Research

***Page Out of Date***

Image Analysis/Machine Learning in Medical Imaging

The aims of the Sys-MIFTA project are to provide medical practitioners with quantitative measurements of interstitial fibrosis and tubular atrophy through automatic biopsy analysis.

Constrained Clustering

Gathering ground truth data can be expensive and defining classes can be difficult in time-series, unsupervised methods may not offer results that align with a user's requirements. Constraints allow the injection of user knowledge into an unsupervised process to mitigate these problems.

Machine Learning within Remote Sensing Images

The aims of the FOSTER project are to provide geologists with information related to soil erosion and land usage (both manmade, for example mining activities, and natural, for example earthquake and landslide fissures) and tools to predict their evolution. This project focuses on regions within New Caledonia and the French Alps and is run in collaboration with LIRIS (Lyon), LISTIC (Chambéry), PPME (New Caledonia) and IPGS (Strasbourg).

The HYEP and TOSCA projects focus on the use of hyperspectral satellite imagery in remote sensing (urban planning, etc.) by investigating the properties of the future HYPXIM satellite. The my participation of this project was to implement machine learning pipelines for a hyper spectral sample database.

Automatic EEG Sleep Stage Identification

Working in conjunction with Cybula Ltd. and the Institute of Medical Sciences, University of Aberdeen, we aim to apply novel pattern recognition algorithms to improve the detection of activity stages in Electroencephalography (EEG) time series data. Currently there is a particular problem in identifying the REM sleep, Quiet Awake and Non-Convulsive Seizure stages. The analysis is performed using four channels of data, three EEG and the fourth from an accelerometer mounted to the head. The EEG recordings are taken from the left hippocampus, right hippocampus and prefrontal cortex. The problem is exacerbated by the low observation probability of these particular stages, resulting in scarce training data. Successful research will greatly reduce the time taken to manually label the data which can take up to three days for one twenty-four hour recording.

An implementation of this algorithm is available as a shared service on the Carmen portal, if you would like access please contact me.

Spectrogram Track Detection (Ph.D. Topic)

Remotely sensed time series data is conventionally transformed into the frequency domain using the Fast Fourier Transform. This allows for the construction of a spectrogram image in which time and frequency are the axes and intensity is representative of the series' power at a particular time and frequency. It follows from this that, if a stationary or non-stationary periodic narrowband component is present during some consecutive time frames a track, or line, will be present within the spectrogram.

The problem of automatic detection of these tracks has drawn increasing attention in the literature. Existing applications cover a wide range and include meteor detection, speech formant tracking, identifying and tracking marine mammals via their calls and identifying noise radiated by mechanical devices. In the broad sense this "problem arises in any area of science where periodic phenomena are evident and in particular signal processing" (Quinn 1994). In practical terms the problem can form a critial stage in the detection and classification of sources in passive sonar systems, the analysis of vibration data and the analysis of any noisy time series data. Possible future applications include identifying trends in temperature variation, sea level rise and fall in altimeter data, light frequency analysis through spectroradiometry and chemical detection through spectroscopy.

The problem is compounded not only by the low Signal-to-Noise Ratio (SNR) in spectrogram images which contain weak periodic signals but also the variability of the observed track structure.

The Active Contour model proposed by Kass et al. allows for non-parametric feature detection within an image - an ideal property in remote sensing environments where generally a priori shape information is not strictly defined. The active contour is constrained by internal energy forces, which ensure that its shape follows certain criteria; these are typically defined as curvature and connectivity. It is guided by potential energy which attracts it towards features by following local changes in energy gradient. As these gradients are calculated on a local basis the active contour needs to be initialised close to the desired feature to ensure correct convergence. The active contour converges on a minimum of the weighted combination of its internal and potential energies. The potential energy constraints translate this convergence to be a local gradient maxima in the image. However, the dependence on defining image features by gradient inhibits its applicability in low signal-to-noise ratio (SNR) conditions.

We have extended the model and developed a framework for spectrogram track detection which overcomes the outlined limitations. In particular we boost detection performance by integrating information from harmonic locations within the spectrogram, enhance the potential energy with spatial information to improve detection rates at low SNRs and alter the internal energy constraints to better fit the track shape.

Two data sets for testing spectrogram track detection algorithms are available here. The Matlab code developed during my PhD, and used to collate the results presented in my thesis and publications, is available here.