| Summary: | The multifocal electroretinogram (mfERG) provides spatial and temporal information on the retina’s function in an objective manner, making it a valuable tool for monitoring a wide range of retinal abnormalities. Analysis of this clinical test can however be both difficult and subjective, particularly if recordings are contaminated with noise, for example muscle movement or blinking. This can sometimes result in inconsistencies in the interpretation process. An automated and objective method for analysing the mfERG would be beneficial, for example in multi-centre clinical trials when large volumes of data require quick and consistent interpretation. The aim of this thesis was therefore to develop a system capable of standardising mfERG analysis. A series of methods aimed at achieving this are presented. These include a technique for grading the quality of a recording, both during and after a test, and several approaches for stating if a waveform contains a physiological response or no significant retinal function. Different techniques are also utilised to report if a response is within normal latency and amplitude values. The integrity of a recording was assessed by viewing the raw, uncorrelated data in the frequency domain; clear differences between acceptable and unacceptable recordings were revealed. A scale ranging from excellent to unreportable was defined for the recording quality, first in terms of noise resulting from blinking and loss of fixation, and secondly, for muscle noise. 50 mfERG tests of varying recording quality were graded using this method with particular emphasis on the distinction between a test which should or should not be reported. Three experts also assessed the mfERG recordings independently; the grading provided by the experts was compared with that of the system. Three approaches were investigated to classify a mfERG waveform as ‘response’ or ‘no response’ (i.e. whether or not it contained a physiological response): artificial neural networks (ANN); analysis of the frequency domain profile; and the signal to noise ratio. These techniques were then combined using an ANN to provide a final classification for ‘response’ or ‘no response’. Two methods were studied to differentiate responses which were delayed from those within normal timing limits: ANN; and spline fitting. Again the output of each was combined to provide a latency classification for the mfERG waveform. Finally spline fitting was utilised to classify responses as ‘decreased in amplitude’ or ‘not decreased’. 1000 mfERG waveforms were subsequently analysed by an expert; these represented a wide variety of retinal function and quality. Classifications stated by the system were compared with those of the expert to assess its performance. An agreement of 94% was achieved between the experts and the system when making the distinction between tests which should or should not be reported. The final system classified 95% of the 1000 mfERG waveforms correctly as ‘response’ or ‘no response’. Of those said to represent an area of functioning retina it concurred with the expert for 93% of the responses when categorising them as normal or abnormal in terms of their P1 amplitude and latency. The majority of misclassifications were made when analysing waveforms with a P1 amplitude or latency close to the boundary between normal and abnormal. It was evident that the multilayered system has the potential to provide an objective and automated assessment of the mfERG test; this would not replace the expert but can provide an initial analysis for the expert to review.
|
|---|
