Innovative signal processing and data mining techniques for aquatic animal health

Abstract

Problem: Aquatic animal health data is often stored in unstructured formats like text and medical images, making large-scale analysis challenging due to the complexity of processing such data. Objectives: In this thesis, we aim to develop text mining, signal processing, image processing, and machine learning techniques to analyse unstructured data effectively. These methods will enable the aggregation of information across large datasets of unstructured aquatic animal health data. Methodology: • For text analysis, we have designed an ontology-based framework for extracting and storing information from aquatic animal post-mortem reports, with a focus on gross pathology reports. While we initially applied this framework to marine mammal stranding reports, it can be adapted for various species and report types. • For medical image analysis, we have created methods for identifying and analysing lesions in whole-slide images (WSIs) of Atlantic salmon gills. Our approach includes a novel feature extraction technique utilising the empirical wavelet transform, and we enhance context-awareness by employing a variational autoencoder to identify regions of interest within histology images. Achievements: The research resulted in the development of an ontology-based framework for systematic text extraction and storage from marine mammal gross pathology reports. We showcased our framework’s performance by using it to analyse bottlenose dolphin attacks on harbour porpoises. Additionally, we created innovative methods for lesion detection in Atlantic salmon gill whole-slide images, incorporating advanced techniques such as the empirical wavelet transform, deep learning, and a variational autoencoder for context-awareness. These achievements collectively advance the analysis of unstructured aquatic animal health data, enabling more comprehensive and efficient data processing. At the time of writing, the project is the only one to apply data-driven approaches to marine mammal post-mortem reports and gill WSIs.