Classifying neurological disease status using deep learning
Inventors
Feng, Xinyang • PROVENZANO, Frank • Small, Scott A.
Assignees
Columbia University in the City of New York
Publication Number
US-12400321-B2
Publication Date
2025-08-26
Expiration Date
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
A method for classifying neurological disease status is described. The method includes acquiring, by a data preprocessor logic, patient image data. The method further includes generating, by a trained artificial neural network (ANN), a classification output based, at least in part, on the patient image data. The classification output corresponds to a neurological disease status of the patient. The trained ANN is trained based, at least in part, on longitudinal source data.
Core Innovation
A method is provided for classifying neurological disease status that includes acquiring, by a data preprocessor logic, patient image data, and generating, by a trained artificial neural network (ANN), a classification output based, at least in part, on the patient image data, the classification output corresponding to a neurological disease status of the patient. The trained ANN is trained based, at least in part, on longitudinal source data. The method optionally includes preprocessing the patient image data to yield input image data, and generating the classification output based, at least in part, on the input image data.
The disclosure further describes identifying, by localization logic, a most predictive image region based, at least in part, on an ANN parameter associated with the trained ANN, and partitioning longitudinal source data that includes a plurality of source image data sets from a same selected patient into training, validation and test data with the partitioning occurring at a level of the same selected patient. The ANN can be selected from CNN, VGGNet, ResNet, and DenseNet and may be implemented as a three-dimensional convolutional neural network (3D CNN) having N CNN stages coupled in series, with each stage including convolutional layers, a batch normalization layer, an activation layer, and a pooling layer, and followed by a flattening layer, a fully connected layer and a sigmoid activation function layer.
The problem addressed is early and accurate classification and progression prediction of neurological diseases (including Alzheimer's disease and prodromal stages such as mild cognitive impairment) where early detection may improve treatment impact; challenges include the degenerative nature of diseases, the need for techniques for early diagnosis, the use of widely available structural MRI modalities, and technical challenges in learning from three dimensional scans such as the risk of overfitting given typically one label per scan and interpretability issues commonly described as the ANN "black box."
Claims Coverage
This section identifies four independent claims (claims 1, 12, 21, and 22) and extracts sixteen main inventive features across those claims.
Acquiring patient image data
acquiring, by a data preprocessor logic configuration, patient image data;
Enriching patient image data with longitudinal source data
electronically preventing overfitting by an artificial neural network (ANN) by enriching the patient image data through an inclusion of a plurality of transformations of the patient image data, the plurality of transformations including longitudinal source data comprising a time series of image data for a patient;
Training the ANN with supervised comparison
training the ANN on the enriched patient image data, the training comprising: inputting a training data set from the enriched patient image data; generating an output score based on the training data set; comparing the output score to a corresponding cognitive status indicator; and upon a determination that a loss function for the output score does not meet a target threshold, adjusting at least one ANN parameter based on the comparison;
Generating classification output from trained ANN
generating, by the trained ANN, a classification output based, at least in part, on testing data from the patient image data, the classification output corresponding to a neurological disease status of the patient.
Preprocessing to yield input image data
preprocessing, by the data preprocessor logic configuration, the patient image data to yield input image data, the classification output generated based, at least in part, on the input image data.
Localization to identify most predictive region
identifying, by a localization logic configuration, a most predictive image region based, at least in part, on an ANN parameter associated with the trained ANN.
Partitioning longitudinal source data at patient level
the longitudinal source data comprises a plurality of source image data sets from same selected patient, wherein the longitudinal source data is partitioned into training data, validation data and test data, and wherein the partitioning occurs at a level of the same selected patient.
Supported imaging modalities
the image data comprises at least one of magnetic resonance imaging (MRI) image data, MRI T1-weighted image data, MRI T2-weighted image data, computed tomography (CT) image data, cerebral blood volume (CBV) image data, cerebral blood flow (CBF) image data, mean transit time (MTT) image data, positron emission tomography (PET) image data, or single-photon emission computerized tomography (SPECT) image data.
Covered neurological disease types
the neurological comprises neurodegenerative diseases and non-neurodegenerative diseases, wherein each neurodegenerative disease comprises at least one of Alzheimer's disease, amyotrophic lateral sclerosis (ALS), frontotemporal dementia (FTD) or Parkinson's disease, and wherein each of the non-neurodegenerative diseases comprises cerebrovascular disease, epilepsy, or stroke.
Classification outputs for Alzheimer's disease
the neurological disease corresponds to Alzheimer's disease and the classification output is comprises at least one of Alzheimer's disease (AD), prodromal AD, mild cognitive impairment (MCI) or cognitively normal (CN).
