Adaptive replanning based on multimodal imaging

Abstract

Systems and methods for adaptive radiation therapy planning are provided. In some aspects, the provided systems and methods include generating a synthetic image from magnetic resonance data using relaxation mapping. The method includes applying corrections to the data and generating a relaxation map therefrom. In other aspects, a method for adapting a radiation therapy plan is provided. The method includes determining an objective function based on a dose gradient from an initial dose distribution and generating an optimized plan based on updated images using an aperture deformation and gradient maintenance algorithm without contouring of organs at risk. In still other aspects, methods are provided for obtaining 4D MR imaging using temporal shuffling of data acquired during normal breathing, methods for deformable image registration using a semi-physical model regularization method applied sequentially to multimodal images, and methods for generating a 4D plan based on 4D CT or 4D MR imaging using an aperture deformation algorithm.

Description

[0001] This invention application is a divisional application of the invention patent application with international application number PCT / US2014 / 072645, international application date December 30, 2014, application number 201480076742.0 entering the Chinese national phase, and titled “Adaptive Replanning Based on Multimodal Imaging”.

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] This application claims priority to U.S. Provisional Patent Application Serial No. 61 / 922,343, filed on December 31, 2013, entitled “SYSTEMS AND METHODS FOR ADAPTIVE REPLANNING BASED ON MULTI-MODALITY IMAGING.” Background Art

[0004] The present disclosure relates generally to tools, systems, and methods for use in radiation therapy treatment delivery and, more particularly, to systems and methods for adaptive planning based on multimodality imaging.

[0005] Radiation therapy (RT) has undergone a series of technological revolutions over the past few decades. With intensity-modulated RT (IMRT), it has become possible to produce highly conformal dose distributions, thereby delivering most of the radiation dose within the confines of the tumor. These techniques utilize 3D anatomical information and biological information extracted from various types of images (e.g., CT, MRI, PET) acquired a few days before the first treatment. However, it has been found that the position, shape, size, and biological properties of tumor(s) and normal anatomy change during the course of treatment, primarily due to day-to-day positioning uncertainties of various ROIs and anatomical, physiological, and / or clinical factors. The latter include tumor shrinkage, weight loss, volume changes in normal organs, and non-rigid changes in different bone structures. The traditional assumption that the anatomy identified from 3D CT images acquired for planning purposes before the treatment process is applicable to each segment may not fully account for inter-segment variations and may limit the ability to fully exploit the potential of highly conformal treatment modalities (such as IMRT). This improved capability in dose conformality necessitates better localization of targets and organs at risk (OARs), as well as the ability to accommodate inter- and intra-fragmentary variations in both the design and delivery of therapy.

[0006] Accurate delivery of radiation therapy necessitates high-fidelity, 3D, anatomical images of the patient. For nearly two decades, RT planning has utilized CT images as the basis for radiation dose calculations. Due to the linear relationship between Hounsfield cells and underlying tissue electron density, CT images allow for voxel-based correction of radiation dose for differences in tissue attenuation, as well as the generation of visually observed digitally reconstructed radiographs (DRRs) of beams for treatment verification. Image contrast in CT is primarily due to photoelectric interactions. This effect results in high contrast between tissues with significantly different densities (e.g., bone, lung, and soft tissue). However, adjacent soft tissue (e.g., in the brain or abdominal region) does not possess substantial density differences. The resulting poor soft tissue contrast on CT images makes delineation of both targets and critical structures extremely challenging. This inability to reliably delineate tumor targets and adjacent critical structures on CT images has significant clinical consequences, requiring the use of large margins, which hinders the ability to safely escalate radiation doses due to toxicity constraints of critical structures.

[0007] Magnetic resonance imaging (MRI) offers powerful imaging capabilities for cancer diagnosis and treatment. In contrast to CT, MRI is nonionizing, provides superior soft-tissue contrast, and offers a wide range of functional contrast-forming mechanisms for characterizing tumor physiology. However, unlike CT, the MRI signal is not directly related to electron density, which has prevented MRI from being used as a basis for dose calculations. Furthermore, MR images can be confounded by spatial distortions caused by gradient nonlinearity and off-resonance effects. These distortions can be severe, ranging up to several centimeters. Furthermore, MR images often exhibit regions of nonuniform image intensity, which can introduce inaccuracies into image registration and segmentation algorithms frequently employed during radiotherapy planning and delivery. When phased array coils are used for signal reception, these nonuniform image densities arise from inhomogeneities in the B1+ (RF transmit) and B1- (RF receive) fields. Therefore, despite the significant soft-tissue contrast advantages offered by MRI, these current limitations (e.g., lack of electron density information, sources of geometric distortion, and signal inhomogeneity) have hindered the establishment of MRI as a primary imaging modality in radiation oncology.

[0008] Adaptive radiation therapy (ART) is a state-of-the-art approach that uses feedback processing to account for patient-specific anatomical and / or biological changes during treatment, thereby delivering highly personalized radiation therapy to cancer patients. The basic components of ART include: (1) detection of anatomical and biological changes, typically facilitated by multimodal images (e.g., CT, MRI, PET), (2) treatment plan optimization to account for patient-specific spatial morphological and biological changes while taking into account radiation response, and (3) techniques for accurately delivering the optimized plan to the patient. ART intervention can consist of both online and offline approaches.

[0009] Inter- and intra-fragment variations, if not accounted for, can lead to suboptimal dose distributions and significant deviations from the original plan, with potential negative consequences in terms of treatment efficiency. Recently, image-guided RT (IGRT) has been widely used to correct (eliminate or reduce) the deteriorating effects of inter-fragment variations. A large number of correction strategies have been developed based on the available IGRT technology. These correction methods can be generally classified as "online" or "offline" schemes. Corrections to patient treatment parameters that are performed just after daily patient information is acquired and before daily treatment doses are delivered are classified as "online" corrections. This is in contrast to "offline" corrections, in which the correction actions are performed after daily treatment has been delivered, thereby affecting treatment on subsequent days. Therefore, when online corrections are applied, the daily dose delivered will be the corrected one using the most recent patient setup and anatomical information.

[0010] When the patient is set up for each fraction, the anatomy may be different from that used for the initial treatment plan. Typically, the most harmful deviations are so-called "systematic" deviations, which are also relatively easier to account for by offline or online correction strategies. The random component of the deviations, although sometimes less harmful than systematic deviations, is often difficult to fully account for and requires an online correction strategy. One of the advantages of an online correction strategy over an offline approach is that the online strategy can correct for both systematic and random variations. In addition, offline correction may not be applicable to treatment procedures with a small number of treatment fractions, such as hypo-fractionation or stereotactic body RT (SBRT).

[0011] The primary challenge for online correction strategies is that they need to be performed within an acceptable time frame while the patient is lying on the treatment table in the treatment position. This requirement limits the variety of corrective actions that can be used as online strategies in today's technology. Therefore, online correction of inter-segment changes by repositioning the patient based on images acquired immediately before treatment delivery is the current standard practice for IGRT. The online repositioning strategies practiced in most clinics to date are limited to correcting only translational movements and fail to account for rotational errors, volume changes and deformations of the target and OARs, and independent motion between different targets / OARs. This is despite the fact that current technology provides enough information to perform modifications to daily treatments that are much more detailed than simple translational movements. In principle, the data required to generate a new treatment plan for that day is available in today's CT-based IGRT practice, but by using this data only to move the patient, the full potential of IGRT is not being exploited.

[0012] There has been ongoing research to extend online correction beyond stage movement to modifications that can correct for variations in anatomy, such as organ rotation and deformation. For example, correction for rotation in addition to translation has been implemented, at least in part, in several techniques (e.g., helical tomotherapy). However, other inter-segment variations may not have been considered and remain a major problem. These variations can be addressed by online replanning methods, including rapid online plan modification based on the day's anatomy and full-blown plan reoptimization. For example, rapid adjustment of beam aperture shape and weights based on the day's CT scan (the day's anatomy) is an online plan modification method facilitated by advances in computer technology, which allows computationally intensive operations to be performed within a reasonable timeframe. Other examples include GPU-based full-blown reoptimization and adaptation based on precomputed plan libraries. Consequently, online correction schemes that are more comprehensive than online repositioning are beginning to enter the clinic. However, such online replanning schemes remain challenging because they require delineation of targets and OARs, a time-consuming and difficult process to fully automate. As such, manual delineation or even manual contour verification on single-modality or multi-modality imaging is a major bottleneck for an online re-planning process to be performed within minutes.

[0013]

[0006] Therefore, in view of the above, there is a need for systems and methods that employ multimodal images for adaptation and delivery of radiation therapy treatments in clinically feasible time frames. Summary of the Invention

[0014] The present disclosure overcomes the aforementioned shortcomings by providing systems and methods for adaptive radiation therapy planning.

[0015] According to one aspect of the present disclosure, a system for developing a radiotherapy treatment is provided. The system includes a data storage device configured to maintain MR image data acquired by an MRI system, and at least one processor configured to: receive the MR image data from the data storage device, apply correction to the MR image data to generate a series of corrected image data, and aggregate the series of corrected image data to generate a set of relaxation maps. The at least one processor is also configured to use the set of relaxation maps to perform segmentation of a plurality of regions of interest, classify the plurality of regions of interest using the set of relaxation maps to generate a plurality of classified structures, and assign electron density values ​​to the classified structures using an assignment process. The at least one processor is further configured to generate a set of corrected synthetic electron density images using the electron density values ​​of the classified structures, and perform dose calculations using the corrected synthetic electron density images to develop a radiotherapy treatment plan.

[0016] According to another aspect of the present disclosure, a method for generating a synthetic image for use in a radiotherapy treatment is provided. The method includes directing a magnetic resonance imaging (MRI) system to acquire a plurality of image data for use in a radiotherapy process, receiving the plurality of acquired image data from the MRI system, and applying correction to the MR image data to generate a series of corrected image data. The method also includes aggregating the series of corrected image data to generate a set of relaxation maps, using the set of relaxation maps to perform segmentation of a plurality of regions of interest, and using the set of relaxation maps to classify the plurality of regions of interest to generate a plurality of classified structures. The method further includes using an assignment process to assign electron density values ​​to the classified structures, generating a set of corrected synthetic electron density images using the electron density values ​​of the classified structures, and optionally using the corrected synthetic electron density images to perform dose calculations to develop a radiotherapy treatment plan.

