Determination of brain age using abnormality suppression

Abstract

A method and system that provides for the automated detection of abnormalities in subject brains from medical imagery of the subject brains, and then applies the known local brain age determination techniques to obtain a local brain age for different image blocks of the subject brain is described. The output from the automatic abnormality detection method and system is used to suppress or negate the output of the local brain age determination method and system, by causing the local brain ages that have been found for image blocks corresponding to subject brain regions for which an abnormality has been detected to be disregarded and / or discarded. Thus, by filtering the output of the local brain age detection method and system with the output of the method and system for automated detection of abnormalities, then only those local brain ages which relate to image blocks of the parts of the subject brain that are more structurally normal are output as local brain ages. Thus, using such a system a clinician can have confidence that the brain age data that is being output is correct, and is not being rendered incorrect by undetected abnormalities in the subject brain.

Description

[0001] DETERMINATION OF BRAIN AGE USING ABNORMALITY SUPPRESSION

[0002] Technical Field

[0003] Embodiments of the present invention described herein relate to methods of determining brain age from medical imagery of a subject brain by suppressing the contribution of any abnormal areas of the subject brain to the brain age determination.

[0004] Background to the Invention

[0005] Brain-predicted age, or simply "Brain-age", is an established and popular paradigm for assessing brain health. The underlying idea is quite straightforward; a machine learning model, typically a convolutional neural network, is trained to predict chronological age from structural MRI brain scans in a cohort of healthy individuals. When applied in independent clinical samples, deviations between brain-predicted age and chronological age - the so-called 'brain-predicted age difference' (brain-PAD), also known as the brainage gap, or delta - can be used to quantify deviations from healthy ageing. Having a brain that more closely resembles that of an older healthy person (i.e., positive brain-PAD) has been associated with a number of neuropsychiatric diseases, including Alzheimer's disease (Franke and Gaser, 2012), schizophrenia (Koutsouleris et al., 2014) and epilepsy (Pardoe et al., 2017); a positive brain-PAD has also been associated with cognitive impairment following traumatic brain injury (Cole et al., 2015), an increased risk of subsequent dementia (Biondo et al., 2020), and a greater risk of mortality (Cole et al., 2018a). These findings support the use of MRI-derived brain-age measures as a screening tool (opportunistically or otherwise) for these and other diseases to identify people at higher risk of poor health outcomes. An example "whole-brain" brain age determination method was described in the inventors recent paper Wood et al, Accurate Brain-Age Models for routine clinical MRI examinations NeuroImage Vol.249, Article 118871, 5 January 2022.

[0006] Potentially, brain-age could be predicted during routine clinical examinations to detect deviations from healthy ageing that may represent early-stage neurodegeneration. This in turn could help improve the patient pathway by expediting care (including intervention where available), and possibly accelerate the development of disease-modifying drugs through improved clinical trial enrolment. Brain-age could also feasibly guide adaptive MRI sequence acquisition (Cole et al., 2018b), enabling patients with older-appearing brains to undergo additional, more targeted, imaging (e.g., 'dementia protocol') whilst still in the scanner. However, a fundamental limitation of the brain-age paradigm currently precludes clinical adoption, namely that models can only reliably predict brain-age when no structural abnormalities are present. This is because the question that brain-age models can answer is "what healthy age does this brain resemble?" Clearly, when a patient has a tumour, stroke, or other underlying structural abnormality, there is no meaningful answer to this question. Nonetheless, it may still be of considerable interest to determine how healthy the rest of the brain is. For example, it may be desirable to know the impact of radiotherapy on the surrounding brain, or how the presence of a structural abnormality affects the health of the rest of the brain.

[0007] Current brain-age models are unable to answer these questions. Instead, when abnormal brains are passed to current brain-age models, they fail i.e., they return an erroneously low or high prediction of brain age. Given that the brain-age paradigm relies on the assumption that deviations between brain-age and chronological age are meaningful, this failure mode represents a severe limitation, and makes brain-age unusable in real clinical settings.

