Cognitive Impairment Anomaly Detection Using Metric Learning on Triplet Networks

Abstract

Machine learning is used for many application purposes; some of the common ones being classification, regression, and anomaly detection. This project aims to deal with anomaly detection using metric learning of the data collected by NHATS. This data is usually used to predict cognitive impairment in individuals. This dataset is called the Clock Drawing Test (CDT) and contains drawings from people of different levels of cognitive impairment which have been classified into different classes by a human coder. Since these drawings have been judged by a coder, it is prone to variations from person to person, or even when performed by the same person and hence contain anomalies. We have developed a triplet network based deep metric learning system architecture for anomaly detection and have implemented it to find first find anomalies in a given class, and then find the correct class for the anomalous images. We have compared two different architectures (ResNet101 and EfficientNetB7) to find the best base for transfer learning, and two different metrics (Root Mean Square (RMS) and Euclidean distances) to find the optimal metric for our use case. We have also compared its performance with the more commonly used Siamese Networks to find the most effective metric learning solution for this dataset. Our results showed best performance with a combination of ResNet101 and the RMS metric, outperforming the current Siamese network on this dataset. This report is divided as follows; Chapter 1 introduces the problem followed by descriptions on the method we will use, Chapter 2 provides an extensive literature review, Chapter 3 describes the methodology that is to be followed, Chapter 4 describes the various results obtained, and Chapter 5 closes with conclusion and future work.