[0017] According to yet another aspect of the present disclosure, a method for adapting a radiotherapy treatment plan is provided. The method includes providing an initial radiotherapy plan to be adapted based on an updated image set, the initial radiotherapy plan having a radiation dose distribution, determining a plurality of dose gradients using the radiation dose distribution, and defining an optimization target using the dose gradients. The method also includes receiving the updated image set, generating an updated contour set representing an updated target volume using the updated image set, and forming a set of partial rings using the updated contour set, the set of partial rings being arranged around the updated contour set representing the updated target volume. The method further includes performing plan optimization using the optimization target and the set of partial rings, and generating a report representing the adapted radiotherapy plan obtained using the plan optimization.

[0018] The foregoing and other advantages of the present disclosure will be apparent from the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flow chart setting forth the steps associated with a process of providing radiation therapy treatment to a patient is shown.

[0020] Figure 2 is a schematic diagram of an example MRI system for use in accordance with the present disclosure.

[0021] Figure 3 This is to illustrate the Figure 2 A flow chart of the steps associated with the operation of an MRI system for generating a synthetic electron density image.

[0022] Figure 4 The use according to the present disclosure Figure 2 A block diagram of an exemplary image acquisition and post-processing system.

[0023] Figure 5 Shown are examples of brain T1 parameter maps from healthy volunteers generated with and without B1+ correction according to the present disclosure.

[0024] Figure 6 is a graphical illustration of an electron density to Hounsfield unit conversion table from a clinical radiation therapy planning system for converting a synthetic electron density image to a synthetic CT image in accordance with the present disclosure.

[0025] Figure 7 Exemplary synthetic electron density and synthetic CT images obtained from the pelvis of a healthy volunteer in accordance with the present disclosure are shown.

[0026] Figure 8Exemplary volumetrically modulated arc therapy plans calculated using synthetic CT images with and without heterogeneity correction in accordance with the present disclosure are shown.

[0027] Figure 9 A diagram illustrating a deformable transformation between a fixed image and a moving image according to the present disclosure.

[0028] Figure 10 An example of bending energy finite diffeomorphism (BELD)-regularized deformable image registration (DIR) according to the present disclosure is shown for the breast cancer case where the patient is in a supine position.

[0029] Figure 11 An example of BELD-regularized DIR according to the present disclosure is shown for the breast cancer case where the patient is in the prone position.

[0030] Figure 12 Exemplary images showing multimodal contoured structures resulting from application of deformation fields in accordance with the present disclosure are shown.

[0031] Figure 13 A graphical example demonstrating the effect of a variable kernel smoothing technique on registration accuracy according to the present disclosure is shown.

[0032] Figure 14 Shown are image results from registration using variable kernel smoothing applied to a pelvic CT scan in accordance with the present disclosure.

[0033] Figure 15 Shown is an example illustrating registration between CT and high-frequency ultrasound (HFUS) images for a patient with superficial basal cell carcinoma.

[0034] Figure 16 is a flow chart setting forth the steps associated with an operating mode for a gradient-maintaining algorithm according to the present disclosure.

[0035] Figure 17 A diagram comparing online replanning schemes using a conventional method and a gradient-maintaining method according to the present disclosure is shown.

[0036] Figure 18 Exemplary dose distributions from image-guided radiation therapy (IGRT) repositioning, traditional full optimization based on a full contour set, and gradient-maintenance according to the present disclosure are shown.

[0037] Figure 19 Showing inter-segment variation in daily treatment for prostate cancer cases Figure 18 Example dose-volume histograms for the three strategies.

[0038] Figure 20 Offline optimization according to the present disclosure is shown.

[0039] Figure 21 A comparison between image intensities obtained using various MRI scanners and scanning conditions and subjected to correction and normalization according to aspects of the present disclosure is shown.

[0040] Figure 22 A diagram showing a retrospective reconstruction of a phase-resolved 4D MR image according to aspects of the present disclosure is shown.

[0041] Figure 23 An example of partially concentric rings utilized in a fast online replanning method according to aspects of the present disclosure is shown. DETAILED DESCRIPTION

[0042] Current clinical practice associated with the delivery of any one or more radiation therapy treatments involves a complex, carefully coordinated workflow utilizing a multitude of tools and systems designed to achieve maximum patient benefit. Figure 1 An example flow chart is shown that illustrates the general steps associated with such a radiation therapy treatment process. Process 100 typically begins at process block 10, where medical image information is typically generated from a patient using various imaging methods for diagnostic or therapeutic purposes. For example, as in Figure 1 As shown in , this can include the use of computed tomography (CT) devices, magnetic resonance imaging (MRI) devices, positron emission tomography (PET) imaging devices, ultrasound (US) imaging devices, etc. The information obtained by imaging is used for two main purposes in radiation therapy (RT). First, it is used to determine the true three-dimensional position and range of the targeted diseased tissue relative to adjacent critical structures or targets at risk (OARs), which typically have radiation dose toxicity constraints. Second, it is used to locate such targets and OARs, for example during daily treatment setup, so that any treatment adjustments can be made before radiation delivery.

[0043] In general, CT images are the standard imaging modality used for treatment planning. During the simulation phase, the patient is immobilized and imaged using reference markers that establish specific coordinates that can then be reproduced in the treatment system during radiation delivery. The acquired images are then used in the planning phase to generate a treatment plan. In addition to CT images, other imaging modalities provide improved contrast and other useful information regarding the anatomical features and biological processes of normal and diseased tissues or structures. Specifically, MRI is non-ionizing and provides superior soft tissue contrast compared to CT, while also providing a wide range of functional contrast that forms mechanisms for characterizing tumor physiology. However, in contrast to CT images, MR images lack electron density (ED) information, which is essential for radiation dose calculations. Therefore, such non-CT images need to be processed or synthesized into "CT-like" images in order to find use in RT planning and delivery.

[0044] Previously, different approaches have been explored to convert MR images into CT-like images, namely: 1) manual volumetric electron density assignment, 2) deformable image registration, 3) voxel-based classification, 4) use of ultrashort echo time (UTE) images, and 5) use of Dixon images. Each of these approaches has its drawbacks, as described below.

[0045] In manual volume density assignment methods, tissue contours are manually constructed on the patient's MR images, and volume electron density is then assigned to each of the manually constructed structures. A major problem with this approach is that it suffers from large intra- and inter-observer variability. Variability in observer contouring, particularly at tissue interfaces, has been shown to produce significant errors in the calculated dose distribution. In deformable image registration (DIR) methods, a vector displacement field (VDF) is typically estimated by deforming the patient's MRI image to a reference MRI image. This VDF is then applied to deform the "gold standard" CT image back into the vector space of the patient's MRI image. However, the accuracy of this approach is limited by the accuracy of DIR and can be particularly challenging in patients with atypical anatomy. Furthermore, non-uniform MR image intensity introduced by B1+ (RF transmit) and B1- (RF receive) inhomogeneities can challenge the accuracy of intensity-based DIR algorithms. Furthermore, variations in grayscale signal intensity, common in conventional amplitude MR images, can also challenge the accuracy and robustness of intensity-based DIR algorithms. In voxel-based classification methods, each voxel is classified as a specific tissue type, followed by direct conversion of MRI grayscale intensity to Hounsfield units or electron density. However, non-uniform image intensity and variations in grayscale intensity in conventional amplitude MR images can again confound the accuracy of tissue classification using voxel-based methods. Although UTE sequences may offer advantages for imaging bone, they lack contrast for other tissue types and are not yet commercially available on all clinical MRI scanners. Finally, poor registration of multi-echo Dixon images can confound image segmentation, resulting in blurred structural boundaries.

[0046] In addition to having accurate and usable image information, the present disclosure recognizes that CT-like or CT-surrogate images need to meet some additional conditions to find practical use in RT treatment. Specifically, such images necessitate 3D volume coverage consisting of serial slices to allow accurate and complete structural delineation for treatment planning evaluation using dose-volume histograms (DVH's). Moreover, a complete patient cross-section within the field of view (FOV) is required for accurate dose calculations, as such calculations involve the distance between the radiation source and the patient's skin. Furthermore, for accurate treatment verification through improved digitally reconstructed radiograph (DRR) image quality, a slice thickness of 3 mm or less is advantageous, while accurate image registration and image segmentation algorithms can benefit from high spatial resolution (e.g., less than 1 mm 2 ) and image uniformity. For imaging of body regions susceptible to respiratory, peristaltic, and cardiac motion, a suitable acquisition speed is desired for artifact-free images, while accurate anatomical delineation and dose accumulation require high geometric fidelity.

[0047] Conditions imposed by the nature of RT treatment have limited the use of nonstandard images for qualitative assessment. Specifically, despite the advantages in soft tissue contrast provided by MRI, deficiencies such as lack of electron density information, geometric distortion, and signal inhomogeneity have hindered the establishment of MR imaging as the primary imaging modality in RT treatment.

[0048] Thus, one aspect of the present disclosure is to provide systems and methods for successfully implementing non-ionizing MR imaging into RT treatment procedures. Specifically, such systems and methods can be used to generate synthetic ED images that overcome the shortcomings of previous methods and are practical for use in RT treatment planning. As will be described, relaxation parameter mapping can be used to obtain synthetic ED images in which the images are fully corrected for known error sources, including magnetic field (B0) errors, RF transmit field (B1+) inhomogeneity errors, and geometric distortion errors from gradient nonlinearity.

[0049] Specific reference Figure 2 , shows an example of a magnetic resonance imaging (MRI) system 200. The MRI system 200 includes an operator workstation 102, which will typically include a display 104, one or more input devices 106, such as a keyboard and mouse, and a processor 108. The processor 108 may comprise a commercially available programmable machine running a commercially available operating system. The operator workstation 102 provides an operator interface that enables scan prescriptions to be entered into the MRI system 200. Typically, the operator workstation 102 may be coupled to four servers: a pulse sequence server 110; a data acquisition server 112; a data processing server 114; and a data storage server 116. The operator workstation 102 and each of the servers 110, 112, 114, and 116 are connected to communicate with each other. For example, the servers 110, 112, 114, and 116 may be connected via a communication system 117, which may comprise any suitable network connection, whether wired, wireless, or a combination thereof. As examples, the communication system 117 may include private or dedicated networks, as well as open networks, such as the Internet.