[0008] References referred to above include: 1) J.H. Cole, K. Franke. Predicting age using neuroimaging: innovative brain ageing biomarkers. Trends Neurosci., 40 (12) (2017), pp. 681-690; 2) K. Franke, C. Gaser. Longitudinal changes in individual BrainAGE in healthy aging, mild cognitive impairment, and Alzheimer's disease. GeroPsych: J. Gerontopsychol. Geriatr. Psychiatry, 25 (4) (2012), p. 235; 3) N. Koutsouleris, C. Davatzikos, S. Borgwardt, C. Gaser, R. Bottlender, T. Frodl, E. Meisenzahl. Accelerated brain ageing in schizophrenia and beyond: a neuroanatomical marker of psychiatric disorders. Schizophr. Bull., 40 (5) (2014); 4) H.R. Pardoe, J.H. Cole, K. Blackmon, T. Thesen, R Kuzniecky. Structural brain changes in medically refractory focal epilepsy resemble premature brain aging. Epilepsy Res., 133 (2017); 5) J.H. Cole, R. Leech, D.J. Sharp, Alzheimer's Disease Neuroimaging Initiative. Prediction of brain age suggests accelerated atrophy after traumatic brain injury. Ann. Neurol., 77 Ann. Neurol., 77. (2015) and 6) F. Biondo, A. Jewell, M. Pritchard, et al. Brain-age predicts subsequent dementia in memory clinics. Alzheimer's and Dementia, 16. (2021) 7) J.H. Cole, S.J. Ritchie et al. Brain age predicts mortality. Mol Psychiatry, 23. (2018)

[0009] As well as there being known "whole-brain" brain age determination techniques as described above, it is also known to segment the brain imagery into segments, and derive a local brain age for each segment. Popescu et al Local Brain-Age: A U-Net Model, Frontiers in Aging Neuroscience, 13 Dec 2021, Volume 13, Article 761954 describe such an arrangement, which is shown at broad level in Figure 1. In more detail, Figure 1 shows a U-Net architecture for voxel-level brain-age prediction. Raw Tl-weighted MRI scans are pre-processed using SPM12, obtaining modulated grey and white matter volume maps registered to the MNU52 template. 523blocks of both grey and white matter are passed through a U-Net architecture to obtain 12 x 12 x 12 voxel blocks of voxel-level brain-age prediction, with a separate estimate of brain age prediction for each block. Additional auxiliary block-level brain-age loss functions were added at each level of the U-Net to facilitate training. Further details of the operation of the U-Net based local brain age model are given in the Popescu et al paper referenced above, which is expressly incorporated herein in its entirety.

[0010] It is thus possible to obtain both "whole-brain" brain age measurements which give an estimated age for the whole of the subject brain, or to divide the brain imagery up into block-type segments and then derive a "local-brain" brain age measurements for each block of imagery. This results in multiple different possible local brain age measurements, with a brain age measurement being generated for each imagery block segment - in the prior art 12 x 12 x 12 voxel blocks were defined and brain age measurements obtained, with at least one for each block. In a healthy brain, these different measurements should be almost the same - however for a brain with abnormalities, for example caused by disease or injury, there is no mechanism for the prior art arrangement to know that a local block is being affected by any such abnormality, and hence that the brain age that is being derived will be erroneous. The local brain age determination technique therefore suffers from the same problem as the "whole-brain" brain age technique described earlier - both techniques are incapable of accurately deriving brain age for brains with abnormalities, for example tumours or lesions, caused for example by disease or injury. In both cases an erroneous output is obtained, without any indication that in fact the calculated brain age is erroneous, or by how much. Essentially, because the error is generally unknown and cannot be compensated for, neither the whole-brain brain age, nor local-brain brain age methods can be relied upon to give accurate results in the present of brain abnormalities.

[0011] This has important implications for the usefulness of the brain age determination techniques. This is because whilst deriving brain age of healthy test volunteers for the purpose of research is useful and helps to develop the technology, in practice in clinical settings a large majority of brain MRI scans will exhibit some sort of abnormality - there is generally some reason why the subject is having the scan in the first place, for example they have already started to exhibit the early stages of cognitive decline, and hence their brain already shows some disease-led abnormality. Cognitive decline is generally nonspecific and there are a wide range of causes including those associated with a range of structural abnormalities. In such cases the presently developed brain age techniques, trained as they are on generally healthy volunteers, fail to accurately predict the brain age of the actual patient subject, due to the existing brain abnormalities that will likely be present in such a patient subject.

[0012] Further development of the brain age determination techniques is therefore required, to render them more applicable in real-life clinical settings, where the brains that are being imaged may already possess some sort of abnormality. of the Disclosure

[0013] Embodiments of the present disclosure address the above problem by the provision of a method and system that provides for the automated detection of abnormalities in subject brains from medical imagery of the subject brains, and then applies the known local brain age determination techniques to obtain a local brain age for different image blocks of the subject brain. The output from the automatic abnormality detection method and system is then used to suppress or negate the output of the local brain age determination method and system, by causing the local brain ages that have been found for image blocks corresponding to subject brain regions for which an abnormality has been detected to be disregarded and / or discarded. Thus, by filtering the output of the local brain age detection method and system with the output of the method and system for automated detection of abnormalities, then only those local brain ages which relate to image blocks of the parts of the subject brain that are more structurally normal are output as local brain ages. Thus, using such a system a clinician can have confidence that the brain age data that is being output is correct, and is not being rendered incorrect by undetected abnormalities in the subject brain.

