Assess the quality of image segmentation into different tissue types for treatment planning using tumor treating fields (TTField)

Abstract

To plan tumor treating field (TTField) therapy, a model of the patient's head is typically used to determine where the transducer arrays will be positioned during treatment. The accuracy of this model depends largely on accurate segmentation of the MRI images. The quality of the segmentation can be improved by presenting it to a previously trained machine learning system. The machine learning system generates a quality score for the segmentation. Revisions to the segmentation are accepted, and the machine learning system scores the revised segmentation. The quality score is optionally used to determine which segmentation provides better results by running simulations of the model corresponding to each segmentation for multiple different transducer array layouts.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U.S. Provisional Application No. 62 / 789,660, filed January 8, 2019, which is incorporated herein by reference in its entirety. Background Art

[0003] Tumor Treating Fields (TTField) is an FDA-approved therapy for the management of glioblastoma multiforme (GBM) and is being studied for multiple additional indications. See, for example, Stupp, R. et al.: Effect of Tumor-Treating Fields Plus Maintenance Temozolomide vs Maintenance Temozolomide Alone on Survival in Patients With Glioblastoma: A Randomized Clinical Trial. JAMA. 318, 2306-2316 (2017), which is incorporated herein by reference. TTFields are delivered to the tumor through the placement of a transducer array (TA) placed on the patient's scalp and the delivery of the electric field to the tumor area. Greater electric field strength or power density within the tumor is associated with improved treatment outcomes. Both parameters are directly dependent on the location of the TA.

[0004] One method for determining where to place a TA on a particular patient's head is to create an accurate model of the patient's head that includes the tissue type (e.g., white matter, gray matter, CSF, etc.) for each voxel in the image; position a simulated TA on the model; apply a simulated voltage to the TA; and calculate the electric field or power density at each tumor voxel in the model. However, until now, creating an accurate model of a patient's head that includes the tissue type of each voxel has been an extremely labor-intensive process.

[0005] There are automatic and semi-automatic methods for creating a model of a patient's head, and these methods are not very labor-intensive. However, the accuracy / quality of the models generated using these methods can vary dramatically from image to image. And when a model with poor accuracy / quality is used to calculate the electric field or power density at each tumor voxel in the model, the calculated electric field or power density can diverge significantly from the actual electric field or power density. And this divergence can have clinical implications because when the actual electric field or power density is significantly lower than predicted, the patient may not receive the full benefit of his or her TTField therapy. Summary of the Invention

[0006] One aspect of the present invention is directed to a first method for estimating the quality of an image segmentation. The first method includes training a machine learning system to estimate the quality of an image segmentation based on a plurality of reference images and at least one quality score assigned to each of the reference images. The first method also includes presenting a new image and a segmentation of the new image to the machine learning system; receiving at least one first quality score for the segmentation of the new image from the machine learning system; and outputting the at least one first quality score for the segmentation of the new image.

[0007] In some instances of the first method, a machine learning system is trained to estimate the quality of image segmentation based on (a) the quality of affine registration, (b) the quality of deformable registration, (c) input image properties, and (d) geometric properties of the segmented tissue.

[0008] In some instances of the first method, the quality of the deformable registration is determined based on the field deviation, directional variability, and mean per-axis variability of the deformation. In some instances of the first method, the input image attributes include the shortest axis length and signal-to-noise ratio of each tissue, respectively. In some instances of the first method, the geometric attributes of the segmented tissue include the volume of the shape and the number of connected components, each of which is calculated per tissue.

[0009] In some instances of the first method, the machine learning system is trained to estimate the quality of the segmentation of the new image based on at least one global quality feature, at least one local quality feature, and the shortest axis length of the intracranial tissue. In these instances, the machine learning system is trained to estimate the quality of the segmentation of the new image based on image quality and tissue shape properties of the extracranial tissue.

[0010] Some examples of the first method further include automatically adjusting the segmentation; presenting the adjusted segmentation to the machine learning system; receiving at least one second quality score for the adjusted segmentation from the machine learning system; and outputting an indication when the at least one second quality score indicates improved quality relative to the at least one first quality score.

