A system and method for determining a treatment to be applied to the brain

Abstract

Disclosed is a method for determining a treatment to be applied to a brain of a patient, such as determining a transcranial magnetic stimulation (TMS) treatment location(s) and treatment strength to be applied to a brain of a patient suffering from depression. A search of the brain, based on coarse-scale parcellations, treatment metric and one or more anatomical metrics, is conducted to select the treatment location(s). In one embodiment, this involves progressively varying a value of a parameter associated with a gyrus threshold and for each value, identifying candidate treatment locations based on a functional connectivity of a subgenual anterior cingulate cortex during resting-state functional magnetic resonance imaging (fMRI) recorded over a plurality of multimodal brain imaging sessions, and selecting as the respective treatment location the candidate location that is closest to all other candidate treatment locations.

Description

Technical Field The present invention relates, in general terms, to a system and method for determining a treatment to be applied to the brain of a patient. More particularly, the invention relates to, but is not limited to, the automated development of an individualised treatment for treating disorders that might benefit from brain stimulation. Background The reference in this specification to any prior publication (or information derived from it), or to any matter which is known, is not, and should not be taken as an acknowledgment or admission or any form of suggestion that that prior publication (or information derived from it) or known matter forms part of the common general knowledge in the field of endeavour to which this specification relates. It has been found to be beneficial to some patients, to receive stimulation of the brain - e.g. through transcranial magnetic stimulation (TMS) - to treat varies mental disorders including treatment resistant depression. Recent developments have focussed on automated identification of locations of the brain, to which to apply stimulation. Approaches for selecting brain stimulation target locations require users to set one or more parameter values (e.g., functional connectivity threshold). However, the optimal parameter values are highly variable across individuals and imaging acquisition / preprocessing pipelines. Brain stimulation typically involve two components: stimulation target locations and stimulation strength / dose. Some treatment design methodologies employ - 2 - coarse-scale individualized brain networks for selecting individualized stimulation locations for actual treatment. However, such approaches do not focus treatment where it would be most beneficial, or overcompensate for lack of accuracy in treatment location by increasing dosage. Increased dosage can lead to patient discomfort and risk of adverse effects (e.g. seizures). Every brain stimulation device has its own physical limitations, which interact with the anatomy of the organ being stimulated. For example, in the case of TMS, the stimulation strength decreases with distance from the device, which means that if the target stimulation location is deep in a brain sulcus, the stimulation strength / dose has to be much higher to achieve a desired stimulation strength at the stimulation location. This leads to various problems. The higher stimulation strength leads to unintended stimulation of off-target superficial brain tissue. Also, the higher stimulation strength makes the stimulation less tolerable for the patient (e.g., causing more headaches during the treatment) and increases risk of adverse effects such as seizures. Other existing brain stimulation techniques utilize a one-spot-fits-all anatomical approach to determine a prefrontal stimulation location, or other brain location - e.g. a brain location dependent on the type of disease being treated. However, anatomical localization does not account for inter-individual variability in brain connectome. It would be desirable to overcome or reduce the drawbacks with prior art methods for developing and / or applying treatments to the brain of a patient for treating a mental disorder (or potentially other disorders not traditionally classified as brain disorders but that are treatable by brain stimulation). These disorders are typically mental disorders, but disorders not traditionally classified as brain disorders can also benefit from brain stimulation. Furthermore, brain stimulation can also be used to improve conditions (e.g., cognitive performance or mood) in healthy individuals. - 3 - Summary Disclosed herein are methods, and systems that execute those methods, for affording a novel brain imaging based - e.g. Functional Magnetic Resonance Imaging (fMRI) based - connectome targeting localization approach for brain stimulation of patients with neurological disorders or psychiatric disorders or potentially other disorders not traditionally classified as brain disorders. Furthermore, brain stimulation can also potentially be used to improve conditions (e.g., cognitive performance or mood) in healthy individuals. It has been found that using individualized functional connectome guided stimulation locations can yield dramatically improved TMS efficacy for treatment-resistant depression. Disclosed is a system for determining a treatment to be applied to a brain of a patient, the treatment comprising a treatment strength and one or more treatment locations, the system comprising: memory storing instructions; a receiver; a parcellation module; an anatomical metric module; a search module; a dosage module; an applicator system; and a processor, the processor executing the instructions to cause the system to: receive at the receiver, multimodal brain imaging data of the brain, the multimodal brain imaging data comprising functional imaging data and anatomical imaging data; - 4 - process the functional imaging data to produce coarse brain parcellations, fine brain parcellations and a treatment metric predictive of a treatment outcome; process the anatomical imaging data to derive one or more anatomical metrics associated with the brain; apply a threshold-free search of the brain, using the search module, to select the one or more treatment locations, the search being based on the coarse scale parcellations, treatment metric and one or more anatomical metrics; and determine the treatment strength, using the dosage module, by identifying one or more dosage stimulation locations using the fine-scale parcellations and one or more anatomical metrics, and stimulating each dosage stimulation location using the applicator system to measure the stimulation effect, and deriving the treatment strength from the stimulation effect. Throughout this specification and the claims which follow, unless the context requires otherwise, the word "comprise", and variations such as "comprises" and "comprising", will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps. As used herein, the term "multimodal brain imaging