Diagnosing Alzheimer's with Machine Learning
Automating Alzheimer’s Diagnosis
Thomas Jefferson High School for Science and Technology
This article was originally included in the 2018 print publication of Teknos Science Journal.
I breathe a sigh of relief. It’s 10 a.m. on a Tuesday, and my work for the week at the NIH Center for Biomedical Communications is complete. I’m not running some conventional program that will finish in a few seconds, or even minutes. The convolutional neural network currently training on millions of Alzheimer’s brain scans will take days to complete.
Alzheimer’s disease is an incurable neurodegenerative brain disorder that progressively destroys neurons and synapses in the brain. Alzheimer’s causes memory loss, impaired reasoning, speaking, reading, and writing, as well as decreased spatial abilities. Alzheimer’s is also widespread: affecting 5.5 million Americans alone and accounting for 60-70% of all cases of dementia. No treatment has been shown to slow the progression of the disease (Alzheimer’s Association, 2017).
There is no single diagnostic test for Alzheimer’s disease. Instead, doctors obtain family, medical, and psychiatric history, consult family members, conduct cognitive tests, and perform neurological exams. Patients who show symptoms of Alzheimer’s also undergo blood tests and brain imaging to rule out other forms of dementia. The process of data collection and physician interpretation can take several weeks per diagnosis (Alzheimer’s Association, 2017). However, within the past few years, machine learning algorithms have been created for Alzheimer’s diagnosis from fMRI data, a collection of MRI datasets. These algorithms could soon replace the long and expensive process—hundreds of hours and thousands of dollars—required for manual diagnosis with a few seconds of computation on an ordinary desktop computer.
Machine learning is a subfield of computer science focused on programming computers to learn from data without being explicitly programmed; machine learning can accomplish tasks impossible with previous algorithms. For example, how could a traditional algorithm distinguish between images of cats from images of dogs? Dogs and cats come in different sizes, colors, and shapes. An image could have the animal in any orientation: facing towards or away from the camera, or even partially obscured. For a traditional algorithm, even this simple binary classification problem has too many variables and is hopelessly complex. And yet, any toddler can distinguish animals from one another. In real life, parents show children pictures of animals and after a while, children pick up on the differences and defining features of dogs and cats.
Machine learning algorithms are basically extraordinarily fast toddlers. Rather than explicitly telling the algorithm which features of dogs are distinguishable from features of cats, thousands of photos of each are inputted so that the algorithm can determine these features by itself. The algorithm learns by slightly altering the weights in its neural network for every photo that is processes. In a matter of hours, a neural network can learn the differences between dogs, cats, or any other type of object.
Machine learning algorithms learn and recognize patterns in seen data (training data) and use these patterns to predict characteristics of unseen data (real data). In particular, deep learning refers to large, complex machine learning models and convolutional networks are the most potent deep learning algorithm (LeCun, Bengio, & Hinton, 2015). Operating solely on image data, convolutional networks have been revolutionary for object detection, image classification, and instance segmentation (Szegedy, Ioffe, Vanhoucke, & Alemi, 2016).
Applying convolutional networks for an Alzheimer’s diagnosis is relatively straightforward. First, thousands of images of fMRI brain scans, half of brains with Alzheimer’s and the other half of brains with normal cognitive function, will be collected and preprocessed. Then, a large convolutional network will be trained with the data for several weeks. If the network fails to learn properly, the parameters are adjusted and the network will be retrained. However, since there are significant visual differences between healthy brains and those afflicted with Alzheimer’s disease, Sarraf, DeSouza, Anderson, and Tofighi (2017) were able to achieve near-perfectly accurate results using a 2D convolutional network for detecting binary Alzheimer's (diseased brains) vs. Normal classification (healthy brains).
However, binary classification lacks applicability to clinical settings. Alzheimer’s and normal cognition represent two opposite ends of the cognitive impairment spectrum, when in reality, many patients lie in between these two extremes. These patients have Mild Cognitive Impairment (MCI), a condition of noticeable cognitive decline that is not severe enough to impact everyday activities. Approximately 15-20% of people over 65 have MCI and 32% of people with MCI develop Alzheimer's within five years (Alzheimer’s Association, 2017). Diagnosing individuals with MCI before they develop Alzheimer's is crucial for maximizing the effectiveness of potential treatments, since patients at the MCI stage are not yet afflicted with the extensive brain damage that Alzheimer's patients have in relation (Wu et al., 2012). Since networks trained for Alzheimer's vs. Normal classification are not trained or tested on mild cognitive impairment (MCI) subjects, such as those in Sarraf et al. (2017), they cannot provide patients with a diagnosis. Thus, for real-world applicability, machine learning models must be trained on a broad range of data encompassing the cognitive impairment spectrum, and they must achieve multiclass accuracy, able to identify multiple degrees of cognitive decline and disease progression.
Multiclass classification (Alzheimer’s vs. MCI vs. Normal) is a far more difficult task than binary classification (Alzheimer’s vs. Normal). The recent CaMCCo model required physiological, proteomic, genomic, and image data in order to reach above 80% accuracy in multiclass classification. Although CaMCCo represents a leap in multiclass classification, it does not have the efficiency or practically of solely image-based models (Singanamalli, Wang, & Madabhushi, 2017). Only bleeding-edge imaging solutions have seen success in multiclass diagnosis.
My research at the NIH focused on multiclass classification using only imaging data. After preprocessing millions of MRI scans from thousands of subjects, I trained and re-trained models for weeks, testing different convolutional network models and parameters.
The neural network I developed achieved 85.1% accuracy. Eighty-five percent accuracy for multiclass diagnosis exceeds previous image-based algorithms (Korolev, Safiullin, Belyaev, & Dodonova, 2017). Although not yet quite comparable to the current doctors’ diagnosis accuracy, it is much most cost and time effective. Furthermore, since convolutional networks for multiclass classification of brain damage are constantly being improved, it does not seem impossible that, in the near future, these algorithms may replace the need for trained experts in the field.
Alzheimer's Association. (2017). Early detection and diagnosis of Alzheimer's disease [Pamphlet]. Retrieved from http://act.alz.org/site/DocServer/Policy_Brief_-_Early_Detection_and_Diagnosis_Brief__Assn.pdf
Korolev, S., Safiullin, A., Belyaev, M., & Dodonova, Y. (2017). Residual and plain convolutional neural networks for 3D brain MRI classification. arXiv. Retrieved from https://arxiv.org/pdf/1701.06643.pdf
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521, 436-444. https://doi.org/10.1038/nature14539
Sarraf, S., DeSouza, D. D., Anderson, J., & Tofighi, G. (2017). DeepAD: Alzheimer′s disease classification via deep convolutional neural networks using MRI and fMRI. bioRxiv. https://doi.org/10.1101/070441
Singanamalli, A., Wang,, H., & Madabhushi, A. (2017). Cascaded multi-view canonical correlation (CaMCCo) for early diagnosis of Alzheimer’s disease via fusion of clinical, imaging and omic features. Scientific Reports, 7(1). https://doi.org/10.1038/s41598-017-03925-0
Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. (2016). Inception-v4, Inception-ResNet and the impact of residual connections on learning. arXiv. Retrieved from https://arxiv.org/abs/1602.07261
Wu, L., Rowley, J., Mohades, S., Leuzy, A., Dauar, M. T., Shin, M., . . . Rosa-Neto, P. (2012). Dissociation between brain amyloid deposition and metabolism in early mild cognitive impairment. PLoS One, 7(10). https://doi.org/10.1371/journal.pone.0047905