Systems and methods for accelerated online adaptive radiation therapy

Abstract

Systems and methods for accelerated online adaptive radiation therapy ("ART") are described. Improvements to online ART are generally provided based on the use of texture analysis and machine learning algorithms implemented using hardware processors and memory. The described systems and methods enable more efficient and accurate online adaptive replanning ("OLAR") that can also be achieved in a clinically acceptable time period. For example, OLAR can be reduced from 10-30 minutes to 5-10 minutes.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U.S. Provisional Patent Application Serial No. 62 / 690,289, filed on June 26, 2018, entitled “SYSTEMS AND METHODS FOR ACCELERATED ONLINE ADAPTIVE RADIATION THERAPY,” which is incorporated herein by reference in its entirety. Background Art

[0003] The anatomy, biology, or both of a tumor or organ at risk ("OAR") can change substantially and randomly during radiation therapy ("RT") delivery. Online adaptive radiation therapy ("ART") is a state-of-the-art approach that typically involves imaging the patient's specific anatomy and / or biology the day before each RT delivery (i.e., fraction), creating a new RT plan adapted to the anatomy / biology of the tumor and OAR, and delivering the new adaptive RT plan to the patient. These operations should be completed in a practical time period, as the patient lies on the table and waits to receive treatment during online adaptive replanning.

[0004] The lengthy and potentially clinically impractical time required to segment anatomical structures in images acquired on the day of treatment and evaluate plans is a bottleneck in practicing online ART. Therefore, it is desirable to provide a faster and more robust method for segmenting anatomical structures and evaluating RT plans within an online ART framework. It is also desirable to provide a robust technique to rapidly identify the need for online ART. Summary of the invention

[0005] The present disclosure addresses the above-mentioned shortcomings by providing methods for improved online adaptive radiation therapy ("ART") and online adaptive replanning ("OLAR").

[0006] In one aspect of the present disclosure, a method is provided for controlling a radiation therapy planning system to identify whether it would be advantageous to perform OLAR and optionally initiate an OLAR procedure.

[0007] In another aspect of the present disclosure, a method for correcting a radiation therapy planning profile and thereby generating corrected radiation therapy planning profile data is provided.

[0008] In yet another aspect of the present disclosure, a method for verifying the accuracy of radiation therapy planning contour data is provided.

[0009] In yet another aspect of the present disclosure, a method for assessing the quality of a radiation therapy plan is provided.

[0010] In another aspect of the present disclosure, a method for generating a synthetic computed tomography (CT) image from a magnetic resonance image is provided.

[0011] The above and other aspects and advantages of the present disclosure will become apparent through the following description. In the description, reference is made to the accompanying drawings which form a part of the description, in which preferred embodiments are shown by way of illustration. However, this embodiment does not necessarily represent the full scope of the invention, and therefore, reference is made to the claims and this document to interpret the scope of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a flowchart setting forth the steps of a method for generating an indication of whether to initiate online adaptive replanning ("OLAR") based on analysis of a metric computed from a Jacobian histogram generated from a deformable image registration between a reference image and daily images of a patient.

[0013] Figure 2 is an example of a decision tree that can be obtained from using a machine learning algorithm to determine JDH metrics and criteria for generating an indication of whether it is advantageous to initiate OLAR. In this example, a decision tree has been obtained for identifying whether it would be advantageous to perform OLAR during radiation therapy for prostate cancer.

[0014] Figure 3 is a block diagram of an example system that analyzes medical image data to assess whether to initiate or otherwise implement OLAR according to embodiments described in the present disclosure.

[0015] Figure 4 It is achievable Figure 3 A block diagram of example components of the OLAR initiated assessment system.

[0016] Figure 5 It is a flow chart of the automatic contour correction method based on texture features.

[0017] Figures 6A-6C A representative example of a texture-based active contour model is shown. Fig. 6A The ground truth (green) contours and the automatically generated contours are shown in FIG. Figure 6B The GLCM cluster shadow feature map is shown in . Figure 6C The contour evolution diagram is shown in . The evolving contour is shown using a cyan line, while the live contour, the initial contour, and the corrected contour are highlighted using different colors (ie, green, blue, and red).

[0018] Figure 7 is a flow chart of an example method for feature-based automated profile quality control and profile correction.

[0019] Figure 8 is an example of a decision tree model that may be implemented for texture-based contour quality verification using the methods described in this disclosure.

[0020] Fig. 9A is an example of a hierarchical clustering tree of computed texture features across the core region of the live pancreatic head contour. In this example, a dissimilarity distance of 0.5 is used to identify clusters. The selected features are highlighted in yellow.

[0021] Fig. 9B is an example of a hierarchical clustering tree of computed texture features across the core region of the ground truth duodenum contour. In this example, a dissimilarity distance of 0.5 was used to identify clusters. The selected features are highlighted in yellow.

[0022] Figures 10A-10D A representative case of texture feature selection is shown. Fig. 10A and Fig. 10B Axial and coronal views of daily CT images are shown, respectively. Fig. 10C and Fig. 10D For the original duodenum ( Fig. 10C ) and pancreatic head ( Fig. 10D )'s contour inwards and outwards for various shell volumes with different characteristic values. Fig. 10C and Fig. 10D The x-axis in is the interval where the contour expands (positive) or contracts (negative). All eigenvalues ​​are normalized to one eigenvalue from the original organ volume.

[0023] Fig.11 is a block diagram of an example automatic contour correction and verification system that may implement embodiments described in the present disclosure.

[0024] Fig.12 It is achievable Fig.11 A block diagram of example components of an automatic contour correction and verification system.

[0025] Fig.13 A dose volume histogram ("DVH") comparison between an example adaptive plan and an example repositioned plan is shown.

[0026] Fig.14 Shown are thermal comparison plots of the dose level co-occurrence matrix ("DLCM") of an example adaptive plan and an example repositioned plan.

[0027] Fig.15Dose level size zone matrix ("DLSZM") plots of a planning target volume ("PTV") in an example adaptive plan and an example repositioned plan are shown.

[0028] Fig.16 is a flow chart of an example method for evaluating radiation therapy plan quality based on dose texture features, spatial dose metrics, or a combination thereof.

[0029] Fig.17 is a block diagram of an example automated OLAR plan quality assessment system that may implement embodiments described in the present disclosure.

[0030] Fig.18 It is achievable Fig.17 Block diagram of example components of an automated OLAR program quality assessment system.

[0031] Fig.19 is a flow chart setting forth the steps of an example method for generating a computed tomography ("CT") image from a magnetic resonance image.

[0032] Fig. 20 Depicted is a schematic diagram of a method for classifying air voxels from non-air voxels for use in synthetic CT generation to generate air region volumes.

[0033] Fig.21 is a flow chart setting forth the steps of another example method for generating a CT image from a magnetic resonance image.

[0034] Fig. 22 is a block diagram of an example synthetic CT image generation system that may implement embodiments described in the present disclosure.

[0035] Fig.23 It is achievable Fig. 22 A block diagram of example components of a synthetic CT image generation system.

[0036] Fig.24 is a block diagram of an example computer system that may implement the methods described in this disclosure.

[0037] Fig.25 is a block diagram of an example radiation therapy planning system that may implement the methods described in the present disclosure. DETAILED DESCRIPTION

[0038] Systems and methods for improved online adaptive radiation therapy ("ART") are described herein. Improvements to online ART are generally provided based on the use of texture analysis and machine learning algorithms implemented using hardware processors and memory. The systems and methods described in the present disclosure enable significant improvements to existing online ART technology and the development of improved online ART planning systems. Specifically, the systems and methods significantly improve the efficiency and accuracy of online ART, thereby improving radiation therapy delivery. As a result, safer and more effective radiation therapy plans can be generated and used to deliver treatment to patients. In addition, the systems and methods described in the present disclosure enable OLAR to be achieved in a clinically acceptable time period. For example, OLAR can be reduced from 10-30 minutes to 5-10 minutes.

[0039] For example, the systems and methods described in the present disclosure can quickly identify the need for online adaptive replanning ("OLAR"). In an example implementation, this can be accomplished by analyzing features or other metrics from a Jacobian histogram of a deformable registration of a reference image with an image of the day. The systems and methods described in the present disclosure can also automatically and quickly identify good quality contours generated by automatic segmentation by using image texture analysis. The systems and methods described in the present disclosure also provide automated plan quality assessment based on 3D dose texture features. The systems and methods described in the present disclosure can also quickly create synthetic computed tomography ("CT") images from magnetic resonance images. As an example, this process can be achieved by automatically segmenting air and bone regions in a magnetic resonance image.

[0040] Thus, in some aspects, the systems and methods described in the present disclosure provide for rapid identification of those situations where online ART is necessary or otherwise beneficial. As described above and in detail below, this assessment can be based on analyzing features or other metrics from a Jacobian histogram of a deformable registration of a reference image with a patient image obtained on the day of treatment.

[0041] In some other aspects, the systems and methods described in the present disclosure provide contour correction, contour verification, or both. As described above, and described in detail below, contours can be automatically identified as having good quality, and those contours identified as inaccurate can be automatically corrected.

[0042] In yet other aspects, the systems and methods described in the present disclosure provide automatic radiation therapy plan evaluation.As described above and in detail below, radiation therapy plan evaluation can be achieved based on analysis of 3D dose texture features.

[0043] As another example, in some aspects, the systems and methods described in the present disclosure provide synthetic CT. As described above and in detail below, a synthetic CT image can be generated from a magnetic resonance image. As an example, a synthetic CT image can be generated from a magnetic resonance image that has been processed to accurately segment air and bone regions.

[0044] OLAR can take into account the changes between doses during RT, but this is generally a time-consuming and labor-intensive process compared to other methods such as repositioning methods. Repositioning is sufficient for minimal deformation between doses. Therefore, an indication of when OLAR is expected is determined. This article describes a method for quickly determining whether OLAR is needed or whether OLAR will be beneficial in other ways. In an example, the present technology includes analyzing a Jacobian histogram obtained from a deformable image registration between a reference image (e.g., a planning image) and a daily image (e.g., an image obtained on the day of treatment).

[0045] Reference now Figure 1 , a flow chart is shown to set forth the steps for determining an indication of whether to initiate OLAR to generate a new radiation therapy plan. The method includes accessing reference images (as indicated at step 102) and daily images (as indicated at step 104) using a computer system. As described above, reference images may also be referred to as planning images, and are typically patient images acquired prior to the day of treatment. Daily images are typically patient images acquired on the day of treatment. Accessing these images may include retrieving previously acquired images from a memory or other data storage, or in the case of daily images, may include acquiring daily images using a suitable medical imaging system (such as a CT system or an MRI system).

[0046] As indicated in step 106, a shell structure is generated from the reference images and is also accessed using a computer system. The shell structure can be, for example, a shell structure with a specific thickness created around a tumor, OAR, or other critical structure depicted in one or more of the reference images. As an example, the shell structure can have a thickness of 10 mm. As another example, the shell structure can have a thickness of 5 mm. More specifically, the shell structure can have a thickness greater than or less than 10 mm.

[0047] As indicated in step 108, a deformable image registration ("DIR") is performed between the reference image and the daily image. For example, the DIR may be performed based in part on the shell structure generated from the reference image. As indicated in step 110, a Jacobian determinant histogram ("JDH") is calculated based on the deformable image registration. For example, the JDH may be calculated based on the determinant of the Jacobian matrix of the deformation field estimated in the DIR. As indicated in step 112, one or more metrics of the JDH are calculated. As an example, the metric may include a minimum value ("min"), a maximum value ("max"), a full width at half maximum ("FWHM"), a peak value, a standard deviation, a slope, or a combination thereof. An indication of whether an OLAR should be initiated is generated based on one or more JDH metrics at step 114. For example, if one of the JDH metrics fails to meet a pre-established standard, an indication that an OLAR should be initiated is generated. The indication may be generated and provided to a user, such as by generating a display element displayed to the user on a graphical user interface ("GUI"). In other implementations, the indication may prompt the user to initiate an OLAR.

[0048] As an example for prostate cancer, if the minimum JDH is less than 0.47, an indication that OLAR is favorable and optionally an indication that OLAR should be initiated may be generated. As another example, if the minimum JDH is greater than 0.47 and the slope is greater than 1.45, but the FWHM is less than 0.037, an indication to reposition a dose sufficient for a radiation therapy session for prostate cancer may be generated.

[0049] When an indication that OLAR should be initiated is generated at decision block 114, a new treatment plan is generated using OLAR and the new treatment plan is used to provide radiation therapy to the patient, as indicated at step 116. Otherwise, the current treatment plan is accessed and used to provide radiation therapy to the patient, as indicated at step 118.

[0050] In some cases, a machine learning algorithm may be implemented using a hardware processor and memory of a computer system to identify a metric of JDH, a metric standard, or both, which may be used to provide an indication of whether OLAR is advantageous and, optionally, an indication that OLAR should be initiated. As an example, a machine learning algorithm may be used to train a complex decision tree using an appropriate JDH metric that may predict whether OLAR will be advantageous and, optionally, an indication of whether to initiate OLAR. Figure 2 An example of such a decision tree is shown in .

[0051] In an example study, the method described above is presented on the daily CT obtained using indoor CT during the CT-guided RT for prostate cancer. DIR is performed between daily CT and reference CT. JDH is extracted from the prostate and the uniform 10mm extension around the prostate. Using a machine learning algorithm that can include a supervised machine learning algorithm, a complex decision tree is trained using the appropriate JDH metric that can predict the needs of daily CT sets for OLAR. Sixty daily CTs from 12 randomly selected prostate cases are used as the training data set for the classifier, wherein the dosimetric (dosimetric) plan for both OLAR and repositioning is used to determine the classification. The resulting classification tree is tested using an independent data set of 45 daily CTs from nine other patients with 5 CTs each. The cumulative doses of 5 split dose plans from different strategies are compared. The cumulative dose with mixed OLAR and repositioning (using a classification tree) is comparable to all OLAR split doses.