[0014] In further embodiments, the different local brain age data outputs that are output for the respective "normal" blocks of subject brain imagery can be combined or fused together, for example by averaging or clustering data processing operations to give a single subject brain age. This single subject brain age is then analogous to the "whole brain" brain age measurements of the prior art, but the utilising clinician can have confidence that it has been generated using only healthy "normal" parts of the subject brain i.e. the parts of the subject brain that are more structurally normal, and hence should be correct and indicative of the subject's global brain age across the healthy or apparently healthy parts of the subject brain. In view of the above, from one aspect the present invention provides a computer implemented method of determining brain age of brains with abnormalities using medical imaging, comprising: receiving medical imagery of a subject brain; detecting abnormal areas of the subject brain corresponding to disease or injury within the medical imagery; segmenting the medical imagery of the subject brain into a plurality of volumes; determining a separate brain age for at least a subset of the plurality of volumes; suppressing the brain age determination for a particular volume of the subject brain if the particular volume at least partially comprises detected abnormal areas; and providing for output the separate brain age determinations for those volumes for which the brain age determination has not been suppressed.

[0015] Further features and aspects will be apparent from the appended claims.

[0016] Brief Description of the Drawings

[0017] Embodiments of the invention will now be further described by way of example only and with reference to the accompanying drawings, wherein like reference numerals refer to like parts, and wherein:

[0018] Figure 1 is a diagram of a prior art arrangement for determining local brain age of local "blocks" of a subject brain and that is deployed as part of embodiments of the present disclosure;

[0019] Figure 2 is a diagram of a first embodiment of the present disclosure, showing the data and processing blocks that are applied to the data;

[0020] Figure 3 is a diagram of a prior art arrangement for detecting abnormalities in whole brain images of a subject brain and that is deployed as part of embodiments of the present disclosure;

[0021] Figure 4 is a system block diagram illustrating a computer implemented embodiment of the present disclosure; and

[0022] Figure 5 is a flow diagram illustrating the processing steps involved in an embodiment of the present disclosure.

[0023] Description of the Embodiments

[0024] Our approach to overcome the limitations of current brain-age models is two-fold. First, we propose a local "patch-wise" training method, such that the model returns a brain-age prediction for each contiguous (e.g. 47 x 47 x 47) cubic region of brain tissue. Second, we propose to "suppress" the brain-age predictions for regions in which an abnormality is present, using interpretability maps generated by a separately trained abnormality detection model. The upshot is that only those regions of the brain which do not contain a structural abnormality will have a brain-age prediction. The predictions from regions which do not contain an abnormality can be aggregated (e.g., using a downstream neural network) to generate a single predicted age which reflects overall brain health; alternatively, the individual predictions could be retained, allowing local deviations from healthy ageing to be detected.

[0025] Two main benefits to our approach can be identified. First, our combined model will be robust to gross structural pathologies, thus enabling its use in clinical settings where such abnormalities are common (our local data shows 70% of brain MRI scans at King's College Hospital (KCH) in London, UK are abnormal and 50% at Guys and St Thomas (GSTT) Hospital, also in London, UK, are abnormal). Second, by using a local "patch-wise" training method, our model will be intrinsically interpretable, generating brain-age predictions at each point in the brain. This will be particularly useful for degenerative diseases characterised by local atrophic changes e.g., Alzheimer's disease which is known to disproportionally shrink the hippocampi, leaving large parts of the surrounding brain intact.

[0026] A further benefit is that in using the embodiment patients will be triaged for the presence or absence of one or more abnormalities as this is intrinsic to the first step of the process. No other known model combines abnormality triage with brain age.