ANN model families
the ANN is comprises at least one of CNN (convolutional neural network), VGGNet (Visual Geometry Group neural network), ResNet (residual network), or DenseNet (densely connected convolutional networks).
3D CNN architectural structure
the ANN is a three-dimensional convolutional neural network (3D CNN), wherein the 3D CNN comprises a number, N, CNN stages coupled in series, wherein each of the CNN stages comprises a first 3D convolutional layer, a second 3D convolutional layer, a batch normalization layer, an activation layer, and a pooling layer, coupled in series, and wherein the 3D CNN further comprises a flattening layer, a fully connected layer and a sigmoid activation function layer.
Identification of most predictive region by trained ANN
identifying, by the trained ANN, a most predictive region in the patient image data, wherein the generated classification output is based on the identified most predictive region.
Three-dimensional volumetric input
the patient image data comprise 3-dimensional volumetric digital image data.
System data preprocessor and training system
a data preprocessor logic configuration configured to acquire patient image data; and a training system configured to: electronically preventing overfitting by an artificial neural network (ANN) by enriching the patient image data through an inclusion of a plurality of transformations of the patient image data, the plurality of transformations including longitudinal source data comprising a time series of image data for a patient; train the ANN on the enriched patient image data, the training comprising: inputting a training data set from the enriched patient image data; generating an output score based on the training data set; comparing the output score to a corresponding cognitive status indicator; and upon a determination that a loss function for the output score does not meet a target threshold, adjusting at least one ANN parameter based on the comparison; and generate by the trained ANN, a classification output based, at least in part, on testing data from the patient image data, the classification output corresponding to a neurological disease status of the patient.
Computer readable storage device with instructions
executable instructions that, when executed by one or more processors, causes a computer processing arrangement to perform procedures comprising: acquiring, by a data preprocessor logic configuration, patient image data; electronically preventing overfitting by an artificial neural network (ANN) by enriching the patient image data through an inclusion of a plurality of transformations of the patient image data, the plurality of transformations including longitudinal source data comprising a time series of image data for a patient; training the ANN on the enriched patient image data, the training comprising: inputting a training data set from the enriched patient image data; generating an output score based on the training data set; comparing the output score to a corresponding cognitive status indicator; and upon a determination that a loss function for the output score does not meet a target threshold, adjusting at least one ANN parameter based on the comparison; and generating, by the trained ANN, a classification output based, at least in part, on testing data from the patient image data, the classification output corresponding to a neurological disease status of the patient.
The independent claims (1, 12, 21, 22) consistently claim acquisition and preprocessing of patient image data, enrichment of training data by inclusion of longitudinal source data to prevent overfitting, supervised ANN training with comparison to cognitive status indicators and parameter adjustment, generation of a classification output corresponding to neurological disease status, and support for specific ANN architectures and three-dimensional volumetric image data.
Stated Advantages
Including longitudinal data expands (i.e., augments) source data and provides a significant increase in the total amount of data available for training, validation and testing.
Partitioning longitudinal data at the level of individual patients is configured to avoid data leakage.
The trained ANN may be utilized to predict whether a selected patient is likely to develop a neurological or other disease of the brain and the result may then be utilized to support treatment planning.
Identification of a most predictive region may be utilized to facilitate understanding of the disease itself and to reduce 'black box' aspects associated with ANNs.
Systems and methods according to the disclosed subject matter demonstrate relatively high classification performance in Alzheimer's disease versus cognitive normal using structural MRI and relatively high accuracy in MCI progression prediction when applying the model trained on AD vs. CN to the MCI subgroup.
The classification and regional analyses methods provide a general framework that may be applied to other disorders and imaging modalities.
A method and/or system may utilize data with relatively less detail or information from a scanner, including data captured from a shorter scan with fewer slices or from an MRI scanner with relatively lower fidelity than that of conventional scanners.
Documented Applications
Diagnosing neurological disease status, including Alzheimer's disease, prodromal AD, mild cognitive impairment (MCI), and cognitively normal (CN).
Using a trained ANN to predict progression from MCI to Alzheimer's disease by applying an AD vs. CN classifier to the MCI subgroup.
Pinpointing most predictive brain regions (e.g., hippocampal formation) via class activation maps and ablation analyses to inform neuroanatomical understanding.
Providing a neurological disease diagnosis framework based on a deep ANN model using structural imaging techniques empowered with inclusion of longitudinal scans.
Applying the framework to other neurological disorders and to various imaging modalities including MRI, CT, PET, SPECT, CBV, CBF, and MTT image data.
Using whole brain volume CNNs as tools to aid in diagnosis of AD and prediction of progression to AD among MCI populations.
Generating individual class activation maps as individual neuroanatomical validity reports while preserving whole-brain prediction power.
Interested in licensing this patent?