learning-based model is one of a Bayesian learning-based model, a classification tree-based model, a random forest-based model, a neural network-based model, and a K-nearest neighbor-based model.

17. A method for image processing and measurement, comprising: Create a fast-moving level set seed on the image; Propagate the fast-moving level set curve from the fast-moving level set seed to locate the boundaries of multiple regions of interest within the image; as well as The measurement of the region of interest is performed in part based on the boundary. The propagation of the fast-moving level set curve from the fast-moving level set seed to locate the boundaries of multiple regions of interest within the image includes: Construct the velocity function image of the image based on the velocity function algorithm; and Propagate the fast-travel level set seed to determine the boundaries of the plurality of regions of interest within the velocity function image; Creating a fast-moving level set seed on the image includes creating a closed seed curve located around the perimeter of the image.

18. A method for image processing and measurement, comprising: Preprocessing the image, wherein preprocessing the image includes: Improve the signal-to-noise ratio of the image; Correct brightness; Enhance contrast; as well as Enhance object boundaries; Create a fast-moving level set seed on the preprocessed image; Propagate the fast-moving level set curve from the fast-moving level set seed to locate the boundaries of multiple regions of interest within the image; as well as The measurement of the region of interest is performed in part based on the boundary; Creating a fast-moving level set seed on the preprocessed image includes creating a closed seed curve located around the perimeter of the image.

19. The method of claim 18, wherein the preprocessing is performed using a deep learning-based model, and wherein the deep learning-based model is a trainable nonlinear reaction-diffusion filter.

20. The method of claim 18, wherein the preprocessed image comprises sharp intensity gradients between different material layers and a smooth, uniform background.

Images