data" refers to functional imaging such as fMRI or EEG, and anatomical imaging, such as Tl-weighted MRI. Also disclosed is a method for determining a treatment to be applied to a brain of a patient, the treatment comprising a treatment strength and treatment location, the method comprising: receiving multimodal brain imaging data of the brain, the multimodal brain imaging data comprising functional imaging data and anatomical imaging data; - 5 - processing the functional imaging data to produce coarse-scale brain parcellations, fine-scale brain parcellations and a treatment metric predictive of a treatment outcome; processing the anatomical imaging data to derive one or more anatomical metrics associated with the brain; applying a threshold-free search of the brain, based on the coarse-scale parcellations, treatment metric and one or more anatomical metrics, to select the treatment location; identifying one or more dosage stimulation locations using the fine-scale parcellations and one or more anatomical metrics; stimulating the one or more dosage stimulation locations and detecting a resulting stimulation effect; and determining the treatment strength from the stimulation effect. Advantageously, embodiments of the invention may employ threshold-free searching of the brain for one or more treatment locations. Thus, proposed therein are approaches that do not require users to manually select thresholds for these parameters. Instead, presently disclosed approaches use other mechanisms (e.g. a tree-based approach) for automatically searching through a range of values across all parameters, to select optimal target location(s). The present threshold-free approaches are highly robust and led to better performance metrics across datasets compared with other approaches. This is highly desirable since it shows the present approaches can be easily translated to new clinics / hospitals / MRI scanners without the need to collect pilot data (from the new clinics / hospitals / MRI scanners) fortuning parameters. Advantageously, the present methods employ multi-resolution parcellations -the simultaneous estimation of coarse-scale (or network-level) and fine-scale (or areal-level) brain parcellations. The coarse-scale brain parcellation is used for selecting stimulation location(s) for actual treatment, while the fine-scale brain parcellation is used to guide the titration of an individualized stimulation - 6 - strength / dose. The use of the fine-scale brain parcellation facilitates the automatic titration of individualized stimulation strength / dose, as opposed to manual trial-and-error in current clinical practice. Consequently, the present approaches speed up the dosage titration procedure and increase the reliability of the procedure because it does not require the clinician / nurse to make subjective decisions. Advantageously, embodiments of the invention take anatomical constraints into account. The organ's anatomy is used to help guide identification of stimulation locations and dose titration. In methods proposed herein, anatomical constraints are used based on the physical limitations of the stimulation device, to improve stimulation location selection and dose titration. Brief description of the drawings Embodiments of the present invention will now be described, by way of nonlimiting example, with reference to the drawings in which: Figure 1 illustrates a multi-resolution, multimodal, threshold-free approach for treatment development for individualised brain stimulation in accordance with some embodiments; Figure 2 illustrates an instantiation of the approach of Figure 1, for TMS treatment of treatment-resistant depression; Figure 3, comprising images (A) and (B) illustrates a threshold-free tree-based approach for selecting individualised brain stimulation locations; Figure 4 is a histogram of optimal sACC correlation threshold for a method described in "Personalized connectivity-guided DLPFC-TMS for depression: Advancing computational feasibility, precision and reproducibility" published in 2021 to Cash et al (hereinafter "Cash2021"), obtained by a leave-one-out cross- - 7 - validation procedure; and Figure 5 shows test-retest reliability (inter / intra-subject distance) and sACC correlation in two datasets and distance-to-scalp in dataset 1, indicating statistical significance after multiple comparisons correction with false discovery rate (FDR.) q<0.05, where distance-to-scalp was not computed in dataset 2 because the analysis was performed in Montreal Neurological Institute (MNI) average space in order to be consistent with Cash2021. Figure 6 illustrates a system for performing the method of Figure 1. Detailed description Disclosed is a novel Functional Magnetic Resonance Imaging (fMRI) connectome targeting localization approach for brain stimulation of patients with neurological disorders or psychiatric disorders or potentially other disorders that are not traditionally classified as brain disorders (but might benefit from brain stimulation). The approach yields more reliable functional network localization than previous fMRI-based targeting approaches, which can further improve treatment efficacy and reliability. While particular embodiments develop treatment regimes for specific stimulation devices that target specific anatomic structures, other stimulation devices might preferentially target sulci or deeper brain structures and anatomical constraints can be adjusted accordingly. Rather than employing an individual-level fMRI targeting approach involving users thresholding one or more parameter values (e.g., functional connectivity threshold), the methods disclosed herein use a threshold-free tree-based approach to automatically search through a broad range of values across all parameters, and select consensus target location(s) across parameter values. The methods disclosed herein yield better performance metrics, relative to - 8 - existing methods, across datasets without any manual tweaking of parameter values / thresholds. The present methods simultaneously estimate coarse-scale (or network-level) and fine-scale (or areal-level) brain parcellations. The coarse-scale brain parcellation is used for selecting stimulation location(s) for actual treatment, while the fine-scale brain parcellation is used to guide the titration of an individualized stimulation strength / dose. Figure 1 illustrates a method 100 for determining a treatment to be applied to the brain of a patient. Notably, in general, the method 100 results in development of a treatment regime (locations for applying treatment, and strength of treatment) but does not actually treat or administer. In some embodiments, method 100 will also be extended to performing treatment. Broadly speaking, the method 100 involves imaging the patient's brain (step 102), coarse-scale (network level; typically comprising 2 to 50 regions of interest) and fine-scale (areal-level; typically comprising 100 regions of interest or more) brain parcellations are developed from functional imaging data (steps 104 and 106, respectively), a