Genetic Algorithm Optimisation of a Technique for Diagnosing Dementing Diseases through the Distribution of Cerebral Atrophy.

Purpose: We previously presented a technique for diagnosing dementing diseases, including Alzheimer's disease, fronto-temporal dementia and vascular dementia, through coarse regional cerebral atrophy measurements derived from CSF volumes. We now present preliminary results from a genetic algorithm optimisation of the technique.Methods and Materials: T2 IR MR volumes were obtained from 161 patients, aged 19 to 85, and coregistered into a standard coordinate system. Binary CSF maps were produced by thresholding, then masking off non-CSF fluid spaces. CSF volumes were measured in twelve regional volumes, then normalised for head size and age-related atrophy. A cross-validated Parzen classifier was applied to this data to produce a diagnosis for each patient. Finally, the diagnoses were used as the cost function for a multi-objective genetic algorithm optimisation, which altered the CSF measurement volumes to optimise the diagnostic capabilities.Results: The original atrophy measurement technique had a correct classification rate of 76% over all groups. Preliminary results from the optimisation indicate that significant improvements, of the order of 10%, could ultimately be achieved. Furthermore, the genetic algorithm has automatically aligned the CSF measurement regions to the underlying biological structure, providing insight into the disease processes on the broadest spatial scale.Conclusion: We conclude that relatively coarse atrophy measurements can be useful for diagnosing dementing diseases, and further that genetic algorithm optimisation can significantly improve the utility of such techniques. Finally, we conclude that the results of such an optimisation can provide insight into the disease processes.