Systems and methods for stereotactic radiosurgery

A machine learning model predicts optimal normal brain tissue dose distribution for multiple brain metastases in SRS, addressing quality variations by providing real-time evaluation and modification, thus enhancing plan quality and reducing healthy tissue exposure.

WO2026236661A1PCT designated stage Publication Date: 2026-11-12THOMAS JEFFERSON UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/027364
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2026-05-08
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Current stereotactic radiosurgery (SRS) plans for multiple brain metastases lack standardization, leading to variations in quality, with excessive doses to healthy brain tissues, and existing metrics like Brain V12Gy are insufficient for evaluating plan quality and are dependent on fractionation schedules.

Method used

A machine learning model, such as a neural network or conditional generative adversarial network, is trained on high-quality single-lesion SRS plans to predict the best achievable normal brain tissue dose distribution (AI-BBD) for multiple lesions, using planning target volume and region of interest, and provides plan quality evaluation scores like SDVH and BSI, independent of radiation dose prescription.

Benefits of technology

The model ensures consistent, high-quality SRS plans across different subjects by predicting optimal dose distributions and providing real-time evaluation and modification guidance, reducing excessive doses to healthy brain tissues and improving plan quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026027364_12112026_PF_FP_ABST
    Figure US2026027364_12112026_PF_FP_ABST
Patent Text Reader

Abstract

Provided herein is a method of training a machine learning model to determine a best achievable normal brain tissue dose distribution (AI-BBD) for stereotactic radiosurgery (SRS) plans. The method includes providing a training set including single lesion SRS plans and training the machine learning model with the training set, the training generating a trained model, wherein the trained model is configured to determine the AI-BBD for multiple lesions SRA plans. Also provided herein is a method of evaluating a clinical SRS plan using the trained machine learning model.
Need to check novelty before this filing date? Find Prior Art

Description

Attorney Docket No. 205961-7137WO1 (00572)SYSTEMS AND METHODS FOR STEREOTACTIC RADIOSURGERYCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The present application claims priority under 35 U.S.C. § 119(e) to U.S.Provisional Patent Application No. 63 / 803,244, filed May 9, 2025, which application is incorporated herein by reference in its entirety.BACKGROUND OF THE INVENTION

[0002] Brain metastases are the most common intracranial cancer in adults, and the incidence thereof is increasing. Currently, the preferred treatment for brain metastasis is Stereotactic radiosurgery (SRS), an advanced radiation technique that uses highly precise and focused radiation to treat tumors or abnormalities. As compared to conventional radiation, SRS provides better sparing of normal brain tissue (z.e., less neurocognitive toxicides without comprising survival).

[0003] SRS delivers high dose to the targets in 1 to 5 fractions. Excessive dose to healthy brain tissues is linked with radionecrosis which can significantly impact quality of life. Although SRS is well optimized for single brain metastasis, it takes a significantly longer time to deliver for multiple targets and there is a lack of standard to evaluate SRS plans for multiple brain metastases. As such, there is often variation in SRS plan quality, where plans having a lower brain dose are considered higher quality. A common metric for evaluating brain dose is the volume of normal brain tissue irradiated by a dose of at least 12 Gy (Brain Vi2Gy). However, Brain Vi2Gy, which is evaluated by absolute volume threshold, is only a plan safety indicator and is related to radionecrosis. Additionally, Brain Vi2Gy changes with fractionation schedule.

[0004] Accordingly, there remains a need in the art for systems and methods that provide improved SRS plan quality. The present invention meets this need.SUMMARY

[0005] In one aspect, a method of training a machine learning model to determine a best achievable normal brain tissue dose distribution (AI-BBD) for stereotactic radiosurgery (SRS) plans includes providing a training set including single lesion SRS plans and training the machine learning model with the training set to generate a trained model. In some embodiments, the trained model is configured to determine the AI-BBD for multiple lesions SRS plans.- 1 - 57737409.4Attomey Docket No. 205961-7137WO1 (00572)

[0006] In some embodiments, the model includes a neural network, a deep learning model, or a conditional generative adversarial networks (cGAN). In some embodiments, the deep learning model comprises a small target model and a large target model. In some embodiments, the small target model is trained on a small lesion dataset and the large target model is trained on a large lesion dataset. In some embodiments, the small lesion dataset includes lesions having a volume of less than 1 cc and the large lesion dataset includes lesions having a volume equal to or greater than 1 cc. In some embodiments, the model includes a random forest (RF) regression model.

[0007] In some embodiments, the single lesion SRS plans of the training set include conformal arc (DCA) plans, volumetric arc therapy (VMAT) plans, Gamma Knife plans, or combinations thereof. In some embodiments, the model is trained based upon a region of interest (ROI) corresponding to each of the brain lesions in the single lesion SRS plans of the training set. In some embodiments, the ROI for each of the brain lesions includes a planning target volume (PTV) of the lesion. In some embodiments, the PTV is determined based upon at least one aspect of the brain lesion selected from diameter (measured in the largest dimension), volume, sphericity, location, depth, distance between lesions, or a combination thereof. In some embodiments, the PTV further comprises a ring or sphere defined around the PTV. In some embodiments, the ring or sphere includes a thickness around the PTV of between 1 cm and 1.5 cm. In some embodiments, the model is configured to merge predicted doses for the lesions based upon a location of the ROIs.

[0008] In some embodiments, the method further includes generating a plan quality evaluation score. In some embodiments, the plan quality evaluation score comprises a dose volume histogram (DVH) score (SDVH) according to the equation:SDVH = x 100''clinicalwhere:AreaV = - .(0.8 - 0.5) / ?%In some embodiments, a SDVHscore of 100% indicates the best brain dose has been achieved. In some embodiments, the plan quality evaluation score is independent of radiation dose prescription of each lesion.- 2 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)

[0009] In another aspect, a method of evaluating a clinical stereotactic radiosurgery (SRS) plan for multiple lesions includes inputting the clinical SRS plan to the machine learning model according to any of the embodiments described herein, receiving an output from the trained model, and evaluating the clinical SRS plan based upon the output. In some embodiments, the method further includes modifying the clinical SRS plan based upon the evaluating. In some embodiments, the method further includes finalizing the clinical SRS plan based upon the evaluating.BRIEF DESCRIPTION OF THE DRAWINGS

[0010] For a fuller understanding of the nature and desired objects of the present invention, reference is made to the following detailed description taken in conjunction with the accompanying drawing figures wherein like reference characters denote corresponding parts throughout the several views.