metric is determined that is used to predict the outcome of treatment (step 108), anatomical metrics are derived from anatomical imaging data (step 110), to constrain stimulation locations, and stimulation target locations are identified (step 112) and dosage titration is determined (step 114). In some, but not all, embodiments, treatment is then administered in accordance with the stimulation target locations and dosage (step 116). In further detail, step 102 involves acquiring functional brain imaging data. For illustration purposes only, acquisition involves the patient undergoing multimodal brain imaging such as Tl-weighted MRI, resting-state functional MRI (fMRI), PET or EEG. It will be appreciated that other functional brain imaging technologies can be used, particularly to measure blood flow and / or metabolism - 9 - characteristics of the patient's brain, and the present disclosure is not intended to be limited to the particular imaging modalities described herein. A system performing the method 100 (e.g., the system shown in Figure 6) may include the imaging apparatus itself, such as an fMRI machine. In other embodiments, the system comprises a receiver that receives multimodal brain imaging data at step 102 from an imaging apparatus, a database or other source. The functional brain imaging data (e.g., fMRI or EEG) is processed at steps 104 and 106, to estimate individualized coarse-scale (or network-level) and fine-scale (or areal-level) brain parcellations. The estimation of both coarse-scale and fine-scale parcellations may occur simultaneously, or one after another. The coarse-scale parcellation involves dividing the brain into multiple regions of interest (ROIs; typically comprising 2 to 50 regions), referred to as brain networks, based on patterns observed in the imaging data. This can be achieved using a neural network trained to identify, or a statistical algorithm for identifying, patterns corresponding to collections of regions across the brain that show functional connectivity, through statistical analysis of the fMRI, EEG or other functional imaging methods. The fine-scale parcellation divides the brain into a high-dimensional set of ROIs, typically exceeding 100 or more regions, without necessarily being constrained by the coarse-scale parcellation boundaries. In addition to parcellating the functional brain imaging data, the imaging data is analyzed at step 108 to derive one or more metrics predictive of treatment outcome. In this regard, step 102 may also involve receiving a diagnosis (e.g., generated by a clinician) and step 108 may involve determining one or more metrics that are measurable using the multimodal imaging data, and measuring each of the one or more metrics. Determining the one or more metrics may itself comprise querying a lookup table or other data structure that lists conditions of the brain together with a metric, or metrics, that can be measured - 10 - using the multimodal imaging modality, to determine effectiveness of treatment of each respective condition. For example, the metrics can be derived as the functional connectivity to one or more ROIs, where an ROI is defined as a set of brain locations, which may or may not be spatially contiguous. The functional connectivity can be defined as the temporal correlation with a pre-specified or predetermined timeseries. The timeseries itself may be derived from a weighted aggregation of the timeseries of brain locations within the ROI. Furthermore, the metrics can include any form of a spatial brain map that identifies regions of the brain considered relevant to the treatment. Relevance may be determined based on functional, structural, responsiveness to a treatment, or other criteria. For example, in the case of treatment resistant depression, the functional connectivity (i.e., Pearson's correlation) between every brain location and subgenual anterior cingulate cortex (sACC) during resting-state fMRI might be used as a metric. In this case, a more negative functional connectivity is associated with improved depression symptoms. As another example, the metric can be defined as the functional connectivity to the timeseries of a brain circuit map (derived using any appropriate neuroimaging technique - e.g., connectograms and others) associated with a specific condition, such as depression. The timeseries is derived as a weighted average of the timeseries of various brain regions, where the weight is determined by the values within the circuit map. These values represent the degree of relevance or contribution of each brain region to the condition, with higher values indicating greater relevance to the associated symptoms. As yet another example, the metric can be also defined as an efficacy brain map for a specific treatment, such as depression therapy. In this case, each value in the map represents the estimated effectiveness of the treating that brain region. Higher values indicate regions that are more effective targets for stimulation or - 11 - other therapeutic interventions. In yet another example, an abnormality brain map of the patient can also be derived. At each spatial location of the brain map, the value indicates deviation of the patient's brain marker from a normative range derived from a healthy group of participants that is demographically matched to the patient. In all the above examples, each metric may be computed for each location across the entire brain or within a pre-defined region of interest (e.g., DLPFC -dorsolateral prefrontal cortex). At step 110, the anatomy of the brain is used to refine, at steps 112 and 114, the locations and dosages of treatment. In particular, anatomical imaging data (e.g., Tl-weighted MRI) is analyzed to derive one or more anatomical metrics associated with the brain. This metric or metrics constrain stimulation locations - i.e., the anatomical metrics are used to limit the locations considered for stimulation, thereby reducing the search field for acceptable stimulation locations. Each metric is computed for the same set of spatial locations as step 108. The anatomical features may further be determined based on the treatment modality or device. Moreover, at step 102, a treatment device type may be received. A lookup table or other data structure may again be used to match anatomical features to treatment modalities or devices. This ensures the anatomical features being analyzed, and thus the metrics derived therefrom, are suitable or optimised for the particular treatment device or modality (i.e., method of delivery). For example, in the case of TMS, stimulation strength decreases with distance from the device. So, it is desirable for stimulation locations to be closer to the surface of the brain. Therefore, brain locations closer to the skull might be given a higher (desirable) anatomical metric value. It is worth noting that other stimulation techniques (such as focused ultrasound, temporal interference electrical field stimulation or deep brain stimulation) might preferentially target sulcus or deeper brain structure, so it's