[0052] Reference now Figure 3 , shows an example of a system 300 for analyzing medical image data to assess whether OLAR is achieved according to some embodiments of the systems and methods described in the present disclosure. Figure 3 As shown, computing device 350 may receive one or more types of data (e.g., medical image data) from image source 302, which may be a medical image source. In some embodiments, computing device 350 may execute at least a portion of OLAR initiation assessment system 304 to determine whether to initiate OLAR to generate an updated radiation therapy plan from the data received from image source 302.

[0053] Additionally or alternatively, in some embodiments, computing device 350 may communicate information about data received from image source 302 to server 352 over communications network 354, which may execute at least a portion of OLAR-initiated assessment system 304. In such embodiments, server 352 may return information indicative of the output of OLAR-initiated assessment system 304 to computing device 350 (and / or any other suitable computing device).

[0054] In some embodiments, computing device 350 and / or server 352 may be any suitable computing device or combination of devices, such as a desktop computer, a laptop computer, a smart phone, a tablet computer, a wearable computer, a server computer, a virtual machine executed by a physical computing device, etc. Computing device 350 and / or server 352 may also reconstruct an image from the data.

[0055] In some embodiments, image source 302 may be any suitable source of image data (e.g., measurement data, images reconstructed from measurement data), such as a CT system, an MRI system, another computing device (e.g., a server storing image data), and the like. In some embodiments, image source 302 may be local to computing device 350. For example, image source 302 may be integrated with computing device 350 (e.g., computing device 350 may be configured as part of a device for capturing, scanning, and / or storing images). As another example, image source 302 may be connected to computing device 350 via a cable, a direct wireless link, and the like. Additionally or alternatively, in some embodiments, image source 302 may be located locally and / or remotely relative to computing device 350, and may communicate data to computing device 350 (and / or server 352) via a communication network (e.g., communication network 354).

[0056] In some embodiments, the communication network 354 may be any suitable communication network or combination of communication networks. For example, the communication network 354 may include a Wi-Fi network (a Wi-Fi network may include one or more wireless routers, one or more switches, etc.), a peer-to-peer network (e.g., a Bluetooth network), a cellular network (e.g., a 3G network, a 4G network, etc. that conforms to any suitable standard (such as CDMA, GSM, LTE, Advanced LTE, WiMAX, etc.)), a wired network, etc. In some embodiments, the communication network 108 may be a local area network, a wide area network, a public network (e.g., the Internet), a private or semi-private network (e.g., a company or university intranet), any other suitable type of network, or any suitable combination of networks. Figure 3 The communication links shown may each be any suitable communication link or combination of communication links, such as a wired link, a fiber optic link, a Wi-Fi link, a Bluetooth link, a cellular link, or the like.

[0057] Reference now Figure 4 , shows an example of hardware 400 that may be used to implement image source 302, computing device 350, and server 354 according to some embodiments of the systems and methods described in the present disclosure. Figure 4As shown, in some embodiments, computing device 350 may include processor 402, display 404, one or more inputs 406, one or more communication systems 408, and / or memory 410. In some embodiments, processor 402 may be any suitable hardware processor or combination of processors, such as a central processing unit ("CPU"), a graphics processing unit ("GPU"), and the like. In some embodiments, display 404 may include any suitable display device, such as a computer monitor, a touch screen, a television, and the like. In some embodiments, input 406 may include any suitable input device and / or sensor that may be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, and the like.

[0058] In some embodiments, the communication system 408 may include any suitable hardware, firmware, and / or software for communicating information over the communication network 354 and / or any other suitable communication network. For example, the communication system 408 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 408 may include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0059] In some embodiments, memory 410 may include any suitable one or more storage devices that can be used to store instructions, values, data, etc., which can be used, for example, by processor 402 to present content using display 404, to communicate with server 352 via communication system(s) 408, and the like. Memory 410 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 410 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, and the like. In some embodiments, memory 410 may have a computer program encoded thereon or otherwise stored therein for controlling the operation of computing device 350. In such embodiments, processor 402 may execute at least a portion of the computer program to present content (e.g., images, user interfaces, graphics, tables), receive content from server 352, send information to server 352, and the like.

[0060] In some embodiments, server 352 may include processor 412, display 414, one or more inputs 416, one or more communication systems 418, and / or memory 420. In some embodiments, processor 412 may be any suitable hardware processor or combination of processors, such as a CPU, GPU, etc. In some embodiments, display 414 may include any suitable display device, such as a computer monitor, a touch screen, a television, etc. In some embodiments, input 416 may include any suitable input device and / or sensor that can be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, etc.

[0061] In some embodiments, the communication system 418 may include any suitable hardware, firmware, and / or software for communicating information over the communication network 354 and / or any other suitable communication network. For example, the communication system 418 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 418 may include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0062] In some embodiments, the memory 420 may include any suitable one or more storage devices that can be used to store instructions, values, data, etc., which can be used, for example, by the processor 412 to present content using the display 414, to communicate with one or more computing devices 350, etc. The memory 420 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, the memory 420 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, etc. In some embodiments, the memory 420 may have a server program encoded thereon for controlling the operation of the server 352. In such embodiments, the processor 412 may execute at least a portion of the server program to send information and / or content (e.g., data, images, user interfaces) to one or more computing devices 350, receive information and / or content from one or more computing devices 350, receive instructions from one or more devices (e.g., personal computers, laptops, tablet computers, smart phones), etc.

[0063] In some embodiments, the image source 302 may include a processor 422, one or more image acquisition systems 424, one or more communication systems 426, and / or a memory 428. In some embodiments, the processor 422 may be any suitable hardware processor or combination of processors, such as a CPU, a GPU, and the like. In some embodiments, the one or more image acquisition systems 424 are generally configured to acquire data, images, or both, and may include a CT system and an MRI system or both. Additionally or alternatively, in some embodiments, the one or more image acquisition systems 424 may include any suitable hardware, firmware, and / or software for coupling to and / or controlling the operation of the CT system and the MRI system or both. In some embodiments, one or more portions of the one or more image acquisition systems 424 may be removable and / or replaceable.

[0064] Note that, although not shown, image source 302 may include any suitable input and / or output. For example, image source 302 may include an input device and / or sensor that may be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, a trackpad, a trackball, etc. As another example, image source 302 may include any suitable display device (such as a computer monitor, a touch screen, a television, etc.), one or more speakers, etc.

[0065] In some embodiments, the communication system 426 may include any suitable hardware, firmware, and / or software for communicating information to the computing device 350 (and, in some embodiments, via the communication network 354 and / or any other suitable communication network). For example, the communication system 426 may include one or more transceivers, one or more communication chips and / or chipsets, and / or the like. In more specific examples, the communication system 426 may include hardware, firmware, and / or software that may be used to establish a wired connection using any suitable port and / or communication standard (e.g., VGA, DVI video, USB, RS-232, etc.), a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, and / or the like.

[0066] In some embodiments, memory 428 may include any suitable one or more storage devices that may be used to store instructions, values, data, etc., which may be used, for example, by processor 422 to: control one or more image acquisition systems 424 and / or receive data from one or more image acquisition systems 424; to image from the data; to present content (e.g., images, user interfaces) using a display; to communicate with one or more computing devices 350, etc. Memory 428 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 428 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, etc. In some embodiments, memory 428 may have a program encoded thereon or otherwise stored therein for controlling the operation of image source 302. In such embodiments, the processor 422 may execute at least a portion of a program to generate images, send information and / or content (e.g., data, images) to one or more computing devices 350, receive information and / or content from one or more computing devices 350, receive instructions from one or more devices (e.g., a personal computer, laptop computer, tablet computer, smart phone, etc.), and so on.

[0067] In some embodiments, any suitable computer-readable medium may be used to store instructions for performing the functions and / or processes described herein. For example, in some embodiments, the computer-readable medium may be transient or non-transient. For example, non-transient computer-readable media may include the following media, such as: magnetic media (e.g., hard disk, floppy disk), optical media (e.g., compact disk, digital video disk, Blu-ray disk), semiconductor media (e.g., random access memory ("RAM"), flash memory, electrically programmable read-only memory ("EPROM"), electrically erasable programmable read-only memory ("EEPROM")), any suitable medium that is not transient or does not have any persistent appearance during transmission, and / or any suitable tangible medium. As another example, transient computer-readable media may include signals on a network, signals on wires, conductors, optical fibers, circuits, or any suitable medium that is transient and does not have any persistent appearance during transmission, and / or any suitable intangible medium.

[0068] As described above, accurate segmentation based on daily images of the patient (i.e., images obtained on the day of radiation therapy (such as immediately before radiation delivery)) is a significant challenge for online ART re-planning. Available automatic segmentation techniques cannot always generate accurate contours. The systems and methods described in the present disclosure provide for automatic correction of inaccurate contours generated from automatic segmentation using image texture information.

[0069] Typically, an initial contour generated by automatic segmentation is evaluated on a slice-by-slice basis against a ground truth contour (e.g., a contour delineated by an experienced radiation oncologist). The initial contour is deemed inaccurate based on one or more similarity criteria compared to the ground truth. An automatic texture-based contour correction method is used to reprocess the inaccurate contour. The process typically includes two steps: computing a voxel-based texture feature map (e.g., a GLCM clustered shaded texture feature map using a 3×3 block size within a region of interest (“ROI”) created from the initial contour), and incorporating the texture feature map as an external force into an active contour algorithm to find the correct structure boundaries.

[0070] Integration of MRI and radiotherapy (e.g., using an integrated linear accelerator, or linac) can provide high-contrast real-time MRI images for daily online treatment adaptation. Intra-day images (i.e., images obtained on the day of treatment) can help take into account anatomical changes between doses and thereby deliver radiation doses more accurately. In the entire chain of online adaptive replanning, contouring of tumors and normal tissues on daily images is one of the most time-consuming processes, which limits the clinical application of online ART.

[0071] In current radiology practice, the interpretation of clinical images mainly relies on visual assessment or some qualitative measures based on image intensity. More quantitative pathophysiological information in the image has not yet been revealed. Quantitative texture feature analysis has been introduced into the field of radiation oncology for comprehensive disease characterization and early prediction of treatment response. Recently, texture information has been used for automatic segmentation algorithms. The active contour model (also known as the "snake" model) is one such algorithm. The adaptation process of the snake model attracts active shapes (e.g., 2D contours or 3D surfaces) to desired image features (such as, specific intensities, image edges or texture patterns).

[0072] A known difficulty with snake models is determining the initial shape and position of the active curve. If the snake is initialized too far from the actual boundary, the result may not be able to cover the object boundary. On the other hand, if the initial contour is close to the actual contour and has a shape similar to the actual contour, the performance of the snake model can be greatly improved in terms of both accuracy and efficiency.

[0073] The systems and methods described in the present disclosure address this problem by providing a fast and semi-automatic contour correction method that incorporates quantitative image texture features into an active contour model. As an example, the contours automatically generated by deformable registration between the doses can be used as initial curves for contour formation algorithms (such as the snake model). In doing so, prior structural knowledge is used to help guide the segmentation of new images.

[0074] Figure 5 The workflow of the automatic contour correction method based on texture of an example is shown. At step 502, a computer system is used to access the actual contour, and at step 504, a computer system is used to access the automatically generated initial contour. At step 506, the initial contour is evaluated on a 2D slice basis according to the actual contour. At decision box 508, the contour is evaluated, and if the similarity standard is not met, the contour is considered to be inaccurate. As an example, compared with the actual situation, a Dice coefficient ("DICE") greater than 0.9 or a Hausdorff distance ("HD") less than 3mm can be used as a similarity standard. If the contour is identified as accurate, then these contours are stored for use at step 510. If the contour is identified as inaccurate, then these contours are stored for additional processing at step 512.

[0075] Subsequently, the identified inaccurate slices are preprocessed at step 514 to prepare the input image and contours for further procedures. The input image can be a magnetic resonance image, a CT image, an ultrasound image, or any image acquired using a suitable imaging modality. For example, heterogeneity in the magnetic field of an MRI machine can cause a bias field that will corrupt the image (such as by causing heterogeneous gray level distributions across the image and blurred edges and boundaries). A preprocessing step can be performed to correct for the effects of the bias field. In addition, a nonlinear bilateral filter can be applied to smooth the MRI image so as to reduce noise while preserving edges.

[0076] The automatic contour correction method 516 begins with a GLCM texture feature map calculation, which can be calculated for the ROI created from the initial contour at step 518. In some implementations, efficiency can be accelerated by extracting texture features of the ROI, which is created by adding a margin (e.g., a 10mm margin) to the bounding box of the initial contour. For each voxel within the ROI, each voxel value is replaced by a GLCM texture feature derived from a block (e.g., a 3×3 block) surrounding the voxel of interest. In some cases, the GLCM features can be GLCM cluster shadow features, which are advantageous due to the ability of the GLCM cluster shadow features to enhance tissue boundary contrast. Block units with grayscale variations across voxels will have larger GLCM cluster shadow feature values, while homogeneous blocks will have smaller feature values. Depending on the number of grayscales used for the GLCM calculation, the shape edges can have very large sizes due to the cubic in the calculation equation of the shape edges.

[0077]

[0078]

[0079] Where p(i,j,θ,d) is a GLCM matrix element, and p(i,j,θ,d) contains the statistical probability of a pair of pixels of gray levels i and j at a specific displacement distance d and a specific angle (θ). In some implementations, the original image may first be rescaled to reduce the number of gray levels. The size of the GLCM matrix may be determined by the number of gray levels. To avoid directional dependence, the matrix may be accumulated along four angles (i.e., θ=0, 45, 90, 135). The distance d between two pixels may be selected to be 1.

