Search

EP-4577978-B1 - SEGMENTATION AND DETECTION OF AMYLOID-RELATED IMAGING ABNORMALITIES (ARIA) IN ALZHEIMER'S PATIENTS

EP4577978B1EP 4577978 B1EP4577978 B1EP 4577978B1EP-4577978-B1

Inventors

  • KLEIN, GREGORY
  • KRISHNAN, ANITHA PRIYA
  • SONG, Zhuang
  • CARANO, Richard Alan Duray

Dates

Publication Date
20260513
Application Date
20230824

Claims (16)

  1. A method for detecting amyloid related imaging abnormalities, ARIA, in a brain of a patient, comprising, by one or more computing devices: accessing (308) a set of one or more brain-scan images (401, 501, 601, 801) associated with the patient; inputting (310) the set of one or more brain-scan images into one or more machine-learning models (500, 600) trained to: generate a segmentation map (403, 503, 603) based on the set of one or more brain-scan images, the segmentation map including a plurality of pixel-wise class labels corresponding to a plurality of pixels in the segmentation map; and generate a classification score (510, 610, 810) based on the segmentation map; and detecting (312) ARIA in the brain of the patient based on the classification score.
  2. The method of Claim 1, wherein at least one of the plurality of pixel-wise class labels comprises an indication of one or more ARIA lesions, and wherein the one or more machine-learning models is further trained to generate the classification score based on the at least one of the plurality of pixel-wise class labels; and/or wherein the one or more machine-learning models comprises a segmentation model (400, 506) and a classification model (508, 608, 800A), optionally wherein the segmentation model comprises: (i) an encoder (402, 502, 602) trained to generate a plurality of down-sampled feature maps based on the set of one or more brain-scan images, and wherein the classification model comprises a decoder (404, 504) trained to receive the plurality of down-sampled feature maps from the encoder; and/or (ii) a bidirectional feature propagation network.
  3. The method of Claim 2, wherein the bidirectional feature propagation network comprises a top-down feature pyramid network FPN (605), and a bottom-up FPN (606), optionally wherein the classification model (608) comprises a decoder trained to receive input data from a plurality of layers of the bottom-up FPN.
  4. The method of Claim 2 or Claim 3, wherein: (i) the classification model is trained by: training the segmentation model; and after training the segmentation model (500, 600), training the classification model (508, 608) based on the trained segmentation model; and/or (ii) the classification model comprises an attention mechanism, optionally wherein: (a) the attention mechanism is configured to pay attention to one or more of the plurality of pixel-wise class labels included in the segmentation map (405, 503, 603); and/or (b) the plurality of pixel-wise class labels comprises a first plurality of pixel-wise class labels, and wherein the attention mechanism is based on a second plurality of pixel-wise class labels included in the segmentation map, the second plurality of pixel-wise class labels being indicative of dilated gray matter; and/or (c) the attention mechanism is based on a plurality of subtraction masks.
  5. The method of any preceding claim, wherein the patient is an Alzheimer's disease patient having been treated with an anti-amyloid-beta, anti-Aβ, antibody, optionally wherein: (i) the method further comprises, in response to detecting the ARIA in the brain of the patient determining a dosage adjustment of the anti-Aβ antibody and/or determining that use of the anti-Aβ antibody in the patient should be terminated or temporarily suspended; and/or (ii) the anti-Aβ antibody is selected from the group consisting of bapineuzumab, solanezumab, aducanumab, gantenerumab, crenezumab, donanemab, and lecanemab.
  6. The method of any preceding claim, further comprising: in response to detecting the ARIA in the brain of the patient, determining one or more anti-ARIA treatments for the patient, optionally wherein the one or more anti-ARIA treatments comprise one or more anti-ARIA antibodies.
  7. The method of any preceding claim, wherein the set of one or more brain-scan images (401, 501, 601, 801) comprises: (i) one or more magnetic resonance imaging images, one or more positron emission tomography images, one or more single-photon emission computed tomography images, one or more amyloid positron emission tomography images, or any combination thereof; and/or (ii) one or more fluid-attenuated inversion recovery images, one or more T2*-weighted imaging images, one or more T1-weighted imaging images, or any combination thereof; and/or (iii) a plurality of volumes corresponding to one or more cross-sectional volumes of the brain of the patient.