not always - 12 - the case that superficial structures are preferentially targeted. Using the above process, the brain parcellations are individualised, and the baseline for the treatment metric baseline are all individualised. This enables the treatment locations and the dosage to be developed for the individual, rather than being selected by a clinician or being generalised over a population. Step 112 involves performing a threshold-free search of the brain with a search module of the system, to select treatment locations. The threshold-free approach described in the illustrative embodiments is tree-based, as described with reference to Figure 2. The threshold-free analysis takes into account the coarse-scale parcellation, treatment outcome metric and anatomical constraints to select stimulation target location(s) for treating the patient. More specifically, depending on the targeted disorder / symptoms, different candidate regions of interest (ROIs), also referred to as brain networks, are selected from the coarse-scale parcellation. The selection of candidate networks is guided by their known functional roles in relation to the targeted disorder / symptoms, as established in the scientific literature and clinical practice. For example, in the case of major depression treatment, the system may focus on the brain networks associated with attention (i.e., attention networks), due to their involvement in cognitive control and emotional processing. Within the candidate networks, different stimulation locations are then derived. Stimulation locations are derived based on treatment outcome metrics and anatomical constraint metrics - the process defined above with reference to steps 108 and 110. These stimulation locations are organized into one or more hierarchical trees, with each node representing a potential stimulation location. Candidate nodes are a subset of tree nodes, chosen based on their hierarchical position within the tree. The selection focuses on two distinct types: leaf nodes, which lack child nodes and represent the most refined locations, and branch nodes, which have multiple child nodes and serve as critical points of convergence or divergence within the hierarchy. From the candidate nodes - 13 - (which may, in some embodiments, be referred to as candidate stimulation locations). Across multiple trees, the final consensus on stimulation location(s) is determined using a predefined selection criterion. This criterion may involve identifying the node that minimizes an average Euclidean distance to other candidate nodes. Alternatively, it may use methods that weigh specific nodes based on factors such as proximity, functional relevance (e.g., relevance to a particular brain network or relevance to a function that is also relevant to other nodes), or other context-specific metrics. The selection process is designed to ensure that the final target location optimally satisfies the requirements of the intended application. More than one stimulation target can be selected depending on the stimulation device. Every patient exhibits unique sensitivity to the stimulation device. An individualized stimulation strength or dosage therefore needs to be determined. At step 114 the fine-scale (or areal-level) parcellation together with the anatomical constraint metric(s) are used by the dosage module to identify a set of candidate stimulation locations for determining stimulation / strength dosage. These stimulation locations are stimulated (generally in a predetermined sequence so that the individual contributions of each stimulation are identifiable) and the dosage metric is computed. Computation of dosage metrics are be performed by clinician observation, patient feedback or an external device. For example, in the case of TMS, the personalized stimulation dosage is often set to be 120% of the resting motor threshold (rMT) of the small hand muscle - this can be automatically set by the system. In other embodiments, a lookup table - e.g., relating dosage metrics (strength, duration and others) to particular conditions and / or proportion of rMT - may be queried to determine the dosage metrics. In current clinical practice, this is achieved by the clinician / nurse manually performing a trial-and-error procedure to find the correct location in motor cortex to stimulate the small hand muscle. In the present method, the dosage module uses the fine-scale parcellation to automatically generate a candidate - 14 - set of stimulation locations, that are then sequentially stimulated with different strength. The dosage then selects the dosage based on an electromyography (EMG) device, which detects the response of the small hand muscle. Thus, a fully automatic and objective procedure is implemented, speeding up the dosage titration procedure and increases the reliability of the procedure because it does not require the clinician / nurse to make subjective decisions. Steps 102 to 114 result in an individualised treatment regime being produced, that comprises both treatment locations that are specific to the patient and their anatomy, along with a dosage that is optimised for those locations. This improves treatment outcomes while lowering the likelihood of over-stimulation resulting in patient discomfort, potential seizures and other side effects. In some embodiments, the method 100 continues after developing the treatment regime to applying the treatment to the patient in accordance with each treatment location and treatment strength (i.e., dosage) specified in the treatment regime (step 116). Treatment is administered by an applicator system. That applicator system is of the same type as that for which the anatomical metrics were developed. The applicator system may, for example, include a robotic arm with a TMS end effector, the robotic arm being movable relative to the patient to contact the treatment locations and deliver the treatment (e.g., TMS) at the determined strength. This robotic arm can be automatically controlled or manually operated by a human. Notably, the strength may vary between treatment locations, and the applicator system accordingly administers treatment at each specific one of the treatment locations at the treatment strength determined for that respective treatment location. Figure 2 illustrates the instantiation of the method 200 for TMS treatment of treatment-resistant depression. In a similar manner to step 102, under step 202 the patient undergoes - 15 - multimodal brain MRI imaging to obtain Tl-weighted MRI and resting-state functional MRI (fMRI), or Tl-weighted MRI and resting-state fMRI fora patient are acquired from another source. Under steps 204 and 206, the fMRI data is pre-processed to denoise the fMRI data. These steps typically include motion correction, distortion correction (if a fieldmap is also acquired), multi-echo denoising (if the fMRI data has multiple echoes), regression of various nuisance variables related to motion and / or physiological artifacts. Parcellation then uses multi-session hierarchical Bayesian models (MS-HBMs) each being