[0050] The pulse sequence server 110 operates in response to instructions downloaded from the operator workstation 102 to operate a gradient system 118 and a radio frequency ("RF") system 120. The gradient waveforms necessary to perform a prescribed scan are generated and applied to the gradient system 118, which energizes the gradient coils in an assembly 122 to generate magnetic field gradients and to position-encode magnetic resonance signals. The gradient coil assembly 122 forms part of a magnet assembly 124, which includes a polarizing magnet 126 and a whole-body RF coil 128.

[0051] The RF waveform is applied to the RF coil 128 or a separate local coil (in the Figure 2 ) in order to perform a prescribed magnetic resonance pulse sequence. The RF coil 128, or a separate local coil ( Figure 2 The responsive magnetic resonance signals detected by the MRI system 120 (not shown) are received by the RF system 120, where they are amplified, demodulated, filtered, and digitized under the guidance of commands generated by the pulse sequence server 110. The RF system 120 includes an RF transmitter for generating the various RF pulses used in the MRI pulse sequence. The RF transmitter generates RF pulses of desired frequency, phase, and pulse amplitude waveforms in response to the scan prescription and the guidance from the pulse sequence server 110. The generated RF pulses can be applied to the whole-body RF coil 128 or one or more local coils or coil arrays (in Figure 2 not shown).

[0052] The RF system 120 also includes one or more RF receiver channels. Each RF receiver channel includes an RF preamplifier that amplifies the magnetic resonance signals received via the coil 128 to which it is connected, and a detector that detects and digitizes the quadrature components of the received magnetic resonance signals. The amplitude of the received magnetic resonance signal at any sampling point can therefore be determined by taking the square root of the sum of the squares of the components:

[0053]

[0054] The phase of the received magnetic resonance signal can also be determined according to the following relationship:

[0055]

[0056] The pulse sequence server 110 also optionally receives patient data from a physiological acquisition controller 130. By way of example, the physiological acquisition controller 130 may receive signals from a number of different sensors connected to the patient, such as electrocardiogram ("ECG") signals from electrodes, or respiration signals from a respiratory bellows or other respiratory monitoring device. Such signals are typically used by the pulse sequence server 110 to synchronize, or "gate," the performance of the scan with the subject's heartbeat or respiration.

[0057] The pulse sequence server 110 is also connected to a scan room interface circuit 132 which receives signals from various sensors associated with the patient's condition and the magnet system. A patient positioning system 134 also receives commands through the scan room interface circuit 132 to move the patient to a desired position during scanning.

[0058] The digitized magnetic resonance signal samples generated by the RF system 120 are received by the data acquisition server 112. The data acquisition server 112 operates in response to instructions downloaded from the operator workstation 102 to receive real-time magnetic resonance data and provides buffer storage so that no data is lost due to data overflow. In some scans, the data acquisition server 112 does little more than transmit the acquired magnetic resonance data to the data storage server 114. However, in scans where information derived from the acquired magnetic resonance data is needed to control further performance of the scan, the data acquisition server 112 is programmed to generate such information and transmit it to the pulse sequence server 110. For example, during a pre-scan, magnetic resonance data is acquired and used to calibrate the pulse sequence executed by the pulse sequence server 110. As another example, navigator signals can be acquired and used to adjust operating parameters of the RF system 120 or gradient system 118, or to control the order in which k-space is sampled. In yet another example, the data acquisition server 112 can also be used to process magnetic resonance signals used to detect the arrival of contrast agents in magnetic resonance angiography (MRA) scans. By way of example, the data acquisition server 112 acquires magnetic resonance data and processes it in real time to generate information used to control the scan.

[0059] The data processing server 114 receives the magnetic resonance data from the data acquisition server 112 and processes it according to instructions downloaded from the operator workstation 102. Such processing may include, for example, one or more of the following: reconstructing a two-dimensional image or a three-dimensional image by performing a Fourier transform on the raw k-space data; performing other image reconstruction algorithms, such as iterative or back-projection reconstruction algorithms; applying filters to the raw k-space data or the reconstructed images; generating functional magnetic resonance images; computing motion images or flow images; and the like.

[0060] The images reconstructed by the data processing server 114 are transmitted back to the operator workstation 102 where they are stored. The real-time images are stored in the database memory cache ( Figure 1 102 ), from which they can be output to the operator display 112 or a display 136 located near the magnet assembly 124 for use by the attending physician. Batch mode images or selected real-time images are stored in a master database on disk storage 138. When such images have been reconstructed and transferred to storage, the data processing server 114 notifies the data storage server 116 on the operator workstation 102. The operator workstation 102 can be used by the operator to archive images, generate movies, or send images to other facilities via a network.

[0061] The MRI system 200 may also include one or more networked workstations 142. By way of example, the networked workstations 142 may include a display 144; one or more input devices 146, such as a keyboard and a mouse; and a processor 148. The networked workstations 142 may be located in the same facility as the operator workstation 102, or in a different facility, such as a different healthcare institution or clinic.

[0062] A networked workstation 142, whether in the same facility as the operator workstation 102 or in a different facility than the operator workstation 102, can obtain remote access to the data processing server 114 or the data storage server 116 via the communication system 117. Thus, multiple networked workstations 142 can have access to the data processing server 114 and the data storage server 116. In this manner, magnetic resonance data, reconstructed images, or other data can be exchanged between the data processing server 114 or the data storage server 116 and the networked workstation 142, so that the data or images can be processed remotely by the networked workstation 142. This data can be exchanged in any suitable format, such as according to the Transmission Control Protocol (TCP), the Internet Protocol (IP), or other known or suitable protocols.

[0063] As described, signals acquired using the MRI system 200, unlike CT image signals, do not bear a direct relationship to electron density. Thus, in some aspects, the MRI system 200, independently or in conjunction with other processing or analysis systems, can be configured to process magnetic resonance data, reconstructed images, and / or other data to produce data, images, or information suitable for use in radiation therapy, as described.

[0064] Now refer to Figure 3 , a flowchart illustrating the steps of a process 300 for generating a synthesized ED image according to aspects of the present disclosure is shown. Specifically, the steps of process 300 can be performed using a variety of suitable systems or devices, including imaging systems, planning systems, processing systems, etc., or a combination thereof.

[0065] Specifically, process 300 may begin at process block 302, whereby a plurality of image data is received, for example, from a data storage device or other computer-readable medium, wherein the image data includes 1D, 2D, 3D, and 4D image data. In some aspects, process block 302 may include directing an MRI system, as described, to acquire a plurality of image data for use in a radiation therapy procedure, and receiving the acquired image data therefrom. By way of example, the received or acquired images may include 3D fast low angle acquisition (FLASH) images using multiple flip angles, 3D balanced steady-state free precession (bSSFP) images using multiple flip angles, 3D GRE images using multiple echo times, 3D actual flip angle (AFI) images using multiple repetition times (TRs), and the like. As described, the acquired images may need to include volume coverage, slice thickness, spatial resolution, geometric fidelity, and image uniformity suitable for use in a radiation therapy procedure. Furthermore, for body regions susceptible to motion, appropriate acquisition speed and / or motion correction may be required to obtain artifact-free images and, therefore, accurate treatment planning. According to some aspects of the present disclosure, retrospective temporal shuffling of k-space data may be used to generate phase-resolved 4D MR image data, as will be described.

[0066] At process block 304 , a number of corrections may be applied to the received or acquired image data, as discussed below.

[0067] Specifically, 3D gradient nonlinear geometric distortion correction can be applied to any or all images by applying various correction algorithms. In some preferred aspects, such correction can be performed using a vector deformation field (VDF) constructed by comparing the inverted gradient image of a distorted phantom, thereby adapting the design image of the phantom on a scanner-by-scanner basis. This approach is more robust than the Legendre polynomial 3D distortion correction algorithm provided by scanner manufacturers, which often leaves residual geometric distortion.

[0068] Likewise, corrections can be performed to account for off-resonance effects, including main field (B0) inhomogeneities, chemical shifts, and magnetic susceptibility effects. Specifically, these can be corrected by generating a magnetic field map of the patient determined using a phase difference map based on 3D gradient-return echo (GRE) images acquired at two echo times. In some aspects, the GRE images used to generate the magnetic field map can be the T2 images used for the T2 images discussed below. * Mapped subsets of multi-echo GRE images, thereby increasing scanning efficiency.

[0069] Furthermore, any or all images can be corrected for RF transmit / receive field inhomogeneities. Specifically, rapid mapping of the B1+ (RF transmit field) can also be performed using the Actual Flip Angle Imaging (AFI) imaging method, which consists of a modified 3D FLASH sequence with two repetition times, a large spoiler gradient, and a specific RF phase cycling schedule. Second- to seventh-order polynomial surfaces can then be fitted to smooth the resulting flip angle (FA) map, which is directly related to the B1+ field. Furthermore, correction for the B1- (RF receive field) can also be performed using phased array coil sensitivities measured immediately prior to data acquisition, for example using a prescan.

[0070] At process block 314, multiple relaxation maps are generated by assembling the corrected image data, including T1, T2, and T2 * Mapping. Specifically, rapid T1 mapping can be performed using a set of multiple 3D FLASH images (e.g., 3 to 5) acquired using different flip angles (e.g., 2, 5, 10, 15, 25 degrees, although other values ​​are possible). As described, the images can be corrected for gradient nonlinearity and off-resonance induced spatial distortion. A "flip angle series" can be formed and a T1 parameter map can be obtained by least squares fitting the flip angle series to the corrected flip angles determined using B1+ mapping. Similarly, rapid T2 mapping can also be performed using a set of several 3D balanced bSSFP images acquired using different flip angles. Each of the multiple flip angle images can be corrected for gradient nonlinearity and off-resonance induced spatial distortion. A flip angle series can be formed and a T2 parameter map can be obtained by least squares fitting the flip angle series to the corrected flip angles determined using B1+ mapping and T1 and B0 parameter maps. In addition, the acquired 3D multi-echo GRE images can be further used to perform T2 mapping using, for example, a set of four to eight echo times. * Mapping, where an "echo series" of phase-corrected real-time images is output. The echo time can be in the range of 0.07 to 30 ms, although other values ​​are possible. As described, the images can be corrected for gradient nonlinearity and off-resonance induced spatial distortion. T2 * The parametric map can then be obtained by least squares fitting of the phase-corrected live images to the echo times.