  8. A method for training a plurality of machine-learning models for detecting amyloid related imaging abnormalities, ARIA, in brains of patients, comprising, by one or more computing devices: accessing (702) a set of brain-scan images (401) associated with one or more patients; training (704) a first machine-learning model (400) of the plurality of machine-learning models, the first machine-learning model being trained to segment one or more ARIA lesions based on the set of brain-scan images; obtaining (706) a first set of weights associated with the trained first machine-learning model; initializing (708) a second set of weights to correspond to the first set of weights, the second set of weights associated with a second machine-learning model (800A) of the plurality of machine-learning models; and training (710) the second machine-learning model to generate a classification score based at least in part on the second set of weights, the classification score corresponding to a detection of ARIA or a severity of ARIA in brains of the one or more patients.
  9. The method of Claim 8, wherein: (i) the first machine-learning model comprises a segmentation model (400), optionally wherein the segmentation model (400) comprises an encoder (402) and a decoder (404); and/or (ii) the second machine-learning model comprises a classification model (800A).
  10. The method of Claim 9, wherein the encoder (402) comprises a harmonic dense neural network, HarDNet, encoder, and/or wherein the decoder (404) comprises a U-Net decoder, and/or wherein the classification model (800A) comprises an attention mechanism configured to pay attention to one or more features of the set of brain-scan images identified based on the segmented one or more ARIA lesions, and/or wherein the classification model (800A) comprises one or more convolutional neural networks.
  11. The method of any of Claims 8 to 10, further comprising: prior to training the second machine-learning model to generate the classification score, inputting the segmented one or more ARIA lesions into the second machine-learning model.
  12. The method of any of Claims 8 to 11, wherein: (i) training (710) the second machine-learning model comprises optimizing the second set of weights along with the classification score; and/or (ii) the first machine-learning model is associated with the second machine-learning model, optionally wherein training (704, 710) the first machine-learning model and the second machine-learning model comprises training the first machine-learning model and the second machine-learning model in accordance with a transfer learning process; and/or (iii) training (710) the second machine-learning model to generate the classification score comprises training the second machine-learning model based on the second set of weights and the segmented one or more ARIA lesions.
  13. The method of any preceding claim, wherein the classification score (510, 610, 810) comprises: (i) a binary value indicative of an absence of ARIA or a presence of ARIA; and/or (ii) a numerical value indicative of a severity of ARIA; and/or (iii) one of a plurality of classification scores, and wherein the plurality of classification scores comprises: a first classification score indicative of mild ARIA; a second classification score indicative of moderate ARIA; and a third classification score indicative of severe ARIA; and/or (iv) a Barkhof Grand Total Score (BGTS) score.
  14. The method of any preceding claim, wherein the ARIA is associated with microhemorrhages and hemosiderin deposits, ARIA-H, in the brain of the patient; and/or the ARIA is associated with parenchymal edema or sulcal effusion, ARIA-E, in the brain of the patient.
  15. A system including one or more computing devices (1300), comprising: one or more non-transitory computer-readable storage media (1304) including instructions; and one or more processors (1302) coupled to the one or more storage media, the one or more processors configured to execute the instructions to implement the method of any preceding claim.
  16. A non-transitory computer-readable medium (1304) comprising instructions that, when executed by one or more processors (1302) of one or more computing devices (1300), cause the one or more processors to implement the method of any of claims 1 to 14.