a probabilistic model developed to parcellate the cerebral cortex using functional brain imaging data. These models utilize a hierarchical structure to group brain regions exhibiting similar functional connectivity patterns into distinct regions of interest (ROIs), while accounting for both within-subject and between-subject variability. The MS-HBMs generate two levels of parcellation: a coarse-scale (network-level) individual-specific parcellation, typically comprising 2 to 50 ROIs, referred to as brain networks (step 204), and a fine-scale (areal-level) individual-specific parcellation, which typically includes 100 or more ROIs (step 206). In this particular instantiation, the fine-scale individual-specific parcellation was computed using MS-HBM by initializing with the publicly available 400-ROI Schaefer population-level parcellation. Schaefer parcellation is a "population-level" parcellation that is not personalized to an individual unlike the MS-HBM individualized parcellation. Following a comparable process at step 108, the pre-processed resting-fMRI is also used to compute a metric predictive of treatment outcome at step 208. In particular, the functional connectivity (i.e., Pearson's correlation) between subgenual anterior cingulate cortex (sACC) during resting-state fMRI is computed within a DLPFC (dorsolateral prefrontal cortex) mask. The DLPFC mask is obtained by combining anatomical landmarks corresponding to key regions of DLPFC that have been established in previous literature. Every brain location within the DLPFC mask has a sACC correlation value. A stronger negative correlation is thought to indicate better treatment outcome for - 16 - depression. Step 210 calculates the anatomical constraint metrics. For treatment-resistant depression, the Tl-weighted MRI is processed to generate a highly accurate segmentation of the white matter and pial surfaces, as well as an estimate of the gyral height (average convexity) at each cortical location. A more negative average convexity value indicates closer proximity to the scalp. Proximity to the scalp make the location more desirable for TMS. Note that instead of using average convexity value, the distance from the scalp can be directly computed and used that as an anatomic metric directly. Despite this, for the purpose of the analysis below, average convexity is used. In the case of treatment-resistant depression, the attention networks have been found to be negatively correlated with sACC. Thus, for step 212, the salience / ventral attention and dorsal attention networks (from step 204) were selected from within the DLPFC mask, and these are collectively referred to as the attentional DLPFC components. Since it is beneficial to select stimulation locations that are close to the scalp and negatively correlated with sACC, within the attentional DLPFC components two parameters are set - namely, the average convexity and sACC negative correlations. If traditional methods were used, different thresholds would need to be manually selected for each parameter - one for average convexity (indicating the closeness to the scalp) and one for sACC negative correlation. A significant issue is that the optimal thresholds vary widely across participants and MRI scanners. In contrast, the present tree-based approach is a threshold-free approach to estimate a consensus location for administration of treatment, across two sets of parameters - in some embodiments, a single parameter (anatomical constraint metric) may be used, or three or more parameters, depending on the most suitable approach for a particular brain condition and / or type of stimulation device. This approach is set out in Figure 3, in which image (A) shows the attentional DLPFC components. The general tree-based method set - 17 - out herein involves constructing a tree by progressively varying a value of one or more parameters associated with the anatomical metric or metrics. For each value, candidate treatment locations are identified based on the treatment metric. Each candidate treatment location is then a coarse-scale parcellation that is identified by that progressive parameter variation (i.e., that is not excluded after the parameter is adjusted to make the threshold higher - more difficult for the anatomy to comply with) and complies with the one or more parameters. With further reference to treatment-resistant depression, the parameter being varied is the negative average convexity of gyrus. A range of gyrus thresholds are considered from 10% to 5% (in intervals of 1%) - other ranges may be selected, as appropriate for a particular treatment device. In this regard, a 10% threshold would select the top 10% of brain locations with the most negative average convexity (closest to the scalp), while a 5% threshold would select the top 5% of brain locations with the most negative average convexity (closest to the scalp). For each gyrus threshold, sACC correlation thresholds are gradually varied from 100% to 5%. A 100% threshold would select the top 100% of brain locations with the most negative correlation with sACC, while a 5% threshold would select the top 5% of brain locations with the most negative correlation with sACC. The tree is thus constructed, based on a centroid (or area) of brain remaining in a closed region from a previous step. For example, at the outset there may only be a single region, the centroid of which corresponds to a root node of the tree. As the thresholds are tightened - e.g., average negative convexity moves from 10% to 9% - one or more regions, within the region corresponding to the root node, are identified and each said region becomes a new node in the tree. Within those new nodes, stills further nodes will be identified and placed on the tree as the thresholds are tightened again - e.g., average negative convexity moves from 9% to 8% - and so on. Thus, the set of brain locations that survives a given gyrus threshold and sACC correlation threshold are extracted, yielding one or more connected components (or clusters). The centroid of each connected component is computed, yielding one centroid for each connected - 18 - component. Each centroid corresponds to a tree node. For a given gyrus threshold, the centroids of the more stringent sACC correlation threshold (less voxels remaining after thresholding) will be the children of the centroids of the less stringent sACC correlation threshold (more voxels remaining after thresholding). There can be one or more trees for a given gyrus threshold. But, for the purpose of illustration, one tree is shown per gyrus threshold (image (A)). For each tree, candidate nodes are selected. Each candidate node corresponds to either a tree node with no children (i.e., a leaf node) or a tree node with multiple children (i.e., a branch node). Leaf nodes correspond to a brain location with local minimum in sACC correlation. Tree nodes with multiple children represent the centroid of a bigger connected component just before splitting into smaller connected components (as the stringency of the sACC correlation