[0080] Subsequently, the feature map is integrated into the active contour algorithm as an external force to find the correct structural boundary, as indicated in step 520. The active contour model can generally include three stages: initialization, deformation, and termination. As an example, the contour automatically generated by deformable registration between the doses can be used as the initial curve of the snake model. During the deformation stage, the snake model moves the contour points by minimizing the energy function. As shown in equation (3), the energy function (E 蛇 ) consists of two items: internal energy (E int ) and external energy (E ext ).

[0081]

[0082] where v(s) is the contour point and N is the total number of contour points.

[0083] As shown in equation (4), the internal energy can be further divided into two energy components: elastic force and bending force, where the elastic force and bending force are the first-order derivative and second-order derivative of the profile, respectively.

[0084]

[0085] α and β represent the weighting factors for each term. The elastic force controls the tension of the snake, while the bending force controls the stiffness of the snake. During the deformation process, the weighting factors can be tuned to adjust the smoothness of the curve to better match the shape of the object boundary. During iterations, the internal force is only related to the shape of the curve, and the internal force is independent of the image features.

[0086] The calculated GLCM cluster shadow feature map is used as an external force to guide the active curve to find the actual boundary, whose eigenvalue is relatively large compared to the homogeneous region. The traditional snake model has the problem that the external force only exists at the edge. For points located inside the homogeneous region of the organ, the active contour may not be able to converge to the boundary. To solve this problem, a "balloon force" can be added to the external energy function, as shown in equation (5). As shown in equation (6), the balloon force is calculated along the normal direction of the curve, where μ is the weighting factor of the balloon force. During snake evaluation, the parameter μ can be regarded as the step size of the snake's movement.

[0087]

[0088]

[0089] Where n(s) is the normal unit vector of the curve at v(s). To avoid contour leakage caused by the balloon force, a binary threshold function ζ(v(s)) may be applied. The balloon force is activated only for contour points with eigenvalues ​​below a predetermined threshold. For high eigenvalue points, the balloon force is equal to zero. As an example, the threshold may be chosen to be a constant, such as a constant of 50.

[0090] As mentioned above, there are a total of four weighting factors for each term of the snake model. These factors have a significant impact on the behavior of the snake and the performance of the deformation process. In addition, the inherent characteristics of MRI images cause large inter-patient differences and even intra-patient differences. Each image set benefits from a different set of parameter values ​​to make the snake perform well. In order to solve this problem, a user interface can be used to interactively tune the key parameters of each image set. The parameters can be tuned based on the contour correction performance of the sample slice. These values ​​can be applied to the remaining slices to improve efficiency. These values ​​can also be tuned for each slice as needed.

[0091] A default maximum number of iterations performed during the deformation process may be set, such as using a value of 300 total iterations. In other implementations, the overlap area change rate ("CR") may also be calculated and stored for each iteration. The iterative process may also terminate if the CR is less than 0.01 for a consecutive number of iterations (e.g., five iterations), meaning that the snake has converged to a stable state.

[0092] After the active contour is formed, morphological operations are performed to refine the contour at step 522. Finally, at step 524, DSC and HD are recalculated on the corrected contour to evaluate the performance of the proposed method. The updated contour can be stored and used for later use.

[0093] Figures 6A-6C Representative examples are shown to illustrate the texture-based contour correction method described above. Fig. 6A The ground truth contour and the contour generated from intra-dose deformable registration are shown. Due to the large deformation between different doses, the contour algorithm used in this example failed to generate an accurate contour of the pancreatic head for this slice. Figure 6B The calculated GLCM cluster shadow feature map is shown, where the boundary between the pancreatic head and the surrounding tissue is enhanced compared to the original MRI image. Figure 6CIn , the evolving profile is shown using a cyan line, while the live profile, the initial profile, and the corrected profile are highlighted with different colors (ie, green, blue, and red). After correction, the profile successfully converges to the live profile.

[0094] The image texture based automatic contour correction method described above can be used to improve the overall contour accuracy and efficiency of contour modification. The method can be used as a part of the MRI guided online re-planning segmentation process.

[0095] As described above, it is desirable to robustly validate automatically segmented contours in online adaptive replanning ("OLAR"). The systems and methods described in the present disclosure provide a fast and automated patient-specific contour validation method using quantitative image features (e.g., image texture features) of medical images. The medical images may be daily images of the subject (i.e., images of the patient obtained on the day of treatment) or previously acquired images, and may be CT images, MRI images, or any other suitable medical images.

[0096] For OLAR, the accuracy of automatic segmentation based on medical images (such as intra-day images) is a challenge. The verification of automatically segmented contours is traditionally performed through a time-consuming manual review process. The systems and methods described in the present disclosure provide a fast and automated patient-specific contour verification method using quantitative image texture features.

[0097] In fractionated RT, due to the variation between fractions, the treatment plan generated using planning CT ("pCT") before treatment may not be optimal for a given fraction. The introduction of indoor imagers (such as MRI, on-track CT and cone beam CT ("CBCT")) has made it possible to visualize the patient's anatomical structure on a daily basis. OLAR can be used to adapt the radiation plan in response to the anatomical structure changes between fractions during the treatment process. Since the re-planning process needs to be completed in a practical time period when the patient lies on the bed and waits, the efficiency of OLAR remains a challenge. Different types of computerized tools have been developed to increase the speed of the re-planning process. An identified bottleneck is the time to circle the target organ and key organs on the image of the day.

[0098] In recent years, new technologies such as automatic image segmentation based on machine learning and contour propagation methods based on deformable image registration ("DIR") have been developed and used to generate contours on daily images. The introduction of daily MRI images is particularly useful in improving the quality of daily contours. Because the contours automatically generated using currently available methods are not perfect, daily contours are usually required to be verified and manually edited during the OLAR process. Currently, the verification of contours is performed by a manual review process, which is time-consuming and therefore impractical for OLAR. It is desirable to have more efficient and robust automatic contour evaluation and verification methods to promote the application of OLAR.

[0099] Medical images contain not only visual depictions of human anatomy, but also data that can be mined (such as quantitative pathophysiological information). Texture analysis can be a very useful image attribute in image analysis and computer vision applications. Texture features are usually calculated from the statistical distribution of pixel intensities in the image. Depending on the number of intensity points (i.e., pixels) in the statistical calculation, texture features can be classified into first-order statistics, second-order statistics, and higher-order statistics. Those higher-order texture features have the potential to reveal information that is difficult to visualize from the original image.

[0100] In the case of the head and neck, the shapes and relative positions of the organs are usually stable with small changes, which makes these organs good candidates for applying geometry-based contour assessment. In contrast, due to a combination of reasons (such as respiratory and intestinal movements and different filling states), the abdominal and pelvic organs can exhibit large displacements and deformations. Therefore, it is difficult to apply geometry-based contour assessment on those organs. However, even in the case of shape and position changes, for specific organs, texture features can be unchanged.

[0101] It is contemplated that each anatomical structure has distinct, quantitative image texture features that may not be visible to a human user. It is also contemplated that the differences in these image textures between the structure and the surrounding tissue can be used to identify whether a contour is valid.

[0102] The systems and methods described in the present disclosure provide a rapid and automated patient-specific contour verification method using quantitative image texture features of intraday images or other medical images of an object. As an example, for each test contour, an inner shell and an outer shell may be generated. For example, the inner shell and the outer shell may be generated by cutting and expanding the contour by a set amount (e.g., 3 mm), respectively. Texture features based on a second-order gray-level co-occurrence matrix ("GLCM") for these inner shells and outer shells may be calculated, and then the texture features may be normalized to the central area (e.g., 20% of the volume) in the corresponding contour. Accurate contours may be identified based on whether the texture values ​​for the volume enclosed by the inner shell are consistent and whether the texture values ​​for the volume enclosed by the outer shell are significantly different. Otherwise, the contour will be identified as inaccurate. An artificial intelligence ("AI")-based model may be constructed and trained to access the texture properties of the core area, the inner shell area, and the outer shell area. As an example, the AI-based model may be a machine learning ("ML")-based model constructed and trained based on supervised learning. As another example, the AI-based model may be a decision tree model. For example, a ML-based model can implement a recursive random forest classification algorithm to rank image features and select the highest ranked features. Supervised classification models can also be implemented for contour verification. Multi-feature criteria can be derived from an accurate contour database to verify the core area. For outer shell and inner shell assessment, suitable texture features can be identified as those with large different feature values ​​between the organ and the surrounding tissue. Receiver operating characteristic ("ROC") curves can be used to determine thresholds for selected features to distinguish accurate contours from inaccurate contours.

[0103] Contours classified as accurate can be confidently used directly for adaptive replanning and treatment, while inaccurate contours require an additional review process by a clinician or correction using the methods described in the present disclosure. During online adaptation, the contour verification method described in the present disclosure can replace the time-consuming manual process of checking contours, thereby facilitating the routine practice of OLAR.

[0104] Reference now Figure 7 , a flow chart is shown to illustrate the steps of an example method for automated contour verification. The method includes a model training phase 702 and a contour evaluation phase 704. The model training phase 702 may be an offline model training phase.

[0105] The contour data is accessed using a computer system, as indicated in step 706. The contour data typically includes an image of the object and corresponding accurate contours and inaccurate contours. For example, the contour data may include automatically generated organ contours and manually generated live contours. As an example, accurate contours may be identified based on multiple criteria, such as having a Dyss similarity coefficient ≥ 0.85, an average distance to agreement ≤ 1.5 mm, and 95 percentile of the distance to agreement ≤ 5 mm.

[0106] The contour data is accessed using a computer system by retrieving the contour data from a memory or other storage device or medium. In some cases, accessing the contour data may include acquiring images using a medical imaging system and communicating or otherwise transferring the images to the computer system. Additionally or alternatively, accessing the contour data may include generating contours from the medical images and communicating or otherwise transferring those contours to the computer system.

[0107] Subsequently, the profile data is divided into training data and test data, as indicated in step 708. The training data will be used to train the AI-based model for evaluating the test data. As an example, the profiles in the profile data may be randomly separated so that a first percentage of the profile data is allocated as training data and a second percentage of the profile data is allocated as test data. For example, two-thirds of the profile data may be used as training data and one-third of the profile data may be used as test data.

[0108] In the model training stage 702, the input images and contours in the training data may be preprocessed at step 710 to facilitate later texture feature calculations. The purpose of the preprocessing step is to prepare the input images and contours for texture extraction and otherwise improve image quality and training efficiency of AI-based models. Because texture is unique and uniform within a given organ, parts with known different textures should be excluded from texture extraction to avoid deteriorating the final result.

[0109] As an example, a CT image may be thresholded within the range [-200HU, 300HU] to include all soft tissues but exclude bones and air. Additionally or alternatively, for pancreatic patients, a threshold of 100–150HU may be used to exclude duodenal stents and calcifications from the pancreatic head, because pancreatic tissue typically has a CT number of less than 150HU, while the CT number of duodenal stents is generally greater than this value. On the other hand, the gas content within the organ may have a serious impact on the calculated texture. Subsequently, a threshold of -30HU may be applied to all selected organs to exclude gas areas. After setting the threshold, a morphological opening and closing filter is used to further smooth the organ contours. Additionally or alternatively, the image intensity values ​​may be rescaled, such as to reduce the number of grayscales. As an example, the image intensity values ​​may be rescaled to 250 grayscales (bin width 2HU), 500 grayscales, and the like. MRI images may be similarly thresholded using appropriate thresholding techniques.

[0110] Only voxels inside the contour can be considered to generate the organ mask to increase computational efficiency. As another example, to increase computational efficiency, the bounding box can be defined as the smallest rectangular box that completely encloses all contour points. Subsequently, the image can be cropped by adding an expansion margin (such as an expansion margin of 25 mm) to the bounding box.

[0111] To reduce image noise, bilateral filtering using two 3×3 Gaussian kernels may be applied to the cropped image. The standard deviation of the kernel in the spatial and intensity domains may be selected to be 1 and 0.5, respectively. This filter type was selected due to its ability to preserve sharp edges. It should be understood that other denoising filters or other denoising algorithms may also be implemented.

[0112] Subsequently, a mask is generated from each contour in the training data, and the mask is subdivided into three sub-regions (e.g., core, inner shell, outer shell), as indicated in step 712. In order to obtain the core region or central region of a given organ contour in the training data, a binary mask can be first constructed by assigning a value of 0 to any pixel inside the contour and assigning a value of 1 to pixels outside the contour. Subsequently, a distance map (e.g., the distance from each pixel contained in the mask to the contour surface) can be generated using a distance metric (such as Euclidean distance) between a pixel and the nearest boundary pixel. The central core region can be determined based on the percentage (e.g., 20%, 30%) of pixels with the maximum value in the distance map. An inner shell and an outer shell can be generated for each test contour. As an example, an inner shell and an outer shell can be generated by cutting and dilating an organ mask by a selected amount (e.g., 4mm) on each 2D slice. A shell thickness of 4mm can be selected to identify contour errors greater than a typical 3mm clinical tolerance.

[0113] Image features (e.g., texture features, shape features) are then calculated or otherwise extracted from the core region, the inner shell region, and the outer shell region, as indicated at step 714. As a non-limiting example, the image features may include some or all of the following image features: image features based on intensity histograms, including mean, standard deviation, range, slope, and kurtosis; image features based on gray-level co-occurrence (GLC); and image features based on gray-level run-length (GLR).

[0114] The GLC and GLR matrices can be calculated by accumulating several displacement angles on each slice to avoid directional dependence and then summing all 2D slices. As an example, four angles (e.g., 0°, 45°, 90°, and 135°) can be used. As another example, nine angles (e.g., 0°, 45°, 90°, 135°, 180°, 235°, 270°, 315°, and 360°) can be used. Image features (e.g., texture features) can be extracted from each of the three sub-regions, respectively. The distance between two pixels used to calculate the GLC matrix can be selected to be 1.