[0011] FIG. 1 shows a schematic illustrating a stereotactic radiosurgery (SRS) Planning Al Dose Evaluation (SPADE) model.

[0012] FIG. 2 shows a schematic illustrating SPADE using conditional generative adversarial network (cGAN).

[0013] FIG. 3 shows a graph illustrating dose-volume histogram (DVH) of target ring.

[0014] FIG. 4 shows a schematic illustrating BSI plan evaluation workflow.

[0015] FIG. 5 shows a graph illustrating BSI by treatment method and target volume.

[0016] FIGS. 6A-C show images illustrating a BSI calculation example. (A) Manual benchmark plan. (B) DCA plan. (C) HyperArc plan.

[0017] FIGS. 7A-C show images illustrating various dose maps. (A) Clinical dose map. (B) Benchmark dose map. (C) Dose difference map between the clinical dose and the benchmark dose.

[0018] FIG. 8 shows a graph illustrating DVH of a target ring.

[0019] FIG. 9 shows a schematic illustrating plan selection according to DVH score.

[0020] FIG. 10 shows plots illustrating true vs predicted DVH score for each target (left) and each patient (right).

[0021] FIG. 11 shows a graph illustrating feature importance in the random forest model.-3 - 57737409.4Attomey Docket No. 205961-7137WO1 (00572)

[0022] FIG. 12 shows a schematic illustrating a model for predicting the benchmark dose for multi-lesion SRS plans.

[0023] FIG. 13 shows graphs illustrating a comparison between Gamma Knife plans and U-Net predicted plans for PTV ring V12Gy, PTV ring V8Gy, and PTV ring V4Gy.

[0024] FIG. 14 shows a schematic illustrating the overall workflow for generating an improved SRS plan.DETAILED DESCRIPTION OF THE INVENTIONDefinitions

[0025] The instant invention is most clearly understood with reference to the following definitions.

[0026] As used herein, the singular form “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise.

[0027] Unless specifically stated or obvious from context, as used herein, the term “about” is understood as within a range of normal tolerance in the art, for example within 2 standard deviations of the mean. “About” can be understood as within 10%, 9%, 8%, 7%, 6%, 5%, 4%, 3%, 2%, 1%, 0.5%, 0.1%, 0.05%, or 0.01% of the stated value. Unless otherwise clear from context, all numerical values provided herein are modified by the term about.

[0028] As used in the specification and claims, the terms “comprises,” “comprising,” “containing,” “having,” and the like can have the meaning ascribed to them in U.S. patent law and can mean “includes,” “including,” and the like.

[0029] Unless specifically stated or obvious from context, the term “or,” as used herein, is understood to be inclusive.

[0030] The terms “proximal” and “distal” can refer to the position of a portion of a device relative to the remainder of the device or the opposing end as it appears in the drawing. The proximal end can be used to refer to the end manipulated by the user. The distal end can be used to refer to the end of the device that is inserted and advanced and is furthest away from the user. As will be appreciated by those skilled in the art, the use of proximal and distal could change in another context, e.g., the anatomical context in which proximal and distal use the patient as reference, or where the entry point is distal from the user.- 4 - 57737409.4Attomey Docket No. 205961-7137WO1 (00572)

[0031] Ranges provided herein are understood to be shorthand for all of the values within the range. For example, a range of 1 to 50 is understood to include any number, combination of numbers, or sub-range from the group consisting 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, or 50 (as well as fractions thereof unless the context clearly dictates otherwise).Detailed Description

[0032] Provided herein are methods of determining stereotactic radiosurgery (SRS) plans for multiple brain lesions. In some embodiments, the method includes training a machine-learning / artificial intelligence (Al) model to predict the best achievable normal brain dose for SRS plans. As used herein, the term “best achievable normal brain dose” refers to an SRS plan dose that provides a clinically effective dose to brain lesions with the lowest possible dose to normal brain tissue which can be delivered by existing radiotherapy treatment machines. Suitable brain lesions include, but are not limited to, brain tumors, meningioma, acoustic neuroma, pituitary adenoma, brain arteriovenous malformation (AVM), recurrent Glioblastoma (GBM), and / or any other tissue formation treatable with SRS.

[0033] As used herein, the term “machine learning” refers to the process by which an algorithm is trained to learn from data, recognize certain patterns within the data, and make decisions or predictions without being explicitly programmed for each specific rule. Machine learning may include training an algorithm to identify certain patterns using training data. The term “training data” refers to data that includes associations which a machine-learning process can use to learn relationships between two or more categories of data elements. Patterns across these data entries in training data can reveal correlations between different categories of data elements; for instance, a higher value in one category may frequently appear with a higher value in another category, suggesting a proportional or otherwise quantifiable connection. Training data can also be arranged by data element categories, in some embodiments, with each element tagged or described to reflect its category. As one example, information can be entered into standardized forms by individuals or systems can associate specific fields with corresponding category descriptors. Data elements in training data may be categorized using tags, identifiers, or similar markers; for instance, the data may be structured in fixed-length formats, in position--5 - 57737409.4Attomey Docket No. 205961-7137WO1 (00572)based formats like CSV, or in self-describing formats like XML or JSON, which allow systems to identify the category of each element.

[0034] Referring to FIG. 1, in some embodiments, the model (also referred to herein as SRS Planning Al Dose Evaluation (SPADE)) is trained with a training set including benchmark SRS plans. The benchmark SRS plans include any suitable high quality SRS plans. In some embodiments, the benchmark SRS plans include single-lesion treatment plans that prioritize brain sparing with no limit on delivery time. Accordingly, in some embodiments, the training set includes multiple single-lesion SRS plans for one or more individual subjects. In some embodiments, the benchmark SRS plans include dynamic conformal arc (DCA) plans, volumetric arc therapy (VMAT) plans, Gamma Knife plans, or combinations thereof. For example, in some embodiments, the training set includes idealized Gamma Knife plans aiming to achieve optimal dose falloff without limiting treatment delivery time. In some embodiments, the training set includes DCA and VMAT plans. As will be appreciated by those skilled in the art, the training set includes any suitable number of subjects (e.g, about 100 or more subjects) and corresponding targets (c.g., about 500 or more targets) for training the model. In some embodiments, additional validation and test datasets (e.g., about 20 subjects) may be included to finetune model structure and hyperparameters and test model performance.