[0011] Another aspect of the present invention is directed to a second method for improving the quality of an image segmentation. The second method includes presenting a new image and a first segmentation of the new image to a machine learning system. The machine learning system has been trained to estimate the quality of the image segmentation based on a plurality of reference images and at least one quality score that has been assigned to each reference image. The second method also includes receiving at least one first quality score for the first segmentation of the new image from the machine learning system; outputting the at least one first quality score for the first segmentation of the new image; and accepting at least one revision to the first segmentation from a user. The second method also includes presenting a second segmentation of the new image to the machine learning system, wherein the second segmentation is based on the at least one revision to the first segmentation; receiving at least one second quality score for the second segmentation of the new image from the machine learning system; and outputting the at least one second quality score for the second segmentation of the new image.

[0012] In some instances of the second method, at least one first quality score of the first segmentation of the new image consists of a single first quality score for the entire new image, and at least one second quality score of the second segmentation of the new image consists of a single second quality score for the entire new image.

[0013] In some instances of the second method, at least one first quality score for the first segmentation of the new image includes a separate first quality score for each of a plurality of regions within the new image, and at least one second quality score for the second segmentation of the new image includes a separate second quality score for each of the plurality of regions within the new image.

[0014] In some instances of the second method, at least one first quality score for the first segmentation of the new image includes a separate first quality score for each of a plurality of tissue types within the new image, and at least one second quality score for the second segmentation of the new image includes a separate second quality score for each of the plurality of tissue types within the new image.

[0015] In some examples of the second method, the at least one modification to the first segmentation includes adjusting a shift in the calculated probability that a given voxel belongs to a particular tissue type.

[0016] In some instances of the second method, at least one revision to the first segmentation includes adjusting, for all voxels in the new image corresponding to tissue, (a) increasing the calculated probability that the given voxel belongs to the first tissue type, and (b) decreasing the calculated probability that the given voxel belongs to the second tissue type.

[0017] Some examples of the second method further include: (a) calculating how the determined quality of the first segmentation may change the average expected power density in the target region of the new image corresponding to the tumor by running a finite element simulation using the tissue type selected based on the first segmentation, and (b) calculating how the determined quality of the second segmentation may change the average expected power density in the target region of the new image by running a finite element simulation using the tissue type selected based on the second segmentation.

[0018] Some examples of the second method further include: (a) calculating how the determined quality of the first segmentation may change the lowest reasonably expected power density in the target region of the new image corresponding to the tumor by running a finite element simulation using the tissue type selected based on the first segmentation, and (b) calculating how the determined quality of the second segmentation may change the lowest reasonably expected power density in the target region of the new image by running a finite element simulation using the tissue type selected based on the second segmentation.

[0019] Some examples of the second method further include maximizing a lowest reasonably expected power density in the tumor bed.Some examples of the second method further include calculating a confidence score for at least one candidate layout, wherein the confidence score indicates a probability that a given layout is an optimal layout.

[0020] Another aspect of the present invention is directed to a third method for determining the quality of a segmentation. The third method includes measuring a Dice coefficient between a calculated head segmentation of a training set and a verified head segmentation; extracting features to account for deformable registration quality; incorporating the shortest axis length and signal-to-noise ratio of each tissue as input image quality; and comparing the measured Dice coefficient between the sets with a prediction of the Dice coefficient.

[0021] In some examples of the third method, features extracted to account for deformable registration quality include: deformed field deviation, directional variability, and mean per-axis variability.

[0022] In some instances of the third method, the volume of a shape and the number of connected components are used to describe the segmented shape. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1A Depicted is an MRI T1w image of the head of a patient with GBM.

[0024] Figure 1B Depicts the use of the reference algorithm Figure 1A MRI segmentation.

[0025] Figure 1C Depicts the corresponding Figure 1A Validated segmentation of MRI.

[0026] Figure 2AThe absolute Pearson correlation between the calculated features and the Dice coefficient for each segmented tissue is depicted.

[0027] Figure 2B Decision tree regressor output predictions depicting the proposed features and dice coefficients.

[0028] Figure 3 is a flowchart depicting a set of steps for using a quality estimation machine learning system to improve positioning planning of a transducer array for TTField therapy.

[0029] Figure 4 An example of a suitable user interface that may be used to make quality presentations to a user and also allow the user to make corrections to the segmentation is depicted.

[0030] Figures 5A-5C Depicted is an example of a user interface that may be used to summarize the estimated field strength (or power density) in a tumor bed for five different candidate transducer array layouts.

[0031] Various embodiments are described in detail below with reference to the drawings, wherein like reference numerals represent like elements. DETAILED DESCRIPTION

[0032] To achieve the desired field strength (or power density) in the tumor, a key step in improving TA placement is to correctly segment the head into tissues with similar electrical properties. Conventional methods of visual inspection of segmentation quality are invaluable but time-consuming.

