×
  • Select the area you would like to search.
  • ACTIVE INVESTIGATIONS Search for current projects using the investigator's name, institution, or keywords.
  • EXPERTS KNOWLEDGE BASE Enter keywords to search a list of questions and answers received and processed by the ADNI team.
  • ADNI PDFS Search any ADNI publication pdf by author, keyword, or PMID. Use an asterisk only to view all pdfs.
Principal Investigator  
Principal Investigator's Name: Elizabeth Fitzgerald
Institution: Stanford University
Department: Computer Science
Country:
Proposed Analysis: For our final project, we want to create a model that can accurately predict not only whether a patient has Alzheimer's, but also the severity of their condition. Specifically, we want to create a binary neural network classifier that determines if a patient has Alzheimer’s, as well as a multi-class neural network classifier that is able to categorize the severity of Alzheimer’s in an afflicted patient, based only on cross-sectional MRI images of the patient’s brain. The very first hurdle to overcome is the difficulty in retrieving data. Alzheimer’s data is well-protected and frequently requires applications to gain access. We believe we are close to overcoming this, as we’ve submitted access applications to ADNI’s repositories as well as found small open-use MRI datasets. The second largest challenge will be utilizing all the datasets that we hope to have acquired for the same goal of predicting the presence and severity of Alzheimer’s. On the topic of data, we intend on using a set of MRI images from a 2019 Kaggle posting (https://www.kaggle.com/tourist55/alzheimers-dataset-4-class-of-images). Each of these images falls into one of four categories: Mild Demented, Moderate Demented, Non Demented, and Very Mild Demented. We will supplement this with MRI data from OASIS1 (https://www.oasis-brains.org/#data), which also has four class categorizations based on a Clinical Dementia Rating. For methodology, we will use a standard CNN architecture (probably extending ResNet) as a baseline model. We hope to read more literature pertaining to the advantages of more advanced CNN architectures (specifically related to disease detection from MRIs) in order to improve this baseline. We also believe we can improve the baseline by incorporating all data into a single model. Our primary idea is to use transfer learning by first creating a model that will produce a binary prediction of whether a patient has Alzheimer’s. Once we have done so, we will extend the predictions to include the severity of the condition. In this way, we hope that the initial binary classification can help increase the performance of the multi-class classification. Based on previous literature, we will also examine the benefit of preprocessing the images into specific numerical features for use by the model (instead of the images directly). To gain a general understanding of the disease, we will rely upon scientific papers describing potential causes of Alzheimer’s. Beyond that, we will attempt to find research that also tackles this same issue. For example, “Applying Deep Learning to Predicting Dementia and Mild Cognitive Impairment” (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7256597/) provides general context for the potential type of model architecture that we can use. However, a more interesting read is “Classification and Visualization of Alzheimer’s Disease using Volumetric Convolutional Neural Network and Transfer Learning” (https://www.nature.com/articles/s41598-019-54548-6). This project is very similar to our own, and we hope to incorporate some of their modelling ideas to our own. At this point, we believe the best way to compare performances and evaluate the results is to use confusion matrices for each of the four predicted classes, as well as a general F1 score comparison when evaluating the binary classification model.
Additional Investigators  
Investigator's Name: Constantine Athanitis
Proposed Analysis: For our final project, we want to create a model that can accurately predict not only whether a patient has Alzheimer's, but also the severity of their condition. Specifically, we want to create a binary neural network classifier that determines if a patient has Alzheimer’s, as well as a multi-class neural network classifier that is able to categorize the severity of Alzheimer’s in an afflicted patient, based only on cross-sectional MRI images of the patient’s brain. The very first hurdle to overcome is the difficulty in retrieving data. Alzheimer’s data is well-protected and frequently requires applications to gain access. We believe we are close to overcoming this, as we’ve submitted access applications to ADNI’s repositories as well as found small open-use MRI datasets. The second largest challenge will be utilizing all the datasets that we hope to have acquired for the same goal of predicting the presence and severity of Alzheimer’s. On the topic of data, we intend on using a set of MRI images from a 2019 Kaggle posting (https://www.kaggle.com/tourist55/alzheimers-dataset-4-class-of-images). Each of these images falls into one of four categories: Mild Demented, Moderate Demented, Non Demented, and Very Mild Demented. We will supplement this with MRI data from OASIS1 (https://www.oasis-brains.org/#data), which also has four class categorizations based on a Clinical Dementia Rating. For methodology, we will use a standard CNN architecture (probably extending ResNet) as a baseline model. We hope to read more literature pertaining to the advantages of more advanced CNN architectures (specifically related to disease detection from MRIs) in order to improve this baseline. We also believe we can improve the baseline by incorporating all data into a single model. Our primary idea is to use transfer learning by first creating a model that will produce a binary prediction of whether a patient has Alzheimer’s. Once we have done so, we will extend the predictions to include the severity of the condition. In this way, we hope that the initial binary classification can help increase the performance of the multi-class classification. Based on previous literature, we will also examine the benefit of preprocessing the images into specific numerical features for use by the model (instead of the images directly). To gain a general understanding of the disease, we will rely upon scientific papers describing potential causes of Alzheimer’s. Beyond that, we will attempt to find research that also tackles this same issue. For example, “Applying Deep Learning to Predicting Dementia and Mild Cognitive Impairment” (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7256597/) provides general context for the potential type of model architecture that we can use. However, a more interesting read is “Classification and Visualization of Alzheimer’s Disease using Volumetric Convolutional Neural Network and Transfer Learning” (https://www.nature.com/articles/s41598-019-54548-6). This project is very similar to our own, and we hope to incorporate some of their modelling ideas to our own. At this point, we believe the best way to compare performances and evaluate the results is to use confusion matrices for each of the four predicted classes, as well as a general F1 score comparison when evaluating the binary classification model.