[0035] The model can be trained in any suitable manner for predicting an Al best achievable normal brain tissue dose distribution (AI-BBD). In some embodiments, the model is trained to predict an AI-BBD for multi-lesion SRS. In some embodiments, the method includes training traditional machine learning and / or deep learning models to predict the AI-BBD. In some embodiments, the model is trained to predict the multi-lesion AI-BBD based upon a planning target volume (PTV) for each of the brain lesions in the single-lesion benchmark SRS plans. The PTV can be determined by any suitable aspect of the brain lesions, such as, but not limited to, diameter (measured in the largest dimension), volume, sphericity, location, depth, distance between lesions, or a combination thereof. For example, in some embodiments, the model is trained based upon PTV diameter, volume, and sphericity.

[0036] In some embodiments, training the model includes generating and / or defining a region of interest (ROI) surrounding the PTV. In some embodiments, a ring or sphere is defined around each PTV prior to generating the ROI. The ring or sphere can be any suitable size, such as, but not limited to, a thickness around the lesion of between 1 cm and 2 cm, between 1 cm and- 6 - 57737409.4Attomey Docket No. 205961-7137WO1 (00572)1.75 cm, between 1 cm and 1.5 cm, between 1.25 cm and 1.75 cm, between 1.25 cm and 1.5 cm, about 1.5 cm, or any suitable combination, sub-combination, range, or sub-range thereof. In some embodiments, the ring or sphere is based upon an exponential falloff outside the PTV. To generate the multi-lesion SRS predicted dose, in some embodiments, the model is trained to merge the predicted doses for individual lesions based upon the ROI locations. In some embodiments, the merging includes averaging of the doses for overlapping lesions. Accordingly, following training, the brain dose prediction from the model can be target-specific local, global, or both. Additionally, the trained model can provide real-time dose prediction in radiation therapy treatment planning system (TPS).

[0037] In some embodiments, to avoid overfitting (z.e., the model does not generalize well on new data), the method includes providing a larger training dataset, regularization, and / or cross-validation. In some embodiments, to avoid issues with accurately predicting dose bridging between several adjacent targets as a result of the model operating on a per target basis, the method includes adding computer-generated phantoms to the training dataset to simulate such scenarios. In such embodiments, the phantoms are procedurally generated to mimic different target distributions, which can improve the robustness of the machine learning model.

[0038] The model trained according to methods disclosed herein can include any suitable type of machine learning model. Suitable models include, but are not limited to, neural networks, deep learning models, generative adversarial networks (GAN) (e.g., conditional GAN (cGAN)), or any other suitable machine learning model. For example, in some embodiments, as illustrated in FIG. 12, the model includes a deep neural network (e.g., 3D U-Net). Additionally or alternatively, in some embodiments, as illustrated in FIG. 2, the model includes a cGAN model where a generator takes a PTV mask as input and predicts a dose distribution, then a discriminator attempts to distinguish between real dose maps and generated predictions. In such embodiments, the generator and discriminator are trained together using adversarial training. Alternatively, in some embodiments, the method includes non-adversarial training.

[0039] In some embodiments, the model includes two distinct models. In some embodiments, the distinct models are trained on separate datasets, such as, but not limited to, a small lesion dataset (e.g., volume <1 cc) and a large lesion dataset (e.g., >1 cc). For example, in some embodiments, a deep learning framework includes two 3D U-Net models with attention mechanisms, where one model is a small target model (i.e., trained on a small lesion dataset) and- 7 - 57737409.4Attomey Docket No. 205961-7137WO1 (00572)the other model is a large target model (i.e., trained on a large lesion dataset). In some embodiments, a loss function is configured to prioritized accuracy in the dose falloff region i.e., the ring or sphere around the target volume). In some embodiments, the dual-model deep learning framework provides statistically significant improvements in predicting patient-specific benchmark dose distributions forbrain SRS relative to single-model alternatives. Additionally or alternatively, in some embodiments, the dual-model framework enables reliable cross-platform plan-quality evaluation and direct utility for routine clinical SRS quality assurance for all lesion sizes.

[0040] The methods described herein are configured to predict the highest quality SRS plans. Since the models provide high resolution, they are also suitable for lesions of all sizes. Accordingly, the methods described herein provide consistently high quality SRS plans across different subjects.

[0041] Also provided herein are methods and scoring systems for SRS plan evaluation. In some embodiments, the method includes generating, or the scoring system generates, an SRS plan quality evaluation parameter. In some embodiments, the plan quality evaluation parameter includes a dose volume histogram (DVH) score (SDVH) based upon the AI-BBD. In some embodiments, the SDVHrepresents the average normal brain volume (V) receiving clinically significant dose for each target. In some embodiments, as illustrated in FIG. 3, generating SDVHincludes evaluating the area under the curve for dose-volume histogram (AUC-DVH) over a relative dose range for each target’s ring or sphere. The relative dose range includes any suitable range, such as, but not limited to, 50-80% Rx. The proposed SRS plan and the AI-BBD are then compared according to the following equation:SDVH =x100"clinicalwhereAreaV = - (0.8 - 0.5) / ?%The higher the SDVHthe better, with an SDVHof 100% indicating the best brain dose has been achieved.

[0042] Additionally or alternatively, in some embodiments, the plan quality evaluation parameter includes a brain sparing index (BS1). Similar to generating SDVH, generating the BS1 includes evaluating the area under the curve for dose-volume histogram (AUC-DVH) over a - 8 - 57737409.4Attorney Docket No. 205961-7I37WOI (00572)relative dose range for each target’s ring or sphere. The relative dose range includes any suitable range, such as, but not limited to, 50-80% Rx. The proposed SRS plan and the AI-BBD are then compared according to the following equation:BSI = ^^AUCpianThe higher the BSI the better, with a BSI of 1 indicating the same brain sparing as the benchmark.

[0043] The plan quality evaluation parameters according to the embodiments disclosed herein are relative to the best achievable dose and independent of radiation dose prescription of each target. Additionally, the plan quality evaluation parameters provide easy to interpret realtime evaluation of the proposed SRS plan. As such, the plan quality evaluation parameters can be used to evaluate / modify the proposed SRS plan prior to administration.