threshold is increased or tightened). Among all the candidate nodes across all trees, a final stimulation target is obtained by finding the candidate that is closest in distance on average to all other candidates (Figure 3 image (B)). Per step 214, in the case of TMS for treatment-resistant depression, an individualized stimulation dosage / strength needs to be determined. This procedure is automated using individualized fine-scale (areal-level) parcellation (from step 206 above). Recall that this fine-scale individual-specific parcellation was computed by initialization with the publicly available Schaefer 400-region brain parcellation. We also recall that the Schaefer parcellation is a "populationlevel" parcellation that is not personalized to an individual unlike the MS-HBM individualized parcellation. By comparing the Schaefer parcellation with and task-fMRI activation from the Human Connectome Project, it is determined that, on average across participants, right hand movement will activate Schaefer parcel 149 in the left motor cortex. Therefore, we infer that parcel 149 of the individualized fine-scale (areal-level) parcellation (which was initialized with the - 19 - Schaefer parcellation) would also correspond to the motor part of the patient's brain that controls the patient's right hand. Thus, the method takes advantage of the correlation between specific fine-scale parcellations and non-cranial anatomical sites through which those parcellations can be stimulated. For example, the method may involve identifying a region of the fine-scale individual-specific brain parcellation that control a non-cranial anatomical site, such as the right hand, left hand, right foot and so on. The identified individualspecific motor hand region (from the individual-specific brain parcellation) is then divided into a grid. One or more, and preferably all, grid locations are sequentially stimulated - e.g. using the applicator system. A response is measured at the corresponding non-cranial anatomical site (using, e.g., an EMG device). The grid node or grid location with highest response (e.g., greatest muscle stimulation as measured using the EMG) is then selected as the targeting location to determine the stimulation dosage. The dosage is then titrated by varying the strength of stimulation application to that targeting location. The stimulation strength resulting in a predetermined change in EMG amplitude, as measured at the non-cranial anatomical site, is taken as the dosage. To illustrate: with reference to the procedure performed at step 206, an individualized fine-scale (areal-level) parcellation was created using a multisession hierarchical Bayesian model (MS-HBM) which accounts for within-subject and between-subject variability. This model utilizes the Schaefer 400-region parcellation as an initialization to generate the individualized parcels. The MS-HBM adapts the initialization parcels to the individual brain imaging data, grouping brain regions with similar functional connectivity patterns (determined from the abovementioned brain networks, statistical analyses and others) into distinct parcels. The MS-HBM preserves the correspondence between the initialization (Schaefer group-level parcellation) and individualized parcels. Therefore, parcel 149 of the individualized fine-scale (areal-level) parcellation represents the individualized candidate parcel for hand stimulation. The centroid of this individualized hand parcel is used to generate an M x M grid of possible stimulation locations for the patient. A robotic arm can then be used to apply - 20 - TMS to each grid location and the stimulation effect is objectively measured using an electromyography (EMG) device attached to the small hand muscle of the patient. We note that the robotic arm can be automatically controlled or manually operated by a human. The grid node with the maximum EMG amplitude is selected as the final individualized Ml targeting location. The stimulation dose is then titrated by varying the TMS strength applied to this final grid node. The individualized resting motor threshold (rMT) of the patient is the stimulation strength that result in an optimal 50pV change in EMG amplitude as measured in the patient. In practice, a different resolution parcellation might be used. Per step 216, the patient then undergoes brain stimulation treatment based on individualized brain location and a fixed percentage of the (G) individualized hand motor stimulation strength / dosage. The same treatment development methods can be applied to other conditions, such as other mental conditions, mood-related and cognitive conditions, as well as non-mental conditions, where any such condition can benefit from brain stimulation - e.g., through TMS. The above methods can be employed in a system, such as system 600 shown in Figure 6. The system 600 includes memory 602 that stores program code 604, executable by one or more processors 606 to implement the method 100 or 200 as described herein. The system also includes a receiver 608 that receives the multi-modal imaging data from an imaging system or database 610. Once received, the images are processed sequentially by parcellation module 614, to produce coarse-level and fine-level parcellations, an anatomical metric module 612, that generates the one or more anatomical metrics used to constrain the search space covered by the search module 616 when identifying stimulation locations. The system 600 further includes a dosage module 618, - 21 - for determining the stimulation dosages (interchangeably referred to as dosage strength and similar) to be applied at the treatment locations identified by the search module 616. Lastly, some embodiments also include an applicator system 620 for administering the treatment, comprising applying a stimulus, at the relevant dosage at each stimulation location. In view of the present teachings, the skilled person will realise that various modules and components of the system 600 can be implemented in hardware, firmware or software, or a combination thereof. Moreover, the system may be located in a single server or distributed across multiple servers. Experimental setup A retrospective analysis of two datasets of healthy participants was performed. The first dataset comprised 18 participants from Singapore. Each participant had 2 sessions of resting-state fMRI (and Tl) roughly 2 weeks apart. Each session comprised a lOmin long run of fMRI data. The second dataset comprised 32 Human Connectome Project (HCP) participants with 2 sessions of restingstate fMRI (and Tl) roughly 1 year apart. We used a single fMRI run (~15min long) from each session. The two datasets utilized different fMRI acquisition protocols: multiecho multiband for dataset 1 and multiband for dataset 2. Dataset 1 was collected on a Prisma-Fit scanner, while dataset 2 was collected from a highly customized Skyra scanner. Dataset 1 was pre-processed using steps including motion correction, multi-echo denoising, regression of various nuisance variables related to motion and / or physiological