Description

TECHNICAL FIELD The present disclosure relates generally to amyloid-related imaging abnormalities (ARIA), and, more specifically, to segmenting and detecting ARIA in Alzheimer's disease (AD) patients. BACKGROUND Alzheimer's disease (AD) is a progressive neurodegenerative disease that may be characterized by a decline in patient memory, speech, and cognitive skills, as well as by adverse changes in patient mood and behavior. AD may generally result from one or more identified biological changes that may occur in the brain of the patient over many years. For example, leading biological markers (e.g., biomarkers) or hallmarks of AD may include the excessive accumulation of amyloid-beta (Aβ) plaques and tau tangles within the brain of the patient. Specifically, while Aβ proteins and tau proteins may be produced generally as part of the normative functioning of the brain, in patients diagnosed with AD, one may observe either an excessive production of Aβ proteins that may accumulate as plaques around the brain cells or an excessive production of tau proteins that may become misfolded and accumulate as tangles within the brain cells. For example, the Aβ plaques or tau tangles may be typically observed in a patient's brain by performing one or more magnetic resonance imaging (MRI) scans, positron-emission tomography (PET) scans, or computed tomography (CT) scans of the patient's brain, and then these scans may be utilized by clinicians to diagnose patients as having AD. In certain instances, for patients diagnosed with AD, when excessive accumulation of Aβ plaques is the basis for the diagnosis (e.g., as opposed to the accumulation of tau tangles), clinicians may treat the AD patient utilizing an anti-amyloid-beta (anti-Aβ) antibody or other similar anti-Aβ immunotherapy. For example, the anti-Aβ antibody may include one or more anti-Aβ monoclonal antibodies (mAbs) that may be suitable for removing or reducing Aβ plaques in the brain of the AD patient by binding to and counteracting the Aβ plaques. While such anti-Aβ antibody treatments have been found to be effective for treating AD patients, in a small number of instances, an AD patient may be susceptible to certain side effects from the anti-Aβ antibody treatments that may manifest as amyloid-related imaging abnormalities (ARIA) in subsequent scans (e.g., MRI scans, PET scans) of the brain of the AD patient. For example, ARIA may include ARIA-E, which includes parenchymal or sulcal hyperintensities on certain MRI scans (e.g., fluid-attenuated inversion recovery (FLAIR) imaging) indicative of parenchymal edema or sulcal effusions. ARIA may further include ARIA-H, which includes hypointense regions on other particular MRI scans (e.g., gradient recalled-echo imaging, T2*-weighted imaging (T2*WI)) indicative of hemosiderin deposition. It may be thus useful to detect ARIA as early as possible, such that the anti-Aβ antibody treatments may be adjusted and/or temporarily suspended in such instances in which an AD patient shows signs of ARIA. Accordingly, it may be useful to provide techniques for analyzing brain scans to detect and quantify ARIA, which may manifest as contextual changes and/or changes in signal intensities in the brain scans. WO 2022/054711 A1 describes a computer program which makes it possible to provide the same brain diagnosis information as does a PET image without performing a PET exam. KR 2020 0143023 A describes a method for predicting Alzheimer's disease through quantification of brain shape features. WO 2018/023036 A1 describes a method of treating or preventing ARIA in patients receiving one or more course of medical treatment for Alzheimer's Disease including administering a SUR1-TRPM4 channel inhibitor. Pemberton et al. ("Quantification of amyloid PET for future clinical use: a state-of-the-art review", European Journal of Nuclear Medicine and Molecular Imaging, Springer, Berlin/Heidelberg, vol. 49, no.10, 7 April 2022, pages 3508-3528) summarises and discusses several tools and measures available for amyloid PET quantification. Cecchin et al. ("A new integrated dual time-point amyloid PET/MRI data analysis method", European Journal of Nuclear Medicine and Molecular Imaging, Springer, Berlin/Heidelberg, vol. 44, no. 12, 4 July 2017, pages 2060-2072) describes the testing of an integrated dual time-point amyloid PET/MRI data analysis method. SUMMARY The claimed invention is defined by the appended claims. Embodiments of the present disclosure are directed to one or more computing devices, methods, and non-transitory computer-readable media that may utilize one or more machine-learning models (e.g., one or more semantic image segmentation and classification models) for analyzing medical images (e.g., brain-scan images) to segment and detect amyloid-related imaging abnormalities (ARIA) in Alzheimer's disease (AD) patients. A method for detecting ARIA in a brain of a patient is provided. One or more computing devices access a set of one or mo