[0071] exist Figure 4 A block diagram representing example image acquisition and post-processing steps as described is shown in FIG. By way of example, the effect of B1+ inhomogeneity on T1 parameter mapping is shown. Figure 5Comparison of corrected and uncorrected brain images from healthy volunteers. Without correction for B1+ inhomogeneity, T1 mapping shows considerable inhomogeneous image intensity (shadows), as seen in Figure 5 This shading translates into errors in the calculated T1 relaxation values, which must be corrected for accurate tissue segmentation. Applying the B1+ correction results in a highly homogeneous T1 parameter map, as shown in Figure 5 Similar results were obtained for the T2 parameter map.

[0072] Reference again Figure 3 At process block 308, the relaxation maps and other images generated based on the corrected image data as described above can be segmented according to specific regions of interest (ROIs). For example, segmentation can be performed by thresholding MR relaxation parameters. In some aspects, segmentation can be performed using Otsu's method (a method that applies cluster-based image thresholding, where T1, T2, and T2* values ​​are used as input). Other image segmentation methods, either automated or semi-automated, can also be used.

[0073] As in Figure 21 As shown in the example of , MR scanners from various manufacturers produce images that can vary not only in uniformity as described, but also in overall intensity levels, as generally indicated by 2100. Although corrections for off-resonance effects, gradient inhomogeneities, and other artifacts can be performed as described, such corrections do not affect the overall intensity levels, as indicated by 2102. Thus, in some aspects, an image or map can also be corrected for intensity variations by performing a normalization procedure. This can include receiving an indication of the scanner type, or manufacturer, and imaging conditions (such as magnetic field strength), and applying normalization based on the received indications. As indicated by 2104, such normalization can standardize images across manufacturers and imaging conditions, thereby fixing window widths and levels in a manner that can be advantageously used during segmentation and position verification.

[0074] At process block 310, tissue classification of the ROI segmented at process block 308 can be performed based on the image intensity of the set of relaxation maps. For example, atlas-based tissue classification can be utilized. Specifically, the segmented structures or organs can be classified based on published MR relaxation values ​​(e.g., T1, T2, T* relaxation times) acquired at a given magnetic field strength (such as 3 Tesla).

[0075] Subsequently, at process block 312, an assignment process may be used to assign electron density values ​​to the classified structures. Specifically, the structures or organs classified at process block 308 may be assigned tissue-specific electron density values. Such ED values ​​may be determined or calculated in any manner, or may otherwise be obtained from any reference, such as ICRU Report #46.

[0076] At process block 314, the classified structures can then be used to generate a synthetic electron density image, corrected for distortion and inhomogeneity, e.g., in the form of a 3D or 4D image set as described above. In some aspects, the corrected synthetic electron density image can be converted to a synthetic CT image, e.g., by utilizing an inverted CT-ED standard conversion table, commonly available on clinical radiation therapy planning systems. Figure 6 A diagram illustrating the conversion between relative electron density and Hounsfield units (CT number) is shown.

[0077] The electron density values ​​associated with the corrected synthetic ED or CT image can allow any system configured for radiation therapy plan development or capable of performing dose calculations using electron density data to generate a treatment plan, as indicated by process block 316. Such dose calculations help determine a radiation dose distribution that represents the dose received within, around, or generally near any desired or target region of interest. In some aspects, the dose distribution can then be used or modified in an online radiation therapy plan adaptation strategy, as will be described.

[0078] By way of example, Figure 7 Shown are composite ED (left) and composite CT (right) images of the pelvis of a healthy volunteer generated according to the preceding description. Six tissue types were segmented and assigned electron density: air, bone, fat, muscle, water (urine in the bladder), and skin. Figure 7 Additional tissue classifications to the one shown in may still be possible (eg, separating cortical bone from bone marrow). Figure 7 The synthesized CT images of the pelvis shown in FIG were transferred and loaded onto the Monaco treatment planning system (Elekta, Sweeden), where volumetric modulated arc therapy (VMAT) plans were calculated with and without heterogeneity correction turned on, as shown in FIG. Figure 8 As shown in . Figure 8 The visible difference between the solid and dashed lines in the DVH diagram shown in the upper right of FIG. 802 and the Figure 8 The + / -90 cGy dose difference shown in the lower right of 804 demonstrates that the planning system used is able to use the information provided by the synthesized CT images to calculate the dose with heterogeneity correction turned on.

[0079] The above process demonstrates a model-independent, MRI-scanner-independent approach for synthesizing electron density information from MR image data suitable for use in radiation therapy planning and delivery. The relaxation parameter maps from which the synthesized electron density and synthesized CT images are derived are fully corrected for known sources of spatial distortion, non-uniform image intensity, and variations in grayscale signal intensity typically experienced in conventional amplitude MR images. These features, among other things, address current obstacles that have prevented the routine use of conventional amplitude MR images in MR-based radiation therapy and adaptive radiation therapy applications. Furthermore, the acquisition of the base images from which the synthesized electron density and synthesized CT images are derived is fast enough to be performed in a breath hold, thereby allowing the generation of synthesized CT images in challenging body regions prone to motion artifacts. Although the described method does not rely on anatomical models (atlases) or deformable registration, these may be used in conjunction with the systems and methods of the present disclosure if desired.

[0080] As can be understood, knowing the motion trajectory of the target and the organs at risk is key in radiotherapy. Motion can cause ambiguity in the planned dose distribution, especially when steep dose gradients are used to reduce the dose close to the OAR. Many strategies have been introduced to manage motion. However, in general, these methods suffer from one or more of the following defects: the use of motion alternatives (e.g., bellows, reflector chambers, body surface areas), the invasive implantation of RF transponders, the use of ionizing radiation, poor soft tissue contrast, limited penetration depth, the lack of spatial information of time association, the large irradiation volume caused by large margins, the reduced treatment efficiency due to gating.

[0081] Due to its non-ionizing and high soft tissue contrast properties, MR imaging presents an ideal four-dimensional (4D) imaging platform. However, the space-time-contrast resolution trade-off has hampered the acquisition of true, 4D MR images. For example, the average human respiratory cycle is approximately five seconds. To resolve motion similarly to 4D-CT (where the respiratory cycle is decimated into 10 phases), an ideal 4D-MRI approach would require acquiring artifact-free, high-contrast, high-resolution 3D volumes every 0.5 seconds or less. Even with the latest MRI technologies, including parallel imaging, multi-band excitation, and compressed sensing, this requirement remains unattainable. Furthermore, the introduction of MR-guided RT has placed additional demands on fast, volumetric MR imaging. Abnormality gating and real-time tumor tracking necessitate low latency in the image acquisition, reconstruction, and segmentation chain.

[0082] Thus, according to aspects of the present disclosure, true 4D MR images for use in RT treatment can be obtained by performing retrospective temporal shuffling of k-space data acquired while the patient was breathing normally. This approach does not rely on external respiratory surrogates (bellows, reflector chambers, body surface areas, etc.), improves image contrast, and eliminates imaging radiation dose.

[0083] By way of example, data acquisition is described herein. Specifically, a golden angle radial pulse sequence can be utilized that can switch between 2D and 3D cine modes acquired with balanced steady-state free precession (bSSFP) or spoiled gradient echo (SPGR). The DC signal from each acquired radial spoke can be used as a navigator, providing a respiratory surrogate to guide phase-resolved image reconstruction. Degraded maximum gradient amplitudes and slew rates can be used to minimize eddy current effects.

[0084] Coronal 4D-MR images, for example, can be collected using a sequence run in 3D cine mode with the following parameters: FOV = 380 cm, TE = 0.8 milliseconds, and TR = 1.9 milliseconds. As described, the total scan time can be approximately six minutes. Although a specific scan implementation has been described above, those skilled in the art will appreciate that various modifications to the scan parameters can be performed and are considered within the scope of the present disclosure.

[0085] You can then use Figure 22 The method shown in reconstructs an image based on the acquired data as described. In some aspects, the raw k-space data can be transferred and processed offline by any suitable system. The DC navigator signal for each radial spoke (or phase encoding line) can be determined and plotted in time, as shown in Figure 22 As shown in . A lookup table of bin start times is generated by decimating each respiratory cycle obtained from the navigator waveform into assumed 10 time bins (stages) based on amplitude or phase. The lookup table of bin start times can then be used to reshape the raw 4D-MRI k-space into a hypothetical 10-bin (10 stage) hybrid k-space using the timestamps in the head of each acquired radial spoke (or phase encoding line) or knowledge of the pulse sequence cyclic structure. Spokes or lines reshaped to the same position in the hybrid space can then be linearly combined with a weighting based on temporal proximity to the center of the bin. Compressed sense reconstruction can be used to fill in the missing lines following the reshaping. A 3D FFT can then be applied to convert the reshaped hybrid k-space data into an image for each bin (e.g., 0%, 10%, 20%, etc.). The final image for each time stage can then be interpolated to an assumed 1 to 2 mm 3The resolution of the image is converted to DICOM for subsequent image processing. As described, such images can be used to obtain synthetic ED or CT images at each stage of the respiratory cycle.

[0086] The above acquisitions describe switchable spoiler or balanced sequences. This facilitates better visualization of certain tumors with T1 contrast versus mixed T2 / T1 contrast. In addition, for tumors and tissues with slow dynamics (e.g., washout), target visualization on 4D data can be improved by acquiring post-gadolinium 4D data. For example, visualization of hepatocellular carcinoma (HCC) and esophageal cancer will benefit from the T2 / T1 contrast obtained using a bSSFP sequence. However, visualization of liver metastases will benefit from the T1 contrast obtained using an SPGR sequence at a delay following contrast injection. Visualization of lung lesions will benefit from SPGR acquisition with reduced banding artifacts.

[0087] Furthermore, acquiring a DC navigator during 2D cine imaging is a novel approach that combines the advantages of a pencil beam navigator and rapid cine imaging. Similar to a pencil beam navigator, extremely low latency is achieved (no image reconstruction is required). This provides information from the pencil beam navigator simultaneously with imaging. Furthermore, a DC navigator can be acquired for each phase encoding line and does not require image reconstruction. This has advantages for MRI-gRT, as the DC navigator acquired for each phase encoding line has extremely low latency and can be used in real-time tumor tracking prediction algorithms to drive multi-leaf collimator (MLC) systems.