[0033] This application describes methods for estimating the quality or accuracy of a model after it has been generated. The model can be generated using manual, automatic, or semi-automatic methods, or a combination of two or more of these methods. In a number of situations, it may be useful to obtain an estimate of the quality of a given model. In one example, when a particular model is tested and the results indicate that the estimated quality of the model is low, the result provides the user with an understanding that steps should be taken to improve the quality of the model before proceeding to the TA placement / simulation stage. On the other hand, when a particular model is tested and the results indicate that the estimated quality of the model is high, knowledge of the result can provide an indication to the operator that they have spent sufficient time refining the segmentation and can proceed to the TA placement simulation stage.

[0034] Automatic quality assessment can assist in automatically refining segmentation parameters, suggest defect points to the user, and indicate whether the segmentation method is accurate enough for TTField simulation.

[0035] Figure 1A Depicted is a T1w MRI image of the head of a GBM patient; Figure 1B Depicts the use of the reference algorithm Figure 1AMRI segmentation; Figure 1C Depicts the corresponding Figure 1A Validated Segmentation of MRI. In two segmentations, the tumor was semi-automatically pre-segmented.

[0036] The embodiments described herein make it possible to predict the segmentation generated by a given algorithm (e.g. Figure 1B ) and the expected verified segmentation without generating the expected verified segmentation (i.e., if generated, would correspond to Figure 1C (Figure ).

[0037] In one approach to accomplishing this goal, the inventors identified a set of features relevant to atlas-based segmentation and showed that these features were significantly correlated with a similarity metric between validated and automatically computed segmentations (p < 0.05). These features were incorporated into a decision tree regressor to predict the similarity of validated and computed segmentations for 20 TTField patients using a leave-one-out approach. The predicted similarity metric was highly correlated with the actual similarity metric (mean absolute difference 3% (SD = 3%); r = 0.92, p < 0.001). It is therefore reasonable to infer that by combining machine learning and segmentation-related features, quality estimation of segmentations is feasible. Note that although only a single machine learning approach is described in this paragraph, various alternative machine learning approaches can be substituted, as will be appreciated by those skilled in the relevant art.

[0038] An example of a semi-automatic protocol for estimating the electric field within a tumor of a specific GBM patient for different TA layouts includes three steps: 1) head segmentation and assignment of electrical tissue properties (conductivity and / or permittivity) to each voxel in a 3D image; 2) virtual placement of a TA on the outer surface of the head, and 3) simulation of electric field propagation and estimation of dose within the tumor. See, for example, Bomzon, Z. et al.: Using computational phantoms to improve delivery of Tumor Treating Fields (TTFields) to patients: 2016 38th Annual IEEE International Conference on Engineering in Medicine and Biology (EMBC). pp. 6461–6464. IEEE (2016), which is incorporated herein by reference. Head segmentation can be performed semi-automatically, for example, by first using SPM-MARS and fine-tuning its parameters and then manually fixing segmentation errors. See, e.g., Huang, Y., Parra, LC: Fully Automated Whole-Head Segmentation with Improved Smoothness and Continuity, with Theory Reviewed. PLoS One. 10, e0125477 (2015), which is incorporated herein by reference.

[0039] One method suitable for performing the first step (head segmentation) is an atlas-based automatic head segmentation method. In one example, to ensure that the estimate of TTField dose (e.g., in the tumor bed) remains similar and relevant to the results, the inventors developed a process for estimating the quality of atlas-based segmentation methods. The method is specifically designed to evaluate atlas-based segmentation algorithms with the aim of promoting better estimates. Optionally, more general methods can also be incorporated for this purpose. See, for example, Gerig, G. et al.: Valmet: A New Validation Tool for Assessing and Improving 3D Object Segmentation (a new validation tool for assessing and improving 3D object segmentation). Proposed on October 14, 2001; Warfield, SK et al.: Simultaneous truth and performance level estimation (STAPLE): an algorithm for the validation of imagesegmentation (simultaneous truth and performance level estimation (STAPLE): an algorithm for validating image segmentation). IEEE Trans. Med. Imaging. 23, 903-21 (2004); and Commowick, O. et al.: Estimating a reference standard segmentation with spatially varying performance parameters: local MAP STAPLE. IEEE Trans. Med. Imaging. 31, 1593–606 (2012), each of which is incorporated herein by reference.