artifacts, while dataset 2 was preprocessed using ICA-FIX and global signal regression with bandpass filtering. The goal of the analysis is to show that the present approach generalized across two very different datasets with no tweaking of parameters. More specifically - 22 - two approaches were compared: (1) Fox2012 optimal group-average location and (2) Cash2021 connectome-guided individualized locations. (1) The Fox2012 group-average location was obtained by averaging sACC correlation maps across many participants and picking the location with the most negative correlation in the group-average sACC correlation map, corresponding to MNI coordinates [-38 44 26], This was chosen as a baseline because these group-average coordinates have been used in a number of highly influential clinical trials (such as the THREE-D trial). (2) The Cash2021 approach improved on previous individualized connectome-guided approaches, yielding superior reliability and better (more negative) sACC correlations across sessions. The Cash2021 baseline therefore represents a state-of-the-art individualized connectome-guided approach. For each dataset, the present approach, Fox2012 and Cash2021 were applied to the first MRI session and second MRI session of each participant independently. Three evaluation metrics were considered: (1) To measure test-retest reliability, for each approach and a given participant, the Euclidean distance was computed between the participant's individualized target and the targets of every other participant. These are then averaged, yielding a single inter-subject distance per participant. For a given participant, the distance between the participant's individualized targets was computed from the two sessions (intra-subject distance). The ratio of inter-subject distance and intra-subject distance (denoted as inter / intra distance) is used as a final metric. A larger inter / intra distance indicates better test-retest reliability. Note that this metric is ill-defined for the Fox2012 group-average location, so for this metric, the comparison was only made between Cash2021 and the present approach. (2) The sACC correlation was also considered as a metric. In the case of Cash2021 and the present approach, the individualized target estimated from - 23 - session 1 is used to compute correlation with sACC in session 2 (and vice versa), and then averaged across the sessions. In the case of Fox2012 group-average location, the correlation of the group-average location was evaluated with sACC in each session and averaged across the two sessions. A more negative sACC correlation indicates better performance. (3) The distance between target location and the scalp was also calculated and used as a metric. TMS treatment typically increases stimulation strength linearly as a function of the target location from the scalp. Therefore a shorter distance is desirable because this translates to lower simulation intensity, yielding greater patient tolerability and also lower off-target stimulation. Note that this metric was only computed in dataset 1 because the analysis for dataset 2 was performed in MNI-average space to be consistent with the previous study (Cash2021), so the concept of distance-to-scalp is not applicable. Results There is no specific parameter to set for the present tree-based approach, since the process swept through a range of gyri and sACC correlation values to obtain a consensus target location. Similarly, there is no parameter to set for Fox2012 group-average location. On the other hand, the Cash2021 approach involved thresholding the sACC correlation map and then computing the centroid of the largest connected component. In the present experiment, this sACC correlation threshold was determined in a rigorous leave-one-subject out cross-validation procedure to minimize the inter-session sACC correlation. For example, to determine the sACC correlation threshold for the first HCP participant, the procedure swept through a range of sACC correlation thresholds (from 0.1% to 50%) in the remaining 31 HCP participants. The sACC threshold that yielded the best (lowest) inter-session sACC correlation in the 31 HCP participants was applied to the first HCP participant. This procedure was repeated for each participant (in each dataset). As can be seen in Figure 4 (histogram of optimal sACC - 24 - correlation threshold for Cash2021, obtained by a leave-one-out crossvalidation procedure), the optimal sACC correlation threshold varied greatly across participants and datasets, motivating the development of our threshold-free tree-based approach. Figure 5, image (A) shows the test-retest reliability (left panel), sACC correlation (middle panel) and distance between target and scalp in dataset 1. Figure 5, image (B) shows the test-retest reliability (left panel) and sACC correlation (middle panel) in dataset 2. In terms of test-retest reliability, the present approach exhibited numerically better (greater) inter / intra-subject distance than Cash2021 in both datasets, with statistical significance achieved in dataset 2 (but not dataset 1). Furthermore, the present approach exhibited statistically better (more negative) sACC correlation than both Fox2012 group-average location and Cash2021 approach. Finally, the present approach yielded target locations that are closer to the scalp compared with both Cash2021 and Fox2023 group-average approach. Thus, the presently proposed, multi-resolution multi-modal threshold-free approach for individualized brain stimulation yielded stimulation locations that are more (or equally) reliable when compared with existing methods, exhibited better (more negative) sACC correlation and were closer to the scalp than competing approaches. The improvements were achieved across two very different datasets collected from two different countries (North America and Singapore) with different demographics, scanners and preprocessing, without the need to tune any parameter. Overall, this suggests the generalizability of the present approach to different populations. The methodologies disclosed herein, and the system for implementing them, can be used in a wide variety of commercial applications. By targeting brain regions more precisely and accurately, the present method can potentially increase the clinical efficacy of brain stimulation, patient's condition can improve better in fewer sessions, thereby reducing healthcare costs. Since patients can - 25 - get better faster, quality of life improves and the economic cost of mental illness is reduced. Furthermore, the use of anatomical constraints can result in a lower stimulation strength / dosage (while maintaining the same level of treatment efficacy), which might make the treatment more tolerable for patients. It will be appreciated that many further modifications and permutations of various aspects of the described embodiments are possible. Accordingly, the described aspects are intended to embrace all such alterations, modifications, and variations that fall within the spirit and scope of the appended claims.