[0027] Figure 2 shows the process of the present embodiment in more detail. Medical imagery 20 (such as obtained from a magnetic resonance (MR) imaging scanner, or computerised tomography (CT) scanner) of a subject brain is obtained, and is input into two separate parallel processing processes, implemented on a computer. The first such process is an abnormality classifier 22, which is arranged to process the received medical imagery 20 of the subject brain, and to identify pathological anomalies in the brain, such as tumours, lesions, or other anomalous structures caused by disease or injury. The abnormality classifier is a "whole-brain" classifier, which processes images of the whole brain and is capable of identifying and segmenting anomalies anywhere in the brain imagery. Such abnormality classifiers are known already perse in the art, and an example of such will be described later. The output of the abnormality classifier is therefore information relating to the presence, characteristics and location of any detected abnormality in the brain imagery. As noted above, such abnormality classifiers are able to detect and segment abnormalities in the brain imagery on a voxel by voxel basis, and hence the location of any detected abnormalities can be precisely mapped within the imagery to voxel level. This voxel-level abnormality location information of any detected abnormality is passed to a local brain age suppressor function 26, which as described will then use the abnormality location information to suppress local brain age estimates.

[0028] The medical imagery 20 of the subject brain is also passed to a local brain age determination process 24. This local brain age determination process 24 applies the prior art local brain age determination processes of Popescu et al. and which were described earlier to segment the brain imagery into contiguous cubic blocks of brain tissue, and to find a local brain age prediction for each block. Precise details of the local brain age prediction for each block were given earlier and are available in the Popescu et al paper; suffice for the present embodiment that a local brain age prediction is obtained for each block. In this first embodiment a local brain age prediction is produced for each block, irrespective of the presence of disease or other abnormalities in the block, and then the predictions are suppressed for diseased or abnormal blocks. In other embodiments, however, the local brain age determination process can be controlled by the abnormality classifier so as to only produce a local brain age measure for those blocks that are not abnormal. Thus by "suppression" of the local brain age predictions that arise from abnormal blocks, we mean that either the predictions can be generated and then controlled so as to be thrown away or not used, or that the generation of the predictions is directly controlled by the abnormality classifier so as to generate local brain age predictions only for those blocks with no abnormal brain tissue therein.

[0029] Once the local brain age predictions have been obtained (in this example for every local block). The predictions are then fed to a local brain age suppression function 26, which receives the abnormality location data (if an abnormality exists) from the abnormality classifier. The location data will be in the form of a three dimensional region in the brain in which an abnormality is detected, with the voxel coordinates of the region defined. The local brain age suppression function 26 then cross-references the three dimensional abnormality region data with the three dimensional location block data of the local blocks for which a local brain age determination has been made by the local brain age determiner function 24. Where there is an overlap between the location of a local brain age determination block and the location of any detected abnormality, then the local brain age determination block is marked for suppression, and is not output for later use. As will be seen in Figure 2, in this example there are a series of local brain age determinations 27 that are suppressed and are not output by the local brain age suppressor function 26, because their locations overlap with that of a detected anomaly.

[0030] With respect to the degree of overlap required, where a local brain age block is completely located within the region of an abnormality, then clearly it should be suppressed. Where a local brain age block is partially located within the region of an abnormality i.e. the local brain age block encompasses voxels that are within the abnormality and others that are within normal brain tissue, then a tuning parameter can be defined which specifies the percentage of overlap of a block with the abnormality region before it is suppressed. That tuning parameter can be very sensitive, such that any overlap (even 1 voxel) will lead to suppression of the local block brain age, up to requiring at least 50% (or even more, for example 70%, 90% or 100%) of voxels within a block to be within the overlap with the abnormality in order to be suppressed. If more than 50% (or even more, for example 70%, 90% or 100%) of voxels within a block also lies within the region of a detected abnormality, then the local brain age for that block should be suppressed. It will be understood that proportion of voxels within a block that are abnormal that are required for the block to be suppressed can be tuned to be disease specific. For example, for late stage Alzheimers a proportion of more than 50% of voxels may be required.

[0031] With respect to outputs of the system, in one embodiment the set of local brain age predications output from the local brain age suppression process 26 can constitute the output, for example by being displayed on a screen. In this case, the local brain age predictions are exclusively from blocks that do not overlap with an abnormality area in the brain image, or only overlap to a small degree as permitted by the tuning parameter. As such, any reviewing clinician can have confidence that the local brain age predictions are correct, and are not influenced either at all or unduly by the presence of the abnormality.