[0115] GLC-based features can be extracted from the GLC matrix ("GLCM"). The corresponding equations for some texture features are listed in Table 1 below. In the example study, three test cases were randomly selected and inner and outer shells were generated from the original contours using different intervals to include different amounts of surrounding tissue in the feature calculations. Suitable features should have high variance between organs and surrounding tissues. Thereafter, during the model training process, inner and outer shells can be generated for each test contour by cutting and extending the contour by a set amount (e.g., 3mm, 4mm). Subsequently, the selected texture features can be calculated for both the inner and outer shells and used for contour quality verification.

[0116] Table 1 Overview of texture features based on GLCM

[0117]

[0118]

[0119]

[0120] In most texture analysis studies, in order to reduce bias and extract meaningful information during statistical analysis, objects are usually selected according to the same scanning protocol (such as the same scanner, the same scanning tube voltage and current, etc.). However, in the systems and methods described in the present disclosure, a general organ contour quality inspection routine is provided for practical use, regardless of the scanning protocol or even different organ states (e.g., normal or abnormal). Even for the same organ, these factors can affect the texture feature values. In order to eliminate machine effects, scan effects, and patient-specific effects, the feature values ​​calculated from each shell can be normalized. As an example, the feature values ​​calculated from each shell can be normalized to the central core area in the corresponding contour (e.g., 20% of the volume).

[0121] Subsequently, an AI-based model is constructed and used to evaluate the image feature attributes of the core region, the inner shell region, and the outer shell region, as indicated in step 716. As described above, in some cases, the AI-based model can be an ML-based model that implements supervised learning. As a non-limiting example, the AI-based model is a decision tree model that can be trained to access the texture attributes of the core region, the inner shell region, and the outer shell region. Figure 8 An example of a decision tree model is shown in Figure 1. Decision tree models provide an effective method for data mining. In general, decision tree models partition data into subsets by asking a series of questions about attributes associated with the data. Each node on the tree contains a question and has a "yes" child node and a "no" child node. In an example, Figure 8 As shown, a three-level decision tree is established to evaluate the texture feature attributes of the core region, the inner shell region, and the outer shell region. The contours that meet all attribute questions from the three levels (i.e., "yes" nodes) are identified as accurate; otherwise, the contours will be determined to be inaccurate and reported for further verification.

[0122] Figure 8 The first level 802 of the example decision tree model shown verifies the accuracy of the generated core region. The distribution of each feature of the core region from the actual contour can be analyzed to derive the statistical confidence limit of the specific organ. As an example, in order to reduce the feature dimension, a correlation coefficient matrix, hierarchical clustering or both can be generated across all calculated texture features. For example, Ward's minimum variance clustering method can be used to implement hierarchical clustering. Independent or weakly correlated features can be identified and used for core region evaluation.

[0123] In the example where a correlation matrix was generated, the following equation can be used to calculate the correlation coefficient ρ:

[0124]

[0125] The operator E represents the expected (mean) value of its independent variable; μ x =E(x) and μ y = E(y) are the means of X and Y respectively; and σ x and σ y are the standard deviations of X and Y, respectively. Each matrix element represents the linear correlation between two variables. The correlation can vary from -1 (i.e., perfect negative correlation) through 0 (no correlation) to +1 (perfect positive correlation).

[0126] A dendrogram can be generated by hierarchical clustering of features using a correlation coefficient matrix. Fig. 9A and Fig. 9B Examples of system dendrograms or hierarchical cluster trees are shown for the pancreatic head outline and the duodenum outline, respectively. The x-axis in the cluster tree shows the connection distance between two features. The smaller the distance, the closer the association of the two features.

[0127] As a non-limiting example, a dissimilarity distance of 0.5 can be used as a cutoff for identifying clusters. Within each cluster, features with a minimum correlation criterion can be selected. In order to further reduce the dimension, principal component analysis ("PCA") can be used on the reduced feature set to identify a set of principal components that describe 95% of the variance in the data. In these cases, these principal components are accepted as "primary features", which describe the texture properties of the core regions of the pancreatic head and duodenum with high reproducibility and without redundancy. The values ​​of the primary features of the core region describing the actual contour can be calculated, and the range of acceptable values ​​for each feature can be defined based on the calculated minimum and maximum values. The model standard can be set to require that for any new contour seen by the model, the values ​​of all primary features in the core region must fall within the corresponding accepted range of the primary feature. The contour with any primary feature value falling outside the corresponding accepted range is marked for further inspection.

[0128] As another non-limiting example, within a cluster on the system tree diagram, the feature with the smallest variance is selected and a total of six independent (or weakly correlated) features are selected, in this example, these six independent (or weakly correlated) features are entropy, kurtosis, slope, mean, information measurement correlation and homogeneity. For a given profile, the standard for the "correct" core area can be selected so that at least 5 of the 6 selected features should be within the corresponding 95% statistical distribution range derived from the actual profile. The statistical probability (P) of the actual profile through these standards will be:

[0129]

[0130] Where operator B represents the binomial coefficient. This means that this criterion can describe 96.7% of the texture pattern of a particular organ. Profiles that fail to meet these criteria (i.e., at least 2 of the 6 features are outside the 95% statistical distribution range) will be identified as inaccurate.

[0131] Figure 8 The second level 804 and third level 806 of the illustrated example decision tree identify features that can best distinguish the organ of interest from surrounding tissue. To this end, multiple shell volumes can be generated for the ground truth contours, and texture features can be calculated for each shell. As a non-limiting example, a paired T-test can be used to compare the feature values ​​of the inner shells of all ground truth contours with the feature values ​​of the outer shells to identify those features that have significant differences in value (e.g., p<0.01). The feature with the largest absolute difference between the inner shell and the outer shell can be selected for the decision tree model, thereby ensuring that the selected feature has the highest sensitivity to distinguish mislabeled surrounding tissue (e.g., fat or different organs).

[0132] It follows that the criteria for passing the second level 804 and the third level 806 of the decision tree are that the difference between the texture feature value of the inner shell and the texture feature value of the core should be less than a defined threshold (second level 804) and the difference between the texture feature value of the outer shell and the texture feature value of the core should be greater than a defined value (third level 806). The difference in texture feature values ​​ΔTF can be defined as:

[0133]

[0134] Figures 10A-10D A representative test case for texture feature selection is shown in . For both the pancreatic head and duodenum contours, the shell volume contracts inwards at intervals of maximum 10 mm and expands outwards at maximum 20 mm. Fig. 10C and Fig. 10D The x-axis in represents the interval of expansion (positive) or contraction (negative) from the original outline. All eigenvalues ​​are normalized to one eigenvalue extracted from the original organ volume. For the inner shell, all relative eigenvalues ​​are stable and close to 1. For the outer shell, the GLCM cluster prominence shows the most significant differences between all features. Even the 1 mm outer shell shows a significant change in value compared to the original organ volume.

[0135] The equation for the GLCM cluster prominence ("PCLP") calculation is shown in Table 1 above. The GLCM cluster prominence is the fourth power of the change between the matrix values ​​and the intensity-weighted average. Due to the mathematical properties of the GLCM cluster prominence, the GLCM cluster prominence is very sensitive to large deviant voxel pairs with high / low intensity (such as the adipose tissue around the pancreatic head). Therefore, in some cases, the GLCM cluster prominence may be selected as a suitable feature for evaluating texture features in the inner shell and outer shell. For simplicity, the GLCM cluster prominence of the inner shell and outer shell may be referred to as InClusP and OutClusP, respectively.

[0136] Reference again Figure 7 , the parameters of the AI-based model are optimized, as indicated in step 718. For example, ROC curve analysis can be used to select the optimal thresholds for the inner shell and the outer shell. A ROC curve can be created by plotting the sensitivity against (1-specificity) at various threshold settings for selected texture features (such as InClusP and OutClusP). The equations for sensitivity and specificity are as follows:

[0137]

[0138]

[0139] Where TP, TN, FP, and FN are true positives, true negatives, false positives, and false negatives, respectively. Sensitivity is the ratio of correct contours that are correctly identified as correct, while specificity represents the ratio of inaccurate contours that are correctly identified as inaccurate. To evaluate model performance, the false negative rate (“FNR”), false positive rate (“FPR”), and accuracy can also be calculated from the ROC curve:

[0140] FNR = 1 - sensitivity (12);

[0141] FPR = 1-specificity (13);

[0142]

[0143] After core region validation, ROC curves can be obtained to determine the threshold for each selected texture feature. A first ROC curve is generated by discretely increasing the threshold of the inner shell feature value from the minimum value of the live database to the maximum value of the live database. As the threshold increases, a greater number of accurate contours will be correctly identified (i.e., sensitivity increases), and likewise, a greater number of inaccurate contours will be misidentified (i.e., specificity decreases). A peak can be observed in the plot of accuracy (i.e., the average of sensitivity and specificity) versus threshold. This property can be used to automatically identify the optimal threshold T 内 , the optimal threshold T内 can be defined as the threshold value with the greatest accuracy. 内 When the shell is fixed, the threshold can be increased incrementally to obtain the second ROC curve. 内 The same method is used to determine the optimal threshold T 外 .

[0144] As another example, a first ROC curve can be generated by discretely increasing the threshold of the inner shell feature value. Subsequently, the optimal threshold can be determined as the threshold corresponding to the point with the highest achievable sensitivity and specificity of the test. A second ROC curve can be obtained by applying a predefined threshold and discretely increasing the threshold of the outer shell feature value. The optimal threshold of the outer shell feature value can be determined based on maintaining the specificity above a selected percentage (e.g., 90%, corresponding to an FPR of less than 10%) and achieving the highest possible sensitivity. The determined threshold can then be used in a decision tree model to determine the accuracy of any new profile.

[0145] Finally, the model features are prepared for automatic contour quality assurance. At the contour assessment stage 704, the identified clinically acceptable good contours can be immediately used for further replanning and treatment, while the reported inaccurate contours can be sent to the physician for manual modification or correction using the contour correction method described in the present disclosure.

[0146] The above method of using quantitative texture features to evaluate and verify the quality of contours can quickly identify accurate contours or inaccurate contours with high sensitivity and specificity. During online adaptation, this method can replace the time-consuming manual process of checking contours, thereby facilitating the routine practice of OLAR.

[0147] Reference now Fig.11 , shows an example of a system 1100 for automatic radiation therapy plan profile verification and correction according to some embodiments of the systems and methods described in the present disclosure. Fig.11 As shown, computing device 1150 may receive one or more types of data (e.g., image data, contour data) from image and contour data source 1102, which may include CT images, magnetic resonance images, other medical images, or a combination thereof, and a source of contours generated from such images. In some embodiments, computing device 1150 may execute at least a portion of automatic contour verification and correction system 1104 to automatically verify contours, automatically correct contours, or both based on data received from image and contour data source 1102.

[0148] Additionally or alternatively, in some embodiments, computing device 1150 may communicate information about data received from image and contour source 1102 to server 1152 via communication network 1154, which may execute at least a portion of automatic contour verification and correction system 1104. In such embodiments, server 1152 may return information indicative of the output of automatic contour verification and correction system 1104 to computing device 1150 (and / or any other suitable computing device).

[0149] In some embodiments, computing device 1150 and / or server 1152 may be any suitable computing device or combination of devices, such as a desktop computer, a laptop computer, a smart phone, a tablet computer, a wearable computer, a server computer, a virtual machine executed by a physical computing device, etc. Computing device 1150 and / or server 1152 may also reconstruct an image from the data.

[0150] In some embodiments, the image and contour data source 1102 can be any suitable source of image data (e.g., measurement data, images reconstructed from measurement data), such as a CT system, an MRI system, another computing device (e.g., a server storing image data), etc. In some embodiments, the image and contour data source 1102 can be local to the computing device 1150. For example, the image and contour data source 1102 can be integrated with the computing device 1150 (e.g., the computing device 1150 can be configured as part of a device for capturing, scanning, and / or storing images). As another example, the image and contour data source 1102 can be connected to the computing device 1150 via a cable, a direct wireless link, etc. Additionally or alternatively, in some embodiments, the image and contour data source 1102 is located locally and / or remotely from the computing device 1150, and can communicate data to the computing device 1150 (and / or the server 1152) via a communication network (e.g., the communication network 1154).

[0151] In some embodiments, the communication network 1154 may be any suitable communication network or combination of communication networks. For example, the communication network 1154 may include a Wi-Fi network (a Wi-Fi network may include one or more wireless routers, one or more switches, etc.), a peer-to-peer network (e.g., a Bluetooth network), a cellular network (e.g., a 3G network, a 4G network, etc. that conforms to any suitable standard (such as CDMA, GSM, LTE, Advanced LTE, WiMAX, etc.)), a wired network, etc. In some embodiments, the communication network 108 may be a local area network, a wide area network, a public network (e.g., the Internet), a private or semi-private network (e.g., a company or university intranet), any other suitable type of network, or any suitable combination of networks. Fig.11The communication links shown may each be any suitable communication link or combination of communication links, such as a wired link, a fiber optic link, a Wi-Fi link, a Bluetooth link, a cellular link, or the like.

[0152] Reference now Fig.12 , shows an example of hardware 1200 that may be used to implement image and contour data source 1102, computing device 1150, and server 1154 according to some embodiments of the systems and methods described in the present disclosure. Fig.12 As shown, in some embodiments, computing device 1150 may include a processor 1202, a display 1204, one or more inputs 1206, one or more communication systems 1208, and / or a memory 1210. In some embodiments, processor 1202 may be any suitable hardware processor or combination of processors, such as a central processing unit ("CPU"), a graphics processing unit ("GPU"), and the like. In some embodiments, display 1204 may include any suitable display device, such as a computer monitor, a touch screen, a television, and the like. In some embodiments, input 1206 may include any suitable input device and / or sensor that may be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, and the like.