[0044] Accordingly, also provided herein are methods of selecting and / or modifying an SRS plan based upon the AI-BBD and / or plan quality evaluation parameter. The method includes first inputting a proposed clinical SRS plan to the trained model. In some embodiments, the trained model generates a plan quality evaluation parameter score for the proposed clinical SRS plan (score-based guidance). The score can include a global score for entire SRS plan and / or scores for individual targets. In some embodiments, the method includes alerting the planner of target dosing that can be improved based upon the plan quality evaluation parameter score. Additionally or alternatively, in some embodiments, the trained model generates an AI-BBD for the proposed clinical SRS plan (volume-based guidance). In some embodiments, the method includes generating a dosing map showing visualization of 3D dosing and DVH differences between the AI-BBD and proposed clinical SRS plan (FIGS. 7A-C). The dose difference map (FIG. 7C) provides visualization of where the existing plan could be improved.

[0045] After inputting the proposed clinical SRS plan to the trained model, the plan is modified and / or finalized based upon the AI-BBD and / or plan quality evaluation parameter. For example, FIG. 4 illustrates a schematic for modifying and finalizing a proposed clinical SRS plan using the plan quality evaluation parameter according to one or more of the embodiments disclosed herein. Although FIG. 4 includes BSI as the plan quality evaluation parameter, as will be appreciated by those skilled in the art, the disclosure is not so limited and the use of BSI can be replaced and / or supplemented with SDVH In some embodiments, the method includes generating customizable structures and constraints from the dose difference map. In such- 9 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)embodiments, the customizable structures and constraints can be used to modify specific plan aspects and improve overall plan quality. In some embodiments, the methods disclosed herein can be used to select a preferred plan (c. ., DCA vs. VMAT).

[0046] Use of the AI-BBD and / or plan quality evaluation parameter as described herein provides objective planning guidance that can be used to evaluate plans made with different modalities. These metrics provide better and more consistent assessment of plan quality as compared to existing metrics. Additionally, in some embodiments, if the score does not align with a physician’s subjective judgement of plan quality, its definition can be adjusted to suit clinical needs. Alternatively, the manual re-optimization can be replaced with a reinforcement learning agent, which can be deployed in the treatment planning system and tries to improve plan quality scores through trial-and-error.

[0047] Although the invention has been described in terms of exemplary embodiments, it is not limited thereto. Accordingly, the appended claims should be construed broadly to include other variants and embodiments of the invention which may be made by those skilled in the art without departing from the scope and range of equivalents of the invention. This disclosure is intended to cover any adaptations or variations of the embodiments discussed herein.

[0048] It is to be understood that wherever values and ranges are provided herein, all values and ranges encompassed by these values and ranges, are meant to be encompassed within the scope of the present invention. Moreover, all values that fall within these ranges, as well as the upper or lower limits of a range of values, are also contemplated by the present application.

[0049] The following examples further illustrate aspects of the present invention.However, they are in no way a limitation of the teachings or disclosure of the present invention as set forth herein.EXAMPLESEXAMPLE 1 — Geometry-Based Machine Learning Model for Treatment Modality Selection For Single-Isocenter Multi-Lesion Stereotactic RadiosurgeryPurpose / Objectives

[0050] Linac-based single-isocenter multi-lesion SRS patients are planned with dynamic conformal arc (DCA) or volumetric modulated arc therapy (VMAT) in our clinic. The two- 10 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)modalities may achieve different levels of normal tissue sparing depending on the geometry and distribution of the targets, but comparisons can only be made after the plans have been generated. This Example describes the development of a machine learning model to guide treatment modality selection between DCA and VMAT using only target geometry prior to treatment planning for single-isocenter multi -lesion SRS.Materials & Methods

[0051] Dataset; DCA (BrainLab Multiple Metastases Element) and VMAT (Varian HyperArc) plans were generated for 30 SRS patients (target number: 4-10, Rx: 14-24 Gy) on a Varian TrueBeam STx with HD-120 MLC. The dataset was split into 22 and 8 patients for training (155 lesions) and testing (62 lesions).

[0052] Geometric features for model input: Dose volume histogram (DVH) was evaluated for a 1.5-cm ring around each PTV. A DVH score was defined as the normalized difference in area under the PTV ring’s DVH curves (50%-80% Rx) between DCA and VMAT (FIG. 8). For a multi-lesion SRS patient, DVH scores are calculated for each target and summed. The summed total score indicates the better modality for a particular case, with a positive total score indicating a better DVA plan and a negative total score indicating a better VMAT plan (FIG. 9).

[0053] Model training: A random forest (RF) regression model was trained to predict the DVH scores from target geometric features. Feature importance was calculated, and a reduced RF model with top 3 features was trained.

[0054] Model evaluation: Model performance was evaluated by the mean absolute error (MAE) of DVH score and modality prediction accuracy.Results

[0055] DVH score distribution: For all 217 targets, DVH score range is [-7.65, 2.01],

[0056] Summed total score: 17 positive (DCA preferred), 13 negative (VMAT preferred).

[0057] Test set results: DVH score MAE = 0.23, target prediction accuracy 75.8%, patient prediction accuracy 7 out of 8. FIG. 10 illustrates true vs predicted DVH score for each target (left) and each patient (right). Predictions were from the random forest model including all geometric features. Points (blue) in the first quadrant (favor DCA) and third quadrant (favor- 11 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)VMAT / HyperArc) represent a correct prediction, while points (red) in the second and fourth quadrants represent an incorrect prediction.

[0058] Top 3 important features: Feature importance is shown in FIG. 11, with the top three features being PTV diameter (measured in the largest dimension), PTV volume, and sphericity.

[0059] Reduced RF model: DVH score MAE = 0.23, target prediction accuracy 72.6%, patient prediction accuracy 7 out of 8.