[0032] In another embodiment, the individual local brain age predictions output from the local brain age suppression function 26 can be input into a local brain age fusion function 28, which combines them into a single brain age prediction, which is then applicable to the whole brain, other than any detected abnormalities. This combination can be performed in a number of ways, for example by applying the predicted brain ages to an averaging function, which calculates an average of the local brain ages. Alternatively, the averaging function may involve ignoring outlying values, for example by ignoring values more than one standard deviation away from the arithmetic mean, and then calculating the average of the remainder. Other embodiments may involve clustering the different local brain age results into clusters, and then finding the centroid value of the largest cluster. Various other algorithms and calculations for finding a single representative value of a range of values will be known to the skilled person, any of which may be used. Howsoever the multiple local brain ages are combined to give a single whole-brain (minus abnormality) age, the combined value is then output to the clinician user, for example by being displayed on a screen. Again the reviewing clinician can have confidence that the single brain age has been calculated using only the parts of the subject brain that are more structurally normal, and hence is representative of the age of the apparently healthier parts of the subject brain, rather than any abnormal parts.

[0033] With respect to the operation of the abnormality classifier 22, Figure 3 shows further details of the operation of such. Figure 3 is taken from the present inventors own paper Wood et al Deep Learning Models for Triaging Hospital Head MRI Examinations Medical Image Analysis 78 (2002) 102391, the entire contents of which that are necessary for understanding the present invention being incorporated herein by reference. In this respect, in 2022 the present inventors devised a deep learning system which was intended to assist with radiologist workload, by automatically scanning MRI brain images and detecting and classifying any brain abnormalities present in the images. The detailed operation of the system is shown in Figure 3 and described in more detail in the paper referenced above, suffice for present purposes that the system provides an abnormality classifier which receives subject brain medical imagery (such as MRI imagery) and is able to detect and segment within the images any abnormal regions of the subject brain, including a confidence probability of how correct the detection and segmentation are. In the context of the prior art paper describing such, the purpose of such a system was to help manage radiologist workload, by automatically detecting and segmenting abnormalities, and identifying the abnormalities with a degree of confidence, which could then be used with a natural language model to help generate automatic scan reports. The present inventors have realised, however, that the abnormality classifier used in their prior art system can be redeployed for an alternative purpose in the present disclosure, to act as the abnormality classifier in the present method and system. As described previously, in the present system the abnormality classifier identifies abnormal regions of the input brain scans, and then passes information on the locations of the regions to the local brain age suppressor module. In identifying abnormal regions of the brain from the input scans, however, the abnormality classifier operates in the same manner as described previously in the inventor's paper referenced above. In the interests of brevity, therefore, no further details of the operation of such will be provided here. Figures 4 and 5 show further details of the implementation and operation of the method and system of the present disclosure. Referring first to Figure 4, Figure 4 shows a general purpose computer platform comprising a CPU 40, random access memory 42, a graphics card 46 which is connected to an output display 44, and an input / output bus 50 through which input controllers 48 provide inputs to the computer system. The input controllers 48 maybe, for example, a mouse, a keyboard, or the like. The CPU 40, ram 42, graphics card 46, and input output bus 50 are connected via a common communications bus. Also provided is a non-volatile storage medium 52, such as a solid-state drive, hard disc drive, or the like. Stored on the non-volatile storage medium 52 is an operating system programme 54 which controls the general operation of the computer platform. Specific to the present embodiment, however, there is further provided an abnormality classifier programme 22, and a local brain age determination programme 24. Also provided is a local brain age suppression programme 26, and, for use in some embodiments, a local brain age fusion programme 28. Input medical imagery of a subject brain 20 is further stored on the storage medium 52. The input medical imagery 20 is processed by the various programmes previously mentioned, to obtain a set of suppressed local brain age data 60, being local brain age data for all of those regions of the brain in which abnormalities are not found, and also a single abnormality adjusted brain age data 56. The abnormality adjusted brain age data 56 is obtained when the multiple sets of suppressed local brain age data are fused together by the local brain age fusion programme 28.

[0034] Figure 5 illustrates the operation of the system in more detail. Firstly, at step 5.2 the computer platform receives or obtains brain imagery 20 of the subject. This is then stored on the non-volatile storage medium 52, as described previously. Two processes are then performed effectively in parallel, although it will be understood that they can also be performed in series one after the other, in either order, by the general purpose computer. The first process is that of step 5.4, where in the whole brain abnormality classifier 22 is applied to the input brain imagery of the subject to identify brain regions with abnormalities. As described previously, the output of the abnormality classifier is a set of brain location data, which may be defined by reference to the voxels of the images, in which an abnormality has been found. As described previously, the abnormality classifier operates in accordance with the method of the prior art, as described previously with respect to Figure 3.

