55% of people with Alzheimer's die without ever being diagnosed as Alzheimer's patients.
The reason for this is the lack of precise tools, and the extent of time and effort it takes to diagnose Alzheimer's. According to one article, "There is no blood test or brain scan that can conclusively tell doctors that a patient does or does not have the condition...it’s still a diagnosis that doctors make based on reports of the patients’ changing intellectual abilities and on psychiatric tests that aren’t specific for Alzheimer’s".
The purpose of this project was to create an automated tool that could accurately diagnose Alzheimer's. The tool would use neural networks to train itself to diagnose Alzheimer's.
I extracted image features from MRI scans of patients, and combined these with clinical features to input into the neural networks. Then, I tested 3 different structures of classifiers, a single stage neural network, chained neural networks, and hierarchical neural networks. I varied the number of hidden neurons, and tried different training algorithms. I also tested the importance of the image features and the MMSE.
After building the ideal classifier, I built a GUI. A doctor would only have to upload an MRI and enter the patient's information to get a diagnosis (see video in Results).
My tool is an improvement upon current diagnosis methods, and deals with the problem of slow and inaccurate diagnoses of Alzheimer's.
My name is Anika Cheerla. I am a 13 year old attending John F. Kennedy Middle School.
I love volunteering. I am a curriculum developer and teacher at the non-profit MathAndCoding where we teach kids how to code. I also founded a program to raise money to buy learning-software for underprivileged kids. I enjoy playing piano at senior homes, fundraising online, and practicing water polo.
I think that science and engineering go hand in hand. Science gives me the understanding and engineering gives me the ability to create anything based on that understanding. I can change the world with some lines of code, and science answers my questions so I don't end up dazed.
My favorite scientist is Florence Sabin, the "First Lady of American Science". I've been inspired not only by her breakthroughs in brain research and the lymphatic system, but also by her establishment of women in science, and her spunk in changing health standards and implementing the 'Sabin Health Laws' to modernize public health.
I want to be a professor when I grow up, since I like teaching and doing research. Winning anything in the Google Science Fair would mean a lot to me, simply because I'd feel that my hard work paid off. Along with that, winning in the Google Science Fair means my project would gain recognition, making it's use in the medical field within the realm of possibility.
Question / Proposal
With this research, I wanted to answer two questions:
- Previous research on Alzheimer's diagnosis using MRI images reached around an 80% accuracy. Can I build a classifier that can achieve greater than 95% accuracy in diagnosing Alzheimer's with MRI images?
- Even though there is research on classifiers of Alzheimer's diseases, no tool was built to directly aid doctors in diagnosis. Doctor's rely on manual inspection of the MRI images and other clinical data for diagnosis. Can I build a tool using this classifier that enables doctors diagnose the disease from MRI images automatically?
Based on my research on classifier architectures, I hypothesized that multiple artificial neural network (ANN) classifiers connected in either chained or hierarchical configurations have the potential to achieve higher accuracy than other classifier architectures. I also hypothesized that feeding the classifiers with a combination of clinical features and the MRI image features would further improve the accuracy. ANN classifiers are prone to over fitting if there is small sample size of the training and testing data. I proposed to overcome this by doing exhaustive cross validation and also testing the final classifier using unknown images.
There are over 5 million people living with Alzheimer's currently. Alzheimer's is the 6th leading cause of death in America. 1 in 3 seniors die of Alzheimer's or another dementia.
The disease results in memory loss and difficulties with thinking, problem-solving or language. There are treatments that, when taken at an early stage, help patients cope with symptoms.
Alzheimer's occurs when proteins build up in the brain to form structures called 'plaques' and 'tangles'. This breaks the connections between nerve cells, leading to loss of nerve cells and brain tissue, resulting in poorer signaling within the brain and eventually, death.
- An early treatment can boost the chemical messengers in the brain, easing symptoms such as memory loss.
The early diagnosis of Alzheimer's helps loved ones and caregivers adjust to the oncoming disease, aids doctors in giving their patient the best treatment, and gives patients a more comfortable life.
Current Alzheimer's Diagnosis
Alzheimer's is diagnosed largely on a doctor's opinion. There is no set test that determines the condition of a patient, rather, it is up to the doctor to decide whether the patient has Alzheimer's. This means the diagnosis has a low accuracy. The method for diagnosing Alzheimer's takes a long time, with medical scans, patient-doctor meetings, multiple mental tests, blood tests, and so on. This discourages early diagnosis and takes away early treatment and care a patient can get.
The Alzheimer's diagnosis is slow and inaccurate.