[0060] The proposed model can provide clinicians with a quick estimation of expected normal-tissue sparing to compare between DC A and VMAT based on target geometry. An upfront estimate can streamline the plan review process and increase clinical efficiency.Summary / Conchsion

[0061] We developed a geometry-based machine learning model to provide clinical decision-making support for selecting the superior modality for normal-tissue sparing between DCA and VMAT for single-isocenter, multi-lesion SRS.EXAMPLE 2 - Deep learning-based plan quality evaluation for multiple brain lesions stereotactic radiosurgeryBackground

[0062] For stereotactic radiosurgery (SRS) planning, linac-based dynamic conformal arc (DCA) or volumetric modulated arc therapy (VMAT) plan quality is highly dependent on the planner’s experience. Due to its superior normal tissue sparing, Gamma Knife is the gold standard for patients with 1-4 brain lesions, but treatment time could become prohibitively long for patients with many targets.Aim

[0063] To develop a deep learning-based plan quality evaluation tool for linac-based multiple brain lesions SRS plans using the idealized Gamma Knife plans as the benchmark to assist treatment planning.- 12 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)Patients & Methods

[0064] 41 patients with multiple brain lesions (range: 5-22) treated with SRS were included in this retrospective study. Idealized Gamma Knife plans aiming to achieve optimal dose falloff without limiting treatment delivery time were created for all patients as the benchmark. The dataset was split into 25 / 7 / 9 for training / validation / testing.Network Design

[0065] A patch-wise 3D U-Net was developed to predict the benchmark dose for multilesion SRS plans (FIG. 12). The model was trained to predict benchmark dose in a region-of-interest (ROI). The predicted doses were merged based on ROI locations to obtain a partially fdled 3D dose distribution. To take into account dose falloff outside the target and other adjacent targets outside the ROI, a 1-cm wide exponential falloff was added for each target before generating the ROI.Results & Discussion

[0066] The model achieved an averaged voxel dose difference of 0.78 Gy for the test set.

[0067] Individual target’s dose was assessed by creating a 1.5-cm wide ring structure for each PTV. FIG. 13 shows the PTV ring’s V12Gy, V8Gy, V4Gy comparison between the GK plans and the U-Net predicted plans for all 66 targets in the 9 test cases. To assess plan quality improvement with Al guidance, paired T-Test can be performed on commonly used dose metrics (brain V12, CI, GI) and the novel DVH score for original plans and Al-guided plans.Conclusion & Perspectives

[0068] A deep learning model was developed to predict the benchmark dose for multilesion SRS with an averaged voxel-wise dose error of 0.78 Gy. Predicted benchmark doses can be used to assist and evaluate treatment planning for any SRS delivery technique.EXAMPLE 3 - SRS Plan Evaluation

[0069] This Example describes the calculation of BSI / SDVWaccording to one or more of the embodiments disclosed herein. To calculate BSI, dynamic conformal arc (DCA) and HyperArc (HA) plans were generated for each of 30 patients in a sample. The patients were split into two target size groups: small targets (volume < 1 cc, count = 162) and large targets (volume- 13 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)> 1 cc, count = 55). The BSI was then calculated as described herein. The results are shown in FIGS. 5-6C.

[0070] Turning to FIGS. 7A-C, SDVHwas calculated as described herein and the corresponding dosing maps were generated. The calculated scores are shown in Table 1.Table 1Target# Rx (Gy)VcihucaiVAI-BBDSDVII1 21 4.97 2.83 57%2 21 1.87 1.35 72%3 24 0.67 0.32 48%Based upon the dosing map and SDVHscores above, the dosing of the individual targets can be adjusted to improve the quality of the proposed SRS plan.EXAMPLE 4Summary

[0071] Intracranial stereotactic radiosurgery (SRS) plans aim to deliver a conformal dose to the planning target volume (PTV) with steep dose gradients to spare normal brain tissue. It is the preferred treatment for patients with brain lesions due to equivalent overall survival with reduced neurocognitive toxicity. Several treatment machines can be used to deliver SRS plans, including Gamma Knife, linear accelerator (linac) and CyberKnife. For patients with multiple lesions (multi-lesion), treatment time with Gamma Knife could become prohibitively long. It is often preferable to use linac-based single-isocenter planning techniques for these cases, such as dynamic conformal arc (DCA) and volumetric modulated arc therapy (VMAT), which can be delivered in a reasonable amount of time. However, such techniques are highly dependent on the planner’s experience, which can result in inconsistent plan quality. Additionally, the complexity of SRS plans, plan variability, and lack of plan quality evaluators hinder treatment optimization and may lead to excess dose in healthy brain tissue. Excess dose to the healthy brain tissues contributes to unwanted complications following treatment, most notably radionecrosis.- 14 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)

[0072] The quality of an SRS plan can be measured by multiple metrics, including target coverage, conformity index (CI), gradient index (GI), brain V12Gy, and maximum doses to organs-at-risk (OARs). It has been shown that brain V12Gy is linked with radionecrosis and thus should be minimized. Due to the variable geometry of the targets and the complexity of multilesion SRS plans, it is often difficult and time-consuming for an inexperienced planner to assess and improve on suboptimal plans for multi-lesion SRS cases. To facilitate plan quality evaluation and automated treatment planning, artificial intelligence (Al) models have been implemented to predict dose metrics and distributions for various anatomic sites based on prior high-quality plans.

[0073] This Examples describes the development of an Al-assisted plan quality evaluation and modification tool for multi-lesion SRS to evaluate plans and improve the treatment planning workflow. The approach uses a single-lesion technique to create high-quality benchmark plans which minimize dose to the healthy brain tissues. The benchmark plans are then used as training data to train a deep learning model to predict the ideal 3D dose distribution for new cases. The predicted dose can then be used to evaluate existing plans with a novel dose volume histogram (DVH) score for normal tissue sparing and create new optimization structures and constraints. By applying Al techniques to the challenging multi-lesion SRS planning process, the methods described herein can change how plans are evaluated and improve patient care quality.Methods

[0074] Described herein is a machine learning model developed to provide artificial intelligence (Al)-predicted ideal dose distributions which can be used to objectively evaluate multi-lesion SRS treatment plan quality and automatically generate guidance for plan modification. For suboptimal clinical plans, Al guidance can achieve statistically significantly lower doses to healthy brain tissues.

[0075] Machine Learning Model: A deep learning model can be trained from optimized single-lesion plans to predict the best achievable dose (benchmark dose) for multi-lesion SRS plans and achieve a mean voxel dose difference of less than 0.5 Gy.

[0076] The deep learning model predicts 3D dose distribution using target contours as inputs. The model is trained from optimized single-lesion treatment plans to provide the best physically achievable dose to each lesion while considering dosimetric interaction of multiple - 15 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)targets. Due to the high resolution of contours and dose in SRS planning and the sharp dose falloff outside the targets, the model predicts dose only for a limited region of interest around each target to assess the healthy brain tissue dose. The model can achieve a mean voxel dose difference of less than 0.5 Gy from ground truth benchmark plans.

[0077] The benchmark plans are optimized for each individual target to prioritize dose conformity without limiting the treatment delivery time. These plans have conformal and physically achievable dose distribution. In a proof-of-concept study, the dataset included 41 patients in total. The model used a training dataset of 25 patients. The model structure and hyperparameter tuning was validated on 7 patients, and the model performance was tested on 9 patients. It achieved a mean voxel dose difference of 0.78 Gy for the test dataset.