Claims

1. A system for determining a treatment to be applied to a brain of a patient, the treatment comprising a treatment strength and one or more treatment locations, the system comprising: memory storing instructions;a receiver;a parcellation module;an anatomical metric module;a search module;a dosage module; anda processor, the processor executing the instructions to cause the system to:receive at the receiver, multimodal brain imaging data of the brain, the multimodal brain imaging data comprising functional imaging data and anatomical imaging data;process the functional imaging data using the parcellation module to produce coarse-scale brain parcellations, fine-scale brain parcellations and a treatment metric predictive of a treatment outcome;process the anatomical imaging data to derive one or more anatomical metrics associated with the brain;apply a threshold-free search of the brain, using the search module, to select the one or more treatment locations, the search being based on the coarse-scale parcellations, treatment metric and one or more anatomical metrics; anddetermine the treatment strength, using the dosage module, by identifying one or more dosage stimulation locations using the fine-scale parcellations and one or more anatomical metrics, and stimulating each dosage stimulation location using the applicator system to measure the stimulation effect, and deriving the treatment strength from the stimulation effect.- 27 -2. The system of 1, further comprising an applicator system is configured to apply the treatment to the patient in accordance with each treatment location and treatment strength.

3. The system of 2, wherein the applicator system comprises a robotic arm for stimulating the dosage stimulation locations and each treatment location, and sensor for measuring the stimulation effect.

4. The system of any one of 1 to 3, wherein each treatment location is identified by:progressively varying a value of one or more parameters associated with one or more anatomical metrics and, for each said value, identifying candidate treatment locations based on the treatment metric, each candidate treatment location being a said coarse-scale parcellation that complies with the one or more parameters; andselecting as the respective treatment location the candidate treatment location that is closest to all other candidate treatment locations.

5. The system of 4, wherein the one or more anatomical metrics comprise a gyrus threshold, such as negative average convexity gyrus, and varying the value of the one or more parameters comprising varying one or both of a negative average convexity or a Euclidean distance to a scalp around the brain.

6. The system of 4 or 5, wherein the treatment metric comprises a functional connectivity of a subgenual anterior cingulate cortex during resting-state functional magnetic resonance imaging (fMRI) recorded over a plurality of multimodal brain imaging sessions.- 28 -7. The system of any one of 1 to 6, wherein the anatomical imaging data is processed to derive the one or more anatomical metrics for all portions of the brain within a predefined mask.

8. The system of any one of 1 to 7, wherein the stimulation effect is measured at a location of a body of the patient, corresponding to the dosage stimulation location being stimulated.

9. The system of any one of 1 to 8, wherein the functional imaging data comprises one or more of functional magnetic resonance imaging data, electroencephalogram data and positron emission tomography.10.The system of any one of 1 to 9, wherein the anatomical imaging data comprises Tl-weighted magnetic resonance imaging.11.A method for determining a treatment to be applied to a brain of a patient, the treatment comprising a treatment strength and one or more treatment locations, the method comprising:receiving multimodal brain imaging data of the brain, the multimodal brain imaging data comprising functional imaging data and anatomical imaging data;processing the functional imaging data to produce coarse-scale brain parcellations, fine-scale brain parcellations and a treatment metric predictive of a treatment outcome;processing the anatomical imaging data to derive one or more anatomical metrics associated with the brain;applying a threshold-free search of the brain, based on the coarse-scale parcellations, treatment metric and one or more anatomical metrics, to select each treatment location;identifying one or more dosage stimulation locations using the fine-scale parcellations and one or more anatomical metrics;- 29 -stimulating the one or more dosage stimulation locations and detecting a resulting stimulation effect; andderiving the treatment strength from the stimulation effect.12.The method of 11, further comprising applying the treatment to the patient in accordance with each treatment location and treatment strength.13.The method of 12, comprising controlling a robotic arm to stimulate the dosage stimulation locations and each treatment location, and measuring the stimulation effect using a sensor.14.The method of any one of 11 to 13, applying the search of the brain comprises:progressively varying a value of one or more parameters associated with one or more anatomical metrics and, for each said value, identifying candidate treatment locations based on the treatment metric, each candidate treatment location being a said coarse-scale parcellation that complies with the one or more parameters; andselecting as the respective treatment location the candidate treatment location that is closest to all other candidate treatment locations.15.The method of 14, wherein the one or more anatomical metrics comprise a gyrus threshold, such as a negative average convexity gyrus, and varying the value of the one or more parameters comprising varying one or both of a negative average convexity or a Euclidean distance to a scalp around the brain.16.The method of 14 or 15, wherein the treatment metric comprises a functional connectivity of a subgenual anterior cingulate cortex during resting-state functional magnetic resonance imaging (fMRI) recorded over a plurality of multimodal brain imaging sessions.- 30 -17.The method of any one of 11 to 16, comprising processing the anatomical imaging data to derive the one or more anatomical metrics for all portions of the brain within a predefined mask.18.The method of any one of 11 to 17, wherein detecting the stimulation effect comprises measuring the stimulation effect at a location of a body of the patient, corresponding to the dosage stimulation location being stimulated.19.The method of any one of 11 to 18, wherein the functional imaging data comprises one or more of functional magnetic resonance imaging data, electroencephalogram data and positron emission tomography.20.The method of any one of 11 to 19, wherein the anatomical imaging data comprises Tl-weighted magnetic resonance imaging.

Images