[0040] Methods for estimating the quality of a given segmentation rely on a similarity metric between a computed segmentation of the predicted head and a verified segmentation, but verified segmentations are lacking. To measure the quality of a segmentation, the Dice coefficient is measured between the computed and verified head segmentations of the training set. Four categories of features that appear to be relevant to atlas-based segmentation methods are then investigated: 1) the quality of the global (affine) registration; 2) the quality of the local (deformable) registration; 3) input image properties; and 4) geometric properties of the segmented tissue.

[0041] In one example, the global registration quality is estimated using inverse consistency. See, for example, Rivest-Hénault et al.: Robust inverse-consistent affine CT–MR registration in MRI-assisted and MRI-alone prostate radiation therapy. Med. Image Anal. 23, 56-69 (2015), which is incorporated herein by reference. In this example, the following features are extracted to account for the quality of deformable registration: 1) deformed field deviation (mean of all vectors); 2) directional variability (SD of a 3-element vector, i.e., the mean of each axis), and; 3) mean per-axis variability (mean of a 3-element vector, i.e., the SD of each axis). In this example, the shortest axis length and signal-to-noise ratio of each tissue are used as indicators of input image quality, respectively. And although many features describing the segmented shape can be defined, two metrics are selected in this example: the volume of the shape and the number of connected components. These metrics are calculated per tissue, such as Figure 2A As described in .

[0042] The features were combined in a decision tree regressor. See, e.g., Breiman, L.: Classification and Regression Trees. Routledge (2017), which is incorporated herein by reference. A leave-one-out approach was applied to 20 TTField patient head MR-T1 images, their validated segmentations, and their automatically generated counterparts. Finally, the measured Dice coefficients across the sets were compared with the predicted Dice coefficients.

[0043] Figure 2A The absolute Pearson correlation between the calculated features (columns) and the Dice coefficient for each segmented tissue (rows) is depicted. The Dice coefficient was calculated between the validated head segmentation and the head segmentation calculated using the new automated segmentation method (*p < 0.05).

[0044] Figure 2B The output predictions of the decision tree regressor for the proposed features and the Dice coefficient are plotted, and are highly correlated with the actual Dice coefficient (r = 0.92; p < 0.001). Figures 2A-2B In the figure, CSF refers to cerebrospinal fluid; Skin refers to skin and muscle; GM refers to gray matter; WM refers to white matter; SNR refers to signal-to-noise ratio; cc refers to the number of connected components. Figures 2A-2BThe results depicted in Figure 3 show that intracranial tissue is significantly correlated with registration (global and local) quality features and shortest axis length (p < 0.05); and in contrast, extracranial structures are significantly correlated with image quality and shape properties of tissue (p < 0.05). Furthermore, the predicted metrics are similar and highly correlated with the actual metrics (mean absolute difference 3% (SD = 3%); r = 0.92, p < 0.001).

[0045] The results described herein show that by combining machine learning methods and segmentation-related features, quality estimation of segmentation is feasible. Optionally, the simultaneous truth and performance level estimation (STAPLE) method or one of its variants can be incorporated to improve the prediction of segmentation quality. See, for example, Akhondi-Asl et al.: Simultaneous truth and performance level estimation through fusion of probabilistic segmentations. IEEE Trans. Med. Imaging. 32, 1840-52 (2013), which is incorporated herein by reference, and the references of Warfield and Commowick identified above. Segmentation quality estimation can improve TTField planning, and the impact of quantified segmentation error on TTField simulation results is described below.

[0046] Optionally, the system can be programmed to automatically adjust the segmentation to achieve an improved segmentation. For example, the system can change one or more parameters that have a global impact on the boundaries between certain tissue types. An example of such a change is to adjust, for all voxels in the image corresponding to tissue, (a) increase the calculated probability that a given voxel belongs to a first tissue type, and (b) decrease the calculated probability that a given voxel belongs to a second tissue type.

[0047] After performing the automatic adjustments, the quality of the segmentation is reassessed by presenting the adjusted segmentation to a machine learning system. The machine learning system outputs at least one second quality score for the adjusted segmentation. An indication is output when the at least one second quality score indicates improved quality relative to the at least one first quality score. This method can be used to improve a segmentation by automatically adjusting the segmentation.