[0088] In some aspects, motion analysis can be performed using the images obtained as described above. For example, the respiratory phase images of each 4D-MRI data set can be analyzed to determine motion. For example, a gross target volume (GTV) can be contoured on a first time stage and then propagated to the other nine time stages using a deformable image registration process. The procedure can also be repeated for other structures or organs, organs at risk, and so on. Using the center of mass (COM) components along each Cartesian axis of each of the contoured structures, a principal component analysis (PCA) can be performed to determine the eigenvectors and eigenvalues ​​of the GTV and structure COM motion. This motion analysis can be useful for implementation during RT treatment delivery, as will be described.

[0089] Synthetic CT images generated using the described systems and methods exhibit several advantageous features compared to other alternatives to MRI-based radiation therapy, including full compatibility with existing kilovolt (kV) and megavolt (MV) CT-based IGRT techniques, full compatibility with rapidly evolving MRI-based IGRT techniques, and full compatibility with MRI- or CT-based adaptive radiation therapy that accounts for changes in the location, size, and shape of tumors and critical structures and allows for additional reductions in margin size. Furthermore, in some contemplated aspects, the application of the above-described methods can be extended to proton-based radiation therapy planning and delivery methods, as well as systems and applications requiring attenuation correction of positron emission tomography (PET) images, such as combined PET / MRI scanners.

[0090] Return to Figure 1 At process block 12, any desired systems and methods may be used to process or analyze any multimodal image information, such as that described above or acquired in any other manner, to generate image information for use in treatment plan generation and delivery. As will be described, post-acquisition processing of the images may include any number of steps or methods. For example, these may include providing corrections for known noise sources, applying various image transformations, deformations, or registration algorithms, determining optimal display characteristics to facilitate accurate identification of anatomical structures (e.g., windows or levels), outlining targets of interest, generating composite images, and the like.

[0091] Many commercial systems currently in use are capable of generating simultaneously acquired multimodal image data, such as using CT / MRI, PET / CT, or PET / MRI systems. However, in many cases, such systems are not available or are not part of the clinical workflow practice for radiotherapy. Therefore, typically, at process block 12, it may be necessary to combine the individually acquired multimodal images to provide the supplemental information necessary to accurately determine the critical target. Since such images are typically acquired using different scanners and often at different times, a transformation or registration as known in the art is necessary for the images or contoured structures in different data sets to occupy the same coordinate system. Simply put, the process of image registration involves determining a geometric transformation that aligns a point in one view of a target with a corresponding point in another view of the target or another target, where the view can be a two-dimensional or three-dimensional view, or the physical arrangement of the target in space.

[0092] Image registration typically involves inter-modality images, such as CT and MR images, MRI and PET images, PET and CT images, contrast-enhanced CT images and non-contrast-enhanced CT images, ultrasound and CT images, etc. In addition, image registration can also be used for intra-modality imaging, where structures may move or distort between image sequences acquired at different times. Figure 9 An example of registration is shown between a fixed image 902 and a moving image 904. Thus, in the context of adaptive radiation therapy, it may be desirable to adjust or modify the images or contours of targets and / or critical structures identified and employed at the initial radiation therapy planning stage based on routine or current imaging, whose volumes or shapes may have changed above a desired threshold.

[0093] Among many free-form registration methods, the parameterized b-spline deformable image registration (DIR) method offers the benefits of automation, local deformation control, and multimodal capabilities. However, the deformation pattern of this model does not necessarily follow the true motion of organs due to a lack of physical model support, and thus causes artifacts such as bone distortion and inaccurate fine structure correspondence. Therefore, more accurate alternative image registration methods are desired.

[0094] In one aspect of the present disclosure, a new registration method is provided that is designed to alleviate the shortcomings of previous methods. Specifically, a modified parameterized b-spline deformation model is used that is designed to overcome previous limitations by using a new regularization method, namely, bending energy limited diffeomorphism (BELD). Diffeomorphism regularization is typically used to ensure deformation smoothness and is computationally expensive due to the calculation of the Jacobian matrix of all grid points at each iteration step. However, the diffeomorphism regularization result does not necessarily follow the true deformation path. Therefore, in the present disclosure, accurate diffeomorphism is achieved by limiting the difference between the deformation parameters of adjacent grids, rather than the laborious calculation of the Jacobian matrix as is common in traditional methods. A bending energy penalty can be further used to limit the smoothing degrees of freedom, maintain structural topology, and increase registration accuracy. Therefore, the bending energy penalty in different parts of the image can then be conveniently differentiated by automatically segmenting the main components (bone, fluid, tissue, and air) of both the target and source images.

[0095] Specifically, the bending energy can be used as a penalty term in the B-spline deformation model and is calculated as follows:

[0096]

[0097] in, is in position This term is computationally expensive when soft tissue constitutes the majority of the region of interest. If the B-spline transform is written in tensor product form

[0098]

[0099] where l∈{x,y,z} and β n is the nth order B-spline basis, and the penalty function is defined as

[0100]

[0101] where the argument t indicates the difference between two adjacent deformation coefficients. When combined with the B-spline coefficients, the penalty function is

[0102]

[0103] where ζ1 = 0 and ζ2 = 0 would correspond to the volume preservation constraint For any

[0104] The above method can be used to ensure diffeomorphisms, but within this constraint, the deformation still has infinite degrees of freedom, and a better solution will meet the requirement of minimizing the bending energy. In the BELD method, instead of building a detailed realistic model, a simplified semi-physical model can be used, where the diffeomorphisms are constrained by minimizing the bending energy at the grid points, which is much lighter in computation. Using the rigidity map Regularization, BELD constraint and feature point distance, the overall penalty function becomes

[0105]

[0106] in and is the physical coordinate of the corresponding feature point, γ i are empirically determined weighting factors. Although the four additional penalty terms may appear to increase computational complexity, each term is only effective in a limited region and can therefore be optimized in successive stages. Furthermore, due to the nonlinear nature of the penalty function and the mutual information metric, a nonlinear conjugate gradient optimization method can be used. The search direction can be defined as a linear combination of the cost function gradient and the previous search direction.

[0107] In some configurations, the deformable registration method as described can be implemented as a software tool that is integrated or coordinated with any radiation planning system, image analysis or processing system, and software. Such a tool can include various steps and functions, such as, for example, the ability to: (1) accept input of images of different modalities, (2) convert existing contours on any reference image (e.g., MRI) into delineated volumes and adjust the image intensity within the volume to match the intensity distribution of a target image (e.g., CT) for enhanced similarity metrics, (3) register the reference image and the target image using a suitable deformable registration algorithm as described (e.g., b-spline, Demons) and generate deformed contours, (4) project the deformed volume onto the target image and calculate the mean, variance, and centroid as initialization parameters for a continuous fuzzy connectivity (FC) image segmentation on the target image, (5) generate an affinity map from the FC, and (6) generate contours by modifying the deformed contours using the affinity map using a gradient distance weighted algorithm. Furthermore, such software tools can benefit from GPU processing, which can be orders of magnitude faster than using CPU processing due to the highly parallel and efficient processing structure.

[0108] This approach was tested in various clinical settings. Figure 10 An example of BELD-regularized deformable image registration applied to a breast cancer case is shown, where the patient is in a supine position. From left to right, the images in panel (A) show the original MR and CT images, respectively, where the CT image contains contour structures. Subsequently, the image on the left of panel (B) shows the contours transferred from CT using DIR using a conventional unregularized method, and the corresponding transformation field on the left of panel (C). In contrast, the images on the right of panels (B) and (C) show the transferred contours and deformation fields, respectively, using the BELD regularization method as described. Figure 10 Improvements using BELD regularization over conventional (unregularized) methods are demonstrated for breast cancer cases where the patient is in a supine position. Clearly, the unregularized DIR results in an unrealistic deformation field with reduced registration accuracy. In contrast, with the disclosed BELD regularization, both the deformation field and registration accuracy are improved.

[0109] Figure 11An example of BELD-regularized DIR for a breast cancer case in which the patient is in the supine position is shown. The top and bottom images show the original MR (top) and CT (bottom) images, respectively. Panel (A) shows the breast and lumpectomy cavity contours from an MR image overlaid onto a CT (bottom) image using a rigid registration algorithm. In contrast, the top image of panel (B) shows the contours transferred from CT to MR, while the bottom image of panel (B) shows the contours transferred from MR to CT based on BELD-regularized DIR. The blue contours are deformed from the red contours on the original MR image. Compared to the unregularized registration, which results in a difference of up to 10 mm (mean 4.2 mm) between the actual contours and the transformed contours, the BELD-regularized DIR reduces this mean difference to 1.8 mm. This demonstrates that the present disclosure allows the contours of the breast and lumpectomy cavity to be accurately transformed from MR to CT for use in radiation therapy planning.

[0110] Furthermore, this approach was tested on CT and MR images of pancreatic cancer patients acquired at the same respiratory phase to minimize motion distortion. Dice's coefficients were calculated for direct delineation on the target image. Contour generation by various methods, including rigid transfer, automatic segmentation, deformable transfer only and regularized b-spline methods, as described above, was compared. Although fuzzy-connected image segmentation involves careful parameter initialization and user involvement, automatic contour transfer by multimodal deformable registration as described provides up to 10% improvement in accuracy over rigid transfer. Providing two additional steps of adjusting the intensity distribution and modifying the deformed contour using affinity mapping can further advantageously improve the transfer accuracy by up to 14%.