An Artificial Neural Network (ANN) is a processing unit containing an input layer, one or more hidden layers and an output layer. Hidden layers contain hidden neurons (processing elements). The ANN mimics the human brain in its ability to recognize patterns and learn. Just like the human brain, the networks are trained by experience (in this case, pre-classified data). During training, the connections between neurons are strengthened or weakened to correlate to the output. After training, the network becomes skilled at classifying new data, and can be tested for accuracy.
There have been previous projects attempting to diagnose Alzheimer's, but they obtained low accuracy, and therefore did not make it to the doctor's office. The research I've looked at extracted image features and tried to achieve a higher accuracy, without using the patient's basic information, or vice versa. None of the papers I found tested out various architectures by combining multiple ANNs. Many of the previous tools used a different type of machine learning, Support Vector Machines, for their classifiers.
I learned from what previous research did, such as their image features or the methods of machine learning. I learned from what they did not, using database combined with image features, testing various components and structures of classifiers, and using neural networks.
Method / Testing and Redesign
I used the OASIS (Open Access Series of Imaging Studies) database for training and testing the classifier. The database contained:
Cross sectional MRI scans of brains of 233 subjects over the age of 40.
Mini Mental State Examination results of the subjects as well as their diagnosis for Alzheimer’s.
- Clinical information about the patients (such as their age, gender, or size of brain)
My project consisted of 3 steps, feature extraction, creation of structures and training/testing of neural networks.
The database contained these features:
Age and Gender
Mini Mental State Examination (MMSE)
Normalized Whole Brain Volume(nWBV): brain volume when the MRI is normalized
Atlas Scaling Factor (ASF): the factor by which the MRI has to be changed to fit to the normalized value
Estimated Total Intracranial Volume (eTIV)
I extracted these features from the images:
Fractal Dimension is the measure of the complexity of the image. This index tells how the details of a pattern changes with scale at which it is measured.
Entropy is the measure of randomness and disorder in the image.
Area of Atrophy is the amount of brain matter lost.
Here is the full feature set.
Single Stage ANN Classifier
This classifier had one input layer, one hidden layer and one output layer.
Chained ANN classifier
This classifier has multiple single stage classifiers connected in series. The first stage classifier is trained with the input features. All the classifiers from the second stage onward are trained with the input features along with the output of the previous stage.
Hierarchical ANN Classifier
This classifier has two ANN classifiers in its first stage . The first one is trained with only the database features. The second one is trained only with the image features. The classifier in the second stage is trained with the outputs from the first two stages.
Training and Testing
I used a stratification method to divide the samples into a 75 : 25 training to testing ratio. I varied the training algorithm and number of hidden neurons, and split up the feature set while training.
I tested the importance of certain features I was inputting into my neural network, by testing accuracies with and without those features. For example, I compared the accuracy of a neural network with an MMSE feature and without. With each feature set, I used Principal Component Analysis to find the principal components of the inputs (convert possibly connected features to a feature set with uncorrelated inputs).
One example of a neural network I created would be a double stage chained network, with a training algorithm of gradient descent, and only the patient's basic info as features.
There were 1080 different neural networks to be made.
The database contained information about 233 patients. 99 patients were clinically diagnosed with Alzheimer's. With the 75-25 split between training and testing sets, I had 198 images in each training set including 83 images of subjects with Alzheimer’s. I had 32 images in each testing set, including 16 images of subjects with Alzheimer’s. There were 10 trials, and in each trial the training and testing sets were randomly picked from the database.
Overall Accuracy: the fraction of the total samples(including training and testing sets) correctly diagnosed
Testing Accuracy: accuracy of diagnoses of the test samples alone
Sensitivity: portion of the total Alzheimer’s samples correctly diagnosed with Alzheimer’s
The 3-stage chained neural network, with a training algorithm of Bayesian Regulation Back propagation, 12-16 hidden neurons, and the full feature set achieved the highest accuracy. This was a 97% testing accuracy when diagnosing 320 times. In addition to finding the best classifier, I determined the importance of the image features and the MMSE. When creating a network without image features or the MMSE, the network's accuracy was significantly lower.
Here is the data.
The bar diagram above shows the average accuracies and test performances of each structure and type of classifier. In the chart above, you can see the performance of the hierarchical, 2-stage, and single staged classifiers. They did not perform as well as the 3-stage classifiers.
Below, I calculated the overall accuracy, sensitivity, and testing performance of the types
of neural networks I created.
Bayesian Regulation Back propagation (trainbr) as a training algorithm provided the best accuracy.
After finding the ideal classifier, I built a GUI interface where one could upload an image and enter the patient information to get a diagnosis. The GUI interface worked correctly for the MRI scans and subjects in the database.