[0078] The BDP model architecture is a 3D U-Net for patch-wise dose prediction (FIG.12). The network input is a 3D region-of-interest (ROI) of the target contour centered at each planning target volume (PTV). The deep learning model uses a U-Net architecture to predict the dose distribution in this ROI, which is repeated for all PTVs. The predicted ROI doses for all PTVs are fdled back into the patient volume to make up the benchmark dose. For example, a cubic region-of-interest (ROI) (6.4x6.4x6.4 cm3) is extracted for each PTV. The model input is the PTV contour mask with voxels assigned to its prescribed dose in gray (Gy). Considering the physical dose falloff around each target, an exponential falloff of 1-cm width is applied around all targets. Thus, adjacent targets just outside the ROI may contribute to the input contour mask. The patch-wise design uses each individual target as a separate data point, which significantly increases the size of the dataset for patients with multiple lesions. Due to the dose conformity of SRS plans, distant targets do not have significant contribution to the local dose. The predicted ROI benchmark doses can be filled back into the total volume to generate a partial dose distribution. The ROIs encompass high dose regions around all targets, which are more clinically relevant to plan quality evaluation.

[0079] The model performance can be evaluated by comparing the predicted dose with the ground truth benchmark dose and computing the mean voxel dose difference. Prediction accuracy and robustness can be improved by refining model structure and increasing the size of the dataset. One possible approach is to use a generative adversarial network (GAN), which uses two neural networks (a generator and a discriminator), to predict a more realistic dose distribution. Since the model predicts benchmark dose for each target, it does not require a large- 16 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)number of patients, since one patient has several brain lesions. However, it is important to include a variety of target sizes / shapes and arrangements when training the model.

[0080] The ideal model is expected to generate realistic and conformal benchmark dose distribution for any brain lesions within one minute without the need to use the treatment planning system.

[0081] Scoring System: An Al-based scoring system was developed to objectively evaluate healthy brain tissue dose for manually created multi-lesion SRS plans and identify plans with suboptimal quality.

[0082] The predicted benchmark dose represents the best achievable plan quality with current technologies, which forms the basis for the novel scoring system. More specifically, the Al scoring system uses dose distributions predicted by the Al model from the machine learning model, which can be integrated in the treatment planning system, and provides real-time feedback on the level of normal tissue sparing for any given plan. Manually created plans can be compared with the predicted dose distribution to highlight specific regions that can be improved. The Al scoring system can be finetuned to align with physician’s assessment and improve the detectability of suboptimal plans.

[0083] The scoring system compares dose volume histogram (DVH) metrics between clinical plans and benchmark plans to calculate DVH scores. Although the brain V12 is widely accepted, it only evaluates one dose level while SRS dose prescriptions can vary from 14 Gy to 24 Gy depending on a variety of factors including target size and distance to critical organs. There may also be several different dose levels for targets in the same patient, making it difficult to evaluate normal tissue sparing around each target. Therefore, the relative dose range of 50% to 80% of the prescription dose for individual target dose evaluation is used. This dose range (50% to 80%) covers 12 Gy for all commonly used dose prescriptions. V is defined as the average DVH volume in cubic centimeters (cc) in the 50% to 80% dose range. The evaluation volume is a 1.5-cm wide PTV ring in normal brain around each target, excluding any adjacent target volumes. The DVH score S for a target is defined as the ratio of V between a clinical plan and a benchmark plan. Since the benchmark plans have the lowest achievable dose to normal brain, S is larger than 1, and a higher S represents worse brain dose sparing.

[0084] In practice, DVH scores are calculated for all targets, with values typically ranging from 1 to 3. The planner can identify targets with high DVH scores and use additional- 17 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)constraints to reduce brain dose. DVH score calculation can be integrated into the treatment planning system via scripting, providing quick feedback to the planner. The DVH score can be used with existing plan evaluation metrics, such as CI, GI, and V12. It provides an objective evaluation of plan quality while considering the plan difficulty and prescription dose level. For example, it is difficult to create conformal dose for an irregularly shaped target due to physical limitations of treatment machines, and its CI would be larger than spherical targets. The DVH score is less sensitive to the target shape since the benchmark dose is based on the target shape.

[0085] In essence, the DVH score measures how far apart the current plan is from the benchmark plan. The predicted benchmark dose is the ideal choice for the reference plan because it is quick and easy to generate and has consistent quality. The current definition of DVH score only considers dose to normal brain while ignoring other OARs such as brainstem, optic nerve, and optic chiasm, whose maximum dose limits should be considered if they are close to any targets. However, other OAR doses can be included to provide a more comprehensive SRS plan quality evaluation. The scoring system can be evaluated and finetuned by planners and physicians to satisfy clinical needs.

[0086] AI-Based Plan Modification: Healthy brain tissue sparing in suboptimal SRS plans can be improved with guidance provided by the Al-based scoring system and the predicted benchmark dose. Al guided plans can achieve statistically significantly lower doses to healthy brain tissues than manual plans.

[0087] To achieve good dose conformity to the targets and dose sparing to normal tissue, manually generated planning support structures, such as rings and avoidance areas are commonly used in SRS plan optimization. These structures and their dose constraints are based on planner’s experience and may not reflect achievable dose distribution. Using the predicted benchmark dose distribution from the machine learning model and plan scores from scoring system, structures and constraints can be automatically generated to assist plan modification. These structures and constraints are patient-specific and not affected by prescription levels and patient anatomy.

[0088] For example, a dose difference map can be calculated between benchmark dose and clinical dose. While the benchmark dose may not be achievable with single-isocenter linac plans, the dose difference map can still provide useful information on potential areas to improve. Multiple optimization structures and constraints can be generated from the dose difference map and used to re-optimize and minimize the dose differences between clinical and benchmark dose.- 18 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)

[0089] In addition to brain, other OARs can be considered separately if they are close to the PTVs, such as brainstem, optic chiasm, and optic nerve. While dose to these OARs is not explicitly constrained in the benchmark plans, the benchmark dose still represents an ideal dose to the OARs without compromising PTV coverage. If a benchmark plan exceeds OAR dose constraints, it is unlikely that a clinical plan can meet these constraints. In such a case, the clinician should consider dose reduction for the involved PTV to meet the OAR constraints. Discussion

[0090] The quality of an SRS plan can be evaluated in two main aspects: target coverage and normal tissue sparing. As target coverage is typically normalized to more than 99% of PTV receiving the prescription dose, lower dose to organs-at-risk (OARs), especially the brain, indicates a higher quality plan. Radionecrosis is a common toxicity of SRS, which has been linked with brain dose, specifically the volume of brain receiving over 12 Gy (V12). According to QUANTEC, brain VI 2 should be less than 5 to 10 cc to reduce the toxicity rate of radionecrosis to under 20%. Studies have shown that lower brain V12 is associated with lower rates of symptomatic radionecrosis. Therefore, every effort should be made to reduce brain dose even below the 5-cc threshold.

