Efficient level set method for novelty detection

Abstract

Novelty detection is the process of identifying unfamiliar patterns that deviate from typical patterns. In the realm of machine learning, novelty detection techniques aim to establish closed boundaries that encompass most typical patterns, often referred to as normal or target patterns. Patterns falling outside these boundaries are flagged as abnormal data. Novelty detection holds significant relevance in applications where tolerance for anomalous patterns is low. This thesis centres on the practical application and development of novelty detection techniques.

In terms of practical application, the thesis employs novelty detection techniques to detect individuals with mild cognitive impairments (MCI) who are at risk of developing Alzheimer's disease within a two-year timeframe. The research also investigates the most influential modalities for the detection. Experimental findings demonstrate that novelty detection techniques can be effectively employed for the detection, offering comparable performance to binary classification methods while requiring less data. Furthermore, the challenge of imbalanced data can be effectively addressed using novelty detection techniques.

With a focus on the development of novel novelty detection algorithms, the thesis proposes two innovative approaches based on the Level Set Boundary Description (LSBD) method. The LSBD method constructs a versatile boundary tailored to the distribution of training data in the input space by evolving the implicit level set function (LSF) that defines the boundary. To enhance the detection performance of the LSBD method in high-dimensional datasets, a Hybrid Autoencoder Level Set Method (AE-LSM) is proposed. It leverages AEs to reduce data dimensionality into a low-dimensional space. Subsequently, the compressed data are used for training the LSBD model. A novel speed adjustment mechanism that controls the boundary evolution, is then proposed to obtain the final decision boundary.

To optimise the efficiency of the LSBD method, the thesis proposes an Adaptive LSM (ALSM)approach that enhances the LSBD in two keyways. First, a Signed Distance Function (SDF)-based LSF is constructed to simplify solving the Level Set Equation (LSE) for boundary evolution. Second, a reaction-diffusion-based boundary evolution strategy is employed to maintain the LSF (i.e., the SDF) throughout the evolution, eliminating the requirement for reinitialisation of the LSF which is time-consuming. Extensive experimental evaluations demonstrate that the proposed methods consistently outperform their counterparts across a wide range of datasets via several performance analyses.

Thesis is embargoed until 31st March 2026.

Date of Award	Mar 2024
Original language	English
Supervisor	Xuemei Ding (Supervisor), Junxiu Liu (Supervisor) & Damien Coyle (Supervisor)