[0035] In parallel with step 5.4 procedurally, at step 5.6 the local brain age determiner programme 24 first of all segments the brain imagery into local volume blocks, wherein each local volume block is the base input unit for a local brain age determination estimation. That is, a local brain age estimation is made for each local volume block. Once the brain imagery has been segmented into local volume blocks, then each local volume block has a local brain age determined for it, at step 5.8. As described previously, the precise mode of determination of local brain age for a local volume block is performed in the same manner as the prior art document Popescu et al described previously. The various respective local brain age estimations for each local volume block are stored by the local brain age determination programme on the non-volatile storage medium 52.

[0036] Next, at step 5.10, the local brain age suppression programme filters those local volumes in which abnormalities are included to exclude those local volumes from being output. That is, the local brain age estimations that have been calculated for the local volumes that include abnormalities are not output and are instead suppressed. At step 5.12 the local volume brain age estimations for those regions with no abnormalities are then output, for example by being displayed on display 44.

[0037] In some embodiments output in the local volume brain age estimations for those regions with no abnormalities is sufficient. However, in other embodiments the local volume brain age estimations for the regions with no abnormalities can be combined together to give a fusion brain age measure, at step 5.14. This fusion brain age measure is then output at step 5.16, for example by being displayed on the display 44. As described previously, there are various different statistical mechanisms by which the plurality of local volume brain ages can be combined to give a single fusion brain age measure.

[0038] The above described embodiments of the present disclosure provide numerous advantages, as explained further below.

[0039] 1) The embodiments provide a robust screening tool for neurodegenerative disease in routine clinical workflow on all brain MRIs. In contrast, other brain age models are built on selected imaging cohorts which have been obtained in research studies and the models are not suitable for routine clinical imaging. Note: it is very challenging for radiologists to determine brain age and therefore typically this would not be commented on in a report; the exception is if there is very marked atrophy locally or globally. Our local data shows 70% of brain MRI scans at KCH are abnormal and 50% at GSTT are abnormal implying our model which handles abnormalities isare required for a screening programme.

[0040] 2) The benefit of detecting neurodegenerative disease early on is to allow social planning and also to allow drug trial cohort identification in routine clinical imaging. In contrast, other brain age models are built on selected imaging cohorts which have been obtained in research studies and the models are not suitable for routine clinical imaging. The former allows the patient and family to plan their future including financial affairs and power of attorney based on their prognosis. The latter has become increasingly important given progress in the treatment of neurodegenerative diseases such as Alzheimer's Disease (1). The model could be used specifically to identify potential candidates for drug trial cohort identification in a stratified and uniform way. Our local data shows 70% of brain MRI scans at KCH are abnormal and 50% at GSTT are abnormal implying our model which handles abnormalities is required for a drug trial cohort identification programme. The model could also be used as a monitoring biomarker to follow up drug efficacy. It is noted that head imaging software used for patient stratification in clinical stroke studies, where evidence shows stroke treatment benefit, have become essential for the stroke patient pathway (2- 4) with considerable commercial success for the only software vendor associated with this high-level evidence (5).

[0041] 3) Patients with certain diseases where there are always gross structural abnormalities would benefit from longitudinal brain health assessment e.g., high-grade brain tumours undergoing radiotherapy. These patients have already had a craniotomy and a maximal resection of the tumour before starting radiotherapy which can reduce brain volume. The model proposed here has the potential to detect cancer therapy-related changes in the normal brain and correlate with long-term cognitive decline or quality-of-life deterioration

[0042] (6). Personalised treatment regimens could be established to mitigate against such therapy-related changes using model-derived biomarkers. It is noted that head imaging software used for patient stratification in clinical stroke studies, where evidence shows stroke treatment benefit, have become essential for the stroke patient pathway (2-4) with considerable commercial success for the only software vendor associated with this high- level evidence (5).

[0043] 4) Since 2012 there has been a 4.5% increase in brain MRIs per year in the UK. The recruitment of radiologists has not matched this increase in MRIs and there is an everwidening gap of supply and reporting demand (only 2% of NHS departments reaching their targets). A similar pattern is seen in many developed countries. A triage tool for normal and abnormal brain scans has become increasingly important to allow imaging departments to focus on reporting abnormal scans first with downstream clinical benefits

[0044] (7). Because a bonus advantage of the innovation is that it is also a triage tool, hospitals may choose to use the software for triage purposes and not for brain age purposes. This would in turn increase the exposure of the brain age product to the healthcare market such that it likely becomes a known screening tool for neurodegenerative disease, and also a well-placed product to trial the future proliferation of drugs targeted to defeat neurodegenerative disease. There would likely be a commercial trickle-down effect for future users.