[0091] Due to a variety of factors including machine limitations, target geometry, and most importantly, the experience of the treatment planner, the quality of SRS plans varies significantly. Commonly used plan metrics, including the brain VI 2, the conformity index (CI), and gradient index (GI) have their limitations. Although brain V12 thresholds are used to assess the safety of the plan, the V12 metrics give little indication of the plan quality and room for improvement. CI is defined for each target as the ratio of the volume receiving the prescription dose to the target volume, while GI is defined as the ratio of the volume receiving 50% of the prescription dose to the volume receiving 100% of the prescription dose. CI and GI can be influenced by the shape and size of the target and are ill-defined for multiple targets in proximity to each other. Therefore, a more robust plan quality evaluator is of great clinical interest for multi-met SRS.

[0092] To address the lack of robust plan quality evaluation criteria, the Al-based SRS treatment planning tool described herein provides patient-specific scoring of plan quality and plan modification guidance, as shown in FIG. 14. To achieve this, a benchmark plan which represents the best achievable dose with the current technology is predicted for each patient. This - 19 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)benchmark plan is used as the goal to evaluate existing clinical plans and provide further plan modification guidance. Referring again to FIG. 14, for a patient with multiple brain lesions, the ideal dose is predicted by a deep neural network trained from benchmark plans. The manually generated clinical plan is compared against the benchmark, and a DVH score can be calculated for each target to assess brain dose sparing. The predicted ideal dose and the DVH score provide planning guidance to improve the existing clinical plan.

[0093] Deep learning techniques, such as convolutional neural networks (CNN), have been utilized in dose prediction for radiotherapy. Typically, a deep learning model takes image data as input, such as CT images and structure contour masks, and predicts the dose distribution in 2D or 3D matrices. These models are ideally trained with “expert” plans for the same treatment site with consistent and high quality. Deep learning leverages the information embedded in these plans and correlates the dose distribution with patient anatomy. Predicting dose for intracranial SRS has several differences with other radiotherapy treatment sites: tissue heterogeneity is less of a concern; targets are typically small (less than 3 cc) and near spherical; OAR dose constraints are less complicated; dose distribution has a large dependence on the treatment technique.Conclusion

[0094] This Example describes the development of an artificial intelligence (Al)-based plan quality evaluation tool with patient-specific brain dose evaluation for multi-lesion SRS planning. The artificial intelligence (Al)-predicted ideal dose distributions can be used to objectively evaluate multi-lesion SRS treatment plan quality and automatically generate guidance for plan modification. For suboptimal clinical plans, Al guidance can achieve statistically significantly lower doses to healthy brain tissues.

[0095] The proposed Al-based plan quality evaluation and optimization tools will greatly benefit the physician in assessing plan quality and the planner by reducing the workload and time when modifying an existing plan. Existing plan evaluation criteria are mostly patient safety based, i.e., whether certain normal tissue constraints are met. With normal tissue constraints satisfied, the plan can be evaluated by the Al-based scoring system and further modified to keep healthy brain dose as low as reasonably achievable (ALARA). In radiation oncology, multiple treatment delivery techniques are available with varying levels of plan complexity and normal- 20 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)tissue sparing capability. For the same patient, plans made with different machines / techniques can be compared efficiently and objectively using the scoring system.EXAMPLE 5 - Dual-Model Deep Learning Framework for Accurate Benchmark Dose Prediction in Brain Stereotactic Radiosurgery

[0096] Purpose / Objectives: Dose gradient in brain SRS directly impacts normal tissue toxicity, yet no standardized benchmark currently exists for cross-platform plan quality assessment. Described herein is a size-stratified deep learning framework to predict patientspecific benchmark dose, employing appropriately scaled volumes of interest (VOI) to ensure accurate dose gradient prediction across the full range of lesion sizes.

[0097] Materials / Methods: Benchmark Gamma Knife (GK) plans optimized for maximal dose conformality were generated for 461 patients (1,097 targets). A deep learning framework consisting of two 3D U-Net models with attention mechanisms was trained: a small target model (for volume <1 cc, 756 targets, median 0.27 cc [range 0.002-1.0], 64x64x64 mm3VOI, 5-fold cross-validation ensemble) and a large target model (for volume >1 cc, 341 targets, median 2.9 cc [range 1.0-75.3], 128x128x128 mm3VOI). A custom loss function prioritized accuracy in the dose falloff region, i.e., the 1-cm ring around the target volume. Performance was evaluated on an independent test set of 137 patients with 209 targets (103 small, 106 large) and compared against two single-model baselines using paired Wilcoxon signed-rank tests.

[0098] Results: On 209 independent external test targets, the small target model (volume < 1 cc, n=103) achieved a mean absolute error (MAE) of 0.43±0.15 Gy in the dose falloff region, V5o% Dice of 0.866, and |Vi2Gy error| of 0.09±0.13 cc. The large target model (volume > 1 cc, n=106) achieved an MAE of 0.35±0.17 Gy in the dose falloff region, Vso% Dice of 0.909, and |Vi2Gy error) of 0.72±1.57 cc. Compared to single-model baselines, the dual-model approach achieved 30% lower intra-target MAE and significantly better Dice, and V i2Gy accuracy overall (all p<0.001). The large VOI baseline model performed 26% worse in falloff MAE (p<0.001), while the small VOI baseline model showed equivalent falloff accuracy but significantly worse intra-target prediction.