[0111] Figure 12 An example of the results from multimodal contoured structures processed by deformation fields according to the present disclosure is shown. Specifically, PET, and various MRI including T1, T2, DWI, and DCE were rigidly registered with CT images of a representative patient with pancreatic cancer. Gross tumor volume (GTV) and organs at risk (OAR) were delineated on the multimodal images and processed using the deformable multimodal image registration tool as described. Substantial changes in the contours from the multimodal images were observed for both the GTV and the OARs, as shown in FIG. Figure 12 As shown in . Figure 12 The top two images 1202 show large variations in the volume of several structures among the different coverage contours generated using the multimodality images. The derived deformation fields are then applied to deform the corresponding modality contours, which are then overlaid onto the planning CT. Figure 12The next two images 1204 show the increased overlap between the different modality contours after processing with the deformation field. When using the T1-weighted contour volume as the nominator, the delineated volume varies between the image modalities from 0.5 to 1.8 for the GTV and from 0.81 to 1.05 for the OAR. The overlap ratio between the different modalities varies from 0.22 to 0.74 for the GTV and from 0.65 to 0.84 for the OAR. After deformable image registration, the contour volumes are altered by the deformation field by 6% to 11%. Figure 12 The remaining changes observed in DIR after DIR are primarily due to inherent differences between imaging modalities. These changes do not affect the volumetric change between different modalities but significantly improve the overlap between contours from different modalities, for example, from 0.55 to 0.82 for GTV. Thus, deformable image registration of multimodal images increases the consistency between contours from different imaging modalities, improving the accuracy of target and normal structure delineation for radiotherapy planning for pancreatic cancer.

[0112] Registration algorithms often use smoothing kernels to reduce discontinuities in the deformation field applied between iterations. In another aspect of the present disclosure, a novel method is provided for implementation in systems and methods configured to perform image registration, wherein the dimensions of the Gaussian smoothing kernel are adaptively modified during runtime based on a threshold, such as the convergence rate of deformable image registration. Specifically, a large kernel size may be used initially and then reduced with subsequent iterations. With other methods, as the number of iterations increases, the convergence rate begins to decrease until it reaches an asymptotic value. At this point, the accuracy of the registration remains unaffected by further iterations. In contrast, in the present disclosure, varying the kernel size allows asymptotic convergence to be avoided, and the accuracy can then continue to improve towards a new range of possible values. In this way, as the size of the smoothing kernel is gradually reduced, the accuracy continues to improve. This method has several advantages that contribute to improved registration accuracy compared to traditional methods that employ demons-based registration algorithms. In particular, one advantage is that the kernel is robust enough to handle large deformations, such as those found in the bladder and rectum, sensitive enough to handle fine details, and is also typically unaffected by non-physical deformations. In addition, it also provides spatial alignment between common structures in the images, making further registration simpler.

[0113] Figure 13An example showing the effect of a variable kernel smoothing technique on registration accuracy is shown. Typically, accuracy improves after multiple subsequent iterations, as indicated by a decrease in the average pixel difference between images. However, when only a given kernel size is used, the accuracy begins to converge to a constant value (dashed line). Thus, by adaptively adjusting the kernel size (circled area) when specified conditions are met, premature asymptotic convergence can be avoided, and the algorithm is allowed to transition to a new range of possible values. This results in improved registration accuracy, exceeding the possible limits of a constant kernel size.

[0114] This method allows for fast and accurate deformable image registration of cone beam CT (CBCT) and CT images typically used in online adaptive radiation therapy. Tests have shown that the variable kernel method results in improved accuracy that exceeds the accuracy for a constant kernel size. Specifically, the planning CT and daily CBCT images acquired for 6 prostate cancer patients were registered using the above-mentioned technique. Histogram matching was used to compensate for the intensity difference between the two modalities. The Pearson correlation coefficient (PCC) and volume overlap index (VOI) were used to quantify the registration accuracy. The results show that the iterative reduction of the smoothing kernel size allows the algorithm to converge to increasingly accurate solutions, exceeding the asymptotic limit for a constant kernel size. The average VOI was calculated for the bladder, prostate, and rectum, with values ​​of 91.9%, 68.7%, and 78.2%, respectively. The correlation coefficient was calculated for each fraction of the overlapping CT and CBCT scan volumes in each patient data set. The average PCC values ​​for these six patients were 0.9987, 0.9985, 0.9982, 0.9980, 0.9985, and 0.9985. The typical run time for an image volume of 512x512x70 was 4.6 minutes. Thus, the results using the disclosed DIR technique validate its use for deformable CT-CBCT registration, eliminating the need for multi-resolution processing and continuous upsampling of images (which can be computationally intensive).

[0115] To illustrate the robustness of the described algorithm in handling large deformations, as seen, for example, in the case of a typical prostate cancer, Figure 14Figure 2 shows the results of the registration for a pelvic CT scan. Specifically, the planning CT (a) was deformably registered to the daily kV cone-beam CT (c) using the described technique. The resulting image (b) is in excellent agreement with the CBCT image. The included graph shows the correlation coefficient for the entire image volume (70 slices) before and after registration, illustrating the consistency of the method. This illustrates the ability of the technique to handle relatively large deformations, especially with reference to the rectal region, where the algorithm is able to compensate for considerable deformations of the surrounding tissue. A noticeable difference exists between the images before registration, while after registration, the resulting images are in good agreement.

[0116] In addition, intra-modality DIR techniques use the Demons algorithm to provide intensity mapping for conversion of multiple modalities to a CT-like contrast scale. These can operate on different types of CT (e.g., cone-beam CT, MVCT), as well as ultrasound (US) images. For US in particular, such methods can implement correspondence functions designed specifically for US. However, a simple relationship between tissue echogenicity and CT Hounsfield units has yet to be established. As such, current US-CT registration practice has been limited to the use of rigid registration or inter-modality methods, such as thin plate splines, which require manual point selection.

[0117] High-frequency ultrasound (HFUS) images are able to reveal accurate spatial information, for example, in skin lesions, and can be used to plan and guide radiotherapy (RT) for skin cancer. Therefore, in yet another aspect of the present disclosure, a system and method for performing fully automatic, accurate, deformable registration between US images and CT images is provided, which can be beneficial for skin cancer patients requiring radiotherapy treatment. Specifically, a novel DIR technique is introduced based on the symmetric force Demons algorithm, in which the size of the Gaussian smoothing kernel can be adaptively adjusted. A correspondence function can also be used to map attenuation values ​​to tumor elasticity and achieve histogram matching. The Pearson correlation coefficient (PCC) can be used as an indicator for evaluating the accuracy of the registration.

[0118] As described, this approach provides robustness to the algorithm under the large deformations typically encountered during skin tumor shrinkage. Since skin lesions, in particular, have the added benefit of having sharp outer boundaries, the present disclosure also facilitates the use of anisotropic diffusion filters to perform edge-preserving smoothing, thereby reducing noise and increasing registration accuracy. Such advantages may be particularly important for superficial US images, which may suffer from poor image quality and low contrast. Furthermore, intensity matching techniques typically used to convert between different CT image types can also be applied to US images acquired at different frequencies, enabling US-US registration.

[0119] To illustrate the feasibility of this approach, according to the present disclosure, HFUS images of skin lesions and their corresponding CT images were registered for a selected region of interest (ROI). Figure 15 An example of registration between CT and HFUS for a patient with a superficial basal cell carcinoma (BCC) in the thigh is shown. The figure shows CT (a) and high-frequency ultrasound (HFUS) images of the lesion, while c) and d) are magnified views of the rectangular areas in a) and b). e) and f) are histogram-mapped and contrast-enhanced HFUS images, respectively, and g) and h) are registered images of CT and HFUS using the variable kernel smoothing technique described above. PCC values ​​of 0.9794 and 0.9815 (enhanced contrast) were observed, indicating excellent agreement between dynamic (HFUS) and static (CT) images. The registration technique also demonstrated the ability to handle large deformations, such as ROI displacements exceeding 200% of the tumor thickness observed near the lesion edge. It was found that the progressive reduction of the kernel dimension prevented non-physical deformations, thereby improving registration accuracy. This robustness is critical for skin cancer RT, as tumors can shrink significantly during treatment.

[0120] Thus, combined with standard image processing techniques, the disclosed method is sufficiently accurate to achieve deformable registration of HFUS and CT images for skin cancer RT. This can be advantageous in the ability to perform image-guided treatment for skin cancer, reducing geographic loss and sparing healthier tissue. The presented technique is sufficiently accurate and robust to handle large deformations, making it a promising tool for RT planning and delivery guidance for skin cancer.

[0121] Software tools used in multimodal image registration, as described, can achieve physically accurate registration with minimal user intervention and computational cost, thereby making multimodal deformable imaging registration fast and accurate, especially in the context of ART replanning or other real-time applications. Specifically, compared to common methods used in radiation therapy planning, the disclosed method can improve the accuracy of contour transfer between different image modalities under challenging conditions of low image contrast and large image deformation.

[0122] Return to Figure 1At process block 14, the processed multimodal images as described can then be received, for example, from an imaging system or data storage device and transferred to a planning system for generating a radiation therapy plan. Specifically, in the case of planning for a first treatment, this process step typically includes determining the beam or radiation source delivery technique and placement based on selected or determined planning objectives. In certain planning methods, this can include generating an eye-view display of the beam, designing the field shape (blockers, multi-leaf collimators), determining beam modifiers (compensators, wedges), and determining beam or source weights. Using contoured critical structures, performed manually or using automated contouring tools, dose calculations can then be performed based on a selected algorithm or method. Using the set relative and absolute dose normalization and dose prescription, the plan quality is then assessed based on visual coverage comparisons, dose volume histogram analysis, and tumor control and normal tissue complication probability. Automated or semi-automated optimization tools can then allow plan refinement based on planning objectives and tolerances.

[0123] In some cases, the generated plan may also involve accounting for intra-fractional motion of the target and critical structures, such as respiratory motion. Thus, according to some aspects of the present disclosure, a 4D planning method for intra-fractional motion based on 4D CT or 4D MR images divided into, for example, 5 phases (although other values ​​are possible) includes the following aspects: (i) creating an IMRT or VMAT plan for a specific respiratory phase (e.g., end of inspiration) based on the phase images by full-scale optimization, (ii) populating this plan onto the remaining respiratory phases by applying a segmentation aperture deformation (SAM) algorithm to account for anatomical changes between respiratory phases, and (iii) combining all phase-specific plans to form a 4D plan. Essentially, all phase plans have the same number of segmentations, where each segmentation has several different MLC patterns, depending on the number of phases with the same MU, jaw setting, gantry angle, etc. The 4D plan generated in this manner can be delivered using dynamic MLC and an adjustable dose rate so that each segmentation can be delivered within an integer number of respiratory cycles, thereby ensuring that each MLC pattern is delivered with its corresponding respiratory phase. Specifically, the delivery file from the multi-leaf sequencer is generated as a function of the respiratory cycle. At each treatment segment, the delivery file can be rapidly updated with the respiratory cycle obtained immediately before the delivery. If the respiratory cycle changes during the delivery, the delivery file can also be updated as frequently as necessary during the delivery.

[0124] For situations when respiration is substantially different than when the planning images were acquired, online adaptive 4D planning and delivery can be achieved as follows: (1) a reference plan is generated based on a single phase image (reference phase, e.g., end of inspiration) from the planning image set; (2) a full rehearsal QA is performed on the reference plan; (3) at the time of the treatment segment, the reference plan is modified using the SAM algorithm based on anatomical changes on the reference phase image of the 4D image acquired with the patient in the treatment position immediately before treatment delivery; (4) a newly established reference plan (adaptive plan) is populated based on each of the remaining phase images of the day using the SAM algorithm; (5) all phase plans are then sequenced taking into account the most current respiratory information; and (6) the 4D plan is delivered under real-time image guidance (e.g., orthogonal cine MRI) and its delivery can be interrupted if the MLC fails to track the target due to sudden changes in patient positioning or respiratory motion.

[0125] Changes to target and critical structures between treatment segments, or over the course of multiple segments, can produce significant deviations from the original plan, potentially having a negative impact on treatment effectiveness. Thus, providing a radiation therapy plan at process block 14 may, in some cases, include performing an application of techniques for adapting the original or first-time plan, such as those provided by ART, to account for patient-specific anatomical and / or biological changes during the course of treatment using any combination of online and offline methods, as will be described.

[0126] Traditional plan optimization methods are typically not fully automated, as optimization algorithms often deliver undesirable results, requiring additional iterations using different weighted objectives. This is due to the difficulty in estimating how much can be achieved for different planning objectives, typically defined in terms of dose-volume constraints. For example, some objectives may be formulated such that portions of an organ or structure are not allowed to exceed the dose. As an example, one such objective could be that 30% of the rectum's volume cannot receive more than 60 Gy of radiation.

[0127] If the target for an individual OAR is too stringent, such as when the dose volume is set too low, or the relative weight of the target is too high compared to other targets, this can lead to undesirable effects such as insufficient tumor coverage or radiation dose hotspots. Because the volume of OARs often changes from day to day, it may not be possible to know exactly how much OAR radiation sparing is possible, or specifically what minimum percentage of the volume needs to be irradiated with a specific amount of dose. Therefore, the iterative process of finding the correct target weighting is tedious and time-consuming, and requires expertise and experience.

[0128] Previously proposed techniques aimed at overcoming this obstacle attempt to determine the best achievable dose-volume histogram (DVH) by computing OAR / target coverage fragments or utilize multi-parameter criteria optimization, in which each target is optimized individually and then linear interpolation is performed to determine a suitable compromise. Specifically, multi-criteria optimization methods are impractical for online ART because they require multiple optimizations. Furthermore, both methods are designed for initial plan optimization, not replanning, and require the generation of a complete set of OARs.

[0129] Contours obtained by organ contouring are a time-consuming part of any online adaptive replanning, with OARs typically representing the majority of the contours to be contoured. Organs in the digestive tract, such as the liver, rectum, intestines, and stomach, are particularly problematic because they are large and require numerous contours. Furthermore, their size, shape, and content vary significantly and unpredictably from day to day, making it extremely difficult for automated contouring methods to accurately generate contours, typically requiring human contouring / editing.

[0130] As described, fast online replanning methods require efficient methods to modify radiation therapy plans, typically while the patient is lying on the treatment table. To overcome the shortcomings of previous online replanning methods, in another aspect of the present disclosure, systems and methods are provided that use a novel gradient-maintained (GM) algorithm that allows fully automated online ART replanning without the need for OAR contouring.

[0131] As will be described, the GM algorithm requires establishing or adjusting beam or segmentation apertures based on that day's target and optimizing beam or segmentation weights using autonomously or semi-autonomously generated ring structures or isodose contours automatically converted from, for example, isodose lines on that day's images. The algorithm determines dose gradients from the original planned dose distribution for each of the critical structures near the target and initiates a replanning optimization procedure aimed at maintaining the dose gradients of the original plan. This algorithm enhances automation, significantly reduces planning time, and improves the consistency throughput of online replanning.

[0132] Specifically, unlike methods that attempt to determine "what is achievable" in a time-consuming manner so that optimization targets can be set, the method of the present disclosure effectively determines the dose gradients imposed by physical dose deposition constraints, thereby providing a direct measure of "what can be achieved." In this way, the transfer of dose gradients from the original plan to the modified plan can result in the best plan achievable with respect to the allowed physical dose deposition constraints. This approach offers several advantages over prior art techniques. Specifically, it may not be necessary, for example, to generate organs at risk (OARs) daily, thereby reducing the delineation practice to only the treatment target. Furthermore, conventional optimization algorithms strive to meet a set of targets specified according to dose-volume constraints for each of the OARs, which typically requires, for example, the generation of volumes on daily image sets. In contrast, the method of the present disclosure strives to achieve a specific dose gradient from the surface of the target toward each organ at risk, allowing the optimization to be more reproducible and predictable, and less likely to require multiple trial-and-error iterations. In this way, human involvement, along with the time cost of optimization, can be reduced or eliminated, an attractive feature for rapid online replanning.

[0133] The GM algorithm as described can be configured to work in dedicated optimization software and hardware, or can be combined in conjunction with a system configured for fixed-beam or rotating-beam radiation delivery or a combination thereof. In some contemplated implementations, maintaining a dose gradient from the surface of the target as described can be achieved by having the planner communicate the desired planning objective to the optimization algorithm via a weighted sum of objectives referred to as an "objective function," which in some aspects can be generated using a dose gradient. Since many commercially available optimization systems only allow the objective function to be defined in terms of a volume of interest and a dose volume objective, the method of the present disclosure can be adapted to allow user-generated partially concentric ring structures (PCRs) to be used in defining the desired dose gradient for the optimization algorithm. In this way, any desired system or method can be used to generate the PCRs. For example, manual or automatic contouring methods can be used to generate the PCRs. The anatomical sites that will benefit most from the method of the present disclosure are likely to be the prostate and pancreas, which have typically large day-to-day unpredictable volume changes and relatively small target organ sizes given the large number of surrounding OAR structures. Figure 23 An exemplary PCR 2300 for the prostate cancer case is shown in .

[0134] Go to Figure 16, an exemplary online plan adaptation process 1600 is shown that illustrates an operating mode for a GM algorithm. Specifically, this process 1600 can be utilized when any condition that would trigger a plan modification has been met, such as, for example, in the event of reduced target coverage or increased radiation dose hotspots. Process 1600 begins at process block 1602, where a radiation therapy treatment plan is provided as input to any system (such as a treatment planning or delivery station) configured to perform process 1600. This treatment plan input, defined by at least the radiation dose distribution, can be a plan generated at the beginning of treatment scheduling, or can be a plan that is subsequently modified. In some aspects, a plan can be generated at process block 1602 using conventional planning techniques (such as IMRT optimization). Such a plan can use the full set of contours on the planning images to achieve the desired dose volume goals for both the target and the OARs, and can be the best possible plan for the planning technique utilized, assuming that the planner is not necessarily under time constraints.

[0135] At process block 1604, a dose gradient is determined using the radiation dose distribution from the provided or generated plan, the dose gradient defining the change in radiation dose from at least one target structure toward any or all non-target structures or targets at risk. Such a dose gradient can be used to define any number of optimization goals that may be desired during the online plan optimization process as described. In particular, the radiation dose distribution in the treatment plan that has been optimized may potentially represent the best achievable solution for the desired dose constraints and, therefore, may prove to be a good starting point in the optimization process.

[0136] Subsequently, at process block 1606, updated image information is provided, which may include a multimodal image set. For example, the updated image set may include any number of magnetic resonance images, computed tomography images, ultrasound images, positron emission tomography images, and synthetic electron density images, or any combination thereof. The updated image information may then be used to generate any updated contours of the target volume at process block 1608. As described, such contours may be generated autonomously or semi-autonomously.

[0137] The updated target contour can then be used at process block 1610 to generate a PCR structure that is generally arranged about or around the updated contour of the target structure. In some aspects, the PCR structure can be centered about the updated target contour and directed toward any or all non-targets, or targets at risk, and points along the dose volume histogram of the PCR structure can be used to generate an objective function for use in the optimization process. Subsequently, at process block 1612, an optimization process can be performed using the generated PCR structure along with the objectives defined according to the determined dose gradient from process block 1604. During this optimization process, which may require any number of iterations, the objective function can be modified according to the defined objectives to achieve a targeted dose constraint or dose gradient, for example, a dose gradient from the surface of the updated target volume about any or all non-targets or targets at risk.

[0138] Subsequently, at process block 1614, a report is generated representing the adapted radiation therapy plan resulting from the plan optimization process, which may take any shape or form as desired or required by the treatment plan verification or delivery system.

[0139] Since, as described, the contoured structures may be subjected to deformable registration using standard algorithms, which typically do not provide very accurate and reliable contours, the direct use of deformed contours may prove problematic for some routine planning optimization schemes. In contrast, the GM algorithm does not require the contours to be very accurate because only the relative positions of the structures can be used, and changes in organ shape may not affect the accuracy of the PCR. For example, in order to generate a PCR towards a specific critical structure, it may only be necessary to know what part of the ring around the target will be towards the critical structure. Therefore, the PCR area is the projection of the critical structure onto the surface of the target or the ring. Since the 3D volume information of the critical structure is contracted (projected) onto the 2D target surface to determine the part of the ring facing the critical structure, the inaccuracy in the volume of the critical structure is greatly reduced by the contraction. In addition, a margin, for example, a 3 mm margin, can be applied around the ring area. For very large structures such as the intestine, only the part of the critical structure that is within a specific distance (for example, 2 cm) of the target can be included. Furthermore, it is possible to generate PCR in a simpler manner for certain organs, such as the rectum, where the PCR can be made to cover the entire extent from the prostate surface in a posterior direction.

[0140] For some structures, such as the bladder, automatic contouring can achieve daily contours with reasonable accuracy. Thus, using PCR can provide similar time reductions in contouring. However, since bladder volume varies significantly from day to day, this makes the optimal target weight in the objective function difficult to predict. Therefore, PCR-based optimization may still be advantageous in terms of providing a more predictable optimization, as mentioned above.

[0141] In the study, the method of the present disclosure was investigated to quickly generate adaptive plans. Using daily treatment CT, only the target structure (CTV) was outlined. Subsequently, a PCR structure of uniform thickness (i.e., 3 mm) was automatically generated using an internal program. There was a separate PCR structure towards each of the important critical structures surrounding the target. This was done by first generating a ring around each target all the way around and then finding the intersection of the ring with the projection of each OAR. Several ring structures were formed for each OAR, such as PCR #1: from 0 to 3 mm, PCR #2: from 3 to 6 mm, PCR #3: from 6 to 9 mm, and so on.

[0142] Figure 17 A diagram is shown showing an example comparing online replanning schemes using conventional (left) and GM (right) methods in three steps, i.e., daily image acquisition, contour generation based on daily images, and plan reoptimization based on the new contours. In the conventional method, the original objective function, which may be a complex set of dose-volume constraints, is required in step 2 and used in the subsequent plan optimization. In contrast, the GM method of the present disclosure provides a simplified method with significant planning time savings, whereby only target and PCR contours, which can be generated automatically, are required, along with a determined dose gradient.

[0143] Assuming that the initial plan is typically not under the time constraints of online planning, an original plan can be generated in any way, for example, using the original volume and dose volume constraints. Such an original plan may describe the best achievable plan, which will have the steepest achievable dose gradient towards each OAR. Using the original plan with a nominally optimal dose distribution, several points along the DVH of the generated PCR are recorded and used to generate an objective function, which is stored for online re-optimization. The optimization step is then performed using the latest generated PCR together with the stored objective function. In order to quickly reach the dose gradient target, the optimization of the adaptive plan is started from the original plan on the images of that day. Since the dose gradient can be the best achievable, each of the target targets (for each PCR) is set to the exact DVH point obtained from the original plan and is equally weighted.

[0144] This approach as described differs from typical optimization practice in which the DVH targets are almost always set to points that are "better than actually desired", making the optimization process unstable and prone to undesirable results. Since, in the method of the present disclosure, it is known what is achievable from the original plan, the DVH targets can be set to those points. The main advantage of this approach is that it eliminates the necessity to outline the entire set of structures, and only the outline of the targets is required, thereby significantly reducing re-planning time. Moreover, this approach makes daily re-planning more reproducible, eliminating the need for trial and error efforts, because the gradients towards each structure are usually achievable (already achieved in the original plan), so adjusting the optimization criteria is not necessary for daily dissections. On the other hand, standard optimization methods rely on dose-volume based criteria, which can vary significantly from day to day due to changes in the volume of the organ.

[0145] A situation worth considering is when the volume of the OAR is very small over the day and maintaining the original dose gradient would result in an unacceptably high volume of the organ receiving a dose. Since there is a physical limit to the size of the dose gradient, once reached, the gradient cannot be increased any further without sacrificing other plan quality parameters, such as target coverage. If the steepness of the gradient for a day can be increased without sacrificing target coverage, this indicates that the dose gradient in the original plan may not be as high as possible. Therefore, it is beneficial that the original plan quality is indeed optimal in terms of dose gradients towards critical structures. If this is the case, the dose gradient cannot be increased any further without sacrificing other parameters, and maintaining the original gradient will meet the goals of online replanning.

[0146] Figure 18 An example comparing dose distributions from image-guided radiation therapy (IGRT) repositioning, traditional full optimization based on a full contour set, and gradient-preserving replanning according to the present disclosure is shown. Clearly, the GM-based dose distribution, such as the dose to the duodenum, is equivalent to the dose distribution from traditional reoptimization, but the dose distribution obtained using repositioning is substantially improved. Furthermore, the planning time required for the GM approach is only 10% of that required for traditional reoptimization.

[0147] Figure 19 Shown is a comparison of inter-segment changes in 6 daily treatments used to address the prostate cancer setting. Figure 18 Again, based on the six sets of DVHs shown for the rectum, bladder, and planning target volumes (PTVs) (prostate and seminal vesicles), it is clear that the GM-based online replanning scheme is equivalent to the traditional global reoptimization and significantly better than the IGRT repositioning method.

[0148] return Figure 1At process block 16, a number of quality control protocols and procedures may be performed prior to treatment for the provided radiation therapy plan. As is known in the art, such protocols are typically based on a clinical workflow adopted and established by the treatment system, providing verification of the accuracy of delivery as dictated by the treatment plan. Prior to radiation delivery, the patient may typically undergo position setup and immobilization, and the therapist may perform a number of position verifications and adjustments as needed based on pre-treatment imaging. Corrections to patient treatment parameters performed while the patient is on the treatment table (classified as "online" corrections) may include translational or rotational adjustments, as well as modifications to the treatment plan using ART, as described.

[0149] Then at process box 18, the radiotherapy treatment is delivered, followed by the step of evaluating and / or modifying the treatment at process box 20, including the report provided by the recording and verification system. This step can trigger multiple corrective actions performed after the daily treatment has been delivered, thereby affecting the treatment on subsequent days. Therefore, in another embodiment of the present disclosure, a system and method for implementing an offline optimization process is provided. The offline optimization process consists of calculating the cumulative dose delivered using a set of deformably registered images (e.g., daily images (CT, MRI or US)) from previous segments, and generating an optimized plan for subsequent treatment to correct any accumulated residual errors from any number of intra-segment changes and other changes that are not corrected by any online replanning scheme. For example, residual errors can occur during radiotherapy for prostate cancer because the prostate can drift systematically or move suddenly due to gas passing through the rectum. These changes may not be taken into account by online replanning using images acquired immediately before treatment.

[0150] Specifically, the dose delivered at any i-th segment using the online replanning scheme as described can be reconstructed using the parameters of the delivery (e.g., aperture and number of monitoring units (MUs)). The cumulative dose up to the i-th segment can be calculated based on the images of that day using any deformable image registration method. Since the cumulative dose distribution and the expected dose distribution (e.g., gold standard) can typically be generated based on the same image (i.e., CT or MRI of that day), the residual error from the intra-segment variations up to the i-th segment that was not accounted for by the previous online replanning scheme can be calculated by simply subtracting the cumulative dose distribution from the expected dose distribution. This residual error can be provided as input into the planning system to be used as an initial background, the generated background including positive and negative numbers. Adaptive optimization using this background can generate an initial plan for online replanning at subsequent segments. For example, Figure 20 Offline optimization as described is shown to potentially be used in conjunction with an online replanning approach.

[0151] The disclosed method as described can be used to correct residual errors after online replanning of one or more segments. For a subset of patients, it is likely that online replanning as described may leave fewer errors than the reoptimized plan. For such cases, an offline adaptive optimization method can be used less frequently (e.g., on a weekly or biweekly basis) to provide correction for the accumulated residual errors from multiple segments rather than just the previous segment.

[0152] Features suitable for these combinations and subcombinations will be apparent to those skilled in the art upon reviewing this application as a whole.The subject matter described herein and in the claims recited is intended to cover and encompass all suitable changes in the technology.

Claims

1. A system for adapting a radiation therapy treatment plan, the system comprising: a data storage device configured to store image data acquired by the magnetic resonance imaging system; as well as At least one processor configured to: providing an initial radiation therapy plan to be adapted based on the updated image set, the initial radiation therapy plan having a radiation dose distribution, determining a plurality of dose gradients using the radiation dose distribution; defining an optimization target using the dose gradient; The updated image set is generated by the following steps: receiving a plurality of acquired image data from the data storage device, wherein the acquired image data comprises 4D MR image data including time-shuffled k-space data; applying correction to the 4D MR image data to generate a series of corrected image data; assembling the series of corrected image data to produce a set of relaxation maps; performing segmentation of a plurality of regions of interest using the set of relaxation maps; classifying the plurality of regions of interest using the set of relaxation maps to produce a plurality of classified structures; assigning electron density values ​​to the classified structures using an assignment process; as well as generating a set of corrected synthetic electron density images using the electron density values ​​of the classified structures; generating an updated set of contours representing an updated target volume using the updated set of images; forming a set of partial rings using the updated set of contours, the set of partial rings being arranged around the updated set of contours representing the updated target volume; performing plan optimization using the optimization objective and the set of partial loops; as well as A report is generated representing an adapted radiation therapy plan obtained using the plan optimization. 2 . The system of claim 1 , wherein the plurality of dose gradients define a variation in radiation dose from at least one target structure toward each of a plurality of non-target structures.

3. The system of claim 1 , wherein the updated image set is a multimodal image set comprising the 4D MR image data and at least one of: a computed tomography image set, an ultrasound image set, a positron emission tomography image set, and a composite image set.

4. The system of claim 1 , wherein the processor is further configured to combine the set of corrected synthetic electron density images with at least one of a magnetic resonance image set, a computed tomography image set, an ultrasound image set, a positron emission tomography image set, and a synthetic image set using deformable image registration to create the updated image set.

5. The system of claim 4, wherein the deformable image registration comprises a parameterized b-spline deformation model using a bending energy finite diffeomorphism (BELD) regularization technique.

6. The system of claim 4, wherein the deformable image registration comprises a symmetric force Demons technique whereby a Gaussian smoothing kernel is adaptively adjusted according to a threshold.

7. The system of claim 1, wherein the set of partial rings comprises partially concentric rings centered around the updated set of contours representing the updated target volume and directed toward a plurality of targets at risk. 8 . The system of claim 7 , the processor further configured to determine a plurality of points along a dose volume histogram of the set of partial rings, the plurality of points being used to generate the objective function.

9. The system of claim 8, the processor further configured to modify the objective function according to the optimization objective to achieve a plurality of target dose gradients from a surface of the updated target volume with respect to each of the plurality of targets at risk.

10. The system of claim 1 , wherein the initial plan is determined using an offline optimization process that uses accumulated doses from at least one of a plurality of treatment segments to generate a background representing a plurality of residual errors accumulated from at least one of a plurality of intra-segment variations.

11. The system of claim 1, wherein the initial radiation therapy plan and the adapted radiation therapy plan are 4D plans generated using a segmentation aperture deformation (SAM) algorithm based on the 4D MR image set. 12 . The system of claim 1 , wherein the processor is further configured to generate a set of isodose contours based on the radiation dose distribution. 13 . The system of claim 12 , the processor further configured to perform the plan optimization using the isodose contours and a segmentation aperture deformation (SAM) algorithm.

Images