[0045] 5) The embodiments permit robust real-time adaptive sequence adaptation (8). A dementia protocol could be performed where sequences such as volumetric Tl-w sequences would be implemented whilst the patient is in the scanner. This allows diagnostic-level scans which might provide additional information to the routine clinical data obtained during the "screening" scan. This reduces the need for another arranged scan which improves overall patient workflow by reducing administration time to organise a second scan, and reducing patient time to travel, to change into a hospital gown, and to climb on and off the scanner gantry.

[0046] 6) Whilst described above with respect to MR scans, the present embodiments could be implemented in other cross-sectional imaging such as head CT with all the benefits described here.

[0047] In addition, within the above we use the term local blocks, local volume blocks, and local volumes interchangeably to refer to the local volume blocks for which a separate brain age calculation is made.

[0048] Various modifications whether by addition, subtraction or substitution of features can be made to the above described embodiments to provide further embodiments, any and all of which are intended to encompassed by the appended claims.

[0049] References:

[0050] 1) van Dyck CH, Swanson CJ, Aisen P, Bateman RJ, Chen C, Gee M, Kanekiyo M, Li D, Reyderman L, Cohen S, Froelich L, Katayama S, Sabbagh M, Vellas B, Watson D, Dhadda S, Irizarry M, Kramer LD, Iwatsubo T. Lecanemab in Early Alzheimer's Disease. N Engl J Med. 2022 Nov 29. doi: 10.1056 / NEJMoa2212948.

[0051] 2) Albers GW, Marks MP, Kemp S, Christensen S, Tsai JP, Ortega-Gutierrez S, McTaggart RA, Torbey MT, Kirn-Tenser M, Leslie-Mazwi T, Sarraj A, Kasner SE, Ansari SA, Yeatts SD, Hamilton S, Mlynash M, Heit JJ, Zaharchuk G, Kim S, Carrozzella J, Palesch YY, Demchuk AM, Bammer R, Lavori PW, Broderick JP, Lansberg MG; DEFUSE 3 Investigators. Thrombectomy for Stroke at 6 to 16 Hours with Selection by Perfusion Imaging. N Engl J Med. 2018 Feb 22;378(8):708-718. doi: 10.1056 / NEJMoal713973. 3) Campbell BC, Mitchell PJ, Kleinig TJ, Dewey HM, Churilov L, Yassi N, Yan B, Dowling RJ, Parsons MW, Oxley TJ, Wu TY, Brooks M, Simpson MA, Miteff F, Levi CR, Krause M, Harrington TJ, Faulder KC, Steinfort BS, Priglinger M, Ang T, Scroop R, Barber PA, McGuinness B, Wijeratne T, Phan TG, Chong W, Chandra RV, Bladin CF, Badve M, Rice H, de Villiers L, Ma H, Desmond PM, Donnan GA, Davis SM; EXTEND-IA Investigators. Endovascular therapy for ischemic stroke with perfusion-imaging selection. N Engl J Med. 2015 Mar 12;372(11) : 1009-18. doi: 10.1056 / NEJMoal414792.

[0052] 4) Nogueira RG, Jadhav AP, Haussen DC, Bonafe A, Budzik RF, Bhuva P, Yavagal DR, Ribo M, Cognard C, Hanel RA, Sila CA, Hassan AE, Millan M, Levy El, Mitchell P, Chen M, English

[0053] JD, Shah QA, Silver FL, Pereira VM, Mehta BP, Baxter BW, Abraham MG, Cardona P, Veznedaroglu E, Hellinger FR, Feng L, Kirmani JF, Lopes DK, Jankowitz BT, Frankel MR, Costalat V, Vora NA, Yoo AJ, Malik AM, Furlan AJ, Rubiera M, Aghaebrahim A, Olivot JM, Tekle WG, Shields R, Graves T, Lewis RJ, Smith WS, Liebeskind DS, Saver JL, Jovin TG; DAWN Trial Investigators. Thrombectomy 6 to 24 Hours after Stroke with a Mismatch between Deficit and Infarct. N Engl J Med. 2018 Jan 4;378(1) : 11-21. doi:

[0054] IO.1056 / NEJMoal706442.

[0055] 5) CT Perfusion Image Analysis Software for Stroke | RapidAI rapidai.com

[0056] 6) Petr J, Hogeboom L, Nikulin P, Wiegers E, Schroyen G, Kallehauge J, Chmelik M, Clement P, Nechifor RE, Fodor LA, De Witt Hamer PC, Barkhof F, Pernet C, Lequin M, Deprez S, Jancalek R, Mutsaerts HJMM, Pizzini FB, Emblem KE, Keil VC. A systematic review on the use of quantitative imaging to detect cancer therapy adverse effects in normal-appearing brain tissue. MAGMA. 2022 Feb;35(l) : 163-186. doi : 10.1007 / S10334- 021-00985-2.

