Modular Neural Networks for Analysis of Flow Cytometry Data

  • Arnaud Autret

    Student thesis: Doctoral Thesis


    In predicting environmental hazards or estimating the impact of human activities on the marine ecosystem, scientists have multiplied the need for sample analysis. The classical microscopic approach is time consuming and wastes the talent and intellectual abilities of trained specialists. Therefore, scientists developed an automated optical tool, called a Flow Cytometer (FC), to analyse samples quickly and in large quantities. The flow cytometer has
    successfully been applied to real phytoplankton studies. However, analysis of the data extracted from samples is still required. Artificial Neural Networks (ANNs) are one of the tools applied to FC data analysis.

    Despite several successful applications, ANNs have not been widely adopted by the marine biologist community, as they can not possible to change the number of species in the classification problem without retraining of the full system from scratch. Training is time consuming and requires expertise in ANNs. Moreover, most ANN paradigms cannot cope effectively with unknown data, such as data coming from new phytoplankton species or from species outside the scope of the studies.

    This project developed a new ANN technique based on a modular architecture that removes the need for retraining and allows unknowns to be detected and rejected. Furthermore, the Support Vector Machine architecture is applied in this domain for the first time and compared against another ANN paradigm called Radial Basis Function Networks. The results show that the modular architecture is able to effectively deal with new data which can be incorporated into the ANN architecture without fully retraining the system.
    Date of Award2003
    Original languageEnglish
    SupervisorColin Morris (Supervisor) & Paul Angel (Supervisor)

    Cite this