[0048] A common prior art approach for planning TTField therapy involves four steps: creation of a segmentation; visual assessment of the segmentation quality by the user; correction of the segmentation using standard tools; and clinical decision-making by the user based on the TTField estimate.

[0049] Figure 3 1 is a flowchart depicting a set of steps for using a quality estimation machine learning system to improve positioning planning of a transducer array for TTField therapy. Step 10 is training the machine learning system to estimate the quality of the segmentation, and this can be done based on a plurality of reference images and at least one quality score that has been assigned to each reference image, as described above in conjunction with FIGs. 1 and 2.

[0050] If we assume that the machine learning system has been previously trained in step 10, the first step of the process for any new image will be step 20. Figure 3 Steps 20-70 in FIG. 1 depict an example of a TTField planning protocol that improves upon prior art methods by incorporating automatic segmentation quality estimates (e.g., those described above in conjunction with FIG. 1-2 ) and guidance regarding segmentation error and TTField computation quality. In step 20, a user creates a segmentation of a medical image (e.g., using computational methods or manually). Any of a variety of conventional methods that will be apparent to one skilled in the relevant art can be used to create the initial segmentation. In step 30, the segmentation and the associated original image are presented to a machine learning system.

[0051] Next, in step 40, a quality estimate of the presented segmentation is obtained from the machine learning system (eg, as described above in connection with Figures 1-2), and the quality estimate is presented to the user.

[0052] Figure 4 An example of a suitable user interface that can be used to present the quality to the user and also allow the user to make corrections to the segmentation is depicted. The quality presentation can be made in various ways, including but not limited to (a) a single overall score for any given segmentation (e.g., Figure 4 ), (b) recommendations that a specific tissue should be revised due to a low quality score in any given segmentation (e.g., as in Figure 4 ), and / or (c) identification of specific regions on the image where, for any given segmentation, the quality of the computed segmentation is low. One way to indicate the latter is by marking the image with a transparent region (e.g., Figure 4 42) or an opaque area (e.g., Figure 4 The regions 44 in FIG. 4 are used to cover the partitions associated with low segmentation quality.

[0053] Various alternative methods for depicting partitions associated with low segmentation quality will be readily apparent to those skilled in the relevant art, including but not limited to generating a color-coded overlay (similar to a heat map) depicting the estimated quality of the segmentation at each point in the image. In an alternative embodiment, only the image quality of the partitions associated with low segmentation quality are provided to the user. Figure 4In an alternative embodiment, a different user interface is provided to the user for quality presentation.

[0054] exist Figure 3 In the next step of the protocol (i.e., step 50), the user modifies the segmentation based on the quality data presented in step 40. This can be done using a variety of methods. In one method (step 50A), the user performs an automatic correction method. In another method (step 50B), the user changes a parameter that has a global effect on the boundary. This can be done, for example, by making the following adjustments: for all voxels corresponding to tissue in the new image, (a) increase the calculated probability that a given voxel belongs to the first tissue type, and (b) decrease the calculated probability that a given voxel belongs to the second tissue type. Such adjustments can be made to any given segmentation using any suitable user interface, including but not limited to the following: Figure 4 The sliders depicted on the bottom. For example, sliding the first slider 46 labeled "Skull to Muscle" to the left gradually increases the calculated probability that any given voxel is a skull voxel and gradually decreases the calculated probability that the given voxel is a muscle voxel. And sliding the first slider 46 to the right gradually increases the calculated probability that any given voxel is a muscle voxel and gradually decreases the calculated probability that the given voxel is a skull voxel. In yet another method (step 50C), the user uses a standard segmentation tool (such as a brush or polygon marker) to correct local errors. Note that the three methods identified above are merely examples of how to correct the segmentation, and various alternative methods that will be apparent to those skilled in the relevant art may also be used.

[0055] In some preferred embodiments, after the segmentation has been revised in step 50, the system recalculates the quality of the segmentation in step 60 (e.g., using the method described above in conjunction with Figures 1 and 2) and presents the revised result to the user. This can be done in real time in response to the adjustments made by the user in step 50. Alternatively, it can be done upon user request (e.g., by including a "refresh quality estimate" button on the user interface).

