Brain disorders are conditions that have been taken into thorough examination due to high social and economic impact they have. Progressive exploration and fact-finding analysis in regards to brain structure has been made possible with exploitation of new technology and advancements in medicine. MRI (magnetic resonance imaging) in combination with computer processing provides us with detailed breakdown of brain structure, which can be furtherly used to examine disorders such as Alzheimer’s dementia or Multiple sclerosis. Dementia is nowadays often referred to as “21st century plague”, since it impacts numerous people and causes irreparable damage. Rapid nerve cell loss affects patients and can never be undone. That is why its quick diagnosis is crucial, since it enables preventive treatment and brisk actions to postpone or delay disease impact for as long as possible.
Clustering presents a category of methods, that are used to distinguish objects and grouping them in a way, where objects of one group differ from objects in another group. Objects in one particular group should have at least one common trait that would make them unique. Group-wise analysis enables further examination without the necessity of prior knowledge. Prior knowledge is what sets apart classification and clustering, where classification requires a set of output labels that define clustered groups. Clustering based groups on the other hand are created autonomously by the algorithm.
Artificial intelligence (AI) characterizes the ability of a computer system to learn autonomously, by observing a given dataset and labelled output. No further programming is required. Deep learning is an aspect of AI, where computer system learns through neural network. Neural network is a complex set of connections between artificial neuron units, which illustrates a computer version of a human brain. Artificial neuron models a biological neuron while connections between them model axons and dendrites. Neural network learns to complete challenging tasks on its own, without additional rule setting. Prevailing topic of neural networks is computer vision respectively image data analysis.
Autoencoders display a neural network type whose task is to learn compact encoding and decoding of input data through so-called bottleneck representation. Bottleneck represents a dimensionally reduced set of parameters, which is particularly useful for either denoising or dimensionality reduction of input data. Variational autoencoder is a special representative of autoencoder, whose bottleneck values are regularized by statistical distribution. Reasoning behind it is to prevent overfitting to input data. For this manner, the architecture of autoencoder has to be adjusted.
Magnetic resonance image datasets used in this master’s thesis were processed and used in combinations with different types of analysis methods, whose purpose is to effectively recognize brain’s structural patterns. Hypothesis states that it is possible to find specific structural changes in brain imaging that correlate with specific diseases and their phenotypes. Brain images were processed then clustered with either k-means algorithm, affinity propagation algorithm or hierarchical clustering. Clustering methods were applied to either images directly, to dimensionally reduced data or to bottleneck representation of variational autoencoders. Finally output clusters were correlated with patient’s clinical data.
We concluded there is a strong parallel between output clusters and patients clinical and demographic data. We successfully generated a cluster that represented exclusively ill patients according to EDSS (Expanded Disability Status Scale), age and other health markers. Best results were concluded with combination of variational autoencoder with only convolutional layers and k-means clustering. Results for healthy clusters were accurate and specific.
In the end we can wrap up with a guarantee that artificial intelligence, neural networks and deep learning will be the future of medical diagnosis. Any help and upfront alert of patient’s condition is beneficial. Even greater in the case of magnetic resonance imaging due to its non-invasive and harmless nature. Studies conducted in this master’s thesis may someday help with such diagnosis.