[0099] Conclusions: The dual -model deep learning framework accurately predicted patient-specific benchmark dose distributions for brain SRS. It achieved statistically significant- 21 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)improvements relative to single-model alternatives, enabling reliable cross-platform plan-quality evaluation and direct utility for routine clinical SRS quality assurance for all lesion sizes.Metric *Small VOI model *Large VOI model 2-MODEL (Ours)(64-mm Baseline) (128-mm Baseline)Falloff MAE (Gy) 0.39±0.17 0.53±0.21 0.39±0.17Inside PTV MAE (Gy) 3.78±2.15 3.97±2.55 2.65±1.91Correlation 0960±0031 0945±0043 0.964±0.031Dice (50% Rx) 0.881±0.122 0.858±0 122 0.888±0.121|V12Gy Error| (cc) 0.43±1.29 0.75±2.93 0.41± 1.16*A common evaluation grid was achieved by using a symmetric resampling scheme for the baselinemodels. Interpolation artifacts and information loss may modestly affect measured performance.Supplementary DataKey Results SummaryModel N Falloff MAE |V12Gy Error| Correlation Dice (50%) Small (<lcc) 103 043±0 15 Gy 009±0 13 cc 0956±0035 0.866±0.160 Large (>lcc) 106 0.35±0 17 Gy 0.72±1.57 cc 0.971±0.025 0.909±0.054Overall 209 0.39±0.17 Gy 0.41± 1.16 cc 0.964±0.031 0.888±0.121 Comprehensive Baseline Comparison (n=209)Metric Single small VOI model Single large VOI model 2-MODEL (Ours)(64-mm Baseline) (128-mm Baseline)Falloff MAE (Gy) *039±0.17 *0.53±0.21 0.39±0.17Inside PTV MAE (Gy) *3 78±2.15 *3.97±2.55 2.65±1.91 Correlation *0.960±0.031 *0.945±0.043 0.964±0.031Dice (50% Rx) *0881±0.122 *0.858±0.122 0.888±0.121 V12Gy Error (cc) *0.43±1.29 *0.75±2.93 0.41± 1.16* While this symmetric resampling scheme ensures a common evaluation grid, it remains an approximation: interpolation artifacts and information loss from down / upsampling may modestly affectmeasured performance, and results should be interpreted with this caveat in mind.Statistical Significance (Wilcoxon signed-rank, p<0.05)

[0100] 2-MODEL vs BASELINE-001 (64mm single):• Significantly better: Inside PTV MAE, Correlation, Dice, |V12Gy Error| (all p<0.001)• Equivalent: Falloff MAE, Global MAE (expected — same native resolution for small targets)

[0101] 2-MODEL vs BASELINE-002 (128mm single):• Significantly better: All 6 metrics overall (all p<0.001, except Dice large p=0.24) • 26% worse falloff MAE overall; 33% worse inside PTV MAEModel Details- 22 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)

[0102] Small Target Model (SMALL-032)• Architecture: 5-fold cross-validation ensemble, 3D U-Net with attention gates • Input: 64x64x64 mm3ROI at 1mm resolution• Training data: 756 targets from 2 institutions• Key features: Ensemble averaging reduces prediction variance, falloff-weighted loss (2:1)

[0103] Large Target Model (EXP-028)• Architecture: 3D U-Net with attention gates• Input: 128x128x128 mm3ROI at 1mm resolution• Training data: 341 targets from 2 institutions• Key features: Captures extended dose gradients, falloff-weighted loss (3:1), enhanced augmentationEQUIVALENTS

[0104] Although preferred embodiments of the invention have been described using specific terms, such description is for illustrative purposes only, and it is to be understood that changes and variations may be made without departing from the spirit or scope of the following claims.- 23 - 57737409.4

Claims

Attorney Docket No. 205961-7137WO1 (00572)CLAIMSWhat is claimed is:

1. A method of training a machine learning model to determine a best achievable normal brain tissue dose distribution (AI-BBD) for stereotactic radiosurgery (SRS) plans, the method comprising:providing a training set including single lesion SRS plans; andtraining the machine learning model with the training set, the training generating a trained model;wherein the trained model is configured to determine the AI-BBD for multiple lesions SRS plans.

2. The method of claim 1, wherein the model includes a neural network, a deep learning model, or a conditional generative adversarial networks (cGAN).

3. The method of claim 2, wherein the deep learning model comprises a small target model and a large target model.

4. The method of claim 3, wherein:the small target model is trained on a small lesion dataset; andthe large target model is trained on a large lesion dataset.

5. The method of claim 4, wherein:the small lesion dataset includes lesions having a volume of less than 1 cc; and the large lesion dataset includes lesions having a volume equal to or greater than 1 cc.

6. The method of claim 1, wherein the single lesion SRS plans of the training set include conformal arc (DCA) plans, volumetric arc therapy (VMAT) plans, Gamma Knife plans, or combinations thereof.

7. The method of claim 1, wherein the model includes a random forest (RF) regression model.- 24 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)8. The method of claim 1, wherein the model is trained based upon a region of interest (ROI) corresponding to each of the brain lesions in the single lesion SRS plans of the training set.

9. The method of claim 8, wherein the ROI for each of the brain lesions includes a planning target volume (PTV) of the lesion.

10. The method of claim 9, wherein the PTV is determined based upon at least one aspect of the brain lesion selected from diameter (measured in the largest dimension), volume, sphericity, location, depth, distance between lesions, or a combination thereof.

11. The method of claim 9, wherein the PTV further comprises a ring or sphere defined around the PTV.

12. The method of claim 11, wherein the ring or sphere includes a thickness around the PTV of between 1 cm and 1.5 cm.

13. The method of claim 8, wherein the model is configured to merge predicted doses for the lesions based upon a location of the ROIs.

14. The method of claim 1, further comprising generating a plan quality evaluation score.

15. The method of claim 14, wherein the plan quality evaluation score comprises a dose volume histogram (DVH) score (SDVH) according to the equation:VAI_SDVH = - x 100’'clinicalwhere:AreaV~ (0.8 — 0.5) / ?x16. The method of claim 15, wherein a SDVHscore of 100% indicates the best brain dose has been achieved.- 25 - 57737409.4Attorney Docket No. 205961-7137WO1 (00572)17. The method of claim 14, wherein the plan quality evaluation score is independent of radiation dose prescription of each lesion.

18. A method of evaluating a clinical stereotactic radiosurgery (SRS) plan for multiple lesions, the method comprising:inputting the clinical SRS plan to the machine learning model according to claim 1; receiving an output from the trained model; andevaluating the clinical SRS plan based upon the output.

19. The method of claim 18, further comprising modifying the clinical SRS plan based upon the evaluating.

20. The method of claim 18, further comprising finalizing the clinical SRS plan based upon the evaluating.- 26 - 57737409.4