[0153] In some embodiments, the communication system 1208 may include any suitable hardware, firmware, and / or software for communicating information over the communication network 1154 and / or any other suitable communication network. For example, the communication system 1208 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 1208 may include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0154] In some embodiments, memory 1210 may include any suitable one or more storage devices that can be used to store instructions, values, data, etc., which can be used, for example, by processor 1202 to present content using display 1204, to communicate with server 1152 via communication system(s) 1208, and the like. Memory 1210 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 1210 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, and the like. In some embodiments, memory 1210 may have a computer program encoded thereon or otherwise stored therein for controlling the operation of computing device 1150. In such embodiments, processor 1202 may execute at least a portion of the computer program to present content (e.g., images, user interfaces, graphics, tables), receive content from server 1152, send information to server 1152, and the like.

[0155] In some embodiments, server 1152 may include a processor 1212, a display 1214, one or more inputs 1216, one or more communication systems 1218, and / or memory 1220. In some embodiments, processor 1212 may be any suitable hardware processor or combination of processors, such as a CPU, GPU, etc. In some embodiments, display 1214 may include any suitable display device, such as a computer monitor, a touch screen, a television, etc. In some embodiments, input 1216 may include any suitable input device and / or sensor that can be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, etc.

[0156] In some embodiments, the communication system 1218 may include any suitable hardware, firmware, and / or software for communicating information over the communication network 1154 and / or any other suitable communication network. For example, the communication system 1218 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 1218 may include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0157] In some embodiments, the memory 1220 may include any suitable one or more storage devices that can be used to store instructions, values, data, etc., which can be used, for example, by the processor 1212 to present content using the display 1214, to communicate with one or more computing devices 1150, etc. The memory 1220 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, the memory 1220 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, etc. In some embodiments, the memory 1220 may have a server program encoded thereon for controlling the operation of the server 1152. In such embodiments, the processor 1212 may execute at least a portion of the server program to send information and / or content (e.g., data, images, user interfaces) to one or more computing devices 1150, receive information and / or content from one or more computing devices 1150, receive instructions from one or more devices (e.g., personal computers, laptops, tablet computers, smart phones), etc.

[0158] In some embodiments, the image and contour data source 1102 may include a processor 1222, one or more image acquisition systems 1224, one or more communication systems 1226, and / or a memory 1228. In some embodiments, the processor 1222 may be any suitable hardware processor or combination of processors, such as a CPU, GPU, and the like. In some embodiments, the one or more image acquisition systems 1224 are generally configured to acquire data, images, or both, and may include a CT system and an MRI system or both. Additionally or alternatively, in some embodiments, the one or more image acquisition systems 1224 may include any suitable hardware, firmware, and / or software for coupling to and / or controlling the operation of the CT system and the MRI system or both. In some embodiments, one or more portions of the one or more image acquisition systems 1224 may be removable and / or replaceable.

[0159] Note that, although not shown, the image and contour data source 1102 may include any suitable input and / or output. For example, the image and contour data source 1102 may include an input device and / or sensor that can be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, a trackpad, a trackball, etc. As another example, the image and contour data source 1102 may include any suitable display device (such as a computer monitor, a touch screen, a television, etc.), one or more speakers, etc.

[0160] In some embodiments, the communication system 1226 may include any suitable hardware, firmware, and / or software for communicating information to the computing device 1150 (and, in some embodiments, via the communication network 1154 and / or any other suitable communication network). For example, the communication system 1226 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 1226 may include hardware, firmware, and / or software that can be used to establish a wired connection using any suitable port and / or communication standard (e.g., VGA, DVI video, USB, RS-232, etc.), a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0161] In some embodiments, memory 1228 may include any suitable one or more storage devices that may be used to store instructions, values, data, and the like, which may be used, for example, by processor 1222 to: control and / or receive data from one or more image acquisition systems 1224; to generate images from the data; to present content (e.g., images, user interfaces) using a display; to communicate with one or more computing devices 1150; and the like. Memory 1228 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 1228 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, and the like. In some embodiments, memory 1228 may have a program encoded thereon or otherwise stored therein for controlling the operation of image source 1102. In such embodiments, the processor 1222 may execute at least a portion of a program to generate images, send information and / or content (e.g., data, images) to one or more computing devices 1150, receive information and / or content from one or more computing devices 1150, receive instructions from one or more devices (e.g., a personal computer, laptop computer, tablet computer, smart phone, etc.), and so on.

[0162] In some embodiments, any suitable computer-readable medium may be used to store instructions for performing the functions and / or processes described herein. For example, in some embodiments, the computer-readable medium may be transient or non-transient. For example, non-transient computer-readable media may include the following media, such as: magnetic media (e.g., hard disk, floppy disk), optical media (e.g., compact disk, digital video disk, Blu-ray disk), semiconductor media (e.g., random access memory ("RAM"), flash memory, electrically programmable read-only memory ("EPROM"), electrically erasable programmable read-only memory ("EEPROM")), any suitable medium that is not transient or does not have any persistent appearance during transmission, and / or any suitable tangible medium. As another example, transient computer-readable media may include signals on a network, signals on wires, conductors, optical fibers, circuits, or any suitable medium that is transient and does not have any persistent appearance during transmission, and / or any suitable intangible medium.

[0163] As described above, for OLAR, currently available methods of plan quality assessment by visual inspection or automated inspection of dose volume metrics and visual inspection of 3D dose distribution can be time consuming. Conventional plan assessment based on dose metrics from wish lists and dose volume histograms ("DVH") lacks spatial information; therefore, physicians must examine the dose distribution slice by slice to infer plan quality with subjective decisions. The systems and methods described in the present disclosure implement metrics based on dose spatial distribution (e.g., DVH metrics) and texture features to establish an objective method for automatic assessment of plan quality.

[0164] In general, the methods described in the present disclosure include calculating multiple texture features and metrics (e.g., DVH metrics) from each 3D dose distribution to be compared. Four categories of features calculated from the spatial dose distribution can be used in the methods described in the present disclosure. These features include: a dose level size zone matrix ("DLSZM") for emphasizing large continuous areas; a dose level co-occurrence matrix ("DLCM") that defines the dose level relationship between the same spatial voxels of two plans; a mutual objective function distance ("MOFD") that evaluates the distance between the planned dose distribution and the target dose distribution in the treatment planning system; and a neighbor level-dose difference matrix ("neighbor level-dose difference matrix) that is similar to human perception of images to indicate local dose variations. As an example, other metrics that can be calculated from these features include: a large cold spot ("LCS") that can be calculated from the DLSZM; and thermal contrast ("HC"), which can be calculated from the DLCM and evaluates the heat from one plan to another at the same spatial voxel. A superior or otherwise optimal result is determined from the set of plans being compared based on the differences in these metrics and characteristics between the plans being compared.

[0165] Online ART provides optimal therapeutic dose delivery with changes between doses of organs. Compared with IGRT repositioning, OLAR provides superior adaptive plans by improving target coverage and / or OAR preservation. Using current technology, OLAR is implemented using human interaction during the entire workflow. Part of these human operations is plan evaluation. In the OLAR workflow, there are usually two plans from which the physician selects the superior plan. The first plan is an adaptive plan, which is a plan adjusted with or without full optimization of daily anatomical changes. The second plan is a repositioning plan, which is a recalculation of the reference plan based on the registration between the reference image and the daily image. The physician selects a superior plan to achieve the optimal dose benefit for the daily dose of treatment.

[0166] Treatment plan evaluation is a clinically subjective decision-making problem that involves visual search and analysis in a contextually rich environment, including circled structures and isodose lines superimposed on CT data. This labor-intensive task is usually performed by oncologists and physicists and is limited by the experience, skills, and capabilities of the oncologist or physicist.

[0167] Similar to a three-dimensional image, the dose level of each voxel is similar to the intensity of each voxel in a spatial dose distribution (such as a dose map), which can be a 2D dose map or a 3D dose map. The dose texture features of the dose map can be defined and calculated similarly to the dose texture features from the image.

[0168] As an example of a texture feature or metric that may be calculated from a spatial distribution of dose (such as a 3D dose map), LCS may be calculated as:

[0169]

[0170] where M(i,z) is an N x S matrix with elements that are the frequencies of consecutive voxels of size z and dose level i.

[0171] As another example of a texture feature or metric that can be calculated from a spatial distribution of dose (such as a 3D dose map), HC can be calculated from a DLCM. The DLCM for the first plan ("Plan 1") and the second plan ("Plan 2") is:

[0172]

[0173] Where F(i,j) is an N×N matrix with elements that are the frequencies at the first dose level i in the first plan and the frequencies at the second dose level j in the second plan at the same voxel. The thermal contrast between the first plan and the second plan is:

[0174] For i∈Plan1 and j∈Plan2 (17).

[0175] As another example of a texture feature or metric that may be calculated from a spatial distribution of dose (such as a 3D dose map), the inverse covariance (“IC”) may be calculated as:

[0176] For i∈Plan1 and j∈Plan2 (18).

[0177] As another example of a texture feature or metric that may be computed from a spatial distribution of dose (such as a 3D dose map), a neighbor level dose difference matrix (“NLDDM”) may be computed based on defining the local difference between each voxel and the average of its neighbors:

[0178]

[0179] where Z is the set of all neighbors of a voxel at dose level i, and where p j is the probability of dose level j.

[0180] As another example of a texture feature or metric that may be calculated from a spatial distribution of dose (such as a 3D dose map), coarseness (“Co”) may be defined as:

[0181]

[0182] As another example of a texture feature or metric that may be calculated from a spatial distribution of dose (such as a 3D dose map), busyness (Bu) may be defined as:

[0183]

[0184] As an example of a texture feature or metric that can be calculated from a spatial distribution of dose (such as a 3D dose map), MOFD can be calculated as follows. First, define an objective function distance to the target. For prescription doses greater than the target dose (i.e., p j >g j ), the objective function distance can be calculated as:

[0185]

[0186] For prescription doses less than the target dose (i.e., p j <g j ), the objective function distance can be calculated as:

[0187]

[0188] Next, the objective function of OAR in the first plan (eg, adaptive plan “A”) and the second plan (eg, repositioning plan “R”) may be defined as:

[0189]

[0190] The parameter β can be determined heuristically in the treatment planning system 1 -β 6 The MOFD for the first and second plans can then be defined as:

[0191]

[0192] where k and s are the number of targets and OARs.

[0193] The quality of the plan can be assessed based on the texture features and metrics described above and other texture features and metrics calculated from the spatial distribution of dose (e.g., 3D dose maps). To assist in this process, a general quality index ("GQI") can be defined as a quantitative metric for clinical decision making in evaluating a treatment plan:

[0194] GQI=α 1 HC 1→2 +α 2 IC 1→2 +α 3 LCS+α 4 Co+α 5 Bu+α 6 MOFD 1→2 (26);

[0195] All coefficients can be determined from the live data set such that plans with lower GQI have superior plan quality.

[0196] In some implementations, a machine learning algorithm implemented using a hardware processor and memory may be used to assess the quality of a radiotherapy plan based on one or more texture features or metrics calculated from a 3D dose map. As an example, a boosted decision tree algorithm with hyperparameter optimization may be implemented to determine the best dose texture features and a range of those parameter values ​​that may be used to identify better plans between competing plans. In such implementations, several dose texture features obtained from multiple sets of different radiotherapy plan types (e.g., adaptive plans, repositioning plans) may initially be used as predictors along with more commonly used dose volume parameters. An initial classifier including a complex tree, discriminant analysis, support vector machine, logistic regression, nearest neighbor, ensemble classification, boosted decision tree, or a combination thereof may be trained using all dose texture features and cross validation (e.g., leave one out cross validation).

[0197] In an example study, planning CT and daily CT were acquired during RT for prostate cancer patients using on-orbit CT. In this study, 11 split-dose daily images were randomly selected in five patients. For each daily CT set, both IGRT repositioning plans and adaptive plans were generated. Daily contours were created and edited based on the daily CT. The repositioning plan was recalculated from the reference plan based on the new structural positions. Fig.13 Examples of dose volume histograms ("DVH") for two plans (an adaptive plan and a repositioned plan) are shown. Fig.14 The DLCM of all OARs for both plans are shown. The co-occurrence frequency is mainly present in the upper right part, indicating that the repositioned plan has higher dose than the adaptive plan at the same location of the OAR. Fig.15 DLSZM of the planning target volume ("PTV") in adaptive planning and repositioning planning is shown.

[0198] Each metric has a physical meaning of the metric itself that is connected to the plan quality. LCS is an index showing large connected areas with low dose. Larger LCS values ​​indicate more or larger low dose connected areas in the structure. In a PTV, a large LCS value will be associated with a poor quality of the plan.

[0199] HC and IC show the cumulative difference in the same location of two different plans. A larger value indicates that one plan is generally hotter than the other plan in the same voxel. In OARs, this value indicates a lower quality plan. Due to the direct comparison with the same location, there is no ambiguous interpretation of the heat of the structure in the plan.

[0200] Roughness and busyness indicate local dose level variations and give a human vision-like perception of the image. High values ​​correspond to a rougher image in vision. For dose distribution, this can be interpreted as a more island-like spatial distribution with lower plan quality.

[0201] MOFD indicates the difficulty of plan evaluation when the quality of plans varies in space. If the quality of a first plan (e.g., "Plan A") is worse than a second plan (e.g., "Plan B") in a first position (e.g., "Position 1"), but worse than the second plan in a second position (e.g., "Position 2"), MOFD can give an absolute quality value by measuring the distance between the planned dose distribution and the optimal dose distribution.

[0202] The GQI provides a general indicator of plan quality. The GQI can be calculated from the GQI formula in equation (26) using empirical parameters. Through regression analysis, accurate coefficients can be extracted from the live data set.