Conclusion / Report
I was able to build a classifier with a greater than 95% accuracy. It achieved an overall (training and testing) accuracy of greater than 99% and a testing accuracy of 97%. This surpassed the accuracies obtained by any previous tool or method. My experiments established that adding fractal dimension, entropy and the area of the atrophy features improve the classifier performance by 5-7% . I also proved that, a classifier based on image features alone (without the mini mental examination results) performed poorly compared to the classifier that is trained with both image features and the patient’s basic information. To diagnose Alzheimer’s, one would need to look at the full set of features with images as well as patient data.
I succeeded in my second goal of building a GUI for the classifier. I created a diagnostic tool that is easy to use and access.
My data is reliable as I tested it over 10 trials, and had a fairly large database of subjects. I would like to improve one thing in my project. The GUI I built requires a certain software license, so I would like to create one that anyone, not only doctors, can use.
This tool can change how patients are diagnosed and the number of patients diagnosed. It addresses the problem of inaccurate testing and, if used, can raise the percentage of people with Alzheimer's who actually get diagnosed with Alzheimer's.
Bibliography, References and Acknowledgements
I'd like to thank multiple people for assisting me in my project. My brother taught me Matlab before I even started on my project, and helped me out when I hit a couple of road blocks during my project. I'd like to thank my family and friends, even those who didn't know I was working on this project, for supporting me in everything I do. I'd also like to thank my science teacher, Mrs. Iyer, for proof reading and going over my project with me.
The "software" I used throughout the project was Matlab. I used it to process the images, structure and run the neural networks, and build the GUI.
I actually went out and got two books, Digital Image Processing Using Matlab by Rafael C. Gonzales, Richard Eugene Woods, and Steve L. Eddins, as well as Neural Networks for Pattern Recognition by Christopher M. Bishop. I can't say I read all of each book, but these were useful for skimming through the text or looking up commands/information I didn't know.
I also used The Nature of Code and most of the other online sources I could get my hands on.
I used the OASIS database and it's accompanying research papers.
Marcus, Daniel, et al. "Open Access Series of Imaging Studies (OASIS): Longitudinal MRI Data in Nondemented and Demented Older Adults." NCBI 22.12 (2010): 2677-684. J Cogn Neurosci. Web. 13 Nov. 2014. http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2895005/.
To determine the method, testing, and image features I would use for my project, I relied on Medscape for understanding the imaging scans.
I also used various publications to research into features to extract from images and the information necessary to diagnose patients.
- Johnson, Keith A., et al. "Brain Imaging in Alzheimer Disease." Cold Spring Harbor Perspectives in Medicine (2012). PMC US Natural Library of Medicine. Cold Spring Harbor Laboratory Press. Web. 17 Nov. 2014.
- Albert M, Knoefel J, Kemper TL. Neuroanatomical and neuropathological changes in normal aging and in dementia. In: Albert M, Knoefel J, editors. Clinical neurology of aging. New York: Oxford University Press; 1994. p. 3-78.
- Jovicich J, et al, MRI-derived measurements of human subcortical, ventricular and intracranial brain volumes: reliability effects of scan sessions, acquisition sequences, data analyses, scanner upgrade, scanner vendors and field strengths. Neuroimage 2009;46:177-92.
- Folstein M, et al. “Mini-Mental State”. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 1975;12:189-98.
I used multiple research papers to see and compare previous research and previous tools.
Comparison of Methods:
Cuingnet, Rémi, et al. "Automatic Classification of Patients with Alzheimer's Disease from Structural MRI: A Comparison of Ten Methods Using the ADNI Database." NeuroImage (2010): 766-81. Print.
Other Research Papers:
- Bowler JV, Munoz DG, Merskey H, Hachinski V. Factors affecting the age of onset and rate of progression of Alzheimer’s disease. J Neurol Neurosurg Psychiatry. 1998;65(2):184–190.
- Desikan, R. S., H. J. Cabral, C. P. Hess, W. P. Dillon, C. M. Glastonbury, M. W. Weiner, N. J. Schmansky, D. N. Greve, D. H. Salat, R. L. Buckner, and B. Fischl. "Automated MRI Measures Identify Individuals with Mild Cognitive Impairment and Alzheimer's Disease." Brain: A Journal of Neurology 132.8 (2009): 2048-057. Print.
- Ahmet, Ekin. "Automated Diagnosis of Alzheimer's Disease Using Image Similarity and User Feedback." ACM Digital Library 34 (2009). Print.
- Anand, Mahesh. "Automated Diagnosis of Early Alzheimer’s Disease Using Fuzzy Neural Network." 4th European Conference of the International Federation for Medical and Biological Engineering 22 (2009): 1455-458. Print.