|M.Sc Student||Cohen Ido|
|Subject||Unsupervised Anomaly and Target Detection using Manifold|
Learning with Application to Deep Brain
|Department||Department of Electrical Engineering||Supervisor||Professor Ronen Talmon|
|Full Thesis text|
Target and anomaly detection refers to the problem of finding unique patterns in data or patterns that do not conform to the expected behavior. Anomaly and target detection is considered challenging, especially since anomalies and targets can be expressed in various forms and anomalies are usually very rare. Most target and anomaly detection algorithms use prior knowledge such as predefined models and labeled data. This could lead to a significant bias, and the subsequent detection performance greatly depends on the quality of the prior knowledge.
To alleviate such a dependence and bias, we develop a data-driven unsupervised method based on manifold learning. We propose to define features which carry sufficient information that can be computed solely from the measurements. Then, we develop a variant of the Mahalanobis distance between these features and incorporate it into a specific manifold learning method, called diffusion maps. We analyze the proposed method using stochastic calculus. Particularly, we show that the modified Mahalanobis distance between the proposed features allows us to approximate the intrinsic variables that characterize the data, and by that facilitates an accurate target and anomaly detection.
We showcase our method on two applications. First, we apply it to observations of a simple mechanical system. This particular mechanical system was chosen since it has a known model, which serves as a definitive ground truth that can be used for validation. Using our method, we show the recovery of the main properties of the system solely from its observations in a data-driven manner. Indeed, the recovered properties coincide with the ground truth.
Second, we address a target detection task involving Deep Brain Stimulation (DBS). Typically in DBS, a surgery to implant a stimulating device is carried out. This device sends electrical signals to specific brain areas that are responsible for body movements. Once the device is implanted in an appropriate position, DBS can help in reducing the symptoms of tremor, slowness, stiffness, and walking problems caused by several neuronal diseases, such as Parkinson’s disease, dystonia, or essential tremor. During the surgery, one important task is to detect the appropriate area for implanting the device. We focus on Parkinson's disease, for which the target region is the Sub-Thalamic Nucleus (STN) and a sub-region within the STN, called Dorso-Lateral Oscillatory Region (DLOR). An accurate detection of the STN and the DLOR is crucial for adequate clinical outcomes. Based on our method, we develop an unsupervised algorithm for the detection of the STN and the DLOR during a DBS surgery. We show that our algorithm attains detection results that outperform the gold standard. In addition, we show a proof of concept extension of the algorithm to the detection of the Globus Pallidus (GP) region that is of interest for treating dystonia.