[0057] 7) Wood DA, Kafiabadi S, Busaidi AA, Guilhem E, Montvila A, Lynch J, Townend M, Agarwal S, Mazumder A, Barker GJ, Ourselin S, Cole JH, Booth TC. Deep learning models for triaging hospital head MRI examinations. Med Image Anal. 2022 May;78: 102391. doi: 10.1016 / j. media.2022.102391.

[0058] 8) Cole JH, Lorenz R, Geranmayeh F, et al. Active Acquisition for multimodal neuroimaging. Wellcome Open Research. 2018 ;3: 145. DOI: 10.12688 / wellcomeopenres.14918.2.

[0059] 9) Wood, D., Cole, J., & Booth, T. (2019). NEURO-DRAM : a 3D recurrent visual attention model for interpretable neuroimaging classification. arXiv preprint arXiv: 1910.04721.

Claims

Claims1. A computer implemented method of determining brain age of brains with abnormalities using medical imaging, comprising: receiving medical imagery of a subject brain; detecting abnormal areas of the subject brain corresponding to disease or injury within the medical imagery; segmenting the medical imagery of the subject brain into a plurality of volumes; determining a separate brain age for at least a subset of the plurality of volumes; suppressing the brain age determination for a particular volume of the subject brain if the particular volume at least partially comprises detected abnormal areas; and providing for output the separate brain age determinations for those volumes for which the brain age determination has not been suppressed.

2. A method according to claim 1, wherein the brain age determination for a particular volume of the subject is suppressed if it comprises at least 90% of abnormal areas.

3. A method according to claim 1 or 2, wherein the brain age determination for a particular volume of the subject is suppressed if it comprises at least 50% of abnormal areas.

4. A method according to claims 1, 2, or 3, wherein the brain age determination for a particular volume of the subject is suppressed if it comprises at least 30% of abnormal areas.

5. A method according to any of claims 1 to 4, wherein the brain age determination for a particular volume of the subject is suppressed if it comprises at least 1 voxel of abnormal areas.

6. A method according to any of the preceding claims, and further comprising combining the separate brain age determinations for each volume that is not suppressed to give a unified brain age result; and providing the unified brain age result for output.

7. A method according to claim 6, wherein the combining comprises processing the separate brain age determinations into the unified brain age result using a statistical calculation.

8. A method according to claim 7, wherein the statistical calculation is one selected from the group comprising: i) an averaging calculation ; or ii) a clustering calculation.

9. A method according to any of the preceding claims, wherein the medical imaging comprises medical imaging obtained by any one of the following medical imaging modalities: i) Magnetic Resonance (MR) imaging; ii) Computerised Tomography (CT) imaging; and / or iii) ultrasound imaging.

10. A method according to any of the preceding claims, and further comprising outputting the separate non-suppressed brain age determinations, or the unified brain age result, as appropriate, by displaying the determinations or result on a display screen.

11. A computer program or suite of computer programs so arranged such that when executed by a computer system they cause the computer system to perform the method of any of the preceding claims.

12. A computer readable storage medium storing a computer program or at least one of the suite of computer programs according to claim 11.

13. A computer system, comprising: a processor; and a computer readable storage medium storing one or more computer readable programs which when executed by the processor cause the computer system to: receive medical imagery of a subject brain; detect abnormal areas of the subject brain corresponding to disease or injury within the medical imagery; segment the medical imagery of the subject brain into a plurality of volumes; determine a separate brain age for at least a subset of the plurality of volumes volume; suppress the brain age determination for a particular volume of the subject brain if the particular volume at least partially comprises detected abnormal areas; andprovide for output the separate brain age determinations for those volumes for which the brain age determination has not been suppressed.

14. A system according to claim 13, wherein the brain age determination for a particular volume of the subject is suppressed if it comprises at least 1 voxel of abnormal areas.

15. A system according to any of claims 13 or 14, wherein the computer readable storage medium further stores one or more computer readable programs which when executed by the processor cause the computer system to combine the separate brain age determinations for each volume that is not suppressed to give a unified brain age result; and provide the unified brain age result for output.

16. A system according to claim 15, wherein the combining comprises processing the separate brain age determinations into the unified brain age result using a statistical calculation.

Images