[0056] Figure 3The next step of the protocol (i.e., step 70) occurs after the segmentation has been corrected. In this step, the estimated quality of the segmentation that has been selected by the user is used to plan the TTField therapy, taking into account the quality estimate of the selected segmentation. The quality of the TTField simulation is directly related to the quality of the tissue segmentation, as misassigned tissue can lead to inaccurate estimates of the field and suboptimal TTField therapy recommendations. The relationship between segmentation and TTField error can be modeled so that the estimate of segmentation quality will facilitate the estimate of the calculated TTField quality. The TTField quality estimate can be used to assist the caregiver in making treatment decisions. For example, suppose there are two recommended transducer array layouts that have similar TTField doses, but one layout has much better TTField quality than the other. The embodiments described herein provide this information to the user so that they can make more informed clinical decisions (e.g., by selecting a layout with higher quality).

[0057] Taking the previous idea further, errors in segmentation could be propagated to automatically implement the recommended treatment. That is, the software would prefer settings where confidence in the TTField dose is high, among other factors.

[0058] Figures 5A-5C Depicted is an example of a suitable user interface that can be used to summarize the estimated field strength (or power density) in the tumor bed for five different candidate transducer array layouts (TAL1-TAL5), and also shows how the quality of the segmentation affects the field strength estimate. The operator can use this information as a basis for selecting a given candidate transducer array layout over other candidates.

[0059] For each candidate transducer array layout based on a given segmentation, the system calculates how the determined quality of the given segmentation can reduce the average expected field strength in the target region (e.g., the region corresponding to the tumor) by running finite element simulations using the tissue type selected based on the given segmentation. In addition, for each candidate transducer array layout based on the given segmentation, the system calculates how the determined quality of the given segmentation can reduce the minimum expected field strength in the target region by running finite element simulations using the tissue type selected based on the given segmentation.

[0060] More specifically, Figure 5A A table showing two values ​​for each candidate transducer array layout (TAL1-TAL5) being considered is shown. For each transducer array layout, the value on the left is the average estimated TTField power density in the tumor when the quality of the segmentation is ignored; and the value on the right is the lowest (or 5th percentile) average estimated TTField power density in the tumor when the quality of the segmentation is considered.

[0061] In this example, the first candidate transducer array layout (TAL1) is associated with 80 mW / cm 2 However, when taking into account the effect of possible errors in the segmentation, the TTField power density for this same layout (TAL1) can be as low as 30 mW / cm 2 In contrast, when possible errors in segmentation are ignored, the third candidate transducer array layout (TAL3) has a lower mean estimated TTField power density in the tumor (63 mW / cm 2 , see the left column for TAL3). However, when considering the effect of possible errors in the partitioning, the lowest reasonably expected (e.g., 5th percentile) power density for the TAL3 layout would also be 63 mW / cm 2 (Right column for TAL3). Because the quality of the segmentation will not have as great a negative impact on the field power density when the TAL3 layout is chosen, and because the lowest reasonably expected power density for this layout (i.e., 63 mW / cm 2 ) is significantly higher than the lowest reasonably expected power density of the TAL1 layout (30mW / cm 2 ), so the TAL3 layout is more preferred than the TAL1 layout and should therefore be selected.

[0062] In other words, determining which TTField simulation is optimal can be accomplished by maximizing X, where X is the lowest reasonably expected power density in the tumor bed. Alternatively, the system can be programmed to calculate which layout maximizes X and recommend that layout. Optionally, a confidence score can be associated with one or more possible layouts, where the confidence score indicates the probability that a given layout is the optimal layout.

[0063] Optionally, the effect of each candidate transducer array layout on the TTField spatial distribution can be displayed to the user to help the user choose which TAL is best. In the example shown, click Figure 5A The left column of TAL1 will bring up Figure 5B , which shows a spatial map of the estimated field power density (in the middle panel 52B) and a plot of the local minimum power density (in the right panel 54B) overlaid on the MRI when ignoring the quality of the segmentation. Figure 5A The right column of TAL1 will bring up Figure 5C , which shows a spatial map of the lowest (or 5% percentile) mean estimated TTField power density (in the middle panel 52C) and a graph of the local minimum power density (in the right panel 54C) overlaid on the MRI when the quality of the segmentation is taken into account.

[0064] Although the present invention has been disclosed with reference to certain embodiments, many modifications, changes, and variations of the described embodiments are possible without departing from the field and scope of the invention as defined in the appended claims. It is intended, therefore, that the present invention not be limited to the embodiments described, but rather have the full scope defined by the language of the following claims and their equivalents.

Claims

1. A method for estimating the quality of image segmentation, the method comprising: training a machine learning system to estimate the quality of an image segmentation based on a plurality of reference images and at least one quality score that has been assigned to each reference image; Present the new image and the segmentation of the new image to the machine learning system; receiving at least one first quality score for the segmentation of the new image from the machine learning system; outputting the at least one first quality score for the segmentation of the new image; Automatically adjust the segmentation; presenting the adjusted segmentation to a machine learning system; receiving at least one second quality score of the adjusted segmentation from the machine learning system; as well as When the at least one second quality score indicates improved quality relative to the at least one first quality score, an indication is output.

2. The method of claim 1 , wherein the machine learning system is trained to estimate the quality of image segmentation based on (a) the quality of affine registration, (b) the quality of deformable registration, (c) input image properties, and (d) geometric properties of the segmented tissue. 3 . The method of claim 2 , wherein the quality of the deformable registration is determined based on the field deviation, directional variability, and mean per-axis variability of the deformation. The method according to claim 2 , wherein the input image attributes include the shortest axis length and signal-to-noise ratio of each tissue, respectively. The method of claim 2 , wherein the geometric properties of the segmented tissue include a volume of a shape and a number of connected components, each geometric property being calculated per tissue.

6. A method according to claim 1, wherein the machine learning system is trained to estimate the quality of the segmentation of the new image based on at least one global quality feature, at least one local quality feature and the shortest axis length of the intracranial tissue, and wherein the machine learning system is trained to estimate the quality of the segmentation of the new image based on the image quality and tissue shape properties of the extracranial tissue.

7. A method for improving the quality of image segmentation, the method comprising: presenting the new image and the first segmentation of the new image to a machine learning system, wherein the machine learning system has been trained to estimate the quality of the image segmentation based on a plurality of reference images and at least one quality score that has been assigned to each reference image; receiving, from the machine learning system, at least one first quality score for a first segmentation of a new image; outputting the at least one first quality score for the first segmentation of the new image; accepting at least one revision to the first segmentation from a user; presenting a second segmentation of the new image to the machine learning system, wherein the second segmentation is based on the at least one revision to the first segmentation; receiving, from the machine learning system, at least one second quality score for a second segmentation of the new image; as well as The at least one second quality score for the second segmentation of the new image is output.

8. The method of claim 7 , wherein at least one first quality score of a first segmentation of the new image consists of a single first quality score of the entire new image, and wherein at least one second quality score of a second segmentation of the new image consists of a single second quality score of the entire new image.

9. The method of claim 7 , wherein the at least one first quality score for the first segmentation of the new image comprises a separate first quality score for each of a plurality of regions within the new image, and the at least one second quality score for the second segmentation of the new image comprises a separate second quality score for each of the plurality of regions within the new image.

10. The method of claim 7, wherein the at least one first quality score for the first segmentation of the new image comprises a separate first quality score for each of a plurality of tissue types within the new image, and the at least one second quality score for the second segmentation of the new image comprises a separate second quality score for each of the plurality of tissue types within the new image.

11. The method of claim 7, wherein the at least one modification to the first segmentation comprises adjusting a shift in the calculated probability that a given voxel belongs to a particular tissue type.

12. The method of claim 7 , wherein the at least one revision to the first segmentation comprises adjusting, for all voxels in the new image corresponding to tissue, (a) increasing the calculated probability that a given voxel belongs to a first tissue type, and (b) decreasing the calculated probability that a given voxel belongs to a second tissue type.

13. The method of claim 7 , further comprising (a) calculating how the determined quality of the first segmentation may change the average expected power density in the target region of the new image corresponding to the tumor by running a finite element simulation using the tissue type selected based on the first segmentation, and (b) calculating how the determined quality of the second segmentation may change the average expected power density in the target region of the new image by running a finite element simulation using the tissue type selected based on the second segmentation.

14. The method of claim 7 , further comprising (a) calculating how the determined quality of the first segmentation may change a minimum reasonably expected power density in a target region of a new image corresponding to the tumor by running a finite element simulation using a tissue type selected based on the first segmentation, and (b) calculating how the determined quality of the second segmentation may change a minimum reasonably expected power density in a target region of the new image by running a finite element simulation using a tissue type selected based on the second segmentation.

15. The method of claim 7, further comprising maximizing the lowest reasonably expected power density in the tumor bed.

16. The method of claim 7, further comprising calculating a confidence score for at least one candidate layout, wherein the confidence score indicates a probability that a given layout is an optimal layout.

Images