[0203] Reference now Fig.16 , a flow chart is shown setting forth steps of an example method for evaluating the quality of different available radiation therapy plans based on calculating and comparing texture features, metrics, or both from spatial dose distributions associated with the plans. The method includes accessing a first radiation therapy plan using a computer system, as indicated by step 1602. A second radiation therapy plan is also accessed by the computer system, as indicated by step 1604. Accessing the radiation therapy plans may include, for example, retrieving a previously generated radiation therapy plan from a memory or other data storage.

[0204] Subsequently, one or more texture features, one or more metrics, or both are calculated from the spatial dose distributions contained in the first radiation therapy plan and the second radiation therapy plan, as generally indicated at step 1606. As described above, the texture features, metrics, or both may include DLSZM-LCS, DLCM-HC, DLCM-IC, NLDDM-Co, MOFD, GQI, combinations thereof, or other texture features or metrics. Subsequently, the calculated texture features, metrics, or both are analyzed to evaluate the quality of the first plan and the second plan; thereby, the quality of the first plan is evaluated at step 1608 and the quality of the second plan is evaluated at 1610. In some implementations, the quality of the two plans may be evaluated simultaneously.

[0205] Based on the evaluation, a plan with a higher quality is selected, as indicated at step 1612 , and the plan is stored for later use or provided to a radiation therapy planning system to control the operation of the radiation therapy system, as indicated at step 1614 .

[0206] Thus, a method for automatically and objectively evaluating dosimetry plan quality using dose texture features extracted from spatial dose distributions has been described. The optimal dose texture features can be determined based on a machine learning algorithm. This dose texture-based plan evaluation method can be implemented for OLAR to significantly reduce the time of plan evaluation.

[0207] Reference now Fig.17 , shows an example of a system 1700 for automatically evaluating the quality of OLAR plans according to some embodiments of the systems and methods described in the present disclosure. Fig.17 As shown, computing device 1750 may receive one or more types of data (e.g., radiation therapy plan data, image data) from data source 1702, which may be a radiation therapy plan data source. In some embodiments, computing device 1750 may execute at least a portion of OLAR plan quality assessment system 1704 to automatically assess OLAR plan quality based on data received from data source 1702.

[0208] Additionally or alternatively, in some embodiments, computing device 1750 may communicate information about data received from data source 1702 to server 1752 via communications network 1754, which may execute at least a portion of OLAR plan quality assessment system 1704. In such embodiments, server 1752 may return information indicative of output of OLAR plan quality assessment system 1704 to computing device 1750 (and / or any other suitable computing device).

[0209] In some embodiments, computing device 1750 and / or server 1752 may be any suitable computing device or combination of devices, such as a desktop computer, a laptop computer, a smart phone, a tablet computer, a wearable computer, a server computer, a virtual machine executed by a physical computing device, etc. Computing device 1750 and / or server 1752 may also reconstruct an image from the data.

[0210] In some embodiments, the data source 1702 may be any suitable source of radiation therapy planning data (e.g., measurement data, images reconstructed from the measurement data), such as a CT system, an MRI system, or both; another computing device (e.g., a server storing radiation therapy planning data and / or image data); and the like. In some embodiments, the data source 1702 may be local to the computing device 1750. For example, the data source 1702 may be combined with the computing device 1750 (e.g., the computing device 1750 may be configured as part of a device for capturing, scanning, and / or storing images). As another example, the data source 1702 may be connected to the computing device 1750 via a cable, a direct wireless link, and the like. Additionally or alternatively, in some embodiments, the data source 1702 is located locally and / or remotely relative to the computing device 1750, and may communicate data to the computing device 1750 (and / or the server 1752) via a communication network (e.g., the communication network 1754).

[0211] In some embodiments, the communication network 1754 may be any suitable communication network or combination of communication networks. For example, the communication network 1754 may include a Wi-Fi network (a Wi-Fi network may include one or more wireless routers, one or more switches, etc.), a peer-to-peer network (e.g., a Bluetooth network), a cellular network (e.g., a 3G network, a 4G network, etc. that conforms to any suitable standard (such as CDMA, GSM, LTE, Advanced LTE, WiMAX, etc.)), a wired network, etc. In some embodiments, the communication network 108 may be a local area network, a wide area network, a public network (e.g., the Internet), a private or semi-private network (e.g., a company or university intranet), any other suitable type of network, or any suitable combination of networks. Fig.17 The communication links shown may each be any suitable communication link or combination of communication links, such as a wired link, a fiber optic link, a Wi-Fi link, a Bluetooth link, a cellular link, or the like.

[0212] Reference now Fig.18 , shows an example of hardware 1800 that may be used to implement an image source 1702, a computing device 1750, and a server 1754 according to some embodiments of the systems and methods described in the present disclosure. Fig.18As shown, in some embodiments, computing device 1750 may include a processor 1802, a display 1804, one or more inputs 1806, one or more communication systems 1808, and / or memory 1810. In some embodiments, processor 1802 may be any suitable hardware processor or combination of processors, such as a central processing unit ("CPU"), a graphics processing unit ("GPU"), and the like. In some embodiments, display 1804 may include any suitable display device, such as a computer monitor, a touch screen, a television, and the like. In some embodiments, input 1806 may include any suitable input device and / or sensor that may be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, and the like.

[0213] In some embodiments, the communication system 1808 may include any suitable hardware, firmware, and / or software for communicating information over the communication network 1754 and / or any other suitable communication network. For example, the communication system 1808 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 1808 may include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0214] In some embodiments, memory 1810 may include any suitable one or more storage devices that can be used to store instructions, values, data, etc., which can be used, for example, by processor 1802 to present content using display 1804, to communicate with server 1752 via communication system(s) 1808, and the like. Memory 1810 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 1810 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, and the like. In some embodiments, memory 1810 may have a computer program encoded thereon or otherwise stored therein for controlling the operation of computing device 1750. In such embodiments, processor 1802 may execute at least a portion of the computer program to present content (e.g., images, user interfaces, graphics, tables), receive content from server 1752, send information to server 1752, and the like.

[0215] In some embodiments, server 1752 may include a processor 1812, a display 1814, one or more inputs 1816, one or more communication systems 1818, and / or memory 1820. In some embodiments, processor 1812 may be any suitable hardware processor or combination of processors, such as a CPU, GPU, etc. In some embodiments, display 1814 may include any suitable display device, such as a computer monitor, a touch screen, a television, etc. In some embodiments, input 1816 may include any suitable input device and / or sensor that can be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, etc.

[0216] In some embodiments, the communication system 1818 may include any suitable hardware, firmware, and / or software for communicating information over the communication network 1754 and / or any other suitable communication network. For example, the communication system 1818 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 1818 may include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0217] In some embodiments, the memory 1820 may include any suitable one or more storage devices that can be used to store instructions, values, data, etc., which can be used, for example, by the processor 1812 to present content using the display 1814, to communicate with one or more computing devices 1750, etc. The memory 1820 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, the memory 1820 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, etc. In some embodiments, the memory 1820 may have a server program encoded thereon for controlling the operation of the server 1752. In such embodiments, the processor 1812 may execute at least a portion of the server program to send information and / or content (e.g., data, images, user interfaces) to one or more computing devices 1750, receive information and / or content from one or more computing devices 1750, receive instructions from one or more devices (e.g., personal computers, laptops, tablet computers, smart phones), etc.

[0218] In some embodiments, the data source 1702 may include a processor 1822, one or more image acquisition systems 1824, one or more communication systems 1826, and / or a memory 1828. In some embodiments, the processor 1822 may be any suitable hardware processor or combination of processors, such as a CPU, GPU, and the like. In some embodiments, the one or more image acquisition systems 1824 are generally configured to acquire data, images, or both, and may include a CT system, an MRI system, or both. Additionally or alternatively, in some embodiments, the one or more image acquisition systems 1824 may include any suitable hardware, firmware, and / or software for coupling to and / or controlling the operation of a CT system and an MRI system or both. In some embodiments, one or more portions of the one or more image acquisition systems 1824 may be removable and / or replaceable.

[0219] Note that, although not shown, data source 1702 may include any suitable input and / or output. For example, data source 1702 may include an input device and / or sensor that can be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, a trackpad, a trackball, etc. As another example, data source 1702 may include any suitable display device (such as a computer monitor, a touch screen, a television, etc.), one or more speakers, etc.

[0220] In some embodiments, the communication system 1826 may include any suitable hardware, firmware, and / or software for communicating information to the computing device 1750 (and, in some embodiments, via the communication network 1754 and / or any other suitable communication network). For example, the communication system 1826 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 1826 may include hardware, firmware, and / or software that can be used to establish a wired connection using any suitable port and / or communication standard (e.g., VGA, DVI video, USB, RS-232, etc.), a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0221] In some embodiments, memory 1828 may include any suitable one or more storage devices that may be used to store instructions, values, data, and the like, which may be used, for example, by processor 1822 to: control and / or receive data from one or more image acquisition systems 1824; to generate images from the data; to present content (e.g., images, user interfaces) using a display; to communicate with one or more computing devices 1750; and the like. Memory 1828 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 1828 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, and the like. In some embodiments, memory 1828 may have a program encoded thereon or otherwise stored therein for controlling the operation of image source 1702. In such embodiments, the processor 1822 may execute at least a portion of a program to generate images, send information and / or content (e.g., data, images) to one or more computing devices 1750, receive information and / or content from one or more computing devices 1750, receive instructions from one or more devices (e.g., a personal computer, laptop computer, tablet computer, smart phone, etc.), and so on.

[0222] In some embodiments, any suitable computer-readable medium may be used to store instructions for performing the functions and / or processes described herein. For example, in some embodiments, the computer-readable medium may be transient or non-transient. For example, non-transient computer-readable media may include the following media, such as: magnetic media (e.g., hard disk, floppy disk), optical media (e.g., compact disk, digital video disk, Blu-ray disk), semiconductor media (e.g., random access memory ("RAM"), flash memory, electrically programmable read-only memory ("EPROM"), electrically erasable programmable read-only memory ("EEPROM")), any suitable medium that is not transient or does not have any persistent appearance during transmission, and / or any suitable tangible medium. As another example, transient computer-readable media may include signals on a network, signals on wires, conductors, optical fibers, circuits, or any suitable medium that is transient and does not have any persistent appearance during transmission, and / or any suitable intangible medium.

[0223] As described above, in magnetic resonance images, it is often difficult to automatically separate air from other low signal regions (especially bone, liver, etc.) due to noise and artifacts, which can cause errors in the generation of synthetic CT ("sCT") for MRI-based RT-planning. This is problematic for the abdomen where air regions are random and therefore cannot be deformed from a reference CT. The systems and methods described in the present disclosure provide a technique that can accurately and automatically determine air regions in MRI, which in turn enables more robust generation of synthetic CT images from magnetic resonance images.

[0224] MR-Linac technology is expected to improve RT by enhancing tumor targeting. The high-contrast images that can be acquired between and during each treatment delivery using MRI allow for online and real-time plan adaptation. However, dose calculations on patient anatomy use maps of relative electron density ("rED") that cannot be obtained directly from daily MRI. Therefore, synthetic CT (sCT) images are generated based on daily MRI.

[0225] The two main types of methods for sCT generation are voxel-based methods, where only MRI voxel data is used, and atlas-based methods, where the position information as well as the intensity information of the voxels is used to generate rED values.

[0226] Voxel-based methods, in which rED data for each voxel is directly obtained based on the voxel's MR intensity, are useful for situations where no reference CT is available. However, since there is no direct relationship between the MR signal and the CT signal, voxel-based methods require multiple MR sequences to acquire images, which increases the online preparation time and is therefore undesirable for the purpose of OLAR.

[0227] Atlas-based methods involve transferring rED data directly from an existing CT of the same patient, either by rigid or deformable registration. These methods can be applied to any single MR sequence (i.e., multiple sequences are not required). Therefore, atlas-based methods are more suitable for OLAR, where a reference CT can be acquired offline. Another advantage is that atlas-based methods use the patient's own CT, and therefore, can be more accurate than voxel-based methods that apply rED data based on the general population rather than patient-specific rED data.

[0228] The Lorentz force from the magnetic field in MR-Linac requires accurate estimates of rED, especially in the lung where population-based rED data may cause errors in dose calculations. Atlas-based methods for extracranial sites often require deformable image registration ("DIR") to transfer rED data, and there is concern that DIR may be inaccurate for large deformations.

[0229] Even for extracranial sites, the systems and methods described in the present disclosure overcome these drawbacks and provide a rapid atlas-based method to generate rED data. The method can be fully automated and thus can be used for OLAR on the MR-Linac system.

[0230] Typically, the methods described in the present disclosure transfer rED density from a reference CT ("rCT") to a daily magnetic resonance image ("dMR") via DIR. The rCT is typically acquired on an earlier date, and the rCT is used to deform the rED data. The dMRI is acquired on the day of treatment, and the dMRI is used as the target image to which the rCT is deformed.

[0231] Reference now Fig.19 , a flowchart is shown to illustrate the steps of an example method for generating a synthetic CT image from a magnetic resonance image. As indicated at step 1902, the method includes accessing input image data with a computer system. Typically, the input image data includes reference CT images and daily MR images of a patient. Accessing the input image data may include retrieving previously acquired input image data from a memory or other data storage device or medium. In some cases, accessing the input image data may include acquiring images (e.g., rCT images, dMR images) using a suitable medical imaging system, and communicating or otherwise transferring those images to a computer system, which may form part of (multiple) corresponding medical imaging systems.

[0232] In the following steps, the CT intensity (i.e., Hounsfield Units ("HU")) is transferred and converted into a rED map. Individual bones are circled on the rCT image, as indicated in step 1904. In some embodiments, this step can be performed offline. In these cases, the bone contour data can be stored in a memory or other data storage and accessed by a computer system at step 1904.

[0233] The individual bone contours are then transferred from rCT to dMR. Perform the first DIR on the rCT image and the dMR image, as indicated in step 1906. As a result of this step, a deformation vector field ("DVF") is calculated. For each individual bone volume (e.g., the volume enclosed by the bone contour), a rigid registration matrix M (i.e., translation and rotation) is calculated from the DVF of the voxels inside the bone structure of each identification, as indicated in step 1908. This matrix M is the best fit from all points inside the bone volume to the DVF, so that the cumulative error between the DVF and the rigid transfer is:

[0234]

[0235] is minimized over all voxels. This matrix M can be calculated independently for each bone volume. Subsequently, the rigid registration matrix is ​​used to rigidly transfer each bone contour from the rCT to the dMR image, as indicated in step 1910.

[0236] A second DIR is performed between the rCT image and the DMR image, this time using a DIR algorithm that constrains the DVF by the bone structure, as indicated in step 1912. All voxels inside a structure in the reference image are only allowed to deform into voxels of the same structure in the target image; thereby, the structure is forced to deform into the shape circled on the target image. This preserves the shape of the bone structure; thereby, each individual bone maintains its rigidity, minimizing the error from the DIR of the bone.

[0237] The volume of random air regions (e.g., gas in the intestine) in the dMR images can be identified in the dMR images by automatic thresholding, as indicated in step 1914. Formulas related to the MRI intensity of air (outside the phantom / patient) and the surrounding materials can be established and used to automatically threshold the air regions on the MRI. These formulas can be further refined based on the patient MRI data. As an example, an air-containing region in the abdomen (e.g., colon, intestine, stomach) transferred from the reference MRI to the daily MRI and expanded by 1 cm gives a region that can be used for thresholding. The expansion takes into account possible inaccuracies in the DIR. Texture-based methods can be used to further refine the air regions determined inside the target region. Quantitative image texture features (including cross-correlation, contrast, and variance, or a combination thereof) from the gray level co-occurrence matrix ("GLCM") can be extracted from the automatically thresholded region, and a machine learning algorithm (e.g., a naive Bayes classifier) ​​implemented using a hardware processor and memory can be trained using a leave-one-out cross-validation technique to distinguish between air voxels and non-air voxels. The area under the ROC curve ("AUC") can be used to judge the performance of a classifier.

[0238] As a non-limiting example of generating air region volumes, a magnetic resonance image may be analyzed as follows. First, the image may be analyzed to roughly identify regions that enclose the most likely air / gas organs (such as the stomach, duodenum, small intestine, large intestine, colon, etc.). Then, a thresholding equation may be developed based on reference data (such as phantom data). As an example, a threshold may be calculated or otherwise determined from phantom data (such as MR images and CT images of one or more phantoms with embedded air cavities). These air cavities in the phantom are identified in both the CT image and the MRI image. The air cavity contours defined in the CT image may be transferred to the MRI image by rigid image registration to be compared with the air cavity contours delineated on the MRI image. The Dess coefficient ("DC") and the Housdorff distance are used to measure the difference between the CT-based contours and the MRI-based contours and to derive thresholds to identify air / tissue boundaries on the patient's MRI image.

[0239] As an example, the following formula was generated based on phantom studies and can be used to indicate the air area:

[0240] MR 阈值1 =(STD U )+(MAX OA ) (28);

[0241] The low signal areas outside the extended joint structure can be discarded because the probability that these low signal areas are air is very low. In equation (28), STD U is defined as the standard deviation of the strength values ​​of the joint structure, and MAX OA is the maximum intensity value of pure air extracted at a selected distance (e.g., 2 cm) outside the boundary (such as outside the front boundary of the patient). Subsequently, a second threshold formula can be used to fine-tune the air region. As an example, the following formula can be used:

[0242] MR 阈值2 =a·(MAX)+b·(MEDIAN)+c·(STD) (29);

[0243] The MR can be obtained from the best match of randomly picked MR signals to the contoured air region. 阈值2 . Statistics can be obtained from reference data (e.g., data acquired from different patients) and used to solve for coefficients a, b, and c. In some implementations, the coefficients can be solved for offline and stored for later use. The coefficients can be used to calculate the final threshold using equation (29).

[0244] To further refine the air boundaries on MRI, image texture features of air and non-air regions can be extracted on a voxel-by-voxel basis. Several different texture features of the GLCM can be extracted to identify the characteristic values ​​of air and non-air regions. In one example, 23 such features were extracted and used for this purpose. Pure air voxels and non-air voxels can be identified by the thresholding method described above. Fig. 20 The method for extracting air / non-air voxels from the automatically thresholded air region is described in 1 ) to identify the pure air region (r 2 ) to perform texture analysis.

[0245] In an example study, a Student's T test was used to identify features that were significantly different in air regions from non-air regions. A combination of the most significant features was tested using a naive Bayes classifier and a leave-one-out cross-validation technique to determine the features that had the greatest impact in distinguishing air regions from non-air regions. The data was divided into a training set consisting of (5000 air voxels and 5000 non-air voxels) and a test set (from a data set of different patients containing 1000 air / non-air voxels each). Classification was performed using the minimum distance to the centroid of the training class. Classifier performance was judged using the area under the ROC curve to identify the features that had the greatest impact on classification. All air regions determined by a thresholding method and refined by performing texture feature analysis on MRI were assigned air densities generated for synthetic CT.

[0246] Still refer to Fig.19 , a synthetic CT image is then generated using the bone contours and air region volumes by transferring rED values ​​from the rCT image to the deformed dMR image, as indicated at step 1916. In some implementations, rED values ​​are calculated from HU data in the rCT image and then transferred to the deformed dMR image, and in other implementations, HU data is transferred from the rCT image to the deformed dMR image before converting the HU data to rED data.

[0247] Air region volumes can be transferred to the sCT image, and the sCT values ​​in these volumes will be overwritten by the air density (rED=0.01). Gas regions appear randomly; therefore, dMR has the most accurate representation of them. On sCT, the approximate expected locations of these regions (e.g., bowel regions) are known, and only the air regions from dMR in these expected locations will be overwritten. In this way, low signal regions (e.g., bone from dMR) will not be mistakenly identified as air regions. The rED of air regions on sCT outside of the air regions transferred from dMR will be overwritten by tissue density (rED=0.92).

[0248] In some implementations, sCT images may be generated without separating and registering bone contours in the input image data. Fig.21 , a flowchart is shown to illustrate steps of another example method for generating sCT images from magnetic resonance images. As indicated at step 2102, the method includes accessing input image data with a computer system. Typically, the input image data includes reference CT images and daily MR images of the patient. Accessing the input image data may include retrieving previously acquired input image data from a memory or other data storage device or medium. In some cases, accessing the input image data may include acquiring images (e.g., rCT images, dMR images) using a suitable medical imaging system, and communicating or otherwise transferring those images to a computer system, which may form part of (multiple) corresponding medical imaging systems.

[0249] Segmented data are generated from the input image data by segmenting air regions (e.g., air cavities, total air regions) and all organs in the input image data, as indicated by process block 2104. For example, these segmented data may be generated by manually drawing air-containing regions on the rCT image. The manually drawn contours may be drawn with extensions (such as 1 cm extensions). In some cases, generating the segmented data may include retrieving previously generated segmented data stored in a memory or other data storage device or medium.

[0250] Subsequently, the CT contours are registered on the dMR images in the input image data, as indicated by step 2106. Subsequently, the segmented air regions may be refined using MRI-derived thresholds, as indicated by step 2108. Fig.19 Methods for determining and refining the MRI-derived thresholds are described. Subsequently, the refined air regions are completed and / or verified based on image feature (e.g., texture feature) analysis, as indicated in step 2110. Also referenced above Fig.19 Methods for completing and verifying air regions based on analysis of image features (eg, texture features) are described.

[0251] Subsequently, a synthetic CT image is generated by transferring the rED values ​​from the rCT image to the dMR image using the air region volume, as indicated at step 2112. In some implementations, the rED values ​​are calculated from the HU data in the rCT image and then transferred to the dMR image, and in other implementations, the HU data is transferred from the rCT image to the dMR image before converting the HU data to rED data.

[0252] Methods have been described for rapidly creating sCT images by transferring rED from a patient's own CT by deformable registration based on CT and MRI, with special consideration of bone and air regions. The generated sCT is more accurate than that generated from the commonly used bulk density assignment, but does not require multiple MR sequences.

[0253] Using the methods described in the present disclosure, air-tissue boundaries can be identified on magnetic resonance images with an accuracy of less than 1 mm compared to those on CT images using automatic thresholding methods. MRI texture differences between air and tissue are used to improve air cavity delineation. As an example, autocorrelation, variance, and contrast can be used to assess texture differences between air and tissue. The systems and methods described in the present disclosure generally include using intensity-based automatic thresholding and image texture-based voxel classification to automatically and accurately segment air regions in magnetic resonance images, which enables rapid and accurate generation of synthetic CT images for MRI-based RT planning (especially for OLAR).

[0254] Reference now Fig. 22 , shows an example of a system 2200 for generating a synthetic CT image according to some embodiments of the systems and methods described in the present disclosure. Fig. 22 As shown, the computing device 2250 may receive one or more types of data (e.g., rCT image data, dMR image data) from a medical image source 2202, which may be a source of rCT images and dMR images. In some embodiments, the computing device 2250 may execute at least a portion of a synthetic CT image generation system 2204 to generate a synthetic CT image from the rCT image and dMR image data received from the medical image source 2202.

[0255] Additionally or alternatively, in some embodiments, the computing device 2250 may communicate information about data received from the medical image source 2202 to a server 2252 via a communications network 2254, which may execute at least a portion of the synthetic CT image generation system 2204. In such embodiments, the server 2252 may return information indicative of the output of the synthetic CT image generation system 2204 to the computing device 2250 (and / or any other suitable computing device).

[0256] In some embodiments, computing device 2250 and / or server 2252 may be any suitable computing device or combination of devices, such as a desktop computer, a laptop computer, a smart phone, a tablet computer, a wearable computer, a server computer, a virtual machine executed by a physical computing device, etc. Computing device 2250 and / or server 2252 may also reconstruct an image from the data.

[0257] In some embodiments, the medical image source 2202 can be any suitable source of medical image data (e.g., rCT image data, dMR image data), such as a CT system, an MRI system, another computing device (e.g., a server storing image data), etc. In some embodiments, the medical image source 2202 can be local to the computing device 2250. For example, the medical image source 2202 can be combined with the computing device 2250 (e.g., the computing device 2250 can be configured as part of a device for capturing, scanning, and / or storing images). As another example, the medical image source 2202 can be connected to the computing device 2250 via a cable, a direct wireless link, etc. Additionally or alternatively, in some embodiments, the medical image source 2202 is located locally and / or remotely relative to the computing device 2250 and can communicate data to the computing device 2250 (and / or the server 2252) via a communication network (e.g., the communication network 2254).

[0258] In some embodiments, the communication network 2254 may be any suitable communication network or combination of communication networks. For example, the communication network 2254 may include a Wi-Fi network (a Wi-Fi network may include one or more wireless routers, one or more switches, etc.), a peer-to-peer network (e.g., a Bluetooth network), a cellular network (e.g., a 3G network, a 4G network, etc. that complies with any suitable standard (such as CDMA, GSM, LTE, Advanced LTE, WiMAX, etc.)), a wired network, etc. In some embodiments, the communication network 108 may be a local area network, a wide area network, a public network (e.g., the Internet), a private or semi-private network (e.g., a company or university intranet), any other suitable type of network, or any suitable combination of networks. Fig. 22 The communication links shown may each be any suitable communication link or combination of communication links, such as a wired link, a fiber optic link, a Wi-Fi link, a Bluetooth link, a cellular link, or the like.

[0259] Reference now Fig.23 , shows an example of hardware 2300 that can be used to implement a medical image source 2202, a computing device 2250, and a server 2254 according to some embodiments of the systems and methods described in the present disclosure. Fig.23As shown, in some embodiments, computing device 2250 may include processor 2302, display 2304, one or more inputs 2306, one or more communication systems 2308, and / or memory 2310. In some embodiments, processor 2302 may be any suitable hardware processor or combination of processors, such as a central processing unit ("CPU"), a graphics processing unit ("GPU"), and the like. In some embodiments, display 2304 may include any suitable display device, such as a computer monitor, a touch screen, a television, and the like. In some embodiments, input 2306 may include any suitable input device and / or sensor that may be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, and the like.

[0260] In some embodiments, the communication system 2308 may include any suitable hardware, firmware, and / or software for communicating information over the communication network 2254 and / or any other suitable communication network. For example, the communication system 2308 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In a more specific example, the communication system 2308 may include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0261] In some embodiments, the memory 2310 may include any suitable one or more storage devices that can be used to store instructions, values, data, etc., which can be used, for example, by the processor 2302 to present content using the display 2304, to communicate with the server 2252 via the communication system(s) 2308, and the like. The memory 2310 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, the memory 2310 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disk drives, one or more solid-state drives, one or more optical drives, and the like. In some embodiments, the memory 2310 may have a computer program encoded thereon or otherwise stored therein for controlling the operation of the computing device 2250. In such embodiments, the processor 2302 may execute at least a portion of the computer program to present content (e.g., images, user interfaces, graphics, tables), receive content from the server 2252, send information to the server 2252, and the like.

[0262] In some embodiments, server 2252 may include processor 2312, display 2314, one or more inputs 2316, one or more communication systems 2318, and / or memory 2320. In some embodiments, processor 2312 may be any suitable hardware processor or combination of processors, such as a CPU, GPU, etc. In some embodiments, display 2314 may include any suitable display device, such as a computer monitor, a touch screen, a television, etc. In some embodiments, input 2316 may include any suitable input device and / or sensor that can be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, etc.

[0263] In some embodiments, the communication system 2318 may include any suitable hardware, firmware, and / or software for communicating information over the communication network 2254 and / or any other suitable communication network. For example, the communication system 2318 may include one or more transceivers, one or more communication chips and / or chipsets, etc. In more specific examples, the communication system 2318 may include hardware, firmware, and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0264] In some embodiments, the memory 2320 may include any suitable one or more storage devices that can be used to store instructions, values, data, etc., which can be used, for example, by the processor 2312 to present content using the display 2314, to communicate with one or more computing devices 2250, etc. The memory 2320 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, the memory 2320 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid-state drives, one or more optical drives, etc. In some embodiments, the memory 2320 may have a server program encoded thereon for controlling the operation of the server 2252. In such embodiments, the processor 2312 may execute at least a portion of the server program to send information and / or content (e.g., data, images, user interfaces) to one or more computing devices 2250, receive information and / or content from one or more computing devices 2250, receive instructions from one or more devices (e.g., personal computers, laptops, tablet computers, smart phones), etc.

[0265] In some embodiments, the medical image source 2202 may include a processor 2322, one or more image acquisition systems 2324, one or more communication systems 2326, and / or a memory 2328. In some embodiments, the processor 2322 may be any suitable hardware processor or combination of processors, such as a CPU, a GPU, and the like. In some embodiments, the one or more image acquisition systems 2324 are generally configured to acquire data, images, or both, and may include a CT system and an MRI system or both. Additionally or alternatively, in some embodiments, the one or more image acquisition systems 2324 may include any suitable hardware, firmware, and / or software for coupling to and / or controlling the operation of the CT system and the MRI system or both. In some embodiments, one or more portions of the one or more image acquisition systems 2324 may be removable and / or replaceable.

[0266] Note that, although not shown, the medical image source 2202 may include any suitable input and / or output. For example, the medical image source 2202 may include an input device and / or sensor that can be used to receive user input, such as a keyboard, a mouse, a touch screen, a microphone, a trackpad, a trackball, etc. As another example, the medical image source 2202 may include any suitable display device (such as a computer monitor, a touch screen, a television, etc.), one or more speakers, etc.

[0267] In some embodiments, the communication system 2326 may include any suitable hardware, firmware, and / or software for communicating information to the computing device 2250 (and, in some embodiments, via the communication network 2254 and / or any other suitable communication network). For example, the communication system 2326 may include one or more transceivers, one or more communication chips and / or chipsets, and / or the like. In more specific examples, the communication system 2326 may include hardware, firmware, and / or software that may be used to establish a wired connection using any suitable port and / or communication standard (e.g., VGA, DVI video, USB, RS-232, etc.), a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, and / or the like.

[0268] In some embodiments, memory 2328 may include any suitable one or more storage devices that may be used to store instructions, values, data, etc., which may be used, for example, by processor 2322 to: control one or more image acquisition systems 2324 and / or receive data from one or more image acquisition systems 2324; to image from the data; to present content (e.g., images, user interfaces) using a display; to communicate with one or more computing devices 2250, etc. Memory 2328 may include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 2328 may include RAM, ROM, EEPROM, one or more flash drives, one or more hard disk drives, one or more solid-state drives, one or more optical drives, etc. In some embodiments, memory 2328 may have a program encoded thereon or otherwise stored therein for controlling the operation of image source 2202. In such embodiments, the processor 2322 may execute at least a portion of a program to generate images, send information and / or content (e.g., data, images) to one or more computing devices 2250, receive information and / or content from one or more computing devices 2250, receive instructions from one or more devices (e.g., a personal computer, laptop computer, tablet computer, smart phone, etc.), and so on.

[0269] In some embodiments, any suitable computer-readable medium may be used to store instructions for performing the functions and / or processes described herein. For example, in some embodiments, the computer-readable medium may be transient or non-transient. For example, non-transient computer-readable media may include the following media, such as: magnetic media (e.g., hard disk, floppy disk), optical media (e.g., compact disk, digital video disk, Blu-ray disk), semiconductor media (e.g., random access memory ("RAM"), flash memory, electrically programmable read-only memory ("EPROM"), electrically erasable programmable read-only memory ("EEPROM")), any suitable medium that is not transient or does not have any persistent appearance during transmission, and / or any suitable tangible medium. As another example, transient computer-readable media may include signals on a network, signals on wires, conductors, optical fibers, circuits, or any suitable medium that is transient and does not have any persistent appearance during transmission, and / or any suitable intangible medium.

[0270] Reference now Fig.24 , a block diagram of an example of a computer system 2400 that can perform the methods described in the present disclosure is shown. The computer system 2400 generally includes an input 2402, at least one hardware processor 2404, a memory 2406, and an output 2408. Therefore, the computer system 2400 is generally implemented with a hardware processor 2404 and a memory 2406.

[0271] In some embodiments, computer system 2400 may be a radiation therapy planning system ("TPS"). In some examples, computer system 2400 may also be implemented by a workstation, a laptop, a tablet device, a mobile device, a multimedia device, a network server, a mainframe, one or more controllers, one or more microcontrollers, or any other general or special computing device.

[0272] Computer system 2400 may operate autonomously or semi-autonomously, or may read executable software instructions from memory 2406 or a computer-readable medium (e.g., a hard drive, a CD-ROM, a flash memory), or may receive instructions from a user or any other source logically connected to the computer or device (such as another networked computer or server) via input 2402. Thus, in some embodiments, computer system 2400 may also include any suitable device for reading a computer-readable storage medium.

[0273] In general, the computer system 2400 is programmed or otherwise configured to implement the methods and algorithms described in the present disclosure. For example, the computer system 2400 may be programmed to generate an indication of whether to initiate OLAR based on an analysis of the Jacobian histogram metric, automatically correct contours in a radiotherapy plan, verify contours for OLAR, automatically assess the quality of a radiotherapy plan based on texture features and spatial dose metrics, generate a synthetic CT image from a magnetic resonance image, or a combination thereof.

[0274] Input 2402 may take any suitable shape or form as desired for operation of computer system 2400, including the ability to select, input, or otherwise specify parameters consistent with performing tasks, processing data, or operating computer system 2400. In some aspects, input 2402 may be configured to receive data, such as data acquired using a medical imaging system, previously generated radiotherapy plans, and the like. Such data may be processed as described above to generate an indication of whether to initiate OLAR based on analysis of Jacobian determinant histogram metrics, automatically correct contours in radiotherapy plans, verify contours for OLAR, automatically assess radiotherapy plan quality based on texture features and spatial dose metrics, generate synthetic CT images from magnetic resonance images, or a combination thereof. In addition, input 2402 may also be configured to receive any other data or information deemed useful for implementing the above methods.

[0275] In processing tasks for operating the computer system 2400 , the one or more hardware processors 2404 may also be configured to perform any number of post-processing steps on the data received via the input 2402 .

[0276] The memory 2406 may include software 2410 and data 2412, such as data acquired using a medical imaging system or a previously generated radiation therapy plan, and may be configured to store and retrieve processed information, instructions, and data to be processed by the one or more hardware processors 2404. In some aspects, the software 2410 may include instructions for implementing the methods described above.

[0277] Additionally, output 2408 may take any shape or form as desired, and may be configured to display medical images, radiation treatment plans, and other data calculated, derived, or otherwise obtained from such images or plans, as well as other desired information.

[0278] As described above, the systems and methods described in the present disclosure may be implemented using a radiation therapy planning system. Fig.25 , shows an example of such a radiation therapy planning system 2510. The radiation therapy planning system 2510 is preferably in communication with one or more radiation therapy systems 2512, which may include any suitable radiation therapy system.

[0279] The radiation therapy planning system 2510 generally includes a memory 2514 operably coupled to a processor unit 2516. As an example, the processor unit 2516 may be a commercially available computer processor, such as those described above. The processor unit 2516 is configured to perform one or more of the steps of the method described above.

[0280] As an example, the memory 2514 may include multiple memory elements or may include a single memory element. Typically, the memory 2514 is configured to store information about patient data, treatment targets (e.g., tumors located in the patient's body), imaging beam model data, treatment beam model data, dose measurements, and the like.

[0281] Preferably, the radiation therapy planning system 2510 includes or is otherwise in communication with a user interface 2518. As an example, the user interface 2518 provides information to a user, such as a medical physicist. For example, the user interface 2518 may include a display 2520 and one or more input devices, such as a keyboard 2522 and a mouse 2524.

[0282] In some embodiments, any suitable computer-readable medium may be used to store instructions for performing the functions and / or processes described herein. For example, in some embodiments, the computer-readable medium may be transient or non-transient. For example, non-transient computer-readable media may include the following media, such as: magnetic media (e.g., hard disk, floppy disk), optical media (e.g., compact disk, digital video disk, Blu-ray disk), semiconductor media (e.g., random access memory ("RAM"), flash memory, electrically programmable read-only memory ("EPROM"), electrically erasable programmable read-only memory ("EEPROM")), any suitable medium that is not transient or does not have any persistent appearance during transmission, and / or any suitable tangible medium. As another example, transient computer-readable media may include signals on a network, signals on wires, conductors, optical fibers, circuits, or any suitable medium that is transient and does not have any persistent appearance during transmission, and / or any suitable intangible medium.

[0283] This disclosure has described one or more preferred embodiments, and it should be understood that many equivalents, alternatives, variations and modifications besides those explicitly described are possible and within the scope of the invention.

Claims

1. A method for verifying the accuracy of radiation therapy plan contour data, the method comprising: (a) accessing, using a computer system, previously generated contour data, the contour data comprising contours associated with a radiation therapy plan for a subject; (b) accessing an image of the object using the computer system; (c) generating inner shell data by reducing a contour in the contour data by a first amount using the computer system; (d) generating shell data using the computer system by extending contours in the contour data by a second amount; (e) calculating image feature data from the image using the computer system; (f) generating an inner shell image feature map from the image feature data on the contours in the inner shell data using the computer system; (g) generating a shell image feature map from the image feature data on the contours in the shell data using the computer system; as well as (h) verifying the contour data by the following steps: inputting the inner shell image feature map and the outer shell image feature map into a machine learning algorithm to generate output as labeled contour data, wherein the labeled contour data marks the verified contour data as accurate contour data and marks the unverified contour data as inaccurate contour data; (i) generating a binary mask for a given contour in the contour data by assigning first binary values ​​to pixels within the contour and second binary values ​​to pixels outside the contour, and wherein the inner shell data is generated by reducing the binary mask by the first amount and the outer shell data is generated by dilating the binary mask by the second amount; as well as (j) generating core region data from the binary mask, wherein the core region data is generated by the following steps: generating a distance map by calculating a distance metric between each pixel in the binary mask and a boundary pixel in the binary mask that is closest to each pixel; as well as The core region is defined based on a percentage of pixels having a maximum distance metric value in the distance map.

2. The method according to claim 1, characterized in that The distance metric is a Euclidean distance metric.

3. The method according to claim 1, characterized in that The percentage is 20 percent.

4. The method according to claim 1, characterized in that The percentage is 30 percent.

5. The method of claim 1, further comprising generating a core region image feature map from the image feature data on the core region using the computer system.

6. The method according to claim 5, characterized in that The machine learning algorithm implements a decision tree model, which takes as input core area image feature data including the image feature data on the core area, inner shell image feature data including the image feature data on the contour in the inner shell data, and outer shell image feature data including the image feature data on the contour in the outer shell data.

7. The method according to claim 6, characterized in that The decision tree model includes: a first stage in which the core region image feature data is evaluated relative to live data; a second stage in which the inner shell image feature data is evaluated relative to live data; and A third stage in which the housing image feature data is evaluated relative to live data.

8. The method according to claim 7, characterized in that The first stage of the decision tree model evaluates a selected set of image features in the core region image feature data as similar image features in the live data.

9. The method according to claim 8, characterized in that The set of image features is selected based on hierarchical clustering of the image feature data contained in the core region.

10. The method according to claim 7, characterized in that When the image feature data corresponding to a contour passes through the first level, the second level, and the third level of the decision tree model, the contour is verified.

11. The method according to claim 1, characterized in that The inner shell image feature map and the outer shell image feature map are each normalized based on image feature data contained in the core region.

12. The method according to claim 1, characterized in that The first amount and the second amount have the same value.

13. The method according to claim 12, characterized in that The value is 4 mm.

14. The method according to claim 1, wherein: The image feature data includes image features based on an intensity histogram.

15. The method according to claim 14, characterized in that The intensity histogram-based image features include at least one of a mean, a standard deviation, a range, a skew, and a kurtosis.

16. The method according to claim 1, wherein: The image feature data includes texture features calculated based on a gray level co-occurrence matrix.

17. The method according to claim 1, wherein: The image feature data includes texture features calculated based on a grayscale run-length matrix.

18. The method of claim 1, wherein: The machine learning algorithm implements recursive random forest classification to select a set of image features, and the inner shell image feature map and the outer shell image feature map are compared with the selected set of image features to verify the contour data.

19. The method of claim 1, further comprising: The inaccurate contour data in the labeled contour data is input to a contour correction algorithm to generate an output as corrected contour data.

20. The method of claim 19, wherein: The corrected contour data is further verified by repeating steps (c)-(h) using the corrected contour data as input contour data.

21. The method of claim 19, wherein: The contour correction algorithm includes: calculating, using the computer system, second image feature data for a region of interest (ROI) in the image, wherein the ROI is defined by a contour in the inaccurate contour data; generating an image feature map from the second image feature data using the computer system, wherein the image feature map depicts image feature values ​​in the ROI; Generating the corrected contour data by the following steps: inputting the image feature map and the inaccurate contour data into an active contour algorithm, generating an output as the corrected contour data; The image feature map is used by the active contour algorithm as an external force, which guides the active curve to find the corrected